Introduction to Mechanical Engineering

The big picture

Look around the room you are sitting in and start counting the things that move, carry a load, move a fluid, or move heat. The chair holding you up. The hinge on the door. The fan in your laptop, and the heat sink it blows across. The compressor humming in the refrigerator down the hall. The elevator, the water pipes, the air handler pushing conditioned air through a duct above the ceiling tile. The car in the lot, with its roughly two thousand moving parts. Somebody sized every one of those things, chose a material for it, decided how it would be manufactured, and signed off that it would not break. Most of those somebodies were mechanical engineers.

Mechanical engineering is the broadest of the engineering disciplines, and its breadth is not an accident of history. It is the direct consequence of what the field is built on: mechanics (how forces and motion work), thermal science (how energy and heat work), and materials (what stuff can take). Almost every physical product touches at least one of those three. That is why you will find mechanical engineers designing artificial heart valves, wind turbine gearboxes, semiconductor wafer handlers, prosthetic knees, rocket turbopumps, roller coasters, insulin pumps, HVAC systems for hospitals, and the machine that puts the cap on a shampoo bottle six hundred times a minute.

Here is the plan for today. First we map the field honestly, subdiscipline by subdiscipline, so you know what is in the tent. Then we look at what the job actually consists of, hour by hour, which is less glamorous and more interesting than the recruiting posters suggest. Then we name the shared toolkit that makes all these specialties one profession. And then we do the thing this course will do in every single lesson: we compute something real. By the end of this lesson you will have sized the propulsion requirement for the machine we are going to design together over the next fifteen lessons.

A map of the field

Mechanical engineering divides, roughly, into seven clusters. They overlap constantly, and a working engineer usually lives in two or three of them.

Solid mechanics and structures asks whether a part will break, bend too far, or wobble. It runs from statics (what forces are in the members) through mechanics of materials (what stresses those forces produce) to fatigue, fracture, and finite element analysis. Dynamics, vibration, and mechanisms asks how things move: the path a robot arm sweeps, the ratio a gearbox delivers, why a machine at 1800 rpm shakes itself apart at 1790 rpm. Thermal and fluid sciences asks where the energy and the heat go: engines, turbines, pumps, pipes, heat exchangers, cooling of everything from a phone to a data center. Materials and manufacturing asks what a part is made of and how it gets made: casting, forming, machining, welding, additive, and the tolerances that decide whether it assembles. Design and machine elements is the synthesis discipline: bearings, shafts, fasteners, springs, seals, gears, and the judgment to combine them. Mechatronics and controls puts sensors, actuators, and feedback into mechanical systems, which is now most of them. Energy systems works at plant scale: power generation, HVAC, refrigeration, renewables, and efficiency.

This course visits all seven. Three of them have their own full courses on this site, and rather than repeat that material we will build on it and point you there. Engineering Mechanics: Statics (ENGR 210) gets you from external loads to internal forces. Engineering Thermodynamics (ENGR 230) develops the first and second laws and the power and refrigeration cycles properly. Materials Science and Engineering (ENGR 250) explains why materials behave as they do, from atomic bonding to heat treatment. Here we take what those courses produce and use it to design machines.

Key idea: Mechanical engineering is broad because it is founded on mechanics, thermal science, and materials, and nearly every physical product depends on at least one of the three.

What the job actually looks like

Ask a practicing mechanical engineer what they did last week and you will rarely hear about a differential equation. You will hear about a supplier who cannot hold a tolerance, a test that failed in an interesting way, a design review with the manufacturing group, a spreadsheet reconciling three load cases, and about four hours of computer-aided design. The technical core is real and load-bearing, but it sits inside a workflow.

That workflow, in rough order: understand the need and turn it into requirements you can test; generate more concepts than feel necessary; select among them with a defensible method; analyze the chosen concept until you believe it; model it in CAD and produce drawings a shop can read; build a prototype; test it, usually to failure; discover that you were wrong about something; iterate. Then comes the part students never see coming, which is documentation, supplier qualification, cost reduction, and sustaining engineering on a product that is already shipping. Design engineers commonly report that a minority of their week is spent generating new geometry and the majority is spent making sure existing geometry is correct, buildable, and affordable.

The United States Bureau of Labor Statistics tracks mechanical engineering as one of the largest engineering occupations, with employment on the order of three hundred thousand and median pay well above the median for all occupations, spread across machinery, transportation equipment, architectural and engineering services, aerospace, and research. That spread is the point: the degree is a general license to work on physical things, and graduates end up in industries they had never heard of at nineteen.

Key idea: The technical analysis is the core of the job but not the bulk of the hours; most mechanical engineering work is requirements, iteration, manufacturability, and documentation around a defensible technical center.

The shared toolkit

What makes these seven clusters one profession is a toolkit, and it is smaller than you would guess. Five habits carry most of the weight.

Draw a boundary and conserve something. Every analysis in this course starts by isolating a body or a region and writing a balance across it: forces sum to zero (statics), mass in equals mass out (continuity), energy in equals energy out plus storage (the first law). The free-body diagram and the control volume are the same move applied to different quantities.

Convert loads to stresses, and compare against a material number. Statics tells you a shaft carries 60 N m of torque. Mechanics of materials converts that into 38 MPa of shear stress in a 20 mm shaft. A materials handbook says the steel yields in shear near 200 MPa. Divide and you have a factor of safety. That three-step chain is the spine of mechanical design.

Estimate before you calculate. A good engineer knows the answer to within a factor of two before opening the software. If your finite element model says the deflection is 40 mm and your hand estimate said 0.6 mm, one of them is wrong, and you now know to find out which. This habit is what separates engineers from people who operate engineering software.

Look it up in a standard. You do not derive the thread profile of an M10 bolt or the geometry of a spur gear tooth. Those are settled and published. ISO, ASME, ASTM, SAE, and AGMA standards encode enormous amounts of hard-won practice, and using them is a professional obligation, not laziness.

Design for the real distribution of loads, not the nominal one. The cart you are about to design will not spend its life carrying exactly 150 kg on smooth pavement. It will be overloaded, dropped off a curb, left in the rain, and operated by someone in a hurry. Every number in a specification is really a distribution, and factors of safety are how we buy margin against the tail of it.

Key idea: Boundary and balance, load to stress to allowable, estimate first, use the standards, and design for the tail of the load distribution: five habits that unify the whole discipline.

The running project: a small electric utility cart

Abstractions do not teach design. So for the rest of this course we are going to design one machine together, returning to it in every module as your tools get sharper. The project is a small electric utility cart, the kind of low-speed vehicle that moves tools, mail, or landscaping supplies around a campus, a warehouse yard, or a large facility. Call it the Atlas cart. Here is the headline specification, and these numbers will not change; every calculation in this course ties back to them.

Parameter	Value
Payload	150 kg
Empty mass	150 kg
Gross mass (design)	300 kg
Top speed, level ground	20 km/h (5.56 m/s)
Grade capability	10 percent at 8 km/h
Wheel diameter	300 mm (radius 0.15 m)
Motor	48 V, 1.2 kW continuous at 3000 rpm
Battery	48 V, 40 Ah (1.92 kWh)
Cargo bed	1.2 m x 0.8 m
Target parts cost	2500 US dollars at a build of 20 units

Notice what a specification is: a set of numbers that are testable. Not "reasonably fast" but 20 km/h. Not "carries a decent load" but 150 kg. We will spend the next lesson on where numbers like these come from, because getting them wrong is the most expensive mistake in engineering and the hardest one to detect early.

Worked example: how much power does the cart need?

Let us do the first real calculation of the course. To hold a steady speed on level ground, the motor must supply exactly enough force to cancel the resistances. For a wheeled vehicle at low speed there are two that matter: rolling resistance and aerodynamic drag.

Rolling resistance comes from the tire deforming as it rolls, plus bearing and drivetrain drag. It is modeled as a coefficient times the normal load: F(roll) = C(rr) x m x g. Pneumatic tires on pavement give C(rr) around 0.010 to 0.020; take 0.015. With m = 300 kg and g = 9.81 m/s^2:

F(roll) = 0.015 x 300 x 9.81 = 44.1 N

Aerodynamic drag is F(d) = 0.5 x rho x C(d) x A x v^2. Air density rho is about 1.2 kg/m^3, a blunt little cart has a drag coefficient C(d) near 0.6, and the frontal area A of a cart with a standing operator is about 0.9 m^2. At top speed v = 5.56 m/s:

F(d) = 0.5 x 1.2 x 0.6 x 0.9 x (5.56)^2 = 0.324 x 30.9 = 10.0 N

Total resistance is 44.1 + 10.0 = 54.1 N, and the power delivered at the wheels is force times velocity:

P(wheels) = 54.1 N x 5.56 m/s = 301 W

Three hundred watts. That is less than a hair dryer, and it is the whole reason electric utility carts are practical. Look at the split, too: at 20 km/h, drag is only 18 percent of the resistance. Aerodynamics barely matters here. Double the speed to 40 km/h and drag quadruples to 40 N while rolling resistance stays at 44 N, and suddenly the shape of the vehicle is worth arguing about. Knowing which physics dominates at which speed is exactly the judgment this course is trying to build in you.

Now the climb, which is the real sizing case. On a 10 percent grade the component of weight along the slope is approximately m x g x 0.10 = 300 x 9.81 x 0.10 = 294 N. Add rolling resistance (44 N) and you need about 338 N of tractive force. At 8 km/h (2.22 m/s), that is P = 338 x 2.22 = 751 W at the wheels. Divide by the drivetrain efficiency, about 0.90 for two chain stages, and by the motor and controller efficiency, about 0.85, giving an overall 0.77, and the battery must deliver roughly 975 W. Our 48 V pack supplies that at about 20 A, which is comfortable. The 1.2 kW motor covers the climb with margin, and at cruise it loafs along at 301 / 0.77 = 391 W from the battery, about 20 watt-hours per kilometer.

One honest caveat, and it is the kind that a first-pass estimate is supposed to surface. On the climb the cart is moving at 8 km/h, so the motor is turning at only 8/20 of its 3000 rpm rating, about 1200 rpm. Power is torque times angular speed, so producing 975 W at 1200 rpm (126 rad/s) demands about 7.7 N m of motor torque, while the motor's continuous rating of 1.2 kW at 3000 rpm (314 rad/s) corresponds to only 3.8 N m. The climb is therefore a short-duty, above-continuous-rating condition. That is fine for a two minute hill and not fine for a twenty minute one, and finding that out on paper in lesson one is much cheaper than finding it out from a burned motor winding in month nine.

Key idea: A first-pass power estimate takes ten minutes, tells you which physics dominates, and exposes duty-cycle problems long before hardware exists.

Common misconceptions

• Mechanical engineering is about cars and engines. Those are two famous applications. The discipline covers medical devices, semiconductors, buildings, robotics, energy, packaging, aerospace, and consumer products just as fully.
• Real engineers do not estimate, they compute exactly. Backwards. Estimation is the senior skill; exact computation is what you do after the estimate has told you what to compute and roughly what to expect.
• Simulation replaced hand analysis. Simulation replaced hand analysis for the final numbers and made hand analysis more important for checking them. An unvalidated finite element result is an opinion with colors.
• The design is finished when the analysis passes. The analysis passing is roughly the midpoint. Manufacturability, cost, assembly, serviceability, standards compliance, and testing all still stand between you and a product.

Recap

• Mechanical engineering is the broadest engineering discipline because it rests on mechanics, thermal science, and materials, which between them touch nearly every physical product.
• The field clusters into solid mechanics, dynamics and mechanisms, thermal and fluid science, materials and manufacturing, machine design, mechatronics, and energy systems.
• The shared toolkit is small: draw a boundary and conserve something, convert load to stress and compare with an allowable, estimate before computing, use published standards, and design for the tail of the load distribution.
• The course runs on one project, a 300 kg gross electric utility cart with a 20 km/h top speed and a 10 percent grade requirement.
• A ten minute estimate gives 54 N of resistance and about 301 W at the wheels at cruise, 751 W on the design grade, and reveals that the climb is a short-duty condition above the motor's continuous torque rating.

Sources

1. U.S. Bureau of Labor Statistics. (2024). Mechanical engineers. Occupational Outlook Handbook. bls.gov
2. American Society of Mechanical Engineers. (n.d.). About ASME. asme.org
3. Encyclopaedia Britannica. (n.d.). Mechanical engineering. britannica.com
4. OpenStax. (2016). Newton's laws of motion. In University physics volume 1. Rice University. openstax.org

The big picture

A facilities manager stops you in a hallway and says: "We need something to move stuff around campus. The pickup truck is overkill and the hand carts are killing my staff." That sentence is a need. It is not a design problem yet, and if you start sketching now you will build the wrong machine beautifully.

The single most expensive class of error in engineering is not a miscalculated stress. It is solving the wrong problem, discovering it late, and paying for the discovery in tooling, schedule, and reputation. Studies of product development repeatedly find that decisions made in the first ten percent of a project commit the large majority of its eventual cost, while the money itself is not spent until much later. The corollary is uncomfortable: the hours you spend before you have drawn anything are the highest-leverage hours of the entire project.

So today we do the unglamorous work. We take that hallway sentence and turn it into a requirements document you could hand to a stranger. Then we open the concept space wide, using two structured tools that reliably beat unstructured brainstorming: functional decomposition and the morphological chart. Next lesson we will narrow the space back down with a decision matrix, prototype, and iterate. Together those two lessons are the engineering design process, and the process is what you actually get paid for.

The process, in stages

Every organization has its own diagram, and they are all roughly the same diagram. A serviceable version has six stages: define the problem (stakeholders, needs, requirements, constraints); generate concepts (many, deliberately, before judging any); select a concept (with a documented, defensible method); develop and analyze (embodiment, calculations, CAD, tolerance stacks); prototype and test (build it, break it, measure it); and iterate and release (fix what testing found, freeze, document, hand to manufacturing).

Two features of that list matter more than the list itself. First, it is not a line. The arrows run backwards constantly, and a project that never loops backwards is a project whose testing is not honest. Second, the stages are separated on purpose so that you do not judge while you are generating. Mixing generation with evaluation is the most common way engineers accidentally narrow themselves to the first idea that occurred to them, which is almost always the idea they have seen before.

NASA formalizes this into systems engineering, in which requirements flow down from mission objectives to subsystems to components, and verification flows back up. That discipline exists because on a spacecraft you cannot iterate after launch. The same logic applies, at smaller scale, to a product that will be tooled for injection molding, or to a machine that will be installed in a customer's plant.

Key idea: The design process is six looping stages, and the deliberate separation of concept generation from concept selection is what keeps you from building the first idea you had.

From need to requirement

A need is a statement in the stakeholder's language. A requirement is a statement in the engineer's language, and it must satisfy three tests. It must have a metric: a physical quantity you can measure. It must have a target: a number, with a direction and ideally a tolerance. And it must have a verification method: how you will prove it, before anyone asks.

"The cart should be fast enough" fails all three. "Top speed on level pavement shall be 20 km/h plus or minus 1 km/h, verified by three timed runs over a measured 100 m course at gross mass" passes all three. Notice how much the second version constrains: it tells you the load condition, the surface, and the fact that a single lucky run does not count.

Sort your requirements into two kinds. Constraints are pass or fail; you either meet them or the design is disqualified. Overall width must not exceed 900 mm because the cart has to pass through a standard doorway, and 901 mm is a failure no amount of elegance redeems. Objectives are things you want more or less of: lower cost, longer range, lighter weight. Objectives get traded against each other; constraints do not get traded at all. Confusing the two produces designs that are optimized into disqualification.

You should also separate functional requirements (what the machine must do: carry, move, stop, steer) from non-functional ones (what it must be like: safe, serviceable, quiet, affordable, weatherproof, legal). Non-functional requirements are where inexperienced teams get ambushed, because they are easy to leave unwritten and expensive to retrofit. Nobody forgets that the cart must move. Plenty of teams forget that it must be chargeable from a standard outlet by a person wearing gloves in the rain.

Here is a working requirements table for the Atlas cart. This is the document the rest of the course designs against.

Need	Metric	Target	Type	Verification
Carries a useful load	Payload mass	150 kg minimum	Constraint	30 minute laden drive with 150 kg of test weight
Keeps pace with service traffic	Top speed, level, laden	20 plus or minus 1 km/h	Constraint	Three timed runs over 100 m
Handles campus hills	Speed on 10 percent grade	8 km/h minimum	Constraint	Timed climb on a measured ramp, laden
Lasts a work shift	Range, mixed duty	40 km minimum	Objective	Repeated drive cycle to 80 percent depth of discharge
Fits campus paths and doors	Overall width	900 mm maximum	Constraint	Measurement over widest point
Stops safely	Stopping distance from 20 km/h	5 m maximum	Constraint	Laden brake test on dry pavement
Affordable to buy	Parts cost at 20 units	2500 US dollars maximum	Objective	Bill of materials rollup with quoted prices
Usable by untrained staff	Time to first competent use	15 minutes maximum	Objective	Trial with five operators who have never seen it

Check the stopping requirement for internal consistency, which is the sort of thing you should do to your own requirements before a reviewer does it for you. Stopping from v = 5.56 m/s in s = 5 m requires a deceleration a = v^2 / (2s) = 30.9 / 10 = 3.09 m/s^2, which is about 0.32 g. That is well within what rubber on dry pavement can deliver (roughly 0.7 g before sliding), so the requirement is achievable and the brakes, not the tires, will be the limit. Had the arithmetic come out at 1.2 g, the requirement would have been physically impossible and we would have caught it on day one instead of during brake testing.

Key idea: A requirement needs a metric, a target, and a verification method; constraints are pass or fail while objectives are traded, and every requirement should be checked for physical possibility before it is accepted.

Functional decomposition

Now, before naming any parts, describe what the machine must do. Start with the black box: draw a rectangle labeled with the overall function, and list what crosses its boundary. Three kinds of thing flow through any machine: energy, material, and signal. For the cart, energy enters as electrical energy from a wall outlet and leaves as motion and waste heat. Material enters as cargo and an operator and leaves as the same cargo somewhere else. Signal enters as operator intent and leaves as feedback about speed and state of charge.

Then decompose the overall function into sub-functions, each stated as a verb plus a noun, deliberately avoiding any solution language. Not "battery" but "store electrical energy." Not "chain" but "transmit rotary power." That discipline is the entire point: the moment you write "battery," you have quietly eliminated supercapacitors, a swappable pack, a tethered cord, and a fuel cell, without ever considering them.

For the cart the sub-functions are: store energy, convert energy to rotation, transmit rotation to the wheels, support and locate the load, control direction, dissipate kinetic energy on demand, and convey operator intent. Seven verbs. Any machine that does those seven things is a candidate solution, whatever it looks like.

Key idea: Decompose the machine into verb-plus-noun sub-functions with no solution language, because naming a component silently deletes every alternative to it.

Concept generation and the morphological chart

Unstructured brainstorming has a well-documented failure mode: groups converge fast, anchor on the first plausible idea, and produce fewer distinct concepts than the same people working alone would have. The fixes are structural. Generate individually first and share second. Set a quota, because a quota of thirty forces you past the obvious ten. Suspend judgment explicitly and visibly. Benchmark competitors and adjacent industries on purpose. And use a morphological chart.

A morphological chart is one row per sub-function and one column per candidate means of achieving it. Concepts are then paths through the chart, picking one cell per row. Here is the cart's chart:

Sub-function	Option A	Option B	Option C	Option D
Store energy	Lead-acid pack	Lithium iron phosphate pack	Swappable modular packs
Convert to rotation	Brushed DC motor	Brushless DC motor	Two hub motors
Transmit rotation	Roller chain	Toothed belt	Enclosed gearbox	Direct drive at the hub
Support the load	Welded steel tube frame	Bolted aluminum extrusion	Formed sheet steel monocoque
Control direction	Tiller on a single front wheel	Ackermann steering, two front wheels	Differential drive, no steered wheel
Dissipate kinetic energy	Mechanical drum brakes	Hydraulic disc brakes	Regenerative plus mechanical parking brake

Count the paths: 3 x 3 x 4 x 3 x 3 x 3 = 972 distinct concepts, from a chart that took twenty minutes to build. Obviously most are nonsense; a lead-acid pack with hub motors and a monocoque is an odd animal. But the chart makes the size of the space visible, and more importantly it surfaces combinations nobody would have proposed out loud. The team's original idea, almost certainly, was chain drive plus brushed motor plus welded tube frame plus tiller, because that is what a golf cart looks like. The chart puts 971 other things on the table.

In practice you do not evaluate 972 concepts. You choose three to six that are genuinely different in kind, sketch each one properly, and take those forward. For the cart we will carry three: Concept A, welded steel tube frame with a brushless motor and two-stage chain drive to a live rear axle, tiller steering, hydraulic disc brakes. Concept B, bolted aluminum extrusion frame with two hub motors, differential drive, regenerative braking plus a mechanical parking brake. Concept C, formed sheet steel body with a brushed motor and an enclosed worm gearbox, Ackermann steering, drum brakes.

Key idea: A morphological chart turns a handful of sub-functions into hundreds of candidate concepts and forces combinations that group brainstorming reliably misses.

Common misconceptions

• Requirements come from the customer fully formed. Customers state needs. Turning needs into metrics, targets, and verification methods is engineering work, and the customer usually cannot do it.
• More concepts is just slower. Concept generation is cheap and concept correction is expensive. Widening the space early is one of the few genuinely free lunches in engineering.
• A constraint and a goal are the same thing with different wording. Constraints disqualify; objectives trade. A design that misses a constraint by one percent is not a slightly worse design, it is not a design.
• Write requirements once and freeze them. Requirements evolve as you learn, but every change must be deliberate, documented, and re-verified. Uncontrolled drift is how scope creep enters.

Recap

• Early decisions commit most of a project's cost, which makes pre-drawing hours the highest-leverage hours you will spend.
• The design process has six looping stages, and generation is deliberately kept separate from selection.
• Every requirement needs a metric, a target, and a verification method; constraints are pass or fail while objectives are traded.
• Checking a requirement against physics first (the 5 m stop needs only 3.09 m/s^2, about 0.32 g) catches impossible specifications on day one.
• Functional decomposition states sub-functions as verb plus noun with no solution language, because naming a part deletes its alternatives.
• A morphological chart of six sub-functions produced 972 candidate concepts, from which three genuinely different ones were carried forward.

Sources

1. National Aeronautics and Space Administration. (n.d.). NASA systems engineering resources. nasa.gov
2. American Society of Mechanical Engineers. (n.d.). Topics and resources: design. asme.org
3. Encyclopaedia Britannica. (n.d.). Engineering. britannica.com
4. Wikipedia. (n.d.). Morphological analysis (problem-solving). en.wikipedia.org

The big picture

You have three concepts on the table and a room full of people with opinions. The loudest engineer likes Concept B because hub motors are elegant. The purchasing manager likes Concept C because sheet steel is cheap. The shop foreman likes Concept A because he has welded tube frames for twenty years and can build one next week. All three are giving you real information wrapped in preference, and if you decide by consensus in that room you will decide by seniority.

The purpose of a selection method is not to remove judgment. Judgment is unavoidable, and pretending otherwise produces false precision. The purpose is to make the judgment explicit and inspectable: to write down what you valued, how much, and how each concept scored, so that six months later, when the design is in trouble, someone can look at the record and see exactly which assumption failed. That is a much more modest claim than "the matrix picked the winner," and it is the honest one.

Today we select, prototype, and iterate. Along the way you will meet the most useful and least comfortable result a decision matrix can give you, which is a tie that flips when you nudge the weights.

