Skip to content
LumenaryAcademy
Frontier 85 min

Advanced manufacturing

Metal additive manufacturing at production scale, robots and automation on the factory floor, in-process monitoring with machine learning, making things at the micro- and nanoscale, and the facts behind reshoring — as of late 2026.

Code tested · 5AI-draftedAwaiting review by a mechanical engineer

Your guides: Chloe · The Designer, Margaret · The Test Engineer, Gideon · The Checker

MotivationH

A rocket engine from a box of powder

In May 2024 the Chennai start-up Agnikul Cosmos flew a small suborbital rocket from its own launch pad at Sriharikota. Its engine, the company says, was printed in a single piece: injector, chamber, nozzle and the cooling channels inside their walls, with no welds. A decade earlier SpaceX had printed the combustion chambers of its SuperDraco abort engines in a nickel superalloy, and Rocket Lab's Rutherford engine, which has flown on dozens of Electron launches since 2017, was designed from the start around printed parts. In jet engines, the fuel-nozzle tip of the CFM LEAP, in service since 2016, is printed in one piece that GE says replaced twenty brazed and welded parts and is a quarter lighter.

Those are the headlines. Behind them is a broader change in how things are made: printers with a dozen lasers running unattended for days; factories where robots outnumber people on the shop floor; cameras and photodiodes watching every layer and every weld, with machine-learning models flagging anomalies; machines that position a silicon wafer to within a nanometre while accelerating it at many times gravity; and governments spending hundreds of billions to bring production of chips, batteries and medicines back home.

This lesson takes four of these threads in turn (metal additive manufacturing at scale, automation, in-process quality and micro-manufacturing) and then the economics and politics of where things are made. It assumes manufacturing processes and additive manufacturing, and uses the conduction of heat transfer.

DefinitionsH

Metal additive manufacturing at scale

The international vocabulary, ISO/ASTM 52900, sorts additive processes into seven families. Three carry almost all production metal work:

Definition— Laser powder bed fusion (LPBF)

A recoater spreads a layer of metal powder, typically 20–80 µm thick, over a build plate in an argon or nitrogen atmosphere; one or more lasers, steered by galvanometer mirrors, melt the cross-section of the part; the plate drops by one layer and the cycle repeats. It gives the finest detail and the best surface of the metal processes, and is the workhorse for aerospace, medical implants and tooling. Electron-beam powder bed fusion is the same idea in a vacuum, with a hot powder bed that reduces residual stress.

Definition— Directed energy deposition (DED)

Metal is fed as powder or wire into a melt pool made by a laser, an electron beam, a plasma or an electric arc, and the head (often on a robot arm) deposits beads layer by layer. It is coarser but far faster than powder bed fusion, has no build box, and can add material to existing parts, which makes it the process for very large parts and for repair. Wire-arc additive manufacturing (WAAM) uses welding equipment and is the cheapest per kilogram.

Definition— Binder jetting

A print head deposits a liquid binder on a powder bed, layer by layer, to make a fragile "green" part that is then sintered in a furnace, where it shrinks by roughly a fifth. It is fast and needs no lasers, so it suits larger batches of small parts, but the shrinkage must be predicted and compensated.

lasermirrorspowder supplyrecoaterpartoverflowinert gas ingas outplatform dropsone layerbeam
A laser powder bed fusion machine in outline. Galvanometer mirrors steer the laser over the powder bed; the recoater spreads each new layer from the supply side; the build platform drops one layer at a time; and a flow of inert gas carries away the smoke and spatter that the melt pool throws up.

Where it is used in production. Rocket engines are the clearest fit: low volumes, high value, and internal cooling channels and injector passages that are very hard to machine. NASA developed a copper–chromium–niobium alloy, GRCop-42, for printed combustion chambers, and its RAMPT project used blown-powder DED to make large nozzles with cooling channels built into their walls. In gas turbines, Siemens Energy has printed and repaired burner components for years, and GE Aerospace builds fuel nozzles and other parts for aero-engines in series production. Norsk Titanium in Norway makes structural titanium parts for the Boeing 787 by plasma DED, and in 2021 a stainless-steel footbridge printed by robots using wire-arc deposition, designed by the Dutch firm MX3D, was installed over a canal in Amsterdam.