Screening first, scoring second

When you have many concepts, do not score them all. Screen them. A Pugh matrix compares every concept against one chosen datum, usually the incumbent or the most conventional option, and rates each on each criterion as better (plus), worse (minus), or the same (S). No numbers, no weights, just direction. Sum the pluses and minuses. Concepts with a lot of minuses and no compensating strengths drop out, and concepts that beat the datum on something interesting stay in even if their totals are mediocre, because a concept that is uniquely good at one thing can often be fixed on the others.

Screening is fast and deliberately coarse, which is the point: it prevents you from spending an afternoon assigning a 3.5 to something that was never going to survive. Once you are down to a handful, switch to scoring.

The weighted decision matrix

A weighted decision matrix does three things in order. First, list the criteria that actually differentiate the concepts, which is not the same as listing everything you care about; a criterion on which all concepts score identically carries no information and only dilutes the matrix. Second, assign weights summing to 1.00, and argue about the weights before you look at any scores, because arguing about them afterwards is how people reverse-engineer their preferred answer. Third, score each concept on each criterion, typically 1 to 5, with a written justification for anything scored 1 or 5.

Here is the matrix for the Atlas cart's three surviving concepts.

Criterion	Weight	A: steel tube, chain	B: extrusion, hub motors	C: sheet steel, worm gearbox
Parts cost at 20 units	0.25	4	2	3
Range and efficiency	0.20	3	5	2
Serviceability in the field	0.15	4	3	2
Manufacturability at 20 units	0.20	4	3	2
Ride and handling on rough ground	0.10	3	4	4
Braking performance	0.10	5	3	3
Weighted total	1.00	3.80	3.25	2.55

Work Concept A to see the arithmetic: 0.25 x 4 + 0.20 x 3 + 0.15 x 4 + 0.20 x 4 + 0.10 x 3 + 0.10 x 5 = 1.00 + 0.60 + 0.60 + 0.80 + 0.30 + 0.50 = 3.80. Concept B totals 3.25 and Concept C totals 2.55. A wins. Done?

No. Now do the step that separates an engineer from someone filling in a spreadsheet: test the sensitivity of the result to the weights. Those weights were guesses. Suppose the customer turns out to care much more about running cost and much less about purchase price, a common shift when a fleet buyer does the arithmetic over a five year life. Move the cost weight from 0.25 down to 0.10 and the range weight from 0.20 up to 0.35, leaving everything else alone. Recompute:

A = 0.10 x 4 + 0.35 x 3 + 0.15 x 4 + 0.20 x 4 + 0.10 x 3 + 0.10 x 5 = 0.40 + 1.05 + 0.60 + 0.80 + 0.30 + 0.50 = 3.65

B = 0.10 x 2 + 0.35 x 5 + 0.15 x 3 + 0.20 x 3 + 0.10 x 4 + 0.10 x 3 = 0.20 + 1.75 + 0.45 + 0.60 + 0.40 + 0.30 = 3.70

C = 0.10 x 3 + 0.35 x 2 + 0.15 x 2 + 0.20 x 2 + 0.10 x 4 + 0.10 x 3 = 0.30 + 0.70 + 0.30 + 0.40 + 0.40 + 0.30 = 2.40

The winner flipped. B now leads A by 0.05, which is nothing. The honest reading of this matrix is therefore not "A is the best concept." It is: C is robustly eliminated, and the choice between A and B is not an engineering question at all, it is a business question about whether this customer values purchase price or operating efficiency. That conclusion is worth far more to your manager than a confident single answer would have been, and you got it in five minutes of arithmetic.

For the rest of this course we will carry Concept A, the welded steel tube frame with a two-stage chain drive to a live rear axle, tiller steering, and hydraulic disc brakes, because a first build for a campus customer prioritizes purchase price and shop familiarity. That is a documented decision with a documented reason, which is the only kind worth making.

Key idea: A decision matrix does not make the decision; it makes your reasoning inspectable, and a sensitivity check on the weights tells you whether the result is real or noise.

Prototyping: each build answers one question

Beginners build one prototype: a small, complete, non-working version of the final thing. That is almost always the wrong prototype, because it answers no question in particular. A prototype is an experiment, and an experiment needs a hypothesis. So before you cut any material, ask: what is the riskiest thing I currently believe? Build the cheapest, ugliest object that tests it.

Prototypes come in recognizable species. A proof of concept tests whether the physics works at all, and it is allowed to be hideous. A looks-like prototype tests form, ergonomics, and human reaction, and does not need to function. A works-like prototype tests function and does not need to look like anything. An alpha or beta unit is close to the real product and goes to real users in real conditions. Combining looks-like and works-like too early is a classic waste, because you spend weeks making something pretty in order to learn something you could have learned from plywood.

Here is the cart's prototype plan, ordered by risk rather than by how the final assembly is organized.

Build	What it is	Question it answers	Rough cost
P1, the mule	Plywood deck, real motor, real 8.4 to 1 chain drive, real wheels, ballast	Does this drivetrain actually reach 20 km/h and climb 10 percent at 8 km/h, and how hot does the motor get doing it?	About 600 USD, 2 weeks
P2, the structural rig	Welded frame only, no drivetrain, loaded statically	Does the frame deflect less than the limit at 150 kg, and where does it fail at 450 kg?	About 400 USD, 1 week
P3, the alpha	Three complete carts, real bodywork and controls	Will campus staff actually use it for four weeks, and what breaks first?	About 2500 USD each, 8 weeks

Notice that P1 costs a quarter of P3 and retires the largest technical risk. That is the ordering principle: risk first, polish last. Notice also that P2 loads the frame to three times the design payload. You test to failure when you can, because a test that passes tells you only that you are somewhere above the load you applied, while a test that breaks tells you exactly where the margin ended.

Key idea: Every prototype should answer one identified question, built as cheaply as possible, with the riskiest assumption tested first and cosmetics last.

Iteration, reviews, and FMEA

P1 will disappoint you. Ours does: at full load on the ramp the motor case reaches 95 C in eleven minutes, confirming the duty-cycle warning from the very first power estimate in this course. Now you iterate, and iteration has a structure. Diagnose (is it copper loss, poor airflow, or an over-optimistic grade requirement?), choose a change, predict what the change should do quantitatively, implement it, and re-test. Skipping the prediction step is how teams end up making five changes at once and learning nothing about any of them.

Organizations gate iteration with design reviews. A preliminary design review asks whether the concept and the requirements are sound before detailed design begins. A critical design review asks whether the detailed design is ready to be built and tested. Reviews work when they are adversarial in the friendly sense: the reviewers' job is to find the problem now, at the cost of an awkward meeting, rather than later, at the cost of a recall.

The most useful structured tool for anticipating trouble is failure modes and effects analysis. For each part, list how it could fail, what happens when it does, and rate three numbers from 1 to 10: severity of the effect, likelihood of occurrence, and likelihood that you would detect it before it reached the customer. Multiply them into a risk priority number.

Take the cart's drive chain. Failure mode: the chain jumps off the sprocket. Effect: total loss of propulsion, possibly on a grade with a loaded cart behind an operator. Severity 7, occurrence 4, detection 6, so RPN = 7 x 4 x 6 = 168. With an action threshold of 100, that demands a fix. Add a chain guide and a spring tensioner: occurrence falls to 2 and detection to 3, giving RPN = 7 x 2 x 3 = 42. Severity did not change, and it never does, because severity is a property of the consequence, not of your countermeasure. The only ways to reduce severity are to redesign so the failure cannot happen or to redesign so its consequence is smaller.

One rule that experienced teams enforce: any failure mode with severity 9 or 10 gets an action regardless of its RPN. A battery pack thermal event might score severity 9, occurrence 2, detection 5, giving an RPN of 90, comfortably under the threshold. Ignoring it because of the arithmetic would be indefensible. The numbers are a prioritization aid, not a permission slip.

Key idea: FMEA ranks risks by severity times occurrence times detection, countermeasures usually move occurrence and detection but never severity, and high-severity modes get acted on whatever the arithmetic says.

Verification and validation are not the same word

Two questions look alike and are not. Verification asks: did we build the thing right? It compares the hardware against the requirements document, one line at a time, using the verification methods you wrote down in the requirements table. Validation asks: did we build the right thing? It compares the hardware against the original need, out in the world, with real users.

The cart could pass every single requirement, hitting 20.3 km/h and stopping in 4.7 m and costing 2410 US dollars, and still fail validation because grounds staff refuse to use it: the tiller forces an awkward wrist angle over a long shift, or the bed sides are too high to load a mower over, or it will not fit through the one gate everyone actually uses. Every one of those is a requirement that should have been written and was not. Validation failures are almost always requirements failures wearing a disguise, which brings the whole design process back around to where it started.

Key idea: Verification checks the design against the requirements; validation checks it against the need, and validation failures are usually missing requirements discovered late.

Common misconceptions

• The decision matrix picks the winner. It records and exposes your reasoning. When two concepts land within a few percent, the correct output is "these are tied and here is what would break the tie," not a forced ranking.
• Prototypes should look like the product. Only the last ones should. Early prototypes should look like whatever answers the question fastest, which is often plywood, tape, and off-the-shelf parts.
• A successful test proves the design. A passing test proves only that you are somewhere above the load you applied. Testing to failure is how you learn where the margin actually is.
• Lowering the risk priority number lowers the risk. Only if the countermeasure is real. Adjusting a detection score in a spreadsheet without changing the hardware or the inspection is a way of documenting risk away rather than removing it.

Recap

• Screen many concepts coarsely with a Pugh matrix against a datum, then score the survivors with weights.
• Agree the weights before seeing any scores, and always run a sensitivity check: our cart matrix gave A = 3.80, B = 3.25, C = 2.55, but a plausible reweighting flipped A and B while leaving C last.
• The useful conclusion was that C is eliminated and the A versus B choice is a business question about purchase price versus operating efficiency.
• Each prototype answers one question, riskiest first: a 600 USD plywood mule retires more risk than a 2500 USD pretty one.
• FMEA ranks failure modes by severity times occurrence times detection; the chain-derailment mode fell from RPN 168 to 42 with a guide and tensioner, and severity never moves.
• Verification compares hardware to requirements; validation compares it to the need, and validation failures are usually requirements you never wrote.

Sources

1. National Aeronautics and Space Administration. (n.d.). NASA systems engineering and design resources. nasa.gov
2. Wikipedia. (n.d.). Decision-matrix method. en.wikipedia.org
3. Wikipedia. (n.d.). Failure mode and effects analysis. en.wikipedia.org
4. American Society of Mechanical Engineers. (n.d.). Topics and resources. asme.org

The big picture

Statics stops one step short of the answer you need. Work through Engineering Mechanics: Statics (ENGR 210) and you can tell me that the tie rod in the cart's steering linkage carries 15 kN of tension. Useful. But the question the customer is actually asking is: will it break? And that question cannot be answered by a force, because a 15 kN pull is trivial for a 40 mm bar and lethal for a paper clip. Force alone is not the thing that breaks parts.

What breaks parts is force concentrated into area, and the moment you write that sentence you have invented stress. That single move, dividing internal force by the area carrying it, converts the output of statics into a number you can compare against a material property measured in a laboratory. The chain runs: external loads, then internal forces (statics), then stresses on a chosen cross section (this module), then comparison against an allowable derived from material test data. Almost all of mechanical design is that chain, run over and over on different geometry.

Today we build the chain for the simplest case, a member pulled or pushed along its axis, and we learn to read the experiment that supplies the material end of it: the tensile test. Materials Science and Engineering (ENGR 250) explains why a metal yields where it does, in terms of dislocations and grain boundaries. Here we care about something more mercenary: which numbers come off that curve, and what a designer does with each one.

Stress and the unit trick that saves your life

Cut a loaded member with an imaginary plane and look at the exposed face. The internal force distributed over that face, divided by the area of the face, is stress. When the force is perpendicular to the face we call it normal stress and write sigma = P / A. When the force lies in the plane of the face we call it shear stress and write tau = V / A. Tension is positive by convention, compression negative.

The SI unit is the pascal, one newton per square metre, which is comically small: standing on one foot puts roughly 15 kPa under your shoe. Engineering stresses live in megapascals. And here is the trick every practicing engineer uses and no textbook prints in large enough type:

1 MPa = 1 N/mm^2.

Check it: 1 MPa = 10^6 N/m^2, and one square metre is 10^6 square millimetres, so the millions cancel exactly. This means that if you keep forces in newtons and areas in square millimetres, your answer comes out in megapascals with no conversion at all. Since drawings are dimensioned in millimetres and material data is published in megapascals, this identity removes the single most common source of factor-of-a-million errors in mechanical design. Use it constantly.

Strain is the deformation response: engineering strain is epsilon = delta / L, the change in length over the original length. It is dimensionless. Because elastic strains in metals are tiny, they are often quoted in microstrain, where 1000 microstrain equals 0.001. When you pull a steel bar to its yield point you stretch it by less than two parts per thousand, which is why elastic deformation is invisible and why it still governs everything.

Key idea: Stress is internal force per unit area and strain is deformation per unit length; keeping newtons and square millimetres makes every stress come out directly in megapascals.

What the tensile test actually gives a designer

Take a round bar machined to a standard shape, grip it in a testing machine, and pull it apart slowly while recording force and elongation. Divide force by the original area and elongation by the original gauge length and you have the engineering stress-strain curve, the single most information-dense object in mechanical design. Work through it with our specimen: original diameter 12.5 mm, so A(0) = pi x 12.5^2 / 4 = 122.7 mm^2, and original gauge length 50 mm.

The elastic region. At first the curve is a straight line and the deformation is fully recoverable. Its slope is Young's modulus E, and Hooke's law says sigma = E x epsilon. Our specimen at P = 25 kN gives sigma = 25000 / 122.7 = 203.7 MPa, and the extensometer reads 0.051 mm of stretch over 50 mm, so epsilon = 0.051 / 50 = 0.00102. Then E = 203.7 / 0.00102 = 199,700 MPa, about 200 GPa. That is the textbook value for steel, and here is the crucial design fact: essentially all steels have E near 200 GPa, whether they are cheap mild steel or exotic alloy. Heat treatment changes strength enormously and stiffness almost not at all. If your part is too floppy, changing to a stronger steel will not help; you must change the geometry.

Yield. Eventually the curve bends over and deformation stops being recoverable. For steels with no sharp yield point, the convention is the 0.2 percent offset yield strength: draw a line parallel to the elastic slope starting at a strain of 0.002 and take the intersection. Our specimen yields at 42.9 kN, so S(y) = 42900 / 122.7 = 350 MPa. This is the number most machine design is built on, because for most parts permanent deformation is already failure.

Ultimate tensile strength. The curve peaks at 67.5 kN, so S(ut) = 67500 / 122.7 = 550 MPa. Past this point the specimen necks, deforming in one local band, and the engineering stress falls even though the true stress in the neck keeps climbing.

Ductility. The specimen finally breaks with a gauge length of 61 mm and a neck diameter of 8.6 mm. Percent elongation = (61 - 50) / 50 = 22 percent. Reduction of area = (122.7 - 58.1) / 122.7 = 52.6 percent, since the final area is pi x 8.6^2 / 4 = 58.1 mm^2. Ductility is not a strength number, but it is a safety number: a ductile part yields visibly and noisily before it separates, while a brittle one goes without warning.

Toughness and resilience. The area under the whole curve is toughness, energy absorbed per unit volume before fracture. The area under the elastic part alone is resilience, energy stored and returnable. A spring wants high resilience. A bumper wants high toughness. They are different properties and no single number captures both.

One more constant falls out of the same test. As the bar stretches it gets thinner, and the ratio of lateral to axial strain is Poisson's ratio nu, about 0.30 for steel. At our 0.00102 axial strain the diameter shrinks by 0.30 x 0.00102 x 12.5 = 0.0038 mm. Utterly invisible, and yet it is why pressed-in bearing races loosen when a shaft is pulled, and why a rubber block confined in a hole behaves almost like a liquid.

Key idea: One tensile test yields E, yield strength, ultimate strength, ductility, toughness, and Poisson's ratio, and the fact that E is nearly identical for all steels means stiffness problems are geometry problems.

Sizing an axial member, and then resizing it

Back to the cart's steering tie rod carrying 15 kN. Suppose the first sketch used a 10 mm round bar. Its area is pi x 10^2 / 4 = 78.5 mm^2, so sigma = 15000 / 78.5 = 191 MPa. Compare that with the 350 MPa yield and it looks fine, with a factor of safety of 350 / 191 = 1.83.

Except that we have not decided what factor of safety this part deserves. Suppose the design standard for a steered linkage on an operator-carrying vehicle sets a factor of safety of 2.5 against yield, which is not unreasonable for a part whose failure removes steering. Then the design allowable is sigma(allow) = 350 / 2.5 = 140 MPa, and our 191 MPa bar fails the check. It would not break in the parking lot; it would be an audit finding and a liability.

Resize. Required area A = P / sigma(allow) = 15000 / 140 = 107.1 mm^2. The diameter that gives it is d = sqrt(4A / pi) = sqrt(136.4) = 11.68 mm. You do not order 11.68 mm bar. You order the next standard size up, 12 mm, which has A = 113.1 mm^2 and gives sigma = 15000 / 113.1 = 132.6 MPa and a factor of safety of 350 / 132.6 = 2.64. Round up to the stock size, always. Rounding down to save four grams is how engineers end up in depositions.

While the bar is under load it also stretches. Combining Hooke's law with the definition of strain gives the axial deflection formula delta = P L / (A E). For a 12 mm rod 800 mm long: delta = 15000 x 800 / (113.1 x 200000) = 12,000,000 / 22,620,000 = 0.53 mm. In consistent millimetre and megapascal units this comes out directly in millimetres, another payoff from the N per square millimetre habit. Half a millimetre of stretch in a steering link is small but not nothing, and it is exactly the sort of compliance that makes a machine feel vague to its operator.

Key idea: Size a member by dividing the load by the design allowable, not by the yield strength, then round up to a stock size and check the deflection you just accepted.

Shear and bearing at a joint

Members fail at their connections more often than in their middles, and joints load material in two ways that tension formulas miss. Consider the cart's steering pivot: a 12 mm pin passing through a clevis so that the pin is cut by two planes, carrying 8 kN.

In double shear the load splits across two cross sections, so the shear area is 2 x pi x 12^2 / 4 = 2 x 113.1 = 226.2 mm^2 and tau = 8000 / 226.2 = 35.4 MPa. Steel's shear yield is roughly 0.577 times its tensile yield (a result from the distortion energy theory), so about 0.577 x 350 = 202 MPa, giving a factor of safety of 5.7. Comfortable.

Bearing stress is the pressure the pin exerts on the side of its hole, and it is computed on the projected area, diameter times plate thickness, not the curved contact area. Through a 6 mm plate: A(bearing) = 12 x 6 = 72 mm^2 and sigma(b) = 8000 / 72 = 111 MPa. That is well under yield, but bearing is what elongates holes into ovals over years of service, which then introduces slop, which then introduces impact loading, which then breaks the pin. Bearing stress is a wear criterion as much as a strength one.

Key idea: Joints must be checked for pin shear (on the sheared cross sections, doubled if double shear) and for bearing on the projected area, because connections fail more often than members.

Stress concentration: the hole you had to drill

Everything above assumed stress spreads evenly across the section. It does not, anywhere the geometry changes abruptly. A hole, a notch, a sharp shoulder, or a keyway forces the stress lines to crowd around it, and the local peak can be several times the average. The multiplier is the stress concentration factor K(t), a purely geometric quantity read from published charts.

Say a cart bracket is a 40 mm wide, 6 mm thick steel strap with a 10 mm bolt hole, carrying 20 kN. The net section, after the hole, is (40 - 10) x 6 = 180 mm^2, so the nominal stress there is 20000 / 180 = 111 MPa. For a hole whose diameter is a quarter of the strap width, charts give K(t) of about 2.4. The actual peak stress at the edge of the hole is therefore 2.4 x 111 = 267 MPa. Still under the 350 MPa yield, but the margin just fell from 3.2 to 1.3, and it fell because of a hole you had no choice but to drill.

Two honest qualifications. For static loading of a ductile material, the local peak yields slightly, redistributes, and does surprisingly little harm, so designers often ignore K(t) for static ductile cases. For brittle materials or fluctuating loads, K(t) matters enormously, and ignoring it is the single most common cause of fatigue failures in machinery. We will pick that thread up properly in the fatigue lesson, where a generous fillet radius stops being a cosmetic choice and becomes the difference between a part that lasts and one that does not.

Key idea: Geometry changes concentrate stress by a factor Kt that is often 2 to 3; it can be forgiven under static loads in ductile metals and never under fluctuating loads.

Common misconceptions

• A stronger steel will make my part stiffer. No. Young's modulus is about 200 GPa for essentially all steels. Strength varies by a factor of five across steels; stiffness does not vary meaningfully. Stiffness is geometry.
• Stress is a force. Stress is force per unit area. The same 15 kN is nothing in a 40 mm bar and catastrophic in a 2 mm one.
• Ultimate tensile strength is the number to design to. For most machine parts, permanent deformation is already failure, so yield strength governs. UTS matters for fracture and for some brittle or wire applications.
• The engineering stress-strain curve falling after the peak means the material is getting weaker. It means the area is shrinking in the neck while we keep dividing by the original area. True stress continues to rise until fracture.

Recap

• Stress converts the internal forces from statics into something comparable with laboratory material data: sigma = P/A for normal stress, tau = V/A for shear.
• Keep newtons and square millimetres and stresses come out in megapascals directly, since 1 MPa = 1 N/mm^2.
• Our tensile specimen (12.5 mm, 122.7 mm^2) gave E = 200 GPa, yield 350 MPa at 42.9 kN, ultimate 550 MPa at 67.5 kN, 22 percent elongation, and 52.6 percent reduction of area.
• All steels share E near 200 GPa, so a stiffness problem must be solved with geometry, never with a stronger alloy.
• Sizing the 15 kN tie rod against a 140 MPa allowable (350 / 2.5) required 107 mm^2, so the 10 mm bar was replaced by a 12 mm stock bar giving a factor of safety of 2.64 and 0.53 mm of stretch.
• Joints need pin shear checks (35.4 MPa in double shear here) and bearing checks on the projected area (111 MPa), and holes concentrate stress by factors of about 2.4 that matter most under fluctuating load.

Sources

1. OpenStax. (2016). Stress, strain, and elastic modulus. In University physics volume 1. Rice University. openstax.org
2. Wikipedia. (n.d.). Stress-strain curve. en.wikipedia.org
3. National Institute of Standards and Technology. (n.d.). Material Measurement Laboratory. nist.gov
4. Engineering LibreTexts. (n.d.). Mechanics of materials bookshelf. eng.libretexts.org

The big picture

Pull on a bar and every fibre in it carries the same stress. That was the whole story last lesson, and it is the only loading for which the story is that simple. Twist a shaft or bend a beam and the stress varies across the section: zero on one line, maximum at the surface, with everything in between. That variation is not a complication to be tolerated. It is the single most exploitable fact in mechanical design, because it means the material near the middle of a shaft or beam is nearly idle, and if you remove it you lose almost no strength while shedding a great deal of mass. Every tube, every I-beam, every bicycle frame, and every aircraft spar is a consequence of that observation.

Today we do the two remaining internal load types you will meet constantly: torque, which twists, and bending moment, which bends. For the cart that means the live rear axle carrying drive torque and the bed rails carrying cargo. Along the way we will hit a result that surprises most students the first time and that governs a large fraction of real machinery: shafts are usually sized by stiffness, not by strength.

Torsion of a circular shaft

Twist a round shaft and each cross section rotates rigidly relative to its neighbour. Plane sections stay plane, radii stay straight, and the shear strain at any point grows linearly with distance from the centre. Elastic behaviour then makes the shear stress do the same:

tau = T x rho / J, maximum at the surface where rho = c, so tau(max) = T c / J

Here T is the internal torque, rho the radial distance from the axis, and J the polar second moment of area, a purely geometric property. For a solid circle J = pi d^4 / 32. For a hollow circle J = pi (D^4 - d^4) / 32. Notice the fourth power: torsional capacity is extraordinarily sensitive to diameter. Increase a shaft's diameter by 20 percent and its J rises by a factor of 1.2^4 = 2.07.

The twist itself follows from the same derivation: phi = T L / (G J), where phi is the angle in radians over a length L and G is the shear modulus, about 79 GPa for steel. G is not independent of E; for an isotropic material G = E / (2(1 + nu)), which for steel gives 200 / (2 x 1.30) = 77 GPa, close enough to the handbook value.

Important limitation, stated honestly. These formulas are exact for circular sections and wrong for non-circular ones. Twist a square bar and its cross sections warp out of plane; the maximum stress moves to the middle of each flat face, not the corners, and you need a different set of coefficients. Open sections such as a channel or a slit tube are catastrophically bad in torsion, sometimes hundreds of times less stiff than the closed tube they were cut from. If a member must resist torque, close the section.

Key idea: Torsional shear grows linearly from zero at the axis to a maximum at the surface, and both stress and twist depend on the polar second moment J, which scales as the fourth power of diameter.

Worked example: the cart's rear axle

The cart has a live rear axle driven by the second chain stage, 700 mm long between its bearings. From the very first lesson, the wheel torque needed on the 10 percent grade was about 51 N m. Apply a shock factor for hitting a kerb under power and design to T = 60 N m. Start with a 20 mm solid steel axle and work in newtons and millimetres, so T = 60,000 N mm.

J = pi x 20^4 / 32 = pi x 160,000 / 32 = 15,708 mm^4

tau(max) = T c / J = 60,000 x 10 / 15,708 = 38.2 MPa

The shear yield strength of our steel is roughly 0.577 x 350 = 202 MPa, so the factor of safety against yielding is 202 / 38.2 = 5.3. That looks like an over-designed axle. Now check the twist:

phi = T L / (G J) = 60,000 x 700 / (79,000 x 15,708) = 42,000,000 / 1,240,932,000 = 0.0339 rad = 1.94 degrees

A widely used rule of thumb for machine shafts limits twist to about one degree per metre of length. Our 0.7 m axle is allowed roughly 0.7 degrees and is twisting 1.94. The axle passes strength by a factor of 5.3 and fails stiffness by a factor of nearly 3. This is not a quirk of our numbers. It is the normal situation for power transmission shafts, and it is why experienced designers size shafts on deflection first and check stress second.

Resize on stiffness. We need phi no greater than 0.7 degrees = 0.01222 rad, so the required J is:

J = T L / (G phi) = 42,000,000 / (79,000 x 0.01222) = 42,000,000 / 965.4 = 43,500 mm^4

and d = (32 J / pi)^(1/4) = (443,200)^(1/4) = 25.8 mm. Now round up, and here a consideration beginners miss should drive the choice: this shaft runs in ball bearings, and metric ball bearing bores come in 25, 30, 35, and 40 mm. Picking 26 or 28 mm would satisfy the arithmetic and force you into an uncommon, expensive bearing. So we go to 30 mm. Then J = pi x 30^4 / 32 = 79,522 mm^4, the twist becomes 42,000,000 / (79,000 x 79,522) = 0.00669 rad = 0.383 degrees, comfortably inside the limit, and the shear stress drops to 60,000 x 15 / 79,522 = 11.3 MPa for a strength factor of safety of 17.9. The cart's rear axle is 30 mm solid steel, and it is enormously overstrength because stiffness, and then the bearing catalog, demanded it. Every subsequent axle calculation in this course uses that 30 mm.

Could we do better with a tube? Compare our 30 mm solid (J = 79,522 mm^4, area = 707 mm^2) with a 38 mm outside diameter tube of 5 mm wall (inside diameter 28 mm). Its J = pi (38^4 - 28^4) / 32 = pi x 1,470,480 / 32 = 144,363 mm^4, and its area is pi (38^2 - 28^2) / 4 = 518 mm^2. The tube delivers 82 percent more torsional stiffness using 27 percent less material. That is the payoff for deleting the idle core, and it is why drive shafts in cars are tubes. We keep the solid bar for the cart only because a live axle must also carry wheel bearings and a keyed sprocket, and solid stock is far cheaper to machine at a build of twenty units. Recording that reasoning is part of the design.

Key idea: Power transmission shafts are usually governed by allowable twist rather than allowable stress, and hollow sections buy large stiffness gains for less material.

Bending: the flexure formula

Now load a beam sideways. Statics gives you the internal bending moment M at each station along the beam. Inside the section, the material on the concave side is compressed, the material on the convex side is stretched, and one surface in between is neither: the neutral axis, which for a symmetric elastic section passes through the centroid. Strain varies linearly with distance y from that axis, and so does stress:

sigma = M y / I, maximum at the extreme fibre where y = c, so sigma(max) = M c / I

Here I is the second moment of area about the bending axis. For a rectangle of width b and depth h, I = b h^3 / 12. For a solid circle, I = pi d^4 / 64, exactly half the polar J, which is a useful check. For a hollow rectangle, subtract the inner rectangle from the outer.

Designers usually collapse the last two terms into the section modulus Z = I / c, so that sigma = M / Z and, more usefully for design, required Z = M / sigma(allow). Section modulus is tabulated for every standard structural shape, which turns beam sizing into a lookup.

Stare at I = b h^3 / 12 for a moment, because it contains a design lesson worth more than the formula. Depth enters as a cube. Double the depth of a beam and you get eight times the second moment and four times the section modulus, for only twice the material. Width enters linearly and buys you almost nothing. That asymmetry is why beams are deep and thin rather than square, why an I-beam puts nearly all its material in the flanges far from the neutral axis, and why laying a plank flat makes a hopeless beam and standing it on edge makes a good one.

Key idea: Bending stress varies linearly from the neutral axis, sigma = M/Z, and because the second moment goes as depth cubed, material is worth far more when it sits far from the neutral axis.

Worked example: the cart's bed rail

The 1.2 m x 0.8 m cargo bed sits on two longitudinal rails of 50 x 50 x 3 mm square steel tube, simply supported over a 1.2 m span. Treat the worst case as the full 150 kg payload concentrated at midspan, shared equally, so each rail takes 75 kg, or P = 736 N.

For a centrally loaded simply supported beam, M(max) = P L / 4 = 736 x 1200 / 4 = 220,800 N mm.

The section: I = (50^4 - 44^4) / 12 = (6,250,000 - 3,748,096) / 12 = 208,492 mm^4, and c = 25 mm, so Z = 208,492 / 25 = 8340 mm^3.

sigma(max) = M / Z = 220,800 / 8340 = 26.5 MPa

Against a 350 MPa yield that is a factor of safety of 13, which sounds absurd until you remember that this is the static load. A wheel dropping off a 100 mm kerb with a loaded bed easily trebles the instantaneous force. Applying a dynamic factor of 3 gives 79.4 MPa and a factor of safety of 4.4, which is a sensible place for a structural member on a vehicle that will be abused.

Now the same mass-efficiency comparison we made for the shaft. Our tube's cross-sectional area is 50^2 - 44^2 = 564 mm^2. A solid square bar of nearly the same area would be about 24 x 24 mm (576 mm^2), with I = 24^4 / 12 = 27,648 mm^4 and Z = 27,648 / 12 = 2304 mm^3. The tube's Z of 8340 mm^3 is 3.6 times greater for the same weight of steel. Hollow sections are not a styling choice.

Key idea: Size beams with required Z = M / sigma(allow), apply a dynamic factor for real service, and prefer hollow sections, which here gave 3.6 times the section modulus at equal mass.

Transverse shear, and when to care

Bending is not the only stress in a loaded beam. The transverse shear force V also produces shear stress, distributed as tau = V Q / (I t), where Q is the first moment of the area beyond the point of interest. Unlike bending stress, transverse shear is zero at the extreme fibres and maximum at the neutral axis, exactly the opposite distribution. For a rectangular section the peak is 1.5 times the average V/A.

For our bed rail, V = P/2 = 368 N over a web area of roughly 2 x 50 x 3 = 300 mm^2, giving an average of about 1.2 MPa. Utterly negligible beside the 26.5 MPa of bending. That is the usual situation: for long slender beams, bending dominates and you may ignore transverse shear. It stops being ignorable for short deep beams, for beams made of materials weak in shear such as wood along the grain or many composites, and at bolted or welded connections where the shear has to funnel through a small area. Knowing when a term is negligible is as much a part of engineering as knowing how to compute it.

Key idea: Transverse shear peaks at the neutral axis, opposite to bending, and is usually negligible in slender metal beams but governs short deep beams, wood, composites, and connections.

Common misconceptions

• The formulas for round shafts work for square ones. They do not. Non-circular sections warp, the peak stress moves to the flat faces, and open sections such as slit tubes lose almost all their torsional stiffness.
• Passing the stress check means the shaft is adequate. Our axle passed stress by 5.3 and failed the twist limit by nearly 3. Check deflection and twist as a matter of routine, not as an afterthought.
• Wider beams are stronger beams. Depth enters as a cube and width linearly. Adding depth is roughly always the better move if the space allows it.
• Shear stress in a beam is largest where bending stress is largest. The opposite: bending peaks at the outer fibres and transverse shear peaks at the neutral axis.

Recap

• Torsion: tau(max) = Tc/J and phi = TL/(GJ), with J = pi d^4/32 for solid circles and G about 79 GPa for steel.
• The 20 mm axle carried 60 N m at 38.2 MPa (factor of safety 5.3) but twisted 1.94 degrees over 700 mm, failing a one-degree-per-metre stiffness rule.
• Resizing on stiffness required J = 43,500 mm^4 (d = 25.8 mm), and rounding up to a standard bearing bore gave the course's 30 mm rear axle: 0.383 degrees of twist and only 11.3 MPa of shear.
• A 38 x 5 mm tube would give 82 percent more torsional stiffness with 27 percent less material than the 30 mm solid bar.
• Bending: sigma = M/Z with Z = I/c, and since I goes as depth cubed, material far from the neutral axis is worth far more.
• The 50 x 50 x 3 bed rail carries 220,800 N mm at 26.5 MPa static, 79.4 MPa with a dynamic factor of 3, and beats an equal-mass solid bar by 3.6 times in section modulus.
• Transverse shear is maximum at the neutral axis and negligible for slender metal beams, but not for short deep beams, wood, composites, or connections.

Sources

1. OpenStax. (2016). Static equilibrium, elasticity, and torque. In University physics volume 1. Rice University. openstax.org
2. Wikipedia. (n.d.). Torsion (mechanics). en.wikipedia.org
3. Wikipedia. (n.d.). Euler-Bernoulli beam theory. en.wikipedia.org
4. Engineering LibreTexts. (n.d.). Mechanics of materials bookshelf. eng.libretexts.org

The big picture

A machine can be perfectly strong and completely unusable. The bed rails of our cart carry their load at 26.5 MPa against a 350 MPa yield, a factor of safety of 13, and if they sagged 40 mm under a load of tools the customer would return the cart. A machine tool whose column flexes by a tenth of a millimetre cuts parts that are out of tolerance, even though nothing in it is anywhere near yielding. Stiffness is a requirement in its own right, and it is governed by different physics from strength: by E and geometry, not by yield strength.

Worse, there is a third failure mode that does not announce itself at all. A slender member in compression does not gradually bend more and more until you notice. It stands straight, straight, straight, and then at a perfectly definite load it snaps sideways and collapses. That is buckling, and it is the most unforgiving failure in mechanics because it is sudden, it happens at stresses far below yield, and doubling the strength of the material does absolutely nothing to prevent it.

Today we handle both, and then we turn to the question that has been quietly hanging over every calculation in this module: where does a factor of safety actually come from? You have seen me use 2.5 and 3 and 4.4 without justification. That ends now.

Beam deflection

Deflection comes from integrating the curvature that bending produces along the beam. The general relationship is E I d^2y/dx^2 = M(x), and integrating it twice with the right boundary conditions gives the deflected shape. In practice, nobody integrates. Standard cases are tabulated, and real problems are built from them by superposition: because the governing equation is linear, the deflection from several loads is simply the sum of the deflections each would cause alone.

Case	Maximum deflection	Where
Simply supported, central point load P	P L^3 / (48 E I)	Midspan
Simply supported, uniform load w per unit length	5 w L^4 / (384 E I)	Midspan
Cantilever, tip point load P	P L^3 / (3 E I)	Free end
Cantilever, uniform load w per unit length	w L^4 / (8 E I)	Free end

Look at the exponent on L. Bending stress in a simply supported beam goes as L to the first power, but deflection goes as L cubed or L to the fourth. Span is therefore far more punishing for stiffness than for strength. Add 25 percent to a span and the stress rises 25 percent while the deflection rises by 1.25^3 = 95 percent. When a customer asks whether the bed can be a little longer, that is the number to have in your head.

Work our bed rail: P = 736 N, L = 1200 mm, E = 200,000 MPa, I = 208,492 mm^4.

delta = P L^3 / (48 E I) = 736 x 1200^3 / (48 x 200,000 x 208,492)

The numerator is 736 x 1.728 x 10^9 = 1.272 x 10^12, and the denominator is 9.6 x 10^6 x 208,492 = 2.002 x 10^12, so delta = 0.635 mm. Stiffness criteria are usually written as a fraction of span: L/360 is a common building-derived limit (3.33 mm here) and L/500 a tighter machine limit (2.4 mm). We are at L/1890 statically and, with the dynamic factor of 3, at 1.91 mm, which still passes L/500. The bed rail is stiff enough with room to spare.

So could we go thinner? Try 50 x 50 x 2 mm tube: I = (50^4 - 46^4)/12 = 147,712 mm^4, Z = 5908 mm^3. Stress at the dynamic load becomes 3 x 220,800 / 5908 = 112 MPa, a factor of safety of 3.1, and deflection becomes 2.69 mm at the dynamic load, inside L/360. It passes both checks and saves 32 percent of the rail mass. And we are going to keep the 3 mm wall anyway, for three reasons the calculation cannot see: 2 mm wall dents when someone drops a mower deck on the bed edge, it is harder to weld without burn-through by a shop welding manually, and the shop already stocks 3 mm. That is not a failure of analysis. It is a reminder that analysis is one input among several, and that the engineer, not the spreadsheet, decides.

Key idea: Deflection scales with span cubed or to the fourth power while stress scales linearly, so stiffness is far more sensitive to length, and stiffness limits are written as a fraction of span such as L/360.

Buckling: the failure with no warning

Load a slender column in compression and something qualitatively different happens. Below a critical load the straight configuration is stable: nudge it and it springs back. Above that load the straight configuration becomes unstable, and the smallest disturbance sends it sideways without limit. Leonhard Euler worked out the critical load in 1744:

P(cr) = pi^2 E I / (L(e))^2

Three things about that formula deserve your attention. First, the only material property in it is E. Yield strength does not appear. Switching from mild steel to an alloy three times stronger changes the buckling load by exactly nothing, which is the single most counterintuitive fact in elementary mechanics. Second, I is the smallest second moment of area of the section, because the column will buckle about its weakest axis, whatever you had in mind. Third, L(e) is the effective length, which depends on how the ends are held.

End conditions	K	Effective length	Relative capacity
Pinned both ends	1.0	L	1.0
Fixed both ends	0.5	0.5 L	4.0
Fixed one end, pinned other	0.7	0.7 L	2.0
Fixed one end, free other	2.0	2 L	0.25

That table spans a factor of sixteen in capacity for the same piece of steel. How you attach a strut matters more than what you make it from.

Worked example. The cart has an optional canopy supported by 25 mm outside diameter steel tube posts of 2 mm wall, 1.0 m long, pinned at both ends.

I = pi (25^4 - 21^4) / 64 = pi (390,625 - 194,481) / 64 = pi x 3065 = 9630 mm^4

P(cr) = pi^2 x 200,000 x 9630 / 1000^2 = 9.8696 x 1926 = 19,009 N, about 19.0 kN

Now validate the result, because Euler's formula is only true while the material is still elastic. The area is pi (25^2 - 21^2)/4 = 144.5 mm^2, so the critical stress is 19,009 / 144.5 = 131.5 MPa, well below the 350 MPa yield. Euler applies. The formal test uses the slenderness ratio L(e)/r, where r = sqrt(I/A) = sqrt(9630 / 144.5) = 8.16 mm is the radius of gyration. Our slenderness is 1000 / 8.16 = 122.5, and the transition slenderness below which Euler over-predicts is C(c) = sqrt(2 pi^2 E / S(y)) = sqrt(2 x 9.8696 x 200,000 / 350) = sqrt(11,280) = 106. Since 122.5 is greater than 106, we are safely in Euler territory. For stockier columns, below the transition, use the Johnson parabola instead, which blends smoothly into plain yielding at zero slenderness.

With a design load of about 2 kN per post the factor of safety is 9.5, which sounds excessive until you remember two things. Buckling is catastrophic and unannounced, so it earns a larger margin than yielding does. And Euler assumes a perfectly straight, perfectly centred, perfectly elastic column, none of which exists; real columns with initial crookedness or eccentric load fail below the Euler value, which is why column design curves in structural codes sit noticeably under the theoretical line. Note too that if a careless technician welds the post at the bottom and leaves the top free, K becomes 2.0, the capacity falls to 19.0 / 4 = 4.75 kN, and the factor of safety drops from 9.5 to 2.4 without anyone changing a single part.

Buckling has relatives worth knowing by name. Local buckling wrinkles a thin wall before the whole member goes, which is why very thin-walled tubes crumple rather than bow. Lateral-torsional buckling tips a deep, narrow beam sideways as it bends, which is why deep beams need lateral restraint. Shell buckling collapses thin cylinders under axial load or external pressure, as an empty drink can demonstrates when you squeeze it.

Key idea: Buckling depends on E, the smallest I, and the effective length, never on yield strength; always check slenderness to confirm Euler applies, and treat buckling with more margin because it gives no warning.

Where a factor of safety actually comes from

The factor of safety is the ratio of capacity to demand, and it is usually written on yield strength for ductile machine parts and on ultimate strength for brittle materials or where fracture is the concern. Both conventions are in use and the two give very different numbers for the same part, so always state which one you mean. A "factor of safety of 3" is meaningless without the basis.

A more honest name for it would be the ignorance factor, because every tenth of it is bought against a specific uncertainty. Five sources dominate.

• Load uncertainty. Do you know the loads well, or are you guessing? Is there shock, vibration, or a foreseeable misuse case?
• Material uncertainty. Is the stock certified and lot-tested, or bought on price? Does the property degrade with temperature, corrosion, or time?
• Geometry and manufacturing uncertainty. Will the part be made to the drawing? Weld quality, casting porosity, and tolerance stack all eat margin.
• Analysis uncertainty. How good is your model? A hand calculation on an idealized beam deserves more margin than a validated finite element model correlated to a strain gauge test.
• Consequence of failure. Does failure ruin a part, stop a line, or hurt somebody? Is the failure detectable by inspection before it becomes dangerous?

A workable structured method multiplies a factor for each cluster. Take our steering tie rod. Material: certified structural steel stock but no lot testing, 1.1. Loads: operator-induced shock, kerb strikes, foreseeable misuse, 1.5. Consequence: loss of steering on a machine an operator walks beside, 1.5. Multiply: 1.1 x 1.5 x 1.5 = 2.48, so 2.5, which is exactly the number used to size that rod in the axial loading lesson. Now do the bed rail. Material 1.1, loads 1.3 (bounded by the payload rating and already carrying a separate dynamic factor), consequence 1.1 (a sagging bed rail is a visible nuisance, not a hazard). Multiply: 1.1 x 1.3 x 1.1 = 1.57.

Notice the discipline that keeps this honest: the dynamic factor of 3 is a load factor and the 1.57 is a resistance factor, and they are counted once each. Double-counting the same uncertainty in both places is the most common way factors of safety silently inflate to 10 and designs become unnecessarily heavy and expensive. Modern structural codes make the separation explicit and call it load and resistance factor design, applying a factor greater than one to loads and a factor less than one to capacities, each calibrated statistically against real failure data.

In regulated work you do not choose at all. The ASME Boiler and Pressure Vessel Code, lifting equipment standards, and elevator codes prescribe margins, and departing from them is not creative engineering but a code violation. Where nothing is prescribed, the professional obligation is to write down which uncertainties you were buying against, so that the next engineer can reassess when the loads or the material change.

Key idea: A factor of safety is a priced-out response to specific uncertainties in load, material, manufacture, analysis, and consequence, stated on a named basis and never double-counted against a load factor.

Common misconceptions

• A stronger material resists buckling better. Euler contains only E and geometry. All steels have essentially the same E, so a high-strength alloy column buckles at the same load as a mild steel one.
• A part that passes the stress check is a good part. It may still deflect unacceptably, buckle, or vibrate. Strength, stiffness, and stability are three separate checks.
• Bigger factors of safety are always safer. They are always heavier and more expensive, and beyond a point they hide the real uncertainty rather than addressing it. A factor of 10 on an unknown load is not safety, it is a guess with a bigger number.
• The factor of safety covers dynamic loads automatically. Not unless you decided it does. Impacts should be handled as a load factor, then a resistance factor applied on top, each counted once.

Recap

• Deflection is a separate requirement from strength, computed from tabulated standard cases and combined by superposition.
• Deflection scales as L^3 or L^4 against stress's L^1, so span is far more punishing for stiffness; the bed rail deflects 0.635 mm statically, L/1890.
• A thinner 2 mm rail passed both checks and saved 32 percent of mass, but was rejected for denting, weldability, and stock availability, which analysis cannot see.
• Euler buckling gives P(cr) = pi^2 E I / L(e)^2 with no yield strength in it; the canopy post came out at 19.0 kN and 131.5 MPa critical stress.
• Slenderness 122.5 exceeded the transition value of 106, confirming Euler applies; end conditions span a factor of 16 in capacity, and a fixed-free post would fall to 4.75 kN.
• Factors of safety are built from load, material, manufacturing, analysis, and consequence uncertainty: 1.1 x 1.5 x 1.5 = 2.5 for the tie rod, 1.57 for the bed rail, with load factors counted separately.

Sources

1. OpenStax. (2016). Elasticity and plasticity. In University physics volume 1. Rice University. openstax.org
2. Wikipedia. (n.d.). Buckling. en.wikipedia.org
3. Wikipedia. (n.d.). Factor of safety. en.wikipedia.org
4. American Society of Mechanical Engineers. (n.d.). Codes and standards. asme.org

The big picture

Motors do one thing: they spin, at roughly constant speed, in one direction. Almost nothing you actually want a machine to do looks like that. You want a needle to punch down and lift up. You want a wiper to sweep an arc and come back. You want a valve to open for exactly 40 degrees of crankshaft rotation and then shut. You want 3000 rpm of motor turned into 357 rpm of wheel with eight times the torque. Every one of those is a motion transformation, and the machinery that performs them is the oldest and most distinctive part of mechanical engineering.

The study of that machinery is kinematics: motion analyzed without regard to the forces causing it. Separating motion from force is a genuinely powerful simplification. You can determine that a linkage will trace a particular path, or that a gear train will deliver exactly 8.4 to 1, purely from geometry, before knowing anything about loads or materials. Kinetics, which adds forces, comes afterwards.

Today we cover the three great families: linkages, which are rigid bars pinned together; cams, which are shaped surfaces that push followers; and gears, which are toothed wheels that trade speed for torque. The cart uses gears and chains; you will meet linkages and cams everywhere else in your career.