Scale. Large laser powder bed machines from makers in Germany, the United States and China now carry a dozen or more lasers over build areas of a metre or more, and DED cells build parts several metres long. The binding constraint is rarely the machine alone: a printed part needs stress relief, removal from the plate and its supports, often hot isostatic pressing to close internal pores, heat treatment, machining of critical surfaces and inspection, and these steps can cost as much as the printing. The industry also went through a hard consolidation in 2023–25, with several well-known firms merging, restructuring or failing, a reminder that the technology works where its economics do.

ComputationC

Computing it: build rates and melt pools

Two numbers summarise an LPBF parameter set. The volumetric energy density Ev=P/(v h t)E_v = P/(v\,h\,t), in J/mm³, combines laser power PP, scan speed vv, hatch spacing hh (the distance between neighbouring scan tracks) and layer thickness tt. The melting rate of one laser is v h tv\,h\,t, the volume swept per second. The program below estimates the build time of a small combustion chamber in a nickel superalloy.

python
# Build time of a laser powder bed fusion job: scanning shares out between lasers, recoating does not.
v, h, t = 1.0, 0.11, 0.04            # scan speed (m/s), hatch spacing (mm), layer thickness (mm): a nickel alloy
rate = v * 1000 * h * t * 3600 / 1000   # cm^3 melted per hour by one laser while it is scanning
print(f"melting rate per laser: {rate:.1f} cm^3/h  (energy density at 370 W: {370 / (v * 1000 * h * t):.0f} J/mm^3)")

density = 8.19                       # g/cm^3, a nickel superalloy such as alloy 718
part_mass, height = 2.0, 250.0       # kg, mm: a small rocket combustion chamber with cooling channels
overhead = 1.20                      # supports and contour passes add about 20 % to the scanned volume
recoat = 9.0                         # seconds per layer to spread powder and settle the gas flow

layers = height / t
recoat_h = layers * recoat / 3600
print(f"{layers:.0f} layers; recoating alone takes {recoat_h:.1f} h\n")
print(" parts on plate  lasers   scan (h)  total (h)  hours per part")
for parts in (1, 10):
    volume = parts * part_mass * 1000 / density * overhead     # cm^3 to scan
    for lasers in (1, 4, 12):
        scan_h = volume / (rate * lasers)                       # ideal sharing between lasers
        total = scan_h + recoat_h
        print(f" {parts:9d}  {lasers:9d}   {scan_h:7.1f}   {total:8.1f}   {total / parts:10.1f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

One laser melts about 16 cm³ an hour, so a 2 kg part takes 34 hours. Twelve lasers cut the scanning from 18.5 hours to 1.5, but the 6250 layers still need 15.6 hours of recoating, so the job takes 17 hours, only twice as fast for twelve times the lasers. This is Amdahl's law from computing, in a factory: the part of the job that cannot be shared limits the speed-up. Fill the plate with ten parts and the recoating is shared among them: twelve lasers then bring the time per part down to 3.1 hours. Multi-laser machines pay off on full plates and tall, bulky parts, and machine makers work as hard on faster recoating as on more lasers.

Whether those layers fuse properly is a heat-transfer question. In 1946 Daniel Rosenthal, studying welding, solved the conduction equation for a point source of heat qq moving at speed vv over a half-space with conductivity kk and diffusivity α\alpha. In coordinates moving with the source, with ξ\xi the distance ahead of it and RR the distance from it,

T−T0=q2πkRexp⁡ ⁣(−v(ξ+R)2α).T - T_0 = \frac{q}{2\pi k R}\exp\!\left(-\frac{v(\xi + R)}{2\alpha}\right).

The region above the melting point is the melt pool. A geometric test by Tang, Pistorius and Beuth (2017) then predicts lack-of-fusion porosity: neighbouring tracks of width WW and depth DD, laid hh apart in layers tt thick, leave unmelted gaps unless

(hW)2+(tD)2≤1.\left(\frac{h}{W}\right)^2 + \left(\frac{t}{D}\right)^2 \le 1.

python
import numpy as np

# Rosenthal's moving point source on a half-space: T - T0 = q / (2 pi k R) * exp(-v (xi + R) / (2 alpha))
k, rho, c = 20.0, 4430.0, 700.0              # Ti-6Al-4V, rough high-temperature values: W/(m K), kg/m^3, J/(kg K)
alpha = k / (rho * c)                        # thermal diffusivity, m^2/s
T0, Tm, eta = 300.0, 1928.0, 0.5             # start and melting temperatures (K); absorbed fraction of the laser
h, t = 0.12e-3, 0.03e-3                      # hatch spacing and layer thickness (m)

def melt_pool(P, v):
    """Length and width (m) of the surface region above the melting point, on a grid around the beam."""
    xi = np.linspace(-1.5e-3, 0.1e-3, 1601)[:, None]    # along the track; negative is behind the beam
    y = np.linspace(1e-7, 0.3e-3, 600)[None, :]          # across the track
    R = np.sqrt(xi**2 + y**2)
    T = T0 + eta * P / (2 * np.pi * k * R) * np.exp(-v * (xi + R) / (2 * alpha))
    molten = T >= Tm
    return np.ptp(xi[molten.any(axis=1), 0]), 2 * y[0, molten.any(axis=0)].max()

print(" power  speed   energy     melt pool (Rosenthal)       lack-of-fusion test")
print("  (W)   (m/s)  (J/mm^3)   length   width   depth       (h/W)^2 + (t/D)^2")
for P, v in [(280, 1.2), (150, 1.2), (100, 1.5), (200, 2.0), (370, 1.0), (280, 0.6)]:
    L, W = melt_pool(P, v)
    D = W / 2                                 # a point source on a half-space melts a half-round pool
    test = (h / W) ** 2 + (t / D) ** 2
    print(f"  {P:4d}  {v:4.1f}   {P / (v * h * t) / 1e9:6.1f}    {L * 1e6:5.0f} um {W * 1e6:5.0f} um {D * 1e6:4.0f} um"
          f"    {test:5.2f}  {'fused' if test <= 1 else 'LACK OF FUSION'}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

At 280 W and 1.2 m/s, a parameter set in the range commonly published for this titanium alloy, the model predicts a melt pool about 145 µm wide and 0.7 mm long (long and thin, because the beam outruns the heat), and the overlap test passes with a little margin (0.85). Cut the power or raise the speed and the pools shrink below what the hatch spacing needs: lack of fusion. The pattern matches experience; the individual numbers deserve less faith.

dense partskeyhole porosity(too much energy, too slow)lack of fusion(too little energy)ballingscan speedlaser power
A schematic process window for laser powder bed fusion. Too little energy for the speed leaves unmelted gaps (lack of fusion); too much power at low speed drives the melt pool into keyhole mode; very fast scanning with high power can break the track into beads (balling). Dense parts come from a band between them, which each alloy, powder and machine must find by experiment.
ComputationC

Watching every layer: in-process monitoring and machine learning

A medium-sized powder bed build produces terabytes of data if you let it: photodiodes and high-speed cameras looking down the laser's own optical path at the melt pool, cameras imaging every freshly spread and freshly melted layer, thermal cameras, acoustic sensors, the laser's power log and the gas-flow sensors. Welding, machining and moulding cells are increasingly instrumented in the same way. The goal is qualification by process data: to know that a part is sound from how it was made, instead of only from destructive tests of sacrificial coupons and costly X-ray computed tomography (CT) of every part.

Machine learning is the natural tool for turning those signals into decisions: convolutional networks that spot recoater streaks, spatter and warped edges in layer images (Oak Ridge National Laboratory's Peregrine software is one openly described example), and models that relate melt-pool signals to pores found later by CT. The trap is in how such detectors are judged. Defects are rare, and a model that never finds one can still be "98% accurate".

The program below makes a synthetic data set (one row per region of a layer, two features from a melt-pool sensor, and a label from CT, with about 2% of regions defective) and compares three detectors on regions they were not fitted to.