Degrees of freedom: does it even move?

Before analyzing a mechanism, ask whether it is a mechanism at all. A rigid bar floating in a plane has three degrees of freedom: two translations and a rotation. Pin it to another link and you remove two of them, leaving one. Counting this up over a whole assembly gives the Kutzbach criterion for planar mechanisms:

M = 3(n - 1) - 2 j1 - j2

where n is the number of links including the fixed frame, j1 is the number of one-degree-of-freedom joints (pins and sliders), and j2 the number of two-degree-of-freedom joints (such as a pin in a slot, or gear teeth in mesh).

Try three cases. A triangle of three links with three pins: M = 3(2) - 2(3) = 0. Zero mobility, so it does not move at all; it is a structure, which is precisely why trusses are triangulated. A four-bar with four links and four pins: M = 3(3) - 2(4) = 1. One degree of freedom, meaning one input determines everything, which is why four-bars are the workhorse of mechanism design. A five-bar with five links and five pins: M = 3(4) - 2(5) = 2, so it needs two inputs, which is how a two-motor drawing machine or a five-bar haptic device works. If you ever build a mechanism and find you have to hold it in two places to make it do one thing, you have accidentally built an M = 2.

Key idea: Kutzbach's count tells you before construction whether an assembly is a structure (M = 0), a controllable mechanism (M = 1), or something needing multiple inputs (M = 2 or more).

The four-bar linkage and the Grashof condition

Four links, four pins, one degree of freedom. Name the parts: the ground or frame is the fixed link, the crank is a link that may rotate fully, the rocker oscillates through a limited arc, and the coupler connects the other two and is the only link with genuinely complicated motion. Points on the coupler trace coupler curves, which can be made astonishingly close to straight lines or figure eights, and this is how mechanisms produced straight-line motion before precision machining existed.

The most useful single fact about four-bars is the Grashof condition. Let s be the shortest link, l the longest, and p and q the other two. If

s + l is less than or equal to p + q

then at least one link can rotate completely relative to the others. Which one depends on where the shortest link sits: if the shortest link is the ground, you get a double crank (both input and output rotate fully, as in a drag link); if the shortest is adjacent to the ground, you get a crank-rocker (continuous input, oscillating output, the most common arrangement); if the shortest is the coupler, you get a double rocker in which neither pivot rotates fully but the coupler does. If the inequality fails, no link fully rotates and you have a non-Grashof double rocker.

Worked example. Design a crank-rocker for a sweeping arm. Try crank 30 mm, coupler 90 mm, rocker 70 mm, ground 100 mm. The shortest is 30 and the longest is 100, so s + l = 130, while p + q = 90 + 70 = 160. Since 130 is less than 160, the linkage is Grashof, and because the shortest link is adjacent to the ground, it is a crank-rocker. A motor on the 30 mm crank will spin continuously and the 70 mm rocker will oscillate. Exactly what a wiper needs.

Now try 80, 90, 85, 100. Here s + l = 80 + 100 = 180 and p + q = 90 + 85 = 175. The inequality fails, so this is non-Grashof: no link rotates fully, and a motor bolted to any pivot will simply stall against a limit. The difference between a working wiper and an expensive paperweight is five millimetres on one link, decided by an inequality you can check in your head.

One more quality measure. The transmission angle is the angle between the coupler and the output link, and it governs how much of the coupler force actually drives the output rather than pushing uselessly into the bearing. It varies through the cycle, and good practice keeps it between roughly 40 and 140 degrees. A linkage that satisfies Grashof but passes through a transmission angle of 10 degrees will bind, wear its bearings, and feel awful, and this is one of the most common beginner mistakes in linkage design.

Key idea: Grashof's inequality plus the position of the shortest link tells you which four-bar type you have, and the transmission angle tells you whether it will work well rather than merely work.

Cams: motion you draw rather than derive

A linkage gives you the motions its geometry allows. A cam gives you any motion you like, because you simply shape the cam surface to match the displacement you want. That freedom is why cams run engine valves, automatic screw machines, and packaging equipment. The price is a contact stress problem, a wear problem, and a part that must be precisely manufactured and cannot be adjusted afterwards.

Design begins with the displacement diagram, a plot of follower position against cam angle, laid out as a sequence of rises, dwells, and returns. Then you choose a motion program for each rise, and the choice matters more than students expect. Constant velocity is the obvious choice and the worst one: the follower must go from zero to full speed instantly at the start, which is infinite acceleration and therefore infinite force. Parabolic motion fixes acceleration but leaves it discontinuous. Simple harmonic motion gives a smooth sinusoidal acceleration but a jump in jerk, the rate of change of acceleration, at the ends of the rise. Cycloidal motion has continuous acceleration and finite jerk throughout and is the standard choice for high-speed cams. Jerk is not an academic nicety: discontinuous jerk excites the follower train's natural frequencies and is what makes fast cam systems noisy and short-lived.

Worked example. A cam lifts a follower h = 12 mm over beta = 120 degrees of cam rotation while running at 300 rpm, using simple harmonic motion.

First convert: omega = 300 x 2 pi / 60 = 31.42 rad/s, and beta = 120 degrees = 2.0944 rad.

v(max) = pi h omega / (2 beta) = 3.1416 x 12 x 31.42 / (2 x 2.0944) = 1184 / 4.189 = 283 mm/s, or 0.283 m/s

a(max) = pi^2 h omega^2 / (2 beta^2) = 9.8696 x 12 x 987.2 / (2 x 4.3865) = 116,921 / 8.773 = 13,327 mm/s^2, or 13.3 m/s^2

Now the part that matters for design. If the follower and its stem weigh 0.15 kg, the inertia force at peak acceleration is F = m a = 0.15 x 13.3 = 2.0 N. The return spring must hold the follower against the cam with more than 2.0 N of force at that instant, or the follower will lift off the cam and slam back down, which is called float and destroys cams quickly. And notice the scaling: acceleration goes as omega squared. Run this same cam at 900 rpm instead of 300 and the acceleration rises ninefold to 120 m/s^2, the required spring force to 18 N, and the contact stresses with it. Cam systems are speed-limited by inertia, not by strength.

Key idea: A cam delivers any displacement program you draw, but acceleration scales with the square of speed, so the spring force needed to prevent follower float, not strength, sets the speed limit.

Gears, module, and gear trains

Gears transmit rotation between shafts at an exact, non-slipping ratio. For the ratio to be constant through the mesh, which is what keeps a gear train quiet and smooth, the tooth profiles must obey the fundamental law of gearing, and the involute curve is the profile that does so while also tolerating small errors in centre distance. That tolerance is why essentially every gear you will ever specify is involute.

In SI practice gears are sized by module, m, in millimetres, defined so that the pitch diameter is d = m N for a gear of N teeth. Two gears mesh only if they share a module. The centre distance between them is C = m (N1 + N2) / 2, and the circular pitch, the tooth-to-tooth distance along the pitch circle, is pi m. So a module 2 pair of 20 and 60 teeth has pitch diameters of 40 mm and 120 mm and sits on 80 mm centres, with 6.28 mm circular pitch. Note that inch practice uses diametral pitch, which is the reciprocal idea, and mixing the two is a classic and expensive error.

The train value for a simple pair is the ratio of tooth counts, and speed and torque trade inversely:

omega(out) / omega(in) = N(in) / N(out), and ideally T(out) / T(in) = N(out) / N(in)

A compound train puts two gears on a shared intermediate shaft so the ratios multiply. Worked example: a 20 tooth pinion drives a 60 tooth gear (3.00 to 1); on that same shaft an 18 tooth pinion drives a 72 tooth gear (4.00 to 1). The overall ratio is 3.00 x 4.00 = 12.0 to 1. Input at 1800 rpm gives 1800 / 12 = 150 rpm out. Input torque of 5 N m gives an ideal 60 N m out, but each mesh is only about 96 percent efficient, so the real output torque is 5 x 12 x 0.96 x 0.96 = 55.3 N m. Efficiency multiplies through the train just as ratio does, which is why a five-stage gearbox with 96 percent stages delivers only 0.96^5 = 82 percent.

An idler gear inserted between two others changes the direction of rotation and the centre distance but has no effect at all on the ratio, because its tooth count cancels. Students routinely and wrongly try to use an idler to change a ratio.

Other gear types trade properties. Helical gears have angled teeth so more than one tooth pair is always engaged, making them quieter and stronger than spur gears at the cost of an axial thrust load the bearings must absorb. Bevel gears turn a corner between intersecting shafts. Worm gears achieve enormous ratios such as 50 to 1 in a single compact stage and are often self-locking, meaning the output cannot back-drive the input, which is exactly what a hoist wants, but they are only 40 to 70 percent efficient because their action is mostly sliding. That efficiency penalty is precisely why Concept C, the worm gearbox cart, scored 2 out of 5 for range in our decision matrix.

Worked example: the cart's drive train. Concept A uses two roller chain stages, and a chain drive follows exactly the same ratio arithmetic as gears, using sprocket tooth counts. Stage one is 13 teeth to 39 teeth, a ratio of 3.00. Stage two is 15 teeth to 42 teeth, a ratio of 2.80. The overall reduction is 3.00 x 2.80 = 8.40 to 1.

Check it against the specification. The motor turns 3000 rpm, so the axle turns 3000 / 8.40 = 357 rpm, which is 357 x 2 pi / 60 = 37.4 rad/s. With a wheel radius of 0.15 m the road speed is v = omega r = 37.4 x 0.15 = 5.61 m/s = 20.2 km/h, satisfying the 20 plus or minus 1 km/h requirement. Now torque: on the design grade the motor produces about 7.7 N m, so with two chain stages at roughly 95 percent each (0.90 overall) the wheel torque is 7.7 x 8.40 x 0.90 = 58.2 N m, comfortably above the 51 N m the 10 percent grade demands. The ratio we chose in lesson one is confirmed by the ratio arithmetic here, which is what internal consistency in a design looks like.

One practical detail: chain sprockets have a pitch diameter of d = p / sin(180 degrees / N), where p is the chain pitch. For 12.7 mm pitch chain on the 13 tooth motor sprocket, d = 12.7 / sin(13.85 degrees) = 12.7 / 0.2393 = 53.1 mm, so the pitch radius is 26.5 mm and the chain tension at 7.7 N m of motor torque is 7700 / 26.5 = 291 N. Keep that number; it loads the motor shaft bearings and we will use it again when we design against fatigue.

Key idea: Ratios and efficiencies both multiply through a train, idlers change direction but never ratio, and the cart's 13:39 and 15:42 chain stages give 8.40 to 1, delivering 20.2 km/h and 58.2 N m of wheel torque.

Common misconceptions

• Adding an idler gear changes the ratio. It cancels out of the arithmetic entirely. Idlers change direction of rotation and bridge centre distance, nothing more.
• Any four bars pinned in a loop make a working crank-rocker. Only if Grashof's inequality holds and the shortest link is adjacent to the ground. Otherwise the motor stalls against a limit position.
• Constant velocity is the smoothest cam motion. It demands infinite acceleration at both ends. Cycloidal motion, with continuous acceleration and finite jerk, is the smooth choice.
• Gears of the same size always mesh. They mesh only if they share a module (or diametral pitch) and pressure angle. Matching diameters is not enough.

Recap

• Kutzbach's M = 3(n-1) - 2j1 - j2 distinguishes a structure (M = 0) from a one-input mechanism (M = 1); a four-bar gives M = 1.
• Grashof's s + l no greater than p + q decides whether any link fully rotates; 30, 90, 70, 100 gives 130 versus 160 and a crank-rocker, while 80, 90, 85, 100 gives 180 versus 175 and no full rotation.
• Transmission angle should stay between about 40 and 140 degrees or the linkage binds and wears even though it satisfies Grashof.
• A 12 mm simple harmonic rise over 120 degrees at 300 rpm peaks at 283 mm/s and 13.3 m/s^2, needing more than 2.0 N of spring force to keep a 0.15 kg follower on the cam.
• Cam acceleration scales with speed squared, so cams are inertia-limited rather than strength-limited.
• Module sets gear size (d = m N, C = m(N1+N2)/2); ratios and efficiencies multiply through a train, so 20:60 plus 18:72 gives 12 to 1 and 55.3 N m from 5 N m.
• The cart's 13:39 and 15:42 chains give 8.40 to 1, producing 20.2 km/h and 58.2 N m of wheel torque, with 291 N of chain tension on the motor shaft.

Sources

1. Encyclopaedia Britannica. (n.d.). Gear. britannica.com
2. Wikipedia. (n.d.). Four-bar linkage. en.wikipedia.org
3. Wikipedia. (n.d.). Cam. en.wikipedia.org
4. OpenStax. (2016). Rotational motion and angular kinematics. In University physics volume 1. Rice University. openstax.org

The big picture

Here is a fact that surprises students and that shapes the working life of a mechanical engineer: you will design comparatively few parts from scratch. Most of a machine is selected rather than designed. The bearings, bolts, springs, seals, chains, belts, retaining rings, and keys in the cart are all catalog items made by the million, and the engineering content is in choosing the right one, mounting it correctly, and understanding what will end its life.

That is not a lesser skill. A bearing selected without regard to contamination will fail in months in a machine that ran for years on paper. A bolted joint tightened by feel will loosen and fatigue. A chain drive sized correctly and lubricated badly will need replacing every season. The failures in real machinery cluster overwhelmingly in these standard components, and almost always because of how they were applied rather than how they were made.

Today we go through the main families: bearings, shafts and their features, threaded fasteners, springs, and flexible drives. For each one, the questions are the same. What load does it carry, in what direction? What ends its life? And what does the catalog actually promise?

Bearings

A bearing supports a load while permitting relative motion, and there are two great families. Plain bearings, also called bushings or journal bearings, are simply a sleeve of a low-friction or sacrificial material such as bronze, graphite-filled polymer, or a babbitt lining. They are cheap, quiet, compact, tolerant of shock and contamination, and they need either lubrication or a self-lubricating material. Rolling element bearings put hardened balls or rollers between two races, replacing sliding with rolling. They have far lower starting friction, they last predictably, and they are less forgiving of dirt, misalignment, and shock.

Within rolling elements, geometry follows the load direction. Deep groove ball bearings take mostly radial load with some axial capacity and are the default choice. Cylindrical roller bearings take much higher radial loads and essentially no axial load. Tapered roller bearings take heavy combined radial and axial load and are used in pairs, which is why they appear in vehicle wheel hubs. Thrust bearings take axial load only. Angular contact bearings take combined loads at a defined contact angle and are used in preloaded pairs in machine tool spindles.

Catalogs give each bearing a basic dynamic load rating C, defined as the load at which 90 percent of a population survives one million revolutions. From it you compute the L10 rating life, the life that 90 percent of bearings will reach:

L10 = (C / P)^p x 10^6 revolutions, with p = 3 for ball bearings and 10/3 for roller bearings

The exponent is the striking part. Life goes as the cube of the load ratio, so halving the load multiplies life by eight, and a 25 percent overload cuts life roughly in half. Bearing life is exquisitely sensitive to load, which is why an inaccurate load estimate matters far more here than in a stress calculation.

Worked example: the cart's rear axle bearings. The 30 mm axle carries about 60 percent of the 300 kg gross mass on the rear, so 1766 N split between two bearings gives 883 N each from the load. It also carries the second-stage chain sprocket. That 42 tooth sprocket on 12.7 mm pitch chain has a pitch diameter of d = 12.7 / sin(180/42 degrees) = 12.7 / 0.0747 = 169.9 mm, so a pitch radius of 85 mm, and the 58.2 N m of axle torque produces a chain pull of 58,200 / 85 = 685 N. Combining the wheel share with the portion of the chain pull reaching the nearer bearing, take a conservative design load of P = 1200 N.

A 6206 deep groove ball bearing has a 30 mm bore, 62 mm outside diameter, and a catalog C of about 19.5 kN. Then:

L10 = (19,500 / 1200)^3 x 10^6 = (16.25)^3 x 10^6 = 4291 x 10^6 = 4.29 x 10^9 revolutions

At the axle's 357 rpm that is 4.29 x 10^9 / 357 = 1.20 x 10^7 minutes, which is about 200,000 hours, or more than twenty years of continuous running. The honest interpretation is not that we have designed a twenty year bearing. It is that this bearing is not load-limited at all. The shaft diameter chose it, and what will actually end its life is a torn seal, water ingress from pressure washing, grease that dries out, or a technician hammering it onto the shaft. That reframing is the real lesson: when a rating life comes out absurdly long, stop optimizing load and start thinking about sealing, lubrication, and mounting.

Key idea: Bearing life goes as the cube of the load ratio, so it is enormously load-sensitive, but when the computed L10 is very long the bearing is dimension-limited and contamination, lubrication, and mounting decide its real life.

Shafts and the features that live on them

We sized the axle by twist in Module 2. A real shaft also has to locate everything mounted on it. Shoulders, machined steps in diameter, provide a positive axial stop and are the preferred method. Retaining rings snap into grooves for light axial location, at the cost of a groove that concentrates stress badly. Keys in keyways transmit torque between shaft and hub; a parallel key is the standard, sized by shaft diameter, and it is designed to be the sacrificial element that shears before something expensive breaks. Press fits and taper-lock bushings transmit torque by friction with no keyway at all, which is often better for fatigue.

Every one of those features is a stress concentration, and a shaft's fatigue life is usually decided at a shoulder or a keyway rather than in its plain sections. The single most valuable habit in shaft design is to specify the largest fillet radius the mating part will accept at every step. It costs nothing and it can double fatigue life, as the next lesson will show quantitatively.

Threaded fasteners and the thing nobody tells you about preload

A metric fastener is specified as M10 x 1.5, meaning a 10 mm nominal diameter and a 1.5 mm thread pitch, plus a property class such as 4.8, 8.8, 10.9, or 12.9. The class is readable: the first number times 100 is the approximate tensile strength in MPa, and the two digits multiplied give roughly the yield strength. So class 8.8 means about 800 MPa tensile and about 640 MPa yield, with a proof strength of 600 MPa.

Now the counterintuitive part. A properly designed bolted joint in tension does not carry the external load through the bolt. Tightening the bolt stretches it and squeezes the clamped members, storing a large preload. When an external tensile load is applied, most of it goes into relieving the compression in the clamped members rather than adding tension to the bolt, because the clamped members are usually much stiffer than the bolt. The bolt sees only a small fraction of the applied load, typically 10 to 30 percent. This is why preload is everything: an adequately preloaded bolt experiences a small stress fluctuation and lasts indefinitely, while a loose bolt takes the full external load cycle directly and fails in fatigue quickly. Most bolt failures are preload failures.

Worked example. Standard practice preloads a reusable fastener to about 75 percent of its proof load. For M10 class 8.8, the tensile stress area is A(t) = 58.0 mm^2 and the proof strength is 600 MPa, so:

F(i) = 0.75 x 600 x 58.0 = 26,100 N, about 26 kN

That is roughly the weight of two and a half tonnes, produced by one modest bolt. Getting it there is the problem. The usual relation is T = K d F(i), where K is a nut factor around 0.20 for plain steel threads:

T = 0.20 x 10 mm x 26,100 N = 52,200 N mm = 52.2 N m

And here is the honest caveat that changes how you specify critical joints: K is not a constant. It varies with surface finish, plating, lubrication, and reuse, and torque control typically delivers preload with a scatter of plus or minus 25 to 30 percent. Roughly 90 percent of the torque you apply goes into overcoming friction under the head and in the threads, and only about 10 percent into stretching the bolt. That is why joints that really matter, such as engine head bolts and structural connections, use turn-of-the-nut angle control, bolt stretch measurement, or load-indicating washers instead of a torque wrench.

Key idea: A preloaded bolt shields itself from most of the external load fluctuation, preload of about 75 percent of proof is standard, and torque control delivers it with 25 to 30 percent scatter because most of the torque fights friction.

Springs

A helical compression spring's rate follows from torsion of the wire wrapped into a helix:

k = G d^4 / (8 D^3 N(a))

where d is wire diameter, D is mean coil diameter, and N(a) is the number of active coils. Once again a fourth power on the wire diameter and an inverse cube on the coil diameter, so small changes in either transform the spring. The spring index C = D/d should sit between about 4 and 12; below 4 the wire is hard to coil and stresses are high on the inside of the coil, above 12 the spring is floppy and tangles in handling.

Worked example: cart suspension. Each corner carries about 75 kg, so 736 N. Try steel wire d = 4 mm, mean coil diameter D = 20 mm, 10 active coils, G = 79,000 MPa:

k = 79,000 x 4^4 / (8 x 20^3 x 10) = 79,000 x 256 / 640,000 = 20,224,000 / 640,000 = 31.6 N/mm

Static deflection is then 736 / 31.6 = 23.3 mm, a sensible amount of sag that leaves travel for bumps. The spring index is 20 / 4 = 5.0, comfortably in range. Now check the ride, because a suspension is really a vibration problem: the natural frequency is f = (1/2 pi) sqrt(k/m) with k = 31,600 N/m and m = 75 kg, giving sqrt(421) = 20.5 rad/s and f = 3.3 Hz. Passenger cars are tuned to 1.0 to 1.5 Hz, so our cart will ride noticeably harshly. That is normal and largely unavoidable for a light vehicle that must not bottom out when loaded to twice its empty mass, and it is worth telling the customer before they discover it.

Key idea: Spring rate goes as wire diameter to the fourth over coil diameter cubed; a spring is selected for rate, index, travel, and the ride frequency it produces, not for rate alone.

Belts and chains

To move power between shafts that are far apart, a flexible drive beats a gear train on cost and simplicity. Three options dominate.

	V-belt	Toothed (timing) belt	Roller chain
Power transfer	Friction, so it slips	Positive, no slip	Positive, no slip
Efficiency	About 95 percent	About 97 to 98 percent	About 97 to 98 percent
Lubrication	None	None	Required, and messy
Shock tolerance	Excellent, it slips	Poor, teeth shear	Moderate
Noise	Quiet	Moderate whine	Loud
Load capacity per width	Low	Moderate	High
Contamination tolerance	Good	Moderate	Poor, grit is abrasive

The cart uses roller chain because it must transmit high torque at low speed in a small space and tolerate a dirty outdoor environment better than a toothed belt's teeth would. The cost is lubrication and noise, both of which appeared as scores in our decision matrix.

One misconception to kill now, because it appears in workshops everywhere. Chains do not stretch. A worn chain is longer than a new one, but not because the steel yielded; the pins and bushings have worn, so each of the hundred-odd joints has grown by a few hundredths of a millimetre and the pitch has increased. That is why the standard replacement criterion is elongation, commonly 2 percent for industrial drives and less for drives with many teeth in mesh, and why running a worn chain destroys sprockets: the chain pitch no longer matches the sprocket pitch, so load piles onto one or two teeth.

Key idea: Belts slip and forgive shock while chains are positive and carry more load per width; chain elongation is pin and bushing wear rather than stretch, and a worn chain rapidly destroys sprockets.

Common misconceptions

• A bearing with a long computed life will last that long. L10 is a fatigue number under clean, well-lubricated, correctly mounted conditions. In the field, contamination and seal failure usually arrive first.
• A tight bolt is a bolt whose threads are strained to the limit. Preload is deliberately set near 75 percent of proof precisely so the bolt stays elastic and keeps clamping; the point is to clamp the joint, not to strain the bolt.
• The torque wrench sets the preload. It sets it with plus or minus 25 to 30 percent scatter, because most of the torque is spent on friction. Critical joints use angle control or stretch measurement.
• Chains stretch under load. They wear at the pins and bushings, increasing pitch. The fix is replacement at a measured elongation limit, not tightening.