python
import numpy as np

rng = np.random.default_rng(2026)
# Synthetic monitoring data: one row per region of one layer. Features: mean melt-pool emission
# (normalised) and its spread within the region. About 2 % of regions hold a defect, found by CT later.
n = 6000
defect = rng.random(n) < 0.02
mean_em = np.where(defect, rng.normal(0.93, 0.05, n), rng.normal(1.00, 0.04, n))
spread = np.where(defect, rng.normal(0.14, 0.03, n), rng.normal(0.10, 0.02, n))
X = np.column_stack([mean_em, spread])
train = np.arange(n) < 4000                                   # first 4000 regions to fit, the rest to test
print(f"defective regions: {defect.sum()} of {n} ({defect.mean():.1%})")

def report(name, flag, truth):
    tp, fp = np.sum(flag & truth), np.sum(flag & ~truth)
    fn, tn = np.sum(~flag & truth), np.sum(~flag & ~truth)
    acc = (tp + tn) / len(truth)
    prec = f"{tp / (tp + fp):6.1%}" if tp + fp else "     -"
    rec = tp / (tp + fn)
    print(f"{name:34s} accuracy {acc:6.1%}  precision {prec}  recall {rec:6.1%}  false alarms {fp:3d}")

test, y = ~train, defect[~train]
print(f"test set: {test.sum()} regions, {y.sum()} defective\n")
report("1. call everything good", np.zeros(test.sum(), bool), y)

# 2. A control chart on the mean emission: limits from regions known to be good
mu, sd = mean_em[train & ~defect].mean(), mean_em[train & ~defect].std()
report("2. control chart, mean - 3 sigma", mean_em[test] < mu - 3 * sd, y)

# 3. Logistic regression on both features, fitted by gradient descent
m, s = X[train].mean(axis=0), X[train].std(axis=0)
Z = (X - m) / s
w, b = np.zeros(2), 0.0
for _ in range(3000):
    p = 1 / (1 + np.exp(-(Z[train] @ w + b)))
    err = p - defect[train]
    w -= 0.5 * Z[train].T @ err / train.sum()
    b -= 0.5 * err.mean()
prob = 1 / (1 + np.exp(-(Z[test] @ w + b)))
for thr in (0.5, 0.2, 0.05):
    report(f"3. logistic regression, p > {thr}", prob > thr, y)
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Calling everything good scores 98.3% accuracy and finds no defects at all. A classic control chart, flagging regions whose mean emission falls more than three standard deviations below normal, finds a fifth of them with few false alarms. Logistic regression on both features does better, and its threshold sets the trade-off: at p>0.5p > 0.5 every flag is real but only 35% of defects are caught; at p>0.05p > 0.05 it catches 76% at the cost of 52 false alarms, two thirds of all flags. Precision (what fraction of flags are real) and recall (what fraction of defects are flagged) are the numbers to ask for, together with the cost of each kind of mistake: a missed pore in a turbine disc and a needless CT scan are not equally bad.

As of 2026, in-process monitoring is standard on new industrial powder bed machines and widely used to decide which parts get extra inspection; it is not yet accepted on its own to certify flight-critical parts, which still rely on qualified, frozen processes, witness coupons, CT and the requirements of standards such as NASA-STD-6030 for spaceflight hardware. Closing the loop, adjusting the laser layer by layer from what the sensors see, is demonstrated in research and in some commercial machines, and it raises a new question for certification: how do you qualify a process that changes itself?

ExplanationH

Robots and automation on the factory floor

According to the International Federation of Robotics, factories worldwide installed about 540,000 industrial robots a year in both 2023 and 2024, more than double the number a decade earlier, bringing the number in operation to over four million. More than half of recent installations were in China, followed by Japan, the United States, Korea and Germany. Korea has the highest density, roughly one robot for every ten manufacturing employees. Car-making and electronics remain the largest users, with metal products, plastics, food and pharmaceuticals growing.

The machines span a range:

  • Articulated arms for welding, painting, machine tending and assembly, fenced off or guarded by light curtains because they are fast and strong enough to kill; their safety requirements are set by ISO 10218.
  • Collaborative robots (cobots), slower arms with force limits and sensing that can work beside people after a risk assessment, under the technical specification ISO/TS 15066 on force and pressure limits.
  • Autonomous mobile robots moving parts between cells and warehouses, which have changed logistics as much as manufacturing.
  • Humanoid robots, whose makers argue they fit factories built for people. Car makers and logistics firms in the United States, Europe and China ran pilots in 2024–26; as of 2026 the published results are short trials of simple handling tasks, and the open questions are reliability, speed, safety certification and cost. The Lumenary Robotics lesson on humanoid robots examines them in detail.