Recap

• Most machine content is selected from catalogs, and application errors, not manufacturing defects, cause most standard-component failures.
• Bearing L10 = (C/P)^3 x 10^6 revolutions for balls; the cart's 6206 at P = 1200 N gives 4.29 x 10^9 revolutions, about 200,000 hours, so it is dimension-limited rather than load-limited.
• Shaft features (shoulders, keyways, retaining ring grooves) locate parts and concentrate stress; generous fillet radii are free fatigue life.
• A preloaded bolt shields itself from most external load fluctuation; M10 class 8.8 preloaded to 75 percent of proof needs 26.1 kN, requiring about 52.2 N m with a nut factor of 0.20.
• Torque control delivers preload with 25 to 30 percent scatter because roughly 90 percent of applied torque overcomes friction.
• The 4 mm wire, 20 mm coil, 10 active coil spring gives 31.6 N/mm, 23.3 mm of static sag, and a 3.3 Hz ride frequency against a car's 1.0 to 1.5 Hz.
• Belts slip and tolerate shock, chains are positive and carry more load; chain elongation is joint wear, and worn chains destroy sprockets.

Sources

1. Wikipedia. (n.d.). Rolling-element bearing. en.wikipedia.org
2. Wikipedia. (n.d.). Bolted joint. en.wikipedia.org
3. Encyclopaedia Britannica. (n.d.). Machine. britannica.com
4. Engineering ToolBox. (n.d.). Mechanical engineering resources. engineeringtoolbox.com

The big picture

Everything you have computed so far in this course asked one question: is the stress below the allowable? Answer yes and the part survives. That framework is correct, complete, and responsible for a minority of real machine failures. The majority, by most industry estimates somewhere between half and ninety percent of all mechanical failures in service, happen to parts whose stresses were always comfortably below yield, and which nonetheless broke.

The mechanism is fatigue: progressive, localized, permanent damage accumulated under fluctuating loads. A steel bar you could load to 350 MPa all day without harm will break after ten million cycles at 160 MPa. Nothing about the material changed. The load simply went up and down enough times.

The lesson was learned the hard way and repeatedly. Railway axles in the 1840s broke without warning under loads they had carried for years, and August Wohler's systematic tests in the 1860s established that a fluctuating stress has a lower limit for indefinite survival than a steady one does. In 1954 two de Havilland Comets, the world's first jet airliners, broke up in flight after cracks grew from the corners of cutouts in their pressurized fuselages, each flight cycle acting as one pressurization cycle. The popular shorthand blames square windows, which oversimplifies the investigation, but the engineering lesson is exactly right: a sharp corner in a structure loaded thousands of times is a crack waiting for permission.

Today we learn to design against it, and we will do it on the cart, whose alpha units come back from a campus trial with a cracked rear axle.

How a fatigue failure actually happens

Fatigue proceeds in three stages, and understanding them explains every design rule that follows.

Stage one, initiation. Under cyclic load, slip bands form in surface grains, extruding and intruding material until a microscopic crack exists. This happens overwhelmingly at the surface, and preferentially wherever stress is locally raised: a machining mark, a corrosion pit, an inclusion, a keyway, a sharp fillet. Almost everything you can do to improve fatigue life is an intervention at this stage.

Stage two, propagation. The crack grows a tiny amount on each cycle, opening under tension and closing again. It leaves a characteristic pattern of concentric curved lines on the fracture surface, called beach marks, that record pauses and changes in loading. This stage occupies most of the part's life and produces no visible symptom whatsoever: no deflection, no noise, no warning.

Stage three, final fracture. The remaining uncracked section eventually becomes too small to carry even a normal load, and it lets go in one instant, in a rough, crystalline-looking region quite distinct from the smooth propagation zone.

That distinction is a genuinely useful field skill. Pick up a broken shaft and look at the fracture face. A smooth region with curved beach marks, adjoining a rough final-fracture region, is a fatigue failure, and the beach marks point back toward the initiation site so you can see what started it. A wholly rough, fibrous surface with visible necking is an overload failure. A flat, bright, granular face across the whole section suggests brittle fracture. Failure investigators read fracture surfaces before they read reports, and you can learn the basics in an afternoon.

Key idea: Fatigue initiates at surface stress raisers, propagates invisibly for most of the part's life leaving beach marks, and then fractures suddenly; the fracture surface records the whole story.

The S-N curve and the corrected endurance limit

Fatigue data comes from rotating-bending tests: a polished specimen is spun under a bending load so every point sees fully reversed stress, and the number of cycles to failure is recorded against the stress amplitude. Plot stress against cycles on log axes and you get the S-N curve.

Steels have a remarkable feature: below a certain amplitude, called the endurance limit, the curve goes flat and the specimen survives essentially forever. A useful first estimate is S(e)' = 0.5 S(ut) for steels up to about 1400 MPa. Aluminium alloys have no true endurance limit; their curve keeps descending, so aluminium parts are designed for a finite life, which is one reason aircraft structures have mandated retirement lives.

But that 0.5 factor comes from a mirror-polished 7.6 mm laboratory specimen at room temperature, and your part is none of those things. Correction factors, usually called Marin factors, bring it back to reality:

S(e) = k(a) x k(b) x k(c) x k(d) x k(e) x S(e)'

Factor	Accounts for	Typical value
k(a) surface	Finish: ground, machined, hot rolled, as forged	0.85 machined, near 0.4 as forged
k(b) size	Larger sections contain more flaws in the highly stressed volume	0.86 at 30 mm
k(c) load type	Axial and torsional loading differ from rotating bending	1.0 bending, 0.85 axial
k(d) temperature	Elevated temperature reduces strength	1.0 near room temperature
k(e) reliability	Design for better than 50 percent survival	0.81 for 99 percent

For our axle steel with S(ut) = 550 MPa, machined, 30 mm diameter, in bending, at ambient temperature, designed for 99 percent reliability:

S(e)' = 0.5 x 550 = 275 MPa, and S(e) = 0.85 x 0.86 x 1.0 x 1.0 x 0.81 x 275 = 0.592 x 275 = 163 MPa

The laboratory number was 275 MPa and the number you may design to is 163 MPa. Roughly 40 percent of the strength vanished into the difference between a polished coupon and a real part, and none of it is conservatism you can argue away.

Key idea: The 0.5 times ultimate estimate applies to a polished laboratory coupon; surface, size, and reliability corrections cut it to about 163 MPa for our machined 30 mm axle.

Notches, mean stress, and the two checks

Stress concentration, forgiven under static loading in ductile metals, is unforgiving here. The full geometric factor K(t) is softened slightly by notch sensitivity q, giving the fatigue stress concentration factor:

K(f) = 1 + q (K(t) - 1)

For the sprocket keyway in the axle, K(t) is about 2.2 and q about 0.85 for this steel, so K(f) = 1 + 0.85 x 1.2 = 2.02. The keyway doubles the local alternating stress. A sharp bearing shoulder is worse still; a generous fillet is better; and this is why the shaft design rule from the previous lesson, always specify the largest fillet radius the mating part allows, is worth real money.

Second, real loads are rarely fully reversed. Decompose any cycle into a mean stress sigma(m) and an alternating amplitude sigma(a). Tensile mean stress makes fatigue worse because it holds cracks open. The modified Goodman criterion combines them:

sigma(a) / S(e) + sigma(m) / S(ut) = 1 / n

Try a bracket seeing sigma(a) = 60 MPa and sigma(m) = 80 MPa: 60/163 + 80/550 = 0.368 + 0.145 = 0.513, so n = 1.95. Adequate. Note the asymmetry the equation encodes: 80 MPa of mean stress costs much less than 80 MPa of amplitude would, but it is not free, and compressive mean stress actually helps, which is the whole basis of shot peening.

Key idea: Under fluctuating load, apply Kf to the alternating stress and check mean plus alternating stress against a Goodman line, remembering that tensile mean stress hurts and compressive mean stress helps.

Worked example: the cracked axle and Miner's rule

The cart's alpha units come back after a campus trial and one rear axle has a crack at the sprocket keyway, with textbook beach marks. Strain gauges fitted to a second cart give a load spectrum per 1000 km of typical campus service, with K(f) already folded in:

Condition	Alternating stress	Cycles per 1000 km
Normal running on paved paths	25 MPa	1.4 million
Rough ground and grass verges	170 MPa	20,000
Kerb strikes and pothole impacts	260 MPa	500

Damage accumulates by Miner's rule: each stress level uses up a fraction n(i)/N(i) of the total life, and failure is expected when the fractions sum to 1.

Start with the easy one. At 25 MPa the stress is far below the 163 MPa endurance limit, so those 1.4 million cycles do no damage at all. This is why steel's endurance limit is such a gift: the vast majority of cycles a machine experiences can be made free.

For the damaging levels, build the finite-life S-N line between 0.9 S(ut) = 495 MPa at 10^3 cycles and S(e) = 163 MPa at 10^6 cycles. In the standard form S = a N^b, a = (0.9 S(ut))^2 / S(e) = 495^2 / 163 = 1503 MPa and b = -(1/3) log(495/163) = -0.1608, so N = (S/a)^(1/b) with 1/b = -6.22.

At 170 MPa: N = (170 / 1503)^(-6.22) = (0.1131)^(-6.22) = 7.72 x 10^5 cycles

At 260 MPa: N = (260 / 1503)^(-6.22) = (0.1730)^(-6.22) = 5.50 x 10^4 cycles

Damage per 1000 km = 20,000 / 772,000 + 500 / 55,000 = 0.0259 + 0.0091 = 0.0350

Life = 1 / 0.0350 = 28.6 thousand kilometres, so about 28,600 km. Against a 30,000 km service life target, the axle fails, and it fails with no margin whatsoever on a calculation whose inputs carry factor-of-two scatter. That is a real finding requiring real action.

Now evaluate a fix. Shot peening bombards the surface with small steel shot, plastically deforming a thin surface layer and locking in compressive residual stress that must be overcome before any applied tensile stress can open a crack. It typically raises the effective endurance limit of a machined steel part by around ten percent; take S(e) from 163 to 180 MPa. Rebuilding the line: a = 495^2 / 180 = 1361 MPa, b = -(1/3) log(495/180) = -0.1464, and 1/b = -6.83.

At 170 MPa: N = (170 / 1361)^(-6.83) = 1.47 x 10^6 cycles. At 260 MPa: N = (260 / 1361)^(-6.83) = 8.11 x 10^4 cycles.

Damage per 1000 km = 20,000 / 1,470,000 + 500 / 81,100 = 0.0136 + 0.0062 = 0.0198, giving a life of 50,500 km.

A ten percent improvement in endurance limit produced a 77 percent improvement in life. That leverage is not a trick of our numbers; it is a direct consequence of the steep exponent on the S-N curve, and it is why surface treatments are so heavily used in fatigue-critical parts such as gears, springs, and crankshafts. Note also what shot peening did not require: no change in diameter, no change in material, no change in the drawing except a surface specification.

Key idea: Miner's rule sums damage fractions across a load spectrum, cycles below the endurance limit contribute nothing, and because the S-N curve is steep a 10 percent gain in endurance limit bought 77 percent more life.

Designing against fatigue in practice

The practical rules follow directly from the mechanism. Put the surface finish where the stress is: a ground fillet is worth more than a ground shaft. Make every fillet radius as large as the mating part permits. Eliminate sharp corners, and where a corner is unavoidable, drill a relief hole at its root to blunt it. Avoid designing keyways, holes, and shoulders at the point of maximum bending moment. Induce compressive residual stress by shot peening, rolling, nitriding, or carburizing. Protect against corrosion, because corrosion fatigue removes the endurance limit entirely, meaning a steel part in a corrosive environment has no safe stress at all. And watch for fretting at press fits and clamped joints, where micro-motion under vibration generates surface damage that behaves like a pre-existing crack.

Materials Science and Engineering (ENGR 250) treats the fracture mechanics side, including crack growth rates and critical crack size, which is what lets aircraft operators run damage-tolerant inspection intervals rather than retiring parts on a calendar. The two views are complementary: the stress-life approach here answers "how long until a crack starts," and fracture mechanics answers "how long do I have once one has."

Key idea: Because fatigue begins at surface stress raisers, nearly every effective countermeasure is geometric or a surface treatment rather than a change of material.

Common misconceptions

• A part that never exceeds yield is safe. Fatigue routinely breaks parts at a third of yield stress, given enough cycles.
• Fatigue gives warning through visible deflection or noise. The propagation stage produces no symptom at all. The first symptom is usually complete fracture.
• Choosing a stronger steel fixes a fatigue problem. Sometimes a little, since Se scales with Sut, but stronger steels are also more notch-sensitive. Geometry and surface condition usually dominate, and they are cheaper to change.
• All metals have an endurance limit. Steels largely do; aluminium alloys do not, which is why aluminium structures are designed to a finite life with mandated inspection or retirement.

Recap

• Fatigue causes most machine failures, at stresses well below yield, and proceeds by surface initiation, invisible propagation with beach marks, and sudden final fracture.
• The endurance limit estimate 0.5 Sut applies to a polished coupon; Marin corrections for surface, size, and reliability took our axle from 275 MPa to 163 MPa.
• Notch sensitivity converts Kt into Kf = 1 + q(Kt - 1), giving 2.02 for the axle keyway, and Goodman combines mean and alternating stress.
• The load spectrum showed 1.4 million cycles at 25 MPa doing zero damage because they fall below the endurance limit.
• Miner's rule gave damage of 0.0350 per 1000 km and a life of 28,600 km, failing the 30,000 km target with no margin.
• Shot peening raised Se from 163 to 180 MPa and extended life to 50,500 km, a 77 percent gain from a 10 percent change, because the S-N curve is steep.
• Practical countermeasures are geometric and surface-based: large fillets, good finish where stress is high, compressive residual stress, corrosion protection, and avoiding notches at peak moment.

Sources

1. Wikipedia. (n.d.). Fatigue (material). en.wikipedia.org
2. Wikipedia. (n.d.). De Havilland Comet. en.wikipedia.org
3. National Institute of Standards and Technology. (n.d.). Material Measurement Laboratory. nist.gov
4. National Aeronautics and Space Administration. (n.d.). NASA structures and materials research. nasa.gov

The big picture

Roughly half of what mechanical engineers do involves a fluid. Water through pipes, air through ducts, oil through hydraulics, refrigerant through coils, fuel through injectors, blood through a pump, coolant through an engine block. The cart alone has hydraulic brake fluid, cooling air over the motor, and air resisting its motion. Fluid mechanics is not an optional specialization; it is a second language every mechanical engineer needs.

It divides cleanly into two halves. Fluid statics deals with fluids at rest, where nothing moves and the only thing that varies is pressure with depth. It is straightforward, exact, and underpins hydraulics, buoyancy, and every pressure vessel. Fluid dynamics deals with fluids in motion, where the governing equations are among the hardest in physics and where engineers survive by using well-chosen simplifications and honest empirical data.

Today covers statics and the first, idealized layer of dynamics: conservation of mass, which is exact, and Bernoulli's equation, which is exact under assumptions that are frequently violated. Next lesson we add the effects Bernoulli ignores, which is where real pumps and ducts get sized. Engineering Thermodynamics (ENGR 230) develops the general energy equation these results descend from; here we use the mechanical form directly.

Pressure in a fluid at rest

In a static fluid, pressure at a point acts equally in all directions and increases with depth for one simple reason: deeper fluid supports the weight of everything above it. That gives the single most useful equation in fluid statics:

P = rho g h

where h is depth below the free surface. Note what is absent: the shape of the container and the total amount of fluid. Pressure at the bottom of a narrow tube 3 m tall is the same as at the bottom of a swimming pool 3 m deep, a result known as the hydrostatic paradox that repeatedly startles people who have not derived it.

Worked example. Water has a density of about 998 kg/m^3. At 3 m depth: P = 998 x 9.81 x 3 = 29,371 Pa, about 29.4 kPa gauge. Add atmospheric pressure of roughly 101 kPa and the absolute pressure is 130 kPa. Keep the distinction sharp: gauge pressure is what a tyre gauge or a manometer reads, absolute pressure is what thermodynamic property tables demand, and confusing them is a classic error.

Notice a useful benchmark: about 10 m of water is one atmosphere. That is why a suction pump cannot lift water more than roughly 10 m no matter how good it is, and why divers speak of pressure in units of depth.

Pascal's principle states that a pressure change applied to an enclosed fluid transmits undiminished throughout. That is the whole basis of hydraulics, and the leverage it provides is startling.

Worked example: the cart's brakes. Concept A uses hydraulic disc brakes, and we can size the whole system from the stopping requirement. The cart must decelerate 300 kg at 3.09 m/s^2, so the tyres need F = 300 x 3.09 = 927 N of braking force. At a wheel radius of 0.15 m, the total brake torque is 927 x 0.15 = 139 N m, and with discs on both rear wheels each disc must produce 69.5 N m. With pads acting at an effective radius of 0.08 m, the friction force per disc is 69.5 / 0.08 = 869 N. A pad on each side gives two friction faces, so with a friction coefficient of 0.4 the clamping force must be N = 869 / (2 x 0.4) = 1086 N.

A floating caliper with a single 32 mm piston has an area of pi x 32^2 / 4 = 804 mm^2, so the required line pressure is 1086 / 804 = 1.35 MPa. A 16 mm master cylinder has an area of 201 mm^2, so the force on the master cylinder piston is 1.35 x 201 = 272 N, and through a 4 to 1 lever the operator applies about 68 N at the handle. That is a light, comfortable squeeze, and we have just sized an entire brake system from a requirement written in Module 1.

Hydraulics multiplies force, never energy. The master cylinder piston must travel four times as far as the caliper piston does, in the ratio of the areas, and the product of force and distance is conserved. Any hydraulic system that appeared to create energy would be a perpetual motion machine, and thermodynamics has opinions about those.

Key idea: Hydrostatic pressure depends only on depth, not container shape, and Pascal's principle multiplies force by the ratio of piston areas while conserving work.

Buoyancy

Archimedes' principle says a submerged or floating body experiences an upward force equal to the weight of the fluid it displaces. The derivation is pure hydrostatics: pressure is greater on the bottom face than the top, and the difference, integrated over the body, comes out to exactly the displaced weight.

F(b) = rho(fluid) x g x V(displaced)

Worked example. A sealed foam-filled float measures 0.2 x 0.2 x 0.3 m, so V = 0.012 m^3. Fully submerged in water: F(b) = 998 x 9.81 x 0.012 = 117.5 N, which will support a mass of 117.5 / 9.81 = 12.0 kg including the float itself. Multiply that by however many floats a dock needs and you have designed a pontoon.

A floating body sinks until the displaced weight equals its own weight, which is why a steel ship floats: the hull encloses a large volume of air, so the average density of the vessel is well below that of water. Whether it floats upright is a separate and harder question of stability, governed by the metacentre, and it is why naval architecture is its own discipline.

Key idea: Buoyant force equals the weight of displaced fluid, so flotation depends on average density, while remaining upright is a separate stability problem.

Continuity: mass is conserved

Draw a control volume around a length of pipe. In steady flow, whatever mass enters must leave:

rho1 A1 v1 = rho2 A2 v2

For liquids, and for gases below about Mach 0.3 where density changes stay under a few percent, density cancels and this reduces to A1 v1 = A2 v2 = Q, the volumetric flow rate. Squeeze the area and the fluid must speed up. That is the entire reason a thumb over a garden hose produces a jet.

Worked example. Water flows at 1.5 m/s through a 25 mm inside diameter pipe. The area is pi x 0.025^2 / 4 = 4.909 x 10^-4 m^2, so:

Q = A v = 4.909 x 10^-4 x 1.5 = 7.36 x 10^-4 m^3/s = 0.736 L/s, or 44.2 litres per minute

Now narrow the pipe to 15 mm. Since v2 = v1 (D1/D2)^2 = 1.5 x (25/15)^2 = 1.5 x 2.778 = 4.17 m/s. Halving the diameter would have quadrupled the velocity, because area goes as diameter squared. Keep that sensitivity in mind: pipe sizing errors punish you quadratically.

Bernoulli's equation and its fine print

Along a streamline in a steady, incompressible, frictionless flow with no shaft work, the following stays constant:

P + (1/2) rho v^2 + rho g z = constant

The three terms are static pressure, dynamic pressure, and elevation pressure, and the equation says they trade against one another. Divide through by rho g and every term becomes a length, called a head, measured in metres, which is how pump and piping engineers habitually work.

Now the fine print, because Bernoulli is the most misapplied equation in engineering. It assumes steady flow, incompressible fluid, no friction, evaluation along a single streamline, and no energy added or removed by a pump, turbine, or heat transfer. Real flows violate the frictionless assumption always and the others often. Bernoulli is best understood as the special, idealized case of the general energy equation from thermodynamics, and its correct use in engineering is over short distances where friction has not had a chance to matter, as an anchor for measurement devices, and as a first estimate to be corrected.

Worked example: a venturi. Take our pipe narrowing from 25 mm to 15 mm at constant elevation, with v1 = 1.5 m/s and v2 = 4.17 m/s. The pressure drop is:

P1 - P2 = (rho / 2)(v2^2 - v1^2) = 499 x (17.39 - 2.25) = 499 x 15.14 = 7555 Pa, about 7.6 kPa

Measure that pressure difference and you can infer the flow rate, which is precisely how venturi meters, orifice plates, and carburettors work.

Worked example: draining a tank. A tank of water 3 m deep has a small hole near the bottom. At the free surface the velocity is negligible and the pressure is atmospheric; at the hole the pressure is also atmospheric. Everything cancels except the elevation term, giving Torricelli's result v = sqrt(2 g h) = sqrt(2 x 9.81 x 3) = 7.67 m/s. Notice that it is exactly the speed a body would reach falling 3 m, which is Bernoulli telling you it is really an energy statement in disguise.

Worked example: a Pitot tube on the cart. Bring moving air to rest against a forward-facing tube and the dynamic pressure converts to a measurable static rise: P(dynamic) = 0.5 rho v^2 = 0.5 x 1.2 x 5.56^2 = 18.5 Pa. That is about 0.02 percent of atmospheric pressure, and it explains two things at once: why low-speed airspeed measurement demands a sensitive transducer, and why aerodynamic forces on our cart were only 10 N in Module 1.

A misconception worth killing. You have probably been told that an aircraft wing works because air travelling over the longer upper surface must "meet up" with air going under, therefore travels faster, therefore has lower pressure. The equal transit time premise is simply false: air over the top arrives sooner, not simultaneously, as flow visualization shows plainly. Bernoulli is not wrong here; the faster air over the top really does have lower pressure. The wrong part is the reason offered for why it is faster, which involves circulation and the flow turning around the airfoil. NASA maintains explanatory material specifically to correct this, and it is worth reading, because carrying a wrong mechanism in your head eventually produces a wrong design.

Key idea: Bernoulli trades static, dynamic, and elevation pressure along a streamline under strict assumptions, and it is a useful idealization rather than a general law of flow.

Common misconceptions

• Pressure at the bottom of a container depends on how much fluid it holds. It depends only on depth. A thin tube and a wide pool of the same height give identical bottom pressure.
• Hydraulics multiplies energy. It multiplies force and divides distance in the same ratio. Work in equals work out, minus losses.
• Bernoulli applies to any flow. It requires steady, incompressible, frictionless flow along a streamline with no shaft work. Long pipes, viscous fluids, and pumps all break it.
• Wings lift because air must travel over the top in the same time as under. Flatly false. The air over the top arrives earlier, and the real explanation involves the flow being turned by the airfoil.