What decides whether automation pays is usually not the robot. A robot arm is a modest part of a cell's cost next to grippers, fixtures, safety systems, integration and programming; and it works best with parts that arrive in a known position and orientation. Machine vision and learned grasping (see manipulation and grasping) are steadily relaxing that requirement. The usual measure of a cell or line is overall equipment effectiveness (OEE): the product of availability (the fraction of planned time it runs), performance (its speed relative to the ideal) and quality (the fraction of good parts). A line at 90% on each of the three has an OEE of only 73%.

Note

Industrial robots, presses and machine tools are guarded, interlocked and locked out before anyone reaches in, and only trained people work on them. If you want to experiment, a small desktop arm or a hobby robot kit powered by a low-voltage supply is the safe way to learn, under the guidance in Lumenary Robotics.

ExplanationH

Micro- and nano-manufacturing

Shrink a machine and the physics changes. Weight and inertia scale with the cube of size, surface forces (friction, adhesion, surface tension, electrostatics) with the square or less, so at the scale of tens of micrometres surfaces dominate: parts stick together (stiction), and electrostatic actuators that would be feeble at the macroscale become practical.

  • MEMS (micro-electro-mechanical systems) are machines etched from silicon wafers with the tools of chip-making: the accelerometers and gyroscopes in every phone and car, microphones, pressure sensors, inkjet print heads and micromirror projectors. Their signature process is deep reactive-ion etching, which alternates etching and passivation steps to cut near-vertical trenches tens of micrometres deep; its best-known form is named after Bosch, still one of the largest MEMS makers, alongside firms in France and Italy, Japan, the United States and China.
  • Ultra-precision machining with single-crystal diamond tools turns optical surfaces accurate to tens of nanometres, for lenses, mirrors and the moulds that make phone-camera lenses by the billion.
  • Two-photon lithography cures a photoresist only at the focus of a femtosecond laser, so it can write true 3-D structures with features below a micrometre: microfluidic parts, micro-optics, scaffolds for cells. Nanoscribe in Germany and UpNano in Austria sell such systems.
  • Lithography machines are among the most demanding mechanical systems ever built. ASML's extreme-ultraviolet scanners, made in the Netherlands with lasers from Trumpf and optics from Zeiss in Germany, make their light by hitting tens of thousands of molten tin droplets a second with a pulsed laser, and move wafer and mask stages in step to within a nanometre. Nanoimprint lithography, which stamps patterns instead of projecting them, reached its first commercial semiconductor tool from Canon in 2024. The Lumenary EE lessons on transistor scaling and chip design follow what those machines make.
ExplanationH

Where things are made: reshoring and supply chains

For three decades, production of many goods concentrated where it was cheapest or where clusters of suppliers had formed: electronics assembly in China and South-East Asia, leading-edge chips in Taiwan and Korea, active pharmaceutical ingredients in China and India. China accounts for roughly 30% of world manufacturing value added. A run of shocks exposed the fragility of long, lean chains: the 2011 Tōhoku earthquake, which shut a Renesas microcontroller plant in Naka and halted car lines worldwide; the COVID-19 pandemic; the 2021 grounding of the Ever Given in the Suez Canal; and the chip shortage of 2021–22 that idled car factories for lack of parts costing a few dollars.

Governments have responded with industrial policy, described here factually:

  • United States. The CHIPS and Science Act of 2022 provided about USD 52.7 billion for semiconductor manufacturing, research and workforce, and the Inflation Reduction Act of the same year added tax credits for clean-energy manufacturing; TSMC's first fab in Arizona began volume production at the end of 2024. In 2023 Mexico overtook China as the largest source of US imports. In 2025 the United States raised tariffs on imports from most countries, with rates and exemptions changed repeatedly; check the current position before relying on any figure.
  • European Union. The European Chips Act of 2023 aims to mobilise more than EUR 43 billion and to double Europe's share of world chip production to 20% by 2030; in 2025 the European Court of Auditors judged that target very unlikely to be met.
  • Japan, Korea, Taiwan and India. Japan subsidised a TSMC-led fab in Kumamoto, which opened in 2024, and Rapidus, which began a 2 nm pilot line in Hokkaido in 2025; Korea and Taiwan support their own leading firms; India's Semiconductor Mission backs a Tata Electronics fab in Gujarat and a Micron assembly-and-test plant at Sanand.
  • China. China has invested heavily in self-sufficiency in chips, machine tools and aircraft, and has used export controls on materials it dominates: gallium and germanium in 2023, graphite later that year, and several rare earths and permanent magnets in 2025.