Recap

• Hydrostatic pressure P = rho g h depends only on depth: 29.4 kPa gauge at 3 m of water, and roughly 10 m of water equals one atmosphere.
• Pascal's principle sized the cart's brakes end to end: 927 N of tyre force, 139 N m of torque, 1086 N of clamp, 1.35 MPa of line pressure, and 68 N at the operator's lever.
• Buoyant force equals displaced fluid weight; a 0.012 m^3 float carries 117.5 N, or 12.0 kg.
• Continuity A1 v1 = A2 v2 gave Q = 0.736 L/s in a 25 mm pipe at 1.5 m/s, and 4.17 m/s when narrowed to 15 mm.
• Bernoulli requires steady, incompressible, frictionless flow along a streamline with no shaft work; the venturi dropped 7.6 kPa and a 3 m tank drains at 7.67 m/s.
• Dynamic pressure on the cart at top speed is only 18.5 Pa, consistent with the 10 N of aerodynamic drag found in Module 1.
• The equal transit time explanation of lift is false, though the pressure-velocity relationship it invokes is not.

Sources

1. OpenStax. (2016). Fluid mechanics. In University physics volume 1. Rice University. openstax.org
2. Encyclopaedia Britannica. (n.d.). Fluid mechanics. britannica.com
3. National Aeronautics and Space Administration. (n.d.). Aerodynamics and lift resources. nasa.gov
4. Wikipedia. (n.d.). Bernoulli's principle. en.wikipedia.org

The big picture

Bernoulli said a fluid can flow forever without losing anything. Every plumber knows better. Open a tap at the top of a building and the pressure is lower than at the bottom by more than the height alone explains; run water through a hundred metres of hose and it barely trickles. The missing ingredient is viscosity, the internal friction of the fluid, and accounting for it is what turns fluid mechanics from an elegant idealization into an engineering tool that sizes actual pumps.

This lesson is where fluid mechanics starts to pay for itself commercially. Pumping and fan systems consume an enormous share of industrial and building electricity, and the U.S. Department of Energy has run programmes for decades on exactly this, because oversized pumps throttled back with a valve are one of the most common and most expensive mistakes in installed plant. By the end of this lesson you will understand why, and you will know the single equation that explains most of the savings.

Viscosity and the no-slip condition

Push one layer of fluid past another and it resists. Dynamic viscosity mu is the constant of proportionality between shear stress and velocity gradient, tau = mu (du/dy), measured in pascal seconds. Water at 20 C has mu = 1.00 x 10^-3 Pa s, air has 1.8 x 10^-5 Pa s, and a typical engine oil is around 0.2 Pa s, some two hundred times thicker than water. Kinematic viscosity nu = mu / rho, in m^2/s, appears whenever inertia and viscosity are compared; for water it is 1.00 x 10^-6 m^2/s and for air, surprisingly, about 1.5 x 10^-5 m^2/s, fifteen times larger, because air's density is so low.

Two facts about viscosity deserve a moment. First, the no-slip condition: at a solid wall, a real fluid has exactly zero velocity relative to the wall. Not nearly zero, zero. Every velocity profile in every pipe is therefore a curve running from zero at the wall to a maximum in the middle, and every pressure loss you will compute traces back to that. Second, temperature affects liquids and gases in opposite directions. Heat a liquid and its molecules escape each other's attraction more easily, so viscosity falls sharply, which is why cold engine oil is treacle and hot oil is thin. Heat a gas and its molecules cross between layers faster, transporting more momentum, so viscosity rises. Getting that backwards produces confidently wrong answers about hot-air systems.

Key idea: Viscosity is internal fluid friction, the no-slip condition pins fluid velocity to zero at every wall, and viscosity falls with temperature in liquids but rises with it in gases.

Reynolds number and flow regime

Whether viscosity dominates or inertia does is settled by a single dimensionless group, the most important number in fluid mechanics:

Re = rho v D / mu = v D / nu

It is the ratio of inertial to viscous forces. Below about Re = 2300 in a pipe the flow is laminar: smooth, orderly, parallel layers, with a parabolic velocity profile and a friction factor you can derive exactly. Above about 4000 it is turbulent: chaotic, full of eddies, with a much flatter velocity profile, far more mixing, and far more pressure loss. Between the two is a transitional band you should design to avoid, since behaviour there is unpredictable.

Almost every practical water and air flow is turbulent. Our 25 mm pipe carrying water at 1.5 m/s gives Re = 998 x 1.5 x 0.025 / 1.002 x 10^-3 = 37,350, firmly turbulent. Laminar pipe flow in engineering means very small passages, very slow flow, or very viscous fluids: lubrication films, hydraulic control orifices, blood in capillaries, polymer melts.

Darcy-Weisbach and the friction factor

Head loss along a straight pipe follows the Darcy-Weisbach equation:

h(f) = f (L / D) (v^2 / 2g)

where f is the dimensionless friction factor. For laminar flow f = 64 / Re exactly. For turbulent flow f depends on Re and on relative roughness, and is read from the Moody chart or computed from the Colebrook equation. For smooth pipes between Re of 4000 and 100,000 the Blasius correlation f = 0.316 Re^(-0.25) is accurate and easy.

Worked example. Our 25 mm pipe, 20 m long, water at 1.5 m/s, Re = 37,350.

f = 0.316 / 37,350^0.25 = 0.316 / 13.90 = 0.0227

The velocity head is v^2 / 2g = 2.25 / 19.62 = 0.1147 m, so:

h(f) = 0.0227 x (20 / 0.025) x 0.1147 = 0.0227 x 800 x 0.1147 = 2.09 m of head

which as a pressure is rho g h = 998 x 9.81 x 2.09 = 20,400 Pa, about 20.4 kPa. Note the L/D term: 20 m of 25 mm pipe is 800 diameters, and that ratio is what makes long thin pipes so costly to pump through.

Fittings add minor losses, each expressed as a multiple K of the velocity head: h = K v^2 / 2g. Typical values are 0.9 for a standard elbow, 0.2 for a fully open gate valve, 0.5 for a sharp pipe entrance, and 1.0 for the exit into a tank. Suppose our line has four elbows, one open gate valve, an entrance and an exit: sum K = 4(0.9) + 0.2 + 0.5 + 1.0 = 5.3, so the minor loss is 5.3 x 0.1147 = 0.61 m.

Total head loss is 2.09 + 0.61 = 2.70 m. Observe that the so-called minor losses are 23 percent of the total. In short, fitting-rich systems such as machine skids and building plant rooms, minor losses routinely exceed pipe friction entirely, and the name is one of engineering's more misleading pieces of vocabulary.

Key idea: Darcy-Weisbach gives pipe friction as f (L/D) times the velocity head, and fitting losses expressed as K times the velocity head are frequently not minor at all.

Sizing a pump

A pump must supply the total dynamic head: the static lift, plus all friction and minor losses, plus any pressure difference between the end reservoirs. Take our line lifting water 8 m:

H = 8 m static + 2.70 m losses = 10.70 m at Q = 0.736 L/s

The hydraulic power actually delivered to the water is:

P(hydraulic) = rho g Q H = 998 x 9.81 x 7.36 x 10^-4 x 10.70 = 77 W

Now chain the efficiencies, exactly as we did for the cart's drivetrain. A small centrifugal pump might be 60 percent efficient, so the shaft power is 77 / 0.60 = 128 W, and with an 85 percent efficient motor the electrical draw is 128 / 0.85 = 151 W. Half the input energy is lost before the water moves, which is typical for small pumps and improves with size.

Pumps come in two families. Centrifugal pumps accelerate fluid outward with an impeller; they deliver high flow at moderate head, run smoothly, and have a characteristic curve on which head falls as flow rises. Their operating point is where that pump curve intersects the system curve, the parabola of head loss versus flow for your particular pipework. Positive displacement pumps, such as gear, piston, and peristaltic types, trap a fixed volume and move it, delivering essentially constant flow regardless of pressure. They must always have a relief valve, because closing a valve downstream of one does not stop the flow, it simply raises the pressure until something bursts.

One failure mode deserves naming. Cavitation occurs when local pressure at the pump inlet falls below the fluid's vapour pressure, so bubbles form and then collapse violently as they reach higher pressure inside the impeller. The collapses are astonishingly energetic, they sound like gravel in the pump, and they erode metal from impeller surfaces. Manufacturers publish a required net positive suction head, and your installation must supply more than that, which is why pumps sit low and close to their supply and why a pump lifting from a deep sump is an installation problem rather than a pump-selection problem.

Key idea: Total dynamic head is static lift plus all losses; hydraulic power is rho g Q H, and chaining pump and motor efficiencies typically doubles the electrical input.

The affinity laws, and why variable speed drives changed everything

For a centrifugal pump or fan, changing rotational speed scales performance in a predictable way:

Q varies as N, H varies as N^2, and P varies as N^3

The cube on power is the whole story. Slow a fan to 80 percent speed and you get 80 percent of the flow, 64 percent of the pressure, and 51 percent of the power. Slow it to 50 percent and you use one eighth of the power.

Now compare the two ways of reducing flow in an installed system. The traditional method throttles with a valve or damper, which reduces flow by adding resistance: the machine still runs at full speed, and the energy you removed from the fluid is dissipated as heat in the valve. The alternative slows the machine with a variable frequency drive and follows the cube law down. For a system that spends most of its life at part load, which is nearly every building ventilation and process pumping system ever installed, the savings are large enough that the Department of Energy has built entire efficiency programmes around them. If you remember one practical result from this module, make it the cube.

Worked example: the cart's cooling fan. On the design climb the motor draws about 975 W from the battery and rejects roughly 120 W as heat. To carry that away in air while allowing only a 15 K temperature rise, the required mass flow follows from Q = m(dot) c(p) delta T:

m(dot) = 120 / (1005 x 15) = 0.00796 kg/s, and dividing by air density, V(dot) = 0.00796 / 1.2 = 0.0066 m^3/s, about 6.6 litres per second

Against maybe 50 Pa of static pressure through a guard and fin stack, the air power is 0.0066 x 50 = 0.33 W, and even at 25 percent fan efficiency the fan needs only about 1.3 W. A small computer fan does the job for a fraction of a percent of the cart's power budget, which is exactly why forced air cooling is such an easy win in electromechanical design.

Key idea: Affinity laws make power scale with the cube of speed, so slowing a pump or fan saves vastly more energy than throttling it, which merely converts the surplus into heat.

Common misconceptions

• Minor losses are minor. In fitting-rich systems they routinely exceed straight pipe friction. Ours were 23 percent, and that was a simple line.
• Viscosity rises with temperature. True for gases, false for liquids, and the liquid case is the one you meet daily in lubrication and hydraulics.
• Throttling a valve saves pump energy. It reduces flow while the pump keeps consuming nearly full power, dumping the difference as heat. Slowing the machine follows the cube law instead.
• A positive displacement pump can be throttled like a centrifugal one. Closing its discharge does not reduce flow, it raises pressure until a component fails. It must have a relief valve.

Recap

• Viscosity is internal fluid friction and the no-slip condition fixes zero velocity at every wall; liquids thin with heat while gases thicken.
• Reynolds number decides regime: laminar below 2300, turbulent above 4000. Our pipe gave Re = 37,350, firmly turbulent.
• Darcy-Weisbach with Blasius f = 0.0227 gave 2.09 m of pipe friction, and K factors summing to 5.3 added 0.61 m for a total of 2.70 m.
• With an 8 m lift the total dynamic head is 10.70 m, needing 77 W of hydraulic power and 151 W electrically after 60 percent pump and 85 percent motor efficiency.
• Cavitation from inlet pressure falling below vapour pressure erodes impellers and is prevented by supplying sufficient net positive suction head.
• Affinity laws give Q proportional to N, H to N^2, and P to N^3, so 80 percent speed uses 51 percent of the power, which is the case for variable speed drives.
• The cart's motor rejects 120 W, needing only 6.6 L/s of cooling air and about 1.3 W of fan power.

Sources

1. U.S. Department of Energy. (n.d.). Energy efficiency resources. energy.gov
2. Wikipedia. (n.d.). Darcy-Weisbach equation. en.wikipedia.org
3. Encyclopaedia Britannica. (n.d.). Fluid mechanics. britannica.com
4. Engineering ToolBox. (n.d.). Fluid flow and pressure loss resources. engineeringtoolbox.com

The big picture

Thermodynamics tells you how much energy moves and what the limits are. Engineering Thermodynamics (ENGR 230) develops all of that: the first and second laws, entropy, and the power and refrigeration cycles. What thermodynamics does not tell you is how fast. And rate is what sizes hardware. A first-law energy balance says a heat exchanger must transfer 20 kW; only heat transfer analysis tells you whether that takes 1 m^2 of surface or 30.

This is one of the largest employers of mechanical engineers on the planet. Electronics cooling, engine cooling, building HVAC, refrigeration, process heat, solar thermal, cryogenics, and thermal management of batteries are all this subject. The cart alone has three heat transfer problems: cooling the motor, cooling the controller, and getting rid of the brake energy we computed back in Module 1.

Heat moves in exactly three ways, and today we take each in turn, then combine them, then use the combination to size a heat exchanger and a cooling system.

Conduction

Conduction moves heat through matter without bulk motion, by molecular vibration and, in metals, by free electrons. Fourier's law for a plane wall:

Q = k A (delta T) / L

where k is thermal conductivity in W/m K. The range across materials is enormous, which is why material choice dominates thermal design.

Material	k (W/m K)	Material	k (W/m K)
Copper	400	Glass	1.0
Aluminium	237	Water	0.6
Carbon steel	50	Wood	0.15
Stainless steel	16	Fibreglass insulation	0.04
Concrete	1.4	Still air	0.026

Copper conducts fifteen thousand times better than fibreglass. Notice also that still air is one of the best insulators available, which is why almost every insulating material, from fibreglass to down jackets to double glazing, works by trapping air and preventing it from convecting.

The most useful reframing is the thermal resistance analogy. Write R = L / (k A) and Fourier's law becomes Q = delta T / R, which is Ohm's law with temperature playing voltage and heat flow playing current. Resistances in series add, exactly as in Electrical Circuits and Electronics (ENGR 240), and this makes composite walls trivial.

Worked example: a composite panel. The cart's battery enclosure is a 2 m^2 panel of 12 mm plywood (k = 0.13), 50 mm fibreglass (k = 0.04), and a 1 mm steel skin (k = 50), across a 30 K temperature difference.

R(plywood) = 0.012 / (0.13 x 2) = 0.0462 K/W

R(fibreglass) = 0.050 / (0.04 x 2) = 0.625 K/W

R(steel) = 0.001 / (50 x 2) = 0.00001 K/W

R(total) = 0.671 K/W, so Q = 30 / 0.671 = 44.7 W

Look at the shares. The fibreglass provides 93 percent of the resistance and the steel skin provides 0.0015 percent. The steel is structurally essential and thermally invisible. This is the general pattern in series thermal problems: the largest resistance dominates completely, so improving anything else is wasted effort. Doubling the steel thickness changes nothing; halving the insulation nearly doubles the heat loss.

Key idea: Conduction follows Q = kA delta T / L, and casting it as thermal resistances in series shows that the single largest resistance controls the result.

Convection

Convection moves heat between a surface and a moving fluid. It is really conduction into a thin boundary layer plus bulk transport away, and rather than model that in detail, engineering wraps the complexity into one empirical coefficient:

Q = h A (T(surface) - T(fluid))

The convective heat transfer coefficient h is not a material property. It depends on fluid, geometry, orientation, and above all velocity, and its range is even wider than that of conductivity.

Situation	h (W/m^2 K)
Natural convection, air	5 to 25
Forced convection, air	25 to 250
Natural convection, water	50 to 1000
Forced convection, water	100 to 15,000
Boiling or condensing	2500 to 100,000

Worked example: cooling the cart's motor. From the previous lesson the motor rejects about 120 W on the design climb, and its case has roughly 0.3 m^2 of surface. With natural convection alone, take h = 8 W/m^2 K:

delta T = Q / (h A) = 120 / (8 x 0.3) = 50 K above ambient

On a 30 C day that puts the case at 80 C, and the P1 mule test in Module 1 measured 95 C, which is entirely consistent once you allow for internal winding temperature exceeding case temperature. Now add the small fan we sized last lesson, raising h to about 40 W/m^2 K:

delta T = 120 / (40 x 0.3) = 10 K above ambient, so a 40 C case

A 1.3 W fan drops the temperature rise by a factor of five. That is why almost every piece of power electronics and every enclosed motor above a few hundred watts has forced air on it, and it is the single highest-leverage intervention in thermal design.

Key idea: Convection is wrapped into an empirical coefficient h that varies by four orders of magnitude, and forcing the fluid to move is usually the cheapest large improvement available.

Radiation

Every surface above absolute zero radiates electromagnetic energy, and unlike the other two modes this needs no medium at all. The Stefan-Boltzmann law for exchange with large surroundings:

Q = epsilon sigma A (T(s)^4 - T(surr)^4)

with sigma = 5.67 x 10^-8 W/m^2 K^4 and emissivity epsilon between 0 and 1. Temperatures must be absolute, in kelvin, and the fourth power means radiation is negligible at low temperature and overwhelming at high temperature.

Worked example. Our motor case, 0.3 m^2 at 350 K (77 C), matt painted so epsilon = 0.9, in surroundings at 293 K (20 C):

Q = 0.9 x 5.67 x 10^-8 x 0.3 x (350^4 - 293^4) = 1.531 x 10^-8 x (1.501 x 10^10 - 7.370 x 10^9) = 1.531 x 10^-8 x 7.636 x 10^9 = 117 W

That is a genuinely surprising result and worth sitting with: at 77 C, radiation from a matt black surface is comparable to natural convection, and engineers who ignore radiation in still air routinely under-predict cooling by a factor of two. Now change one thing. Leave the motor case as bare polished aluminium, epsilon = 0.05, and radiation collapses to 117 x (0.05 / 0.9) = 6.5 W. Surface finish changes radiation by a factor of eighteen and changes convection by nothing at all. Paint is thermal engineering. This is why spacecraft radiators are carefully coated, why domestic radiators are painted rather than polished, and why a shiny thermos works.

Key idea: Radiation goes as absolute temperature to the fourth power and scales directly with emissivity, so surface finish is a first-order thermal design variable even though it does nothing for convection.

Heat exchangers

A heat exchanger transfers heat between two fluid streams without mixing them. Double-pipe exchangers are the simplest; shell-and-tube dominates process industry; plate exchangers pack enormous area into small volume and are standard in food, refrigeration, and district heating; finned crossflow units are what a car radiator and an air conditioning coil are.

Combine all the resistances between the streams into one overall heat transfer coefficient U, defined so that 1/(U A) is the total resistance from hot fluid to cold, including both convective films and the tube wall. Then the exchanger duty is:

Q = U A (LMTD)

where the driving temperature difference must be the log mean temperature difference, because the difference between the streams varies along the length and the correct average is logarithmic, not arithmetic:

LMTD = (delta T1 - delta T2) / ln(delta T1 / delta T2)

Worked example. A hot stream enters at 80 C and leaves at 50 C; a cold stream enters at 20 C and leaves at 40 C; the duty is 20 kW and U is 500 W/m^2 K.

Counterflow, where the streams run in opposite directions: at one end the hot inlet (80) faces the cold outlet (40), so delta T1 = 40 K; at the other the hot outlet (50) faces the cold inlet (20), so delta T2 = 30 K.

LMTD = (40 - 30) / ln(40/30) = 10 / 0.2877 = 34.76 K, so A = 20,000 / (500 x 34.76) = 1.15 m^2

Parallel flow, both streams entering the same end: delta T1 = 80 - 20 = 60 K and delta T2 = 50 - 40 = 10 K.

LMTD = (60 - 10) / ln(6) = 50 / 1.792 = 27.90 K, so A = 20,000 / (500 x 27.90) = 1.43 m^2

Parallel flow needs 25 percent more surface for identical duty, and it has a harder limitation still: in parallel flow the outlet temperatures must converge, so the hot stream can never be cooled below the cold stream's outlet temperature. Counterflow has no such restriction and can, given enough area, cool the hot stream below the cold outlet entirely. Counterflow is therefore the default, and parallel flow is chosen only when you specifically want to limit the wall temperature or avoid thermally shocking a sensitive fluid.

One practical matter that dominates the operating life of real exchangers: fouling. Scale, biological growth, corrosion products, and particulates build up on surfaces and add thermal resistance, sometimes halving performance. Designers add a fouling allowance, which means every exchanger is deliberately oversized when clean, and operators schedule cleaning. An exchanger that meets its duty exactly on day one is an exchanger that fails its duty by month six.

Key idea: Exchanger area follows Q = U A times the log mean temperature difference; counterflow beat parallel flow by 25 percent here and has no outlet temperature restriction, and fouling allowances mean real units are deliberately oversized.

Where HVAC and refrigeration fit

Heating, ventilation, and air conditioning is one of the largest single fields of mechanical engineering employment, and it is a systems discipline built entirely on the last two lessons plus thermodynamics. The refrigeration cycle itself belongs to ENGR 230; the mechanical engineer's work in HVAC is the surrounding system.

It starts with a load calculation: how much heat enters or leaves a space, from conduction through the envelope, solar gain through glazing, ventilation air, occupants (about 100 W each), lighting, and equipment. Loads split into sensible heat, which changes temperature, and latent heat, which changes humidity, and getting the split wrong produces a system that hits temperature while leaving the space clammy.

Worked example. A 50 m^2 office with a typical combined load of 80 W/m^2 has a cooling load of 4.0 kW, which is 4000 / 3517 = 1.14 refrigeration tons. At a seasonal coefficient of performance of 3.5, the compressor draws 4000 / 3.5 = 1.14 kW of electricity to move 4 kW of heat, and the fact that the output exceeds the input is not a violation of anything, because the machine moves heat rather than creating it. To deliver the cooling with supply air 11 K below room temperature:

m(dot) = Q / (c(p) delta T) = 4000 / (1005 x 11) = 0.362 kg/s, so V(dot) = 0.362 / 1.2 = 0.302 m^3/s, about 300 litres per second

And now you can size the ductwork with the pressure loss methods from the previous lesson, select the fan, apply the affinity laws to part load operation, and choose the coil with the log mean temperature difference method above. The whole module chains together into one design.

Key idea: HVAC is a systems application of heat transfer, fluid mechanics, and thermodynamics: load calculation first, then equipment, then air and water distribution, then controls.

Common misconceptions

• Radiation only matters at red heat. At 77 C a matt surface radiated 117 W, matching natural convection. Ignoring it in still air commonly halves your predicted cooling.
• Emissivity affects all heat transfer. It affects radiation only. A polished surface convects exactly as well as a painted one and radiates eighteen times less.
• Thickening every layer improves insulation. In series, the largest resistance dominates. Our steel skin contributed 0.0015 percent, so doubling it does nothing.
• Parallel and counterflow exchangers are equivalent if the area is the same. Counterflow needed 25 percent less area here, and parallel flow cannot cool the hot stream below the cold outlet temperature at all.

Recap

• Thermodynamics gives how much energy moves; heat transfer gives how fast, and rate is what sizes hardware.
• Conduction Q = kA delta T/L cast as resistances in series showed fibreglass supplying 93 percent of a panel's resistance and steel 0.0015 percent, for 44.7 W of loss.
• Convection Q = hA delta T with h from 5 to 100,000: the cart's motor sits 50 K above ambient in still air and only 10 K above it with a 1.3 W fan.
• Radiation Q = epsilon sigma A (Ts^4 - Tsurr^4) gave 117 W from a matt case at 77 C, falling to 6.5 W if left polished.
• Exchanger sizing uses Q = U A x LMTD: 1.15 m^2 counterflow against 1.43 m^2 parallel flow for the same 20 kW duty.
• Counterflow is the default because it needs less area and can cool the hot stream below the cold outlet temperature, which parallel flow cannot.
• A 50 m^2 office at 80 W/m^2 needs 4.0 kW of cooling, 1.14 kW of compressor power at a COP of 3.5, and about 300 L/s of supply air at an 11 K temperature difference.