For an engineer, these policies show up as design and sourcing decisions. Qualify a second supplier and a second process for critical parts, even at higher cost. Design with standard, widely available components where you can. Hold strategic inventory of parts with long lead times ("just in case" alongside "just in time"). Map the supply chain beyond the first tier, since a shortage two or three levels down stops the line just as surely. And recognise the trade-off: resilience costs money, and duplicated capacity in many regions is less efficient than one concentrated chain. Where the balance should lie is a matter of judgement and policy, not a calculation with one answer.

Common misconceptionsH

Common misconceptions

Misconception

"3D printing will replace machining and casting." Additive processes win where geometry is complex, volumes are low or consolidation saves assembly; casting, forging, stamping and machining remain far cheaper per part for most shapes at volume, and printed metal parts are usually machined on their critical faces anyway. The processes are complementary.

Misconception

"A printed part is only as good as its design." It is only as good as its process: powder quality, parameters, scan strategy, gas flow, heat treatment and inspection all change porosity, residual stress, microstructure and therefore fatigue life (see fatigue, creep and corrosion in Lumenary Materials). Identical files on two machines can give different parts, which is why processes are qualified, not just designs.

Misconception

"A defect detector with 99% accuracy is excellent." When defects are rare, accuracy is dominated by the easy majority. Ask for precision and recall on held-out data, with the number of real defects they are based on.

Misconception

"Reshoring means bringing back the old jobs." New plants in high-wage countries are typically far more automated than those they replace, so they employ fewer people per unit of output, and many of the jobs they create are for technicians and engineers. Their costs, timelines and staffing remain contested and are worth checking case by case.

Exercises

Exercises

1. Advanced manufacturing

Warm-upHuman

A laser powder bed fusion parameter set uses a laser power of 200 W, a scan speed of 0.8 m/s, a hatch spacing of 0.10 mm and a layer thickness of 0.05 mm. What is the volumetric energy density, in J/mm³?

2. Advanced manufacturing

CoreHuman

Single-track experiments show melt pools 150 µm wide and 60 µm deep for a candidate parameter set. The layer thickness is 40 µm. Using the overlap criterion (h/W)2+(t/D)2≤1(h/W)^2 + (t/D)^2 \le 1, what is the largest whole number of micrometres of hatch spacing hh that avoids lack of fusion?

3. Advanced manufacturing

CoreComputational

Write two functions.

build_hours(volume_cm3, height_mm, layer_mm, rate_cm3_h, lasers, recoat_s) returns the build time in hours of a powder bed job: the scanned volume shared ideally between lasers lasers, each melting rate_cm3_h cm³ per hour, plus recoat_s seconds of recoating for every layer.

min_lasers(volume_cm3, height_mm, layer_mm, rate_cm3_h, recoat_s, target_h, max_lasers=64) returns the smallest number of lasers (from 1 to max_lasers) that finishes within target_h hours, or None if no number of lasers can.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

4. Advanced manufacturing

CoreComputational

Write precision_recall(flags, truth), where flags and truth are equal-length sequences of booleans (the detector's flags and the CT result for each region). Return a tuple (precision, recall) as floats. If nothing is flagged, precision is undefined: return None for it. You may assume at least one region is truly defective.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

5. Advanced manufacturing

StretchHuman

A monitoring system flags 90% of defective regions and, wrongly, 2% of good ones. Only 1% of regions are defective. Of all the regions it flags, what percentage are actually defective? Give the nearest whole number.

6. Advanced manufacturing

Warm-upHuman

A robotic welding cell runs 85% of its planned time (availability), at 92% of its ideal cycle rate when running (performance), and 97% of its welds pass inspection (quality). What is its overall equipment effectiveness, as a whole-number percentage?

7. Advanced manufacturing

CoreHuman

A gear is shrunk by a factor of 1000 in every dimension, from 10 mm to 10 µm, keeping its shape and material. Roughly how does the ratio of surface adhesion forces (which scale with area or less) to its weight change?

Choose one

8. Advanced manufacturing

CoreHuman

An engineering team is asked to make a product less vulnerable to supply disruptions. Which measures directly help? Select all that apply.

Choose all that apply

Further reading and resources