Sources

1. Encyclopaedia Britannica. (n.d.). Heat transfer. britannica.com
2. OpenStax. (2016). Heat and heat transfer. In University physics volume 2. Rice University. openstax.org
3. U.S. Department of Energy. (n.d.). Energy Saver: heating and cooling. energy.gov
4. Wikipedia. (n.d.). Heat exchanger. en.wikipedia.org

The big picture

A design that cannot be made is not a design. It is a wish with dimensions on it. And the choice of manufacturing process is not a downstream detail handed to somebody else after the engineering is done; it determines what geometry is possible, what tolerances you can promise, what material properties the finished part will actually have, and, above all, what it costs. A bracket that costs 60 US dollars printed, 18 machined, and 2 die cast is the same bracket. The engineering decision that separates those numbers is process selection, and it is made by the mechanical engineer.

Every manufacturing process on Earth belongs to one of five families, defined by what happens to the material. You either pour it into a shape, push it into a shape, cut a shape out of it, stick shapes together, or build a shape up in layers. Casting, forming, machining, joining, additive. Learn what each family is good and bad at and you can make sensible decisions in an unfamiliar factory on your first day.

Casting: pour it into a shape

Melt the metal, pour it into a cavity, let it solidify. Casting is the oldest process and still the only practical way to make genuinely complex three-dimensional shapes in one piece, including internal passages that no cutting tool could reach. An engine block is a casting for exactly that reason.

Sand casting uses an expendable sand mould around a reusable pattern. It handles almost any size, from a few grams to many tonnes, with cheap tooling, but gives a rough surface and loose tolerances, typically around plus or minus 0.8 mm. Investment casting, or lost wax, gives excellent detail and finish for turbine blades and surgical instruments at higher cost per part. Die casting injects molten aluminium or zinc into a steel die under pressure, producing superb tolerances near plus or minus 0.1 mm and beautiful surfaces at very high rates, but the dies cost tens of thousands and only suit non-ferrous alloys.

Three design rules follow from the physics. Metal shrinks as it solidifies, typically 1 to 2 percent, so patterns are made deliberately oversized and thick sections shrink more than thin ones, causing internal voids unless the design feeds them. Every surface must have draft, a taper of one or two degrees, or the part will not release from the mould. And uniform wall thickness matters enormously: an abrupt thick section cools slowly, creates a shrinkage cavity, and becomes the part's weak point. Castings can also contain porosity, which is why castings in critical applications get radiographed.

Forming: push it into a shape

Deform the material plastically without removing any of it. Rolling makes plate and sheet, forging hammers or presses hot metal into dies, extrusion squeezes it through an opening to make constant-section shapes such as the aluminium the cart's Concept B frame would have used, drawing pulls wire and tube, and stamping cuts and bends sheet at enormous speed.

Forming has one enormous advantage that beginners overlook: it produces grain flow that follows the part's shape. A forged crankshaft or connecting rod has its metallurgical grain wrapped continuously around the contour, giving substantially better fatigue strength than the same shape machined from bar stock, where the cutting tool severs the grain and exposes end grain at every fillet. When a part is fatigue-critical and made in quantity, forging is usually the right answer. There is also no waste: forming moves material rather than discarding it.

The classic forming difficulty is springback. Bend a sheet and the elastic portion of the strain recovers when you release it, so the part opens up a few degrees. You must overbend to land on the target angle, and the required overbend depends on the material's yield strength, thickness, and bend radius. A higher-strength steel springs back more, which is why substituting a stronger material into an existing stamping die produces parts that are suddenly the wrong shape.

Key idea: Casting makes complex shapes cheaply but needs draft, uniform walls, and shrinkage allowance; forming wastes nothing and produces favourable grain flow, but must account for springback.

Machining: cut the shape out

Remove material with a cutting tool. Turning on a lathe makes cylindrical parts, milling makes prismatic ones, drilling makes holes, grinding achieves the finest tolerances and finishes, and electrical discharge machining erodes hardened material with sparks where a cutter could not go. Machining offers the best tolerances of any process, commonly plus or minus 0.025 mm and far better with grinding, works on almost any solid material, and needs no part-specific tooling beyond fixtures.

Its weakness is that it is a rate business: you pay for time on the machine, and time comes from physics.

Worked example. Turn a 50 mm diameter steel bar over a 100 mm length. A carbide tool on steel runs at a cutting speed of about V = 150 m/min. The spindle speed follows from the surface speed at that diameter:

N = 1000 V / (pi D) = 150,000 / (pi x 50) = 150,000 / 157.08 = 955 rpm

With a feed of f = 0.2 mm per revolution, each pass advances 0.2 x 955 = 191 mm/min, so:

t = L / (f N) = 100 / 191 = 0.52 min, about 31 seconds of cutting

Now the honest part. That 31 seconds is the cutting time. The setup, workholding, tool changes, measurement, and part handling might take fifteen minutes. At low volumes, setup dominates so completely that the cutting parameters barely matter, which is why job shops quote by the hour and why reducing the number of separate setups a part requires saves far more money than optimizing feeds and speeds. Design a part that can be finished in one setup on one machine and you have done more for its cost than any amount of clever toolpath.

Joining: stick the shapes together

Welding fuses parts by melting them, with or without filler. Arc welding in its several forms covers most structural work: shielded metal arc for field repair, gas metal arc for speed and easy automation, gas tungsten arc for precision and thin material. Resistance spot welding joins sheet metal in a fraction of a second and is why car bodies exist in their current form. Friction stir welding joins aluminium below its melting point with a rotating tool and gives properties that fusion welding cannot match.

Every fusion weld creates a heat affected zone beside it, where the metal did not melt but was heated enough to change its microstructure: grain growth, loss of work hardening, sometimes hardening and embrittlement in steels. The heat affected zone, not the weld metal, is frequently the weakest part of a welded joint, and it is why welding a heat-treated part usually ruins its heat treatment locally. Materials Science and Engineering (ENGR 250) covers what happens metallurgically; the design consequence is to keep welds away from highly stressed regions and away from features that depend on hardness.

Brazing and soldering join with a filler that melts below the base metal's melting point, so the parts never melt and distortion is far lower. Adhesives distribute load over an area instead of concentrating it at fasteners, seal joints, join dissimilar materials, and add no weight, at the cost of surface preparation, cure time, and temperature limits. Mechanical fasteners are the only family that is designed to come apart again, which is a serviceability decision as much as a structural one.

Key idea: Machining buys the best tolerances but is a rate business dominated by setup at low volume, and welding's heat affected zone is often the weakest region of a joint.

Additive manufacturing

Build the part up layer by layer from a digital model. Fused deposition extrudes thermoplastic filament and is the accessible workhorse. Stereolithography cures liquid resin with light for fine detail. Selective laser sintering fuses polymer powder with no support structures needed. Powder bed fusion in metal melts metal powder with a laser or electron beam and makes genuine load-bearing metal parts, now flying in aerospace hardware.

Additive inverts the usual economics in a way worth stating precisely: geometric complexity is free, and quantity is expensive. A lattice-filled bracket with internal cooling channels costs no more to print than a solid block of the same volume, which is the opposite of every other process. But there is no economy of scale worth the name, because each part takes the same time as the last. A 40 cm^3 part at a deposition rate of 15 cm^3 per hour takes 2.7 hours, and that is 2.7 hours for the second part and the thousandth.

The honest limitations: layer-built parts are anisotropic, often substantially weaker across layers than along them, so build orientation is a load-bearing design decision. Surface finish is stepped and usually needs secondary work. Build volumes are limited. And metal powder bed parts normally require stress relief, support removal, and machining of critical surfaces, so the printed part is a starting point rather than a finished one.

Choosing a process: the break-even calculation

Here is the decision framed as arithmetic. Suppose the cart needs a motor mounting bracket, and you have three quotes: 3D printed at 60 US dollars per part with no tooling; machined from billet at 18 US dollars per part with 1500 dollars of fixturing; die cast at 2 US dollars per part with 20,000 dollars of tooling.

Process	Tooling	Per part	Total cost for N parts
Additive	0	60	60 N
Machining	1500	18	1500 + 18 N
Die casting	20,000	2	20,000 + 2 N

Additive versus machining: 60 N = 1500 + 18 N, so 42 N = 1500 and N = 36 parts.

Machining versus die casting: 1500 + 18 N = 20,000 + 2 N, so 16 N = 18,500 and N = 1156 parts.

So print below about 36, machine between 36 and roughly 1150, and die cast above that. Apply it to our own project: the Atlas cart is a build of 20 units, so printing the bracket costs 1200 dollars against 1860 for machining and 20,040 for casting. Additive wins on cost for our build volume, which is a genuinely modern answer and one that would have been absurd twenty years ago.

But do not stop at the cheapest number, because cost is one requirement among several. The bracket carries the motor, sees vibration continuously, and sits in an outdoor machine. Is a printed polymer part stiff enough, is it strong enough across the layers, will ultraviolet exposure degrade it, and how does it behave at the 80 C the motor case reaches? If any of those answers is unsatisfactory, you print in metal, or you machine it and pay the 660 dollar difference across the whole build, which is 33 dollars per cart. Stating that trade explicitly is the engineering; the arithmetic was only the setup.

Key idea: Process choice is a break-even calculation between tooling cost and per-part cost, giving additive below 36 units, machining to about 1150, and die casting beyond, with the final decision checked against properties and environment.

Tolerance by process

Each process can hold only so much precision, and asking for more than a process can deliver is the most common way drawings become expensive.

Process	Typical tolerance	Typical surface roughness Ra
Sand casting	plus or minus 0.8 mm	12 to 25 micrometres
Die casting	plus or minus 0.1 mm	1.6 micrometres
Forging	plus or minus 0.5 mm	3 to 12 micrometres
Turning and milling	plus or minus 0.025 mm	0.8 to 3.2 micrometres
Grinding	plus or minus 0.005 mm	0.1 to 0.8 micrometres
Fused deposition	plus or minus 0.2 mm	Stepped, 10 to 25 micrometres

The practical instruction is direct: specify the loosest tolerance that works, on the fewest surfaces possible. Every tightened tolerance costs money at an accelerating rate, and a part with three precision surfaces and twenty rough ones is dramatically cheaper than a part machined all over.

Key idea: Every process has a precision ceiling, and cost rises steeply as tolerances tighten, so precision should be specified only where a function actually requires it.

Common misconceptions

• Manufacturing is somebody else's job. Process choice determines geometry, properties, tolerance, and most of the cost. It is a design decision made by the designer, ideally with the shop in the room.
• Machined parts are always stronger than cast or forged ones. A forging's grain flow follows its contour, giving better fatigue strength than the same shape cut from bar, which severs the grain.
• 3D printing is always the expensive option. With zero tooling it wins at low volume; for our 20 unit build it was the cheapest of three quotes.
• Tighter tolerances are safer. They are more expensive, and they can hide the fact that nobody identified which surfaces actually matter. Tolerance every feature tightly and you have communicated nothing.

Recap

• All manufacturing falls into five families: casting, forming, machining, joining, and additive.
• Casting makes complex shapes and internal passages but needs draft, uniform walls, and 1 to 2 percent shrinkage allowance.
• Forming wastes no material and gives grain flow that follows the contour, improving fatigue strength, but suffers springback that grows with material strength.
• Turning a 50 mm bar at 150 m/min needs 955 rpm and 31 seconds of cutting over 100 mm, but setup usually dominates, so reducing setups saves more than optimizing speeds.
• Fusion welding creates a heat affected zone that is often the weakest part of the joint and locally destroys heat treatment.
• Additive makes complexity free and quantity expensive, with anisotropic properties that make build orientation a design decision.
• Break-even gave additive below 36 parts, machining to 1156, and die casting above; at the cart's 20 unit build additive was cheapest at 1200 dollars, subject to a properties check.

Sources

1. Encyclopaedia Britannica. (n.d.). Machine tool. britannica.com
2. National Institute of Standards and Technology. (n.d.). Manufacturing research and resources. nist.gov
3. Wikipedia. (n.d.). Casting (metalworking). en.wikipedia.org
4. Wikipedia. (n.d.). 3D printing. en.wikipedia.org

The big picture

You have finished the analysis. The axle is 30 mm, the bed rails are 50 x 50 x 3, the bracket will be printed. Now you must transmit all of that to somebody who was not in the room, who may be in another country, who will build it months from now, and who will build exactly what you specify rather than what you meant. The engineering drawing, or its modern equivalent the annotated three-dimensional model, is that transmission, and it is a contractual document. If the drawing says 30.0 mm with no tolerance and the shop delivers 30.4 mm, the argument that follows is settled by what the drawing said.

This lesson is about saying exactly what you mean, no more and no less. Say too little and you get parts that meet the drawing and do not work. Say too much, by tolerancing every feature tightly, and you get parts that work and cost four times what they should. Between those two failures lies a real skill, and it has a formal language: dimensioning, tolerancing, and geometric dimensioning and tolerancing. Then, having learned to specify a part precisely, we will ask the better question, which is whether the part needs to exist at all.

The drawing and the model

A traditional detail drawing uses orthographic projection: several two-dimensional views, each looking straight at the part along one axis. There are two conventions for arranging them, third angle in North America and first angle in most of Europe and Asia, and they place the views on opposite sides. Every title block carries a symbol declaring which is used, and misreading it produces a mirror-image part. It has happened to real companies.

Beyond the views, a drawing carries section views that cut through the part to show internal features, detail views that magnify small regions, a title block with part number, material, finish, scale, units, general tolerances, and approvals, and a revision block recording every change. Assembly drawings add a bill of materials and item balloons. Line types matter and are standardized: continuous thick for visible edges, dashed for hidden ones, chain-dash for centrelines.

Modern practice increasingly replaces the drawing with model-based definition, in which tolerances and notes are attached directly to the three-dimensional CAD model and the model itself is the authority. Parametric feature-based CAD, in which sketches and features carry dimensions that can be changed and propagated, is universal, and the model then feeds finite element analysis, computational fluid dynamics, machining toolpaths, and the bill of materials from one source. That single-source property is the real advantage, because the classic error of a drawing that no longer matches the model simply cannot occur.

Tolerances and fits

No dimension can be produced exactly, so every dimension needs a tolerance, either explicitly or through the general tolerance note in the title block. Tolerances are written bilaterally (30 plus or minus 0.1), unilaterally (30 plus 0.2 minus 0), or as limits (30.0 to 30.2).

When two parts must fit together, the ISO 286 system standardizes the choice. A letter designates the position of the tolerance zone and a number designates its width, with capitals for holes and lower case for shafts. Hole basis, in which the hole is H and the shaft varies, is the norm because holes are harder to adjust than shafts.

Worked example: the cart's 30 mm axle. The wheel hub must slide onto the axle for assembly, so specify a H7/g6 sliding fit. From the standard, for the 18 to 30 mm size band, H7 gives the hole a tolerance of 0 to plus 21 micrometres and g6 gives the shaft minus 7 to minus 20 micrometres:

Hole: 30.000 to 30.021 mm. Shaft: 29.993 to 29.980 mm.

Minimum clearance = 30.000 - 29.993 = 0.007 mm. Maximum clearance = 30.021 - 29.980 = 0.041 mm.

The parts always assemble and never rattle badly. Now consider the drive sprocket, which we want permanently located. Specify H7/p6, an interference fit: p6 gives the shaft plus 22 to plus 35 micrometres, so 30.022 to 30.035 mm.

Minimum interference = 30.022 - 30.021 = 0.001 mm. Maximum interference = 30.035 - 30.000 = 0.035 mm.

The sprocket must now be pressed or thermally fitted, and it will not come off in service. Two letters on a drawing decided whether a part slides on by hand or requires a press, and that is the economy of the fit system: it compresses a paragraph of intent into four characters.

Key idea: ISO fits encode intent in a few characters; H7/g6 gave 0.007 to 0.041 mm of clearance for a sliding hub, while H7/p6 gave 0.001 to 0.035 mm of interference for a permanently located sprocket.

Tolerance stack-up

Tolerances accumulate. Stack four spacers each specified at 25 plus or minus 0.1 mm and ask how long the stack is.

Worst case: assume every part is at its extreme in the same direction. The stack is 100 plus or minus 0.4 mm. This is guaranteed never to be exceeded, and it is what you use for safety-critical clearances, for interference that must never occur, and for low-volume production where you cannot rely on statistics.

Statistical, or root sum square: individual variations are random and independent, so they rarely all conspire. The combined tolerance is sqrt(0.1^2 + 0.1^2 + 0.1^2 + 0.1^2) = sqrt(0.04) = 0.2 mm, exactly half the worst case.

That factor of two is enormous in manufacturing cost, and it is why high-volume industries tolerance statistically. But the assumptions must actually hold: the processes must be centred and in control, the variations independent, and the volume high enough for statistics to mean anything. If a single supplier's tooling drifts to one edge of its tolerance band, the independence assumption evaporates and worst case is what you get. Use root sum square where you have process data and worst case where you do not.

Geometric dimensioning and tolerancing

Coordinate tolerancing, where every feature is located by plus-or-minus dimensions from an edge, has two structural weaknesses. First, it produces square tolerance zones for what are usually round features. Second, it does not say what matters: it never tells the inspector which surface to measure from, so two inspectors can measure the same part from different faces and disagree about whether it passes.

Geometric dimensioning and tolerancing, standardized as ASME Y14.5 and ISO 1101, fixes both. Its central idea is the datum reference frame: you explicitly nominate a primary, secondary, and tertiary datum, usually the surfaces that will actually locate the part in its assembly or fixture, and every geometric tolerance is measured from that frame. Ambiguity disappears, because the drawing now says how to hold the part.

Fourteen characteristics fall into five groups. Form tolerances (flatness, straightness, circularity, cylindricity) need no datum, because they describe a surface's own shape. Orientation (perpendicularity, parallelism, angularity) relates a feature to a datum's direction. Location (position, concentricity, symmetry) relates it to a datum's place. Profile (of a line, of a surface) controls a complete contour. Runout (circular, total) controls a rotating surface relative to an axis, and is the natural control for our axle.

The bonus you get for free. Take a hole located at plus or minus 0.1 mm in both x and y. Its tolerance zone is a 0.2 by 0.2 mm square of area 0.04 mm^2. But the true functional requirement is almost always radial: the hole must be within some distance of nominal in any direction. The largest circular zone that fits the square has diameter 0.2, but the correct equivalent is the circle that circumscribes the square, of diameter 0.2 x sqrt(2) = 0.283 mm, with area pi x 0.1414^2 = 0.0628 mm^2. Switching to a diametral position tolerance of 0.283 mm accepts 57 percent more parts while rejecting nothing that would actually have functioned. That is free yield, obtained purely by describing the requirement honestly.

There is a second bonus. Under maximum material condition, a hole at its smallest or a pin at its largest is the worst case for assembly; when the feature departs from that condition, the extra clearance can be added to the position tolerance. A clearance hole made slightly larger than minimum is genuinely easier to line up, and maximum material condition lets the drawing say so, converting a manufacturing variation into extra allowable position error rather than a rejection.

Key idea: Geometric tolerancing replaces square zones and implicit datums with explicit datum reference frames and functional controls, and a diametral position zone accepts 57 percent more good parts than the equivalent square.

Design for manufacture and assembly

Now the higher-leverage question. Having learned to specify a part exactly, ask whether it should exist. Design for manufacture and assembly is the systematic pursuit of that question, and its central result is that reducing part count beats almost every other cost reduction, because eliminating a part eliminates its purchase price, its handling, its insertion, its fastener, its drawing, its inspection, its inventory, its supplier, and its failure mode all at once.

The standard test asks three questions about each part relative to the ones it attaches to. Does it move relative to them in operation? Must it be a different material for a fundamental reason such as insulation or thermal isolation? Must it be separable for assembly or service? If the answer to all three is no, the part is a candidate for elimination by combining it with a neighbour.

Worked example: the cart's control panel. The first design has a steel backing plate, a plastic front bezel, a separate switch bracket, a separate label plate, four standoffs, and fourteen fasteners in three sizes. Apply the three questions and most of those parts fail all three. The redesign integrates the bezel and the switch bracket into one moulded part with the standoffs formed as bosses, replaces the label plate with a printed graphic, and drops to four parts and four fasteners.

Put numbers on it. Estimating roughly ten seconds to handle and drive each fastener and five seconds to handle and place each part, the original panel takes 14 x 10 + 11 x 5 = 195 seconds, and the redesign takes 4 x 10 + 4 x 5 = 60 seconds. That is a saving of 135 seconds per cart. Across a build of twenty carts it is 45 minutes, worth perhaps 45 US dollars in labour, which is honestly not much. Across twenty thousand carts it is 750 hours and around 45,000 dollars. Assembly savings scale with volume, and at our low volume the real wins are different and still substantial: seven fewer parts to design, draw, buy, stock, inspect, and get wrong, and three fewer fastener sizes on the assembly bench.

The remaining rules follow from watching people assemble things. Make parts symmetric so orientation does not matter, or if they must be asymmetric, make the asymmetry obvious and impossible to get wrong. Add chamfers and lead-ins so parts guide themselves into place. Design top-down assembly so gravity holds parts while hands work, rather than requiring a third hand. Avoid parts that tangle or nest in a bin, such as open springs and unflanged cups. Standardize fasteners ruthlessly, since every additional size means another tool change, another bin, and another chance to grab the wrong one. And apply mistake proofing: where a part could be fitted backwards with bad consequences, add a feature that makes the wrong way physically impossible. A connector that only mates one way costs nothing extra and prevents a whole class of field failure.

Finally, design for the rest of the life. What wears out, and can a technician reach it without removing six other things? On our cart the chain will need adjustment and eventual replacement, the brake pads will wear, and the battery will be replaced once in the machine's life. If any of those requires dismantling the frame, the design has failed a requirement nobody wrote down, which is exactly the validation failure we discussed back in Module 1.

Key idea: Part count reduction is the highest-leverage design change because it deletes purchase, handling, fastening, documentation, inspection, inventory, and failure modes together; the three-question test identifies candidates.

Common misconceptions

• Tighter tolerances everywhere make a better part. They make an expensive part and hide which surfaces actually matter. Tolerance the few features that carry function and leave the rest to the general note.
• Statistical stack-up is always the right choice. It halves the apparent tolerance only if processes are centred, in control, and independent. Safety-critical clearances and low volumes call for worst case.
• Geometric tolerancing just adds symbols to a drawing. It changes the tolerance zone shape and names the measurement datums, which typically accepts more good parts and removes inspection disputes entirely.
• Cheaper parts make a cheaper product. A part's real cost includes handling, insertion, fastening, inspection, inventory, and documentation. Deleting a part beats discounting one.

Recap

• The drawing or annotated model is a contractual document; first and third angle projection differ, and the title block declares which is used.
• ISO fits compress intent into a few characters: H7/g6 gave 0.007 to 0.041 mm clearance on the 30 mm axle, H7/p6 gave 0.001 to 0.035 mm interference.
• Four parts at 25 plus or minus 0.1 mm stack to plus or minus 0.4 mm worst case but only plus or minus 0.2 mm by root sum square, valid only when processes are centred, controlled, and independent.
• Geometric tolerancing supplies explicit datum reference frames and functional controls across form, orientation, location, profile, and runout.
• A diametral position zone of 0.283 mm accepts 57 percent more parts than an equivalent 0.2 mm square zone, and maximum material condition adds bonus tolerance.
• The three-question test (does it move, must it differ in material, must it separate) identifies parts to eliminate.
• The control panel redesign went from 11 parts and 14 fasteners to 4 and 4, cutting assembly from 195 to 60 seconds, with savings that scale strongly with volume.

Sources

1. American Society of Mechanical Engineers. (n.d.). Codes and standards. asme.org
2. National Institute of Standards and Technology. (n.d.). Physical Measurement Laboratory. nist.gov
3. Wikipedia. (n.d.). Geometric dimensioning and tolerancing. en.wikipedia.org
4. Wikipedia. (n.d.). Engineering drawing. en.wikipedia.org

The big picture

A century ago a mechanical engineer could design a complete machine without touching a wire. Today that machine is almost extinct. Your washing machine weighs its load and adjusts the cycle. Your car's engine is a mechanical device managed continuously by software. A modern excavator, a surgical robot, a semiconductor wafer handler, and a domestic heat pump are all mechanical systems whose behaviour is defined as much by their control logic as by their linkages. The field that lives at that intersection is mechatronics, and no mechanical engineer graduating now can afford to be illiterate in it.

The second half of this lesson turns to energy, because mechanical engineers design most of the machinery that consumes it and most of the machinery that produces it. And energy leads directly to sustainability, a subject that attracts a great deal of vague enthusiasm and deserves better: it deserves arithmetic. We will do the arithmetic on our own cart, and the answer will not be the one you expect.

The mechatronic loop

Every mechatronic system is the same loop. Sensors measure the physical world and turn it into signals. A controller, usually a microcontroller running fixed-interval code, compares measurement with intent and decides what to do. Actuators convert that decision back into physical action. The plant, meaning the machine itself, responds, and the sensors measure the result. Round and round, typically hundreds or thousands of times per second.

Common sensors in mechanical systems include incremental and absolute encoders and potentiometers for position, Hall effect sensors and tachometers for speed, strain gauge load cells for force and torque, thermocouples, resistance temperature detectors and thermistors for temperature, piezoresistive elements for pressure, and accelerometers and gyroscopes packaged as inertial measurement units. Each is specified by range, accuracy, resolution, and bandwidth, and the last of those catches people out: a sensor that is beautifully accurate but responds in half a second is useless in a loop running at 100 Hz.

Worked example: how well can we measure the cart's speed? Fit a 24 pulse per revolution encoder to the rear axle, which turns at 357 rpm at top speed, or 5.95 revolutions per second. That gives 5.95 x 24 = 142.8 pulses per second, and reading both edges of both quadrature channels multiplies it by four to 571 counts per second. If the controller updates speed every 100 ms, it counts about 57 counts per update, and a plus or minus one count uncertainty is plus or minus 1.75 percent of speed. Acceptable for cruise control, marginal for anything finer. To improve it you either count for longer, which adds lag, or fit a higher resolution encoder, which costs money. That trade between resolution, latency, and cost recurs in every measurement problem you will ever face.

On the actuator side, mechanical engineers work with brushed and brushless DC motors, stepper motors for open-loop positioning, servo motors with integrated feedback, solenoids for binary motion, and hydraulic and pneumatic cylinders where force density matters. Introduction to Robotics (ENGR 360) develops the actuator and kinematics side much further; Electrical Circuits and Electronics (ENGR 240) covers the drive electronics.

Key idea: Mechatronic systems are a sensor, controller, actuator loop, and a sensor must be judged on resolution, latency, and cost together rather than accuracy alone.

Feedback control and PID

An open-loop controller acts without measuring the result: a toaster runs its timer regardless of how brown the bread is. A closed-loop controller measures the output, computes the error between measurement and setpoint, and acts on that error. Closed loop rejects disturbances, tolerates plant variation, and is the reason a cruise control holds speed up a hill.

The overwhelmingly dominant control law in industry is PID, which sums three terms acting on the error. Proportional action responds in proportion to current error: strong, immediate, and it leaves a residual offset. Integral action accumulates past error and keeps pushing until the error is exactly zero, eliminating that offset, at the risk of winding up during saturation. Derivative action responds to the rate of change of error, adding damping and anticipation, at the cost of amplifying sensor noise.

Worked example: cruise control for the cart. Model the cart as mass times acceleration equals drive force minus a resistance proportional to speed. At 5.56 m/s the resistance was 54.1 N, so the effective damping is b = 54.1 / 5.56 = 9.73 N s/m. Open loop, the cart's speed responds with a time constant:

tau = m / b = 300 / 9.73 = 30.8 s

Half a minute to settle after a change in throttle. Sluggish, and exactly what you would feel driving it manually. Now close the loop with proportional control that commands force in proportion to speed error, with a gain of K(p) = 200 newtons per metre per second of error. The closed-loop time constant becomes:

tau = m / (b + K(p)) = 300 / 209.7 = 1.43 s

Twenty times faster, with no change to any hardware. That is what feedback buys. But now check the steady state. With proportional action only, some error must remain in order to produce any force at all, so:

v(steady) / v(setpoint) = K(p) / (b + K(p)) = 200 / 209.7 = 0.954

Ask for 20 km/h and you get 19.1 km/h, a 4.6 percent droop, and it gets worse on a hill because a larger disturbance requires a larger error to fight it. Adding integral action fixes it exactly: the integral term accumulates that persistent small error until the output reaches the setpoint and the error is genuinely zero. This is why almost every real speed and temperature controller is at least PI rather than pure P.

Why not simply raise the proportional gain until the droop is negligible? Because every real system has lag, from sensor filtering, computation, and actuator response. With lag, a large gain means the controller is still pushing hard in response to an error that has already been corrected, and the system overshoots, then over-corrects, and oscillates. Beyond a critical gain it becomes unstable outright. The tension between speed of response and stability is the central problem of control engineering, and no amount of gain gets you out of it.

Key idea: Feedback cut the cart's speed response from 30.8 s to 1.43 s with no hardware change, but proportional action alone leaves a 4.6 percent droop, and raising gain to remove it eventually causes oscillation and instability.

Energy systems

Mechanical engineers design the equipment that converts, moves, and consumes most of the world's energy: turbines, engines, compressors, pumps, fans, boilers, heat exchangers, wind turbines, solar thermal collectors, and the entire building services industry. The analytical toolkit is what you have already built: the first and second laws from ENGR 230, the fluid and heat transfer methods from Module 4, and the habit of chaining efficiencies.

That last habit deserves emphasis because it is where intuition fails. Efficiencies multiply, so a system with five 90 percent stages is 59 percent efficient overall, and the way to improve a system is almost never to polish the best stage. Find the worst one.

Our cart's chain runs: battery to controller and motor at 85 percent, motor to axle through two chain stages at 90 percent, giving 76.5 percent from stored energy to the wheels. At cruise we computed 301 W at the wheels and 391 W from the battery. With a 1.92 kWh pack that is 1920 / 391 = 4.9 hours of continuous running, or about 98 km, derated in practice to perhaps 50 to 70 km once hills, stops, accessories, and a sensible depth of discharge limit are included. Expressed per kilometre it is roughly 20 Wh/km from the battery. For comparison, an electric bicycle uses about 10 to 15 Wh/km, an electric car 150 to 200 Wh/km, and a small petrol pickup roughly ten times the car's figure in primary energy. Low speed and low mass are extraordinarily effective.

Sustainability, done with arithmetic

Now the part that is usually asserted rather than calculated. Is the cart a green product? Do the life cycle assessment, at least at the order-of-magnitude level that a designer can do on one page.

Use phase. Assume 30 km per working day, 250 days a year, over a five year life: 37,500 km. At 20 Wh/km that is 750 kWh of electricity. At roughly 0.4 kg of carbon dioxide per kilowatt hour for an average grid, the use phase is about 300 kg of CO2.

Embodied phase. The cart's 150 kg empty mass is mostly steel, and primary steel carries roughly 1.9 kg of CO2 per kilogram, so around 110 kg of steel accounts for about 210 kg. The 1.92 kWh lithium battery, at very roughly 75 kg of CO2 per kilowatt hour of capacity, adds about 145 kg. Wheels, electronics, and the rest bring the total to roughly 400 kg of CO2.

Read that result carefully, because it is the opposite of the usual story: for this product, manufacturing emits more than a five year working life of use. That is a general pattern for efficient, lightly used equipment, and it has direct design consequences. If embodied carbon dominates, then squeezing another 5 percent from the drivetrain is nearly pointless, while making the cart last ten years instead of five roughly halves its lifetime carbon per kilometre. Durability, repairability, and a replaceable battery become the sustainability features, not efficiency. This is exactly why the fatigue redesign in Module 3 and the serviceability rules in Module 5 were sustainability work even though nobody called them that.

One more comparison keeps the whole thing in proportion. If the cart replaces a petrol pickup used for the same duty at 12 litres per 100 km, that pickup burns 4500 litres over 37,500 km and emits roughly 10,400 kg of CO2, about twenty-six times the cart's entire life cycle footprint. So the honest summary has two parts, and both matter: the substitution is overwhelmingly worthwhile, and within the cart's own footprint, manufacturing rather than electricity is what you should be attacking. Refusing to collapse that into a single slogan is what makes it engineering.

Key idea: Life cycle arithmetic showed 400 kg of embodied CO2 against 300 kg in five years of use, so durability beats efficiency for this product, while replacing a petrol vehicle still saves roughly 10,000 kg.

Common misconceptions

• Higher controller gain is always better performance. Gain speeds response until lag turns it into overshoot, oscillation, and eventually instability. Speed and stability trade against each other.
• A proportional controller can reach the setpoint exactly. It cannot; some error must persist to generate output. Integral action is what drives the error to zero.
• Electric vehicles have no manufacturing footprint worth counting. For an efficient, low-utilization machine, manufacturing can exceed the use phase entirely, as it does for our cart.
• Improving the most efficient component improves the system. Efficiencies multiply, so the worst stage dominates the product. Find the weak link before polishing the strong one.

Recap

• Mechatronic systems close a loop of sensors, controller, and actuators, and sensors must be judged on resolution, latency, and cost together.
• A 24 pulse per revolution encoder with quadrature gives 571 counts per second at top speed, or plus or minus 1.75 percent speed resolution on a 100 ms update.
• Proportional feedback cut the cart's 30.8 s open-loop time constant to 1.43 s, but left a 4.6 percent droop that integral action removes.
• Raising gain to eliminate droop eventually causes overshoot and instability because every real loop contains lag.
• Efficiencies multiply: the cart's 76.5 percent battery-to-wheel chain gives about 20 Wh/km, against 10 to 15 for an e-bike and 150 to 200 for an electric car.
• Life cycle assessment gave roughly 400 kg of embodied CO2 against 300 kg for five years of use, making durability and repairability the dominant sustainability levers.
• Replacing a petrol pickup on the same duty avoids roughly 10,400 kg of CO2, about twenty-six times the cart's whole footprint.

Sources

1. U.S. Department of Energy. (n.d.). Energy efficiency and renewable energy resources. energy.gov
2. Wikipedia. (n.d.). PID controller. en.wikipedia.org
3. Encyclopaedia Britannica. (n.d.). Automation. britannica.com
4. Wikipedia. (n.d.). Life-cycle assessment. en.wikipedia.org

The big picture

On 20 March 1905, a boiler at the Grover Shoe Factory in Brockton, Massachusetts exploded, launched itself through the building, and brought the four-storey structure down in flames. Fifty-eight people died. It was not an unusual event: boiler explosions in that era killed people in the United States at a rate of hundreds per year, because there was no agreed way to design, build, or inspect a pressure vessel. Massachusetts responded with a state boiler law, other states followed with incompatible rules, and the American Society of Mechanical Engineers set out to write one technical document that everyone could adopt. The first edition of the ASME Boiler and Pressure Vessel Code appeared in 1914 and 1915, and boiler explosions became rare.

That story contains most of what you need to know about engineering as a profession. The technical content is real and hard-won. It exists because people died. It works only because it is written down, agreed collectively, adopted into law, and enforced. And the engineers who wrote it were doing something that is not captured anywhere in a stress calculation: they were accepting responsibility to people who would never know their names.

This final lesson is about that frame around the technical work: the standards you will design to, the ethical obligations you carry, the licensure and accreditation system, and the careers this degree opens.

Codes and standards

A standard is a voluntary consensus document specifying how something should be designed, made, tested, or measured. A code is a standard that a government has adopted into law, at which point compliance stops being optional. The same document is often both, depending on jurisdiction.

Standards are why an M10 bolt from any supplier fits an M10 nut from any other, why a pressure vessel built in one country can be sold in another, and why the fits you specified in the last lesson mean the same thing to every shop on Earth. They also encode enormous amounts of experience: when a standard specifies a minimum weld throat or a maximum allowable stress, that number usually descends from a failure.

Body	Covers
ASME	Pressure vessels and boilers, piping, elevators, dimensioning and tolerancing (Y14.5), fasteners
ASTM	Material specifications and test methods
ISO	International standards including limits and fits (286), general tolerances (2768), quality management
SAE	Ground vehicle and aerospace practice
AGMA	Gear design, rating, and quality
AWS	Welding procedures, qualification, and inspection
ASHRAE	HVAC and refrigeration design and building energy
UL and equivalents	Product safety testing and certification

The Atlas cart alone touches many of them. As a low-speed powered industrial vehicle it falls under ANSI/ITSDF B56 family requirements for powered industrial trucks, covering stability, braking, guarding, and operator controls. Its welds would be qualified to an AWS procedure. Its fits came from ISO 286 and its untoleranced dimensions from an ISO 2768 general tolerance class. Its geometric callouts follow ASME Y14.5. Sold in Europe it would need conformity marking against the Machinery Directive. None of this is optional decoration, and discovering it late is one of the most expensive mistakes a young engineer can make. Find out which standards apply before you design, not after.

Key idea: Standards are voluntary consensus documents encoding hard-won experience, codes are standards adopted into law, and identifying which apply is a first-week task rather than a final-week one.

Engineering ethics

Every engineering code of ethics, including those of the National Society of Professional Engineers and of ASME, opens with the same idea, usually as the first and paramount canon: engineers shall hold paramount the safety, health, and welfare of the public. Everything else, obligations to your employer, your client, your colleagues, and your own career, is subordinate to that. The codes then add duties to perform services only in your area of competence, to issue public statements truthfully and objectively, to act as a faithful agent for each employer or client, to avoid deceptive acts, and to disclose conflicts of interest.

Those are easy to agree with in the abstract and genuinely hard in the moment, so we will look at what the moment actually looks like.

Case study: Challenger, 28 January 1986. The Space Shuttle Challenger broke apart 73 seconds after launch, killing all seven crew members including schoolteacher Christa McAuliffe. The technical cause was established by the Rogers Commission: the rubber O-ring seals in a field joint of the right solid rocket booster failed to seal at low temperature, allowing hot gas to escape and burn through toward the external tank. The overnight temperature before launch had fallen to about minus 2 C, far colder than any previous shuttle launch.

The part that matters for you happened the evening before. Engineers at Morton Thiokol, the booster contractor, had been worried about O-ring erosion in cold weather for years and had documented it. That night, Roger Boisjoly, Arnold Thompson, and colleagues argued strongly against launching, and Thiokol initially recommended no launch below 53 F. Under pressure in the teleconference, Thiokol management called an offline caucus, and a senior manager was urged to "take off your engineering hat and put on your management hat." The company reversed its recommendation and signed a launch approval. The engineers who objected did not sign it.

Several failures compound here, and each is instructive.

• The burden of proof inverted. The normal rule is that you must demonstrate a system is safe to fly. That night the engineers were effectively asked to prove it was unsafe, which is a far harder thing to do with sparse data, and which quietly moves the default from caution to launch.
• The data was real but poorly presented. The charts shown did not clearly display the relationship between temperature and O-ring damage across all flights. Presenting evidence so that a decision maker can actually see it is an engineering responsibility, not a communications afterthought.
• Normalization of deviance. Sociologist Diane Vaughan's term for what happens when an anomaly recurs without catastrophe and gradually gets reclassified as normal. O-ring erosion had been seen before and had not yet killed anyone, so it drifted from alarming to expected.
• The engineering judgment was overridden by schedule pressure, and the organization had no channel through which a dissenting engineer could stop a launch.

Richard Feynman, serving on the commission, made the physics undeniable in public by dropping a piece of O-ring material into a glass of ice water and showing that it lost its resilience. His appendix to the report closed with a line worth carrying: reality must take precedence over public relations, because nature cannot be fooled.

A second case, briefly. On 17 July 1981, two suspended walkways in the Kansas City Hyatt Regency hotel collapsed onto a crowded lobby, killing 114 people. The cause was a connection detail: the original design hung both walkways from a single continuous rod, and a change made during fabrication substituted two shorter offset rods, which doubled the load on the fourth-floor walkway's box beam connection. The change was submitted on shop drawings and approved without adequate engineering review. The National Bureau of Standards, now NIST, investigated, and the engineers of record lost their licences. The lesson is unglamorous and permanent: reviewing shop drawings and change requests is engineering work, not paperwork, and a change that looks like a fabrication convenience can be a structural redesign.

What do you actually do? If you believe something is unsafe: put your concern in writing, with the technical basis, and keep a copy. Raise it through your chain of management. Use your organization's formal ethics or safety reporting channel if it has one. Professional societies including NSPE and ASME maintain guidance and, in some cases, assistance for engineers facing these situations. Regulated industries have statutory reporting routes, and whistleblower protections exist in law in many jurisdictions. Be clear-eyed, though: the engineers who objected at Thiokol were right and their careers suffered anyway. Anyone who tells you this is easy has not done it.

Key idea: Codes of ethics hold public safety paramount, and the Challenger and Hyatt Regency cases show the recurring mechanisms of failure: inverted burden of proof, poorly presented data, normalized deviance, and reviews treated as paperwork.

Licensure, accreditation, and the path

In the United States the professional path runs through several gates. Your degree programme should be ABET accredited, meaning it has been reviewed against outcome-based criteria covering the ability to identify and solve engineering problems, design within realistic constraints, communicate, function on teams, recognize ethical responsibilities, and engage in lifelong learning. Graduating from an accredited programme is usually a prerequisite for licensure.

Next comes the Fundamentals of Engineering examination, written by NCEES and typically taken in the final year of study or shortly after. Passing makes you an engineer in training or engineer intern. After roughly four years of qualifying experience under a licensed engineer, you may sit the Principles and Practice of Engineering examination, and passing it, plus meeting your state board's requirements, makes you a licensed Professional Engineer. NCEES writes and administers the examinations; individual state boards grant licences and set their own rules.

A PE licence lets you sign and seal engineering documents and offer engineering services directly to the public. Whether you need one depends heavily on where you work. In building services, HVAC design, structural and civil work, forensic engineering, and independent consulting it is close to essential. In manufacturing, aerospace, automotive, and product development at large companies, the so-called industrial exemption means many excellent mechanical engineers never obtain one. The honest advice is to take the FE examination while the material is fresh regardless, because it costs little and keeps the option open, and options are worth more than they look at twenty-two.

Careers

The Bureau of Labor Statistics tracks mechanical engineering as one of the largest engineering occupations in the United States, with employment on the order of three hundred thousand and median pay well above the all-occupation median. The breadth we started this course with is the career story too.

Sector	What you would work on
Machinery and industrial equipment	Production machines, packaging lines, material handling, automation
Automotive and transportation	Powertrain, chassis, thermal systems, electric vehicle integration
Aerospace and defence	Structures, propulsion, thermal control, mechanisms
Energy and utilities	Turbines, wind, solar thermal, grid-scale storage, plant engineering
Buildings and HVAC	Load calculation, equipment selection, distribution, controls, commissioning
Medical devices	Implants, surgical instruments, diagnostic equipment, regulated design
Semiconductors and electronics	Precision motion, wafer handling, thermal management, cleanroom equipment
Consulting and forensics	Independent analysis, failure investigation, expert testimony

Within any of those, the roles differ more than the industries do. Design engineers create geometry and own requirements. Analysis engineers run finite element and computational fluid dynamics work and correlate it to test. Test engineers build rigs and generate the data everyone else trusts. Manufacturing and process engineers make the design producible and keep production running. Quality engineers own capability, inspection, and root cause analysis. Field service engineers see what actually breaks, which is arguably the fastest education available. Project and programme management and technical sales pull engineers toward the commercial side. And a mechanical degree is a common launchpad into patent law, management consulting, finance, and founding companies.

What a text course cannot give you. Let me be direct, since this is the last lesson. Everything in this course was necessary and none of it is sufficient. Mechanical engineering is a physical discipline, and judgment about physical things comes from handling them. Take a machine shop course and cut metal. Join a student design team, Formula SAE, Baja, a robotics competition, a rocketry club, and live through a build with a deadline. Take things apart. Do an internship, then another. Learn one CAD package properly rather than three badly. When something breaks, keep the pieces and look at the fracture surface, because you now know how to read one. The calculations in this course will make sense of what your hands discover, and your hands will tell you which calculations were worth doing.

The cart. We began with a sentence in a hallway and ended with a machine: 300 kg gross, 20.2 km/h on an 8.40 to 1 chain drive, a 30 mm rear axle sized by twist and by bearing bores, 50 x 50 x 3 bed rails deflecting 0.635 mm, hydraulic brakes needing 68 N at the lever, a shot peened axle good for 50,500 km, a printed motor bracket at 20 units, and a life cycle footprint dominated by its own manufacture. It is not finished, because nothing is: it needs a full FMEA, a wiring design, a stability calculation against tipping, a cost rollup, and a season of abuse by people who did not design it. But you now know how every one of those would be done, and where to look up what you do not know. That is what an introductory course is for.

Key idea: The professional path runs through ABET accreditation, the FE, qualifying experience, and the PE, and the degree opens sectors and roles far broader than the stereotype suggests.

Common misconceptions

• Standards restrict creativity. They remove the need to reinvent settled practice and encode failures you would otherwise repeat. Creativity belongs in the parts no standard has settled.
• Ethics problems are obvious when they happen. Challenger's decision makers were not villains. The failures were an inverted burden of proof, unclear data, and gradually normalized anomalies, all of which feel reasonable from inside.
• Every mechanical engineer needs a PE licence. It is close to essential in building services, consulting, and forensics, and many excellent engineers in manufacturing and aerospace never obtain one. Taking the FE early keeps the option open cheaply.
• Reviewing a supplier's shop drawings is administrative. The Hyatt Regency change arrived as a fabrication convenience and was a structural redesign that killed 114 people.

Recap

• The ASME Boiler and Pressure Vessel Code of 1914 and 1915 followed decades of fatal boiler explosions, including Brockton in 1905; standards encode failures so they are not repeated.
• A standard is voluntary consensus and a code is a standard adopted into law; ASME, ASTM, ISO, SAE, AGMA, AWS, ASHRAE and others cover the mechanical field.
• Codes of ethics hold paramount the safety, health, and welfare of the public, above obligations to employer or client.
• Challenger failed through an inverted burden of proof, poorly presented temperature data, normalized deviance, and schedule pressure overriding engineering judgment.
• The Hyatt Regency walkway collapse came from a fabrication change that doubled a connection load and was approved without adequate review.
• The professional path is an ABET accredited degree, the FE examination, about four years of qualifying experience, and the PE examination administered through NCEES and state boards.
• Careers span machinery, transport, aerospace, energy, buildings, medical devices, semiconductors, and consulting, in design, analysis, test, manufacturing, quality, service, and management roles.

Sources

1. National Aeronautics and Space Administration. (n.d.). Space Shuttle Challenger accident resources. nasa.gov
2. American Society of Mechanical Engineers. (n.d.). Codes and standards. asme.org
3. National Council of Examiners for Engineering and Surveying. (n.d.). FE and PE examinations. ncees.org
4. ABET. (n.d.). Accreditation of engineering programs. abet.org
5. U.S. Bureau of Labor Statistics. (2024). Mechanical engineers. Occupational Outlook Handbook. bls.gov