Lesson 9.2Lesson 9.2 · Computational Workflow
Digital Twins & Feedback
Comparing as-built to as-designed, and closing the loop between model and made object
A model tells you what should be. A digital twin, fed by a scan of the real part, tells you what is - and by how much they differ.
Make a part and a quiet question follows it off the machine: does this actually match what I designed? A drawing cannot answer - it is intent, frozen. A digital twin can, because it is kept live by measurement of the real object or the real machine.
The core move is comparison: register a scan of the as-built part against the as-designed model, colour every point by how far it strayed, and check that against tolerance. Then close the loop - feed the deviation back so the next part comes out right, instead of discovering the error four hundred parts later.
The twin is the loop back to the model. No loop, no twin - just a fancy render.
What a digital twin actually is
A digital twin is a digital counterpart of a physical thing or process that is kept up to date with data from the real world. The phrase is badly overused, so pin it down by what it is not. It is not just a 3D model - a model is a design intent, frozen. It is not just a bill of materials or a BIM file - those describe what should be. A twin earns the name only when a stream of measurements from the real object or the real machine flows back into it, so the digital version tracks the physical one over time. In fabrication that stream is scans, probe points, camera frames, machine telemetry - reality, digitised.
There are really two twins worth separating. A product twin mirrors the made object: is this cast panel, this printed wall, this milled mould the shape we designed? A process twin mirrors the making: what were the spindle load, the nozzle temperature, the layer times while it was being made? Both matter, and they answer different questions - one about the result, one about how you got there. For a designer entering fabrication, the product twin is the natural first tool: it is how you check that the thing you made is the thing you drew.
A model says what SHOULD be; a twin, fed by measurement, says what IS.
As-designed versus as-built
The core operation of a product twin is comparison. On one side is the as-designed geometry - the nominal model, the intent. On the other is the as-built capture - the object as it actually came out, measured by a 3D scanner (structured-light or laser), a CMM (coordinate-measuring machine) touching hundreds of points, photogrammetry from a ring of photos, or a laser tracker for large assemblies. Each tool trades speed for accuracy: a handheld scanner might resolve 0.1 mm over a chair; a CMM reaches microns but slowly; photogrammetry is cheap but noisier.
The measured cloud is then registered to the nominal model - aligned by best-fit or by datum features - and the software colours every point by its signed distance to the design surface. The result is a deviation map: green where the part is on nominal, warming through amber to red where it drifts out. Now you can ask the only question that matters for acceptance: is every deviation inside tolerance? A GD&T (geometric dimensioning and tolerancing) callout might allow the surface to wander plus or minus 0.5 mm but hold a mounting hole to plus or minus 0.1 mm. The twin turns a vague worry - 'does it fit?' - into a coloured, numbered, defensible answer.
Getting there cleanly takes care. A bad registration invents deviations that are not real: align to the wrong datums and a perfectly good part can glow red all over, sending you chasing a ghost. So you align to the features that actually matter for function - the mounting faces, the mating edges - not just a global best-fit that smears the error evenly. And you read the map with judgement: random scatter within tolerance is noise to accept, while a consistent bias in one direction is a signal - a tool that deflected, a setting that drifted - worth acting on.
Closing the loop
A deviation map you look at once and file is not a twin doing its job; it is an autopsy. The value is in the arrow that goes back to the model and changes what you make next. That is closing the loop, and it happens at several timescales. Fastest: a scan-and-compensate step during a run - measure the part mid-process and adjust the remaining toolpath, which is really the subject of the next lesson. Medium: first-article inspection, where you fully measure part one of a batch, correct the definition or the machine offsets, and only then release the rest - so a systematic error is caught on one part, not four hundred. Slowest: feeding a season of deviation data back into how you design and mould, so the next project starts smarter.
A worked case makes it concrete. You CNC-mill a batch of foam moulds for cast panels and the twin shows every one is 0.3 mm shallow in the same place - a sign the tool deflected under load. Because you caught it on the first article, you dial in a 0.3 mm compensation and the rest come out true. Without the loop you would have cast four hundred slightly-wrong panels and discovered it on site. The twin is only worth its cost when someone acts on what it shows.
Notice how this joins the previous lesson. The compensation you dial in does not live in a spreadsheet - it goes back into the parametric definition or the machine offsets that produced the part, so the next batch the pipeline emits is already corrected. Design, make, measure, correct: the loop closes on the same file that opened it. That is the difference between a workflow that merely records its mistakes and one that learns from them between the first part and the second.
First-article inspection: measure part one, correct, THEN release the batch.
Process twins and machine telemetry
Alongside the product twin sits the process twin - a live picture of the making itself. Modern machines emit data you can log and mirror: a CNC reports spindle load, feed override and position; an industrial robot streams joint angles, torque and cycle time; an FDM printer logs nozzle and bed temperature, flow and layer duration; large-format and metal printers add cameras and even thermal imaging of each layer. Watched in real time, this telemetry catches trouble the finished part would only reveal later - a spike in spindle load warning of a dulling tool, a temperature dip hinting at a clog, a layer-time creep that means the machine is struggling.
This is the fabrication face of Industry 4.0: connected machines whose data feeds monitoring, quality records and predictive maintenance. For a design practice it need not be exotic - even logging a print farm's temperatures and failure photos, or a laser's job times, builds a process twin that tells you which machine, material and setting reliably make good parts. The honest caveat: data is only insight if someone reads it. A dashboard nobody watches is theatre. Start by mirroring the few signals that actually predict your failures.
Spindle load, nozzle temp, cycle time - the process twin warns before the part fails.
Value, cost and honest limits
Digital twins are genuinely useful and genuinely oversold, and a maker should hold both truths. The value is real: fewer scrapped parts because errors are caught on the first article; a defensible as-built record for the client, the engineer and the warranty; tighter tolerances held with confidence; and, over time, a feedback trail that makes the next job better. In prefabrication and construction, an as-built twin of a component is increasingly what proves it was made to spec.
The cost is also real. Scanning and registering take time and skill; alignment errors can invent deviations that are not there; storing and plumbing the data is its own project; and the hardest part is closing the loop culturally - the discipline of stopping to measure, and of acting on what the measurement says. A twin nobody feeds or reads is worse than none, because it invites false confidence. And a twin is not a substitute for engineering: structural adequacy, MEP performance and code compliance are judged by qualified professionals, whatever the deviation map is coloured. Use the twin for what it is superb at - telling you, precisely and early, whether the made thing matches the drawn one - and no further.
Digital twin
A model kept live by data from its physical counterpart
Earns the name only with a feedback stream from the real object or machine; distinguishes product twin from process twin.
As-designed vs as-built
Nominal intent versus measured reality
The comparison at the heart of the product twin; the gap between them is the deviation.
3D scanning / CMM / photogrammetry
Ways to capture the as-built geometry
Trade speed for accuracy: handheld scan fast and coarse, CMM slow and micron-precise, photogrammetry cheap and noisier.
Deviation map / tolerance
Signed distance from measured to nominal, judged against limits
Green on-nominal to red out; acceptance is whether every deviation sits inside the GD&T tolerance.
First-article inspection
Fully measuring part one before releasing a batch
Catches a systematic error on one part instead of hundreds; the practical way to close the loop.
Workshop — scan, compare, close the loop
Verify something you made against its model, then act on the result. The point is not the scan - it is the arrow back to the design that changes what you make next.
A fabricated part and its model; a 3D scanner or a phone plus free photogrammetry (e.g. Meshroom); inspection/CAD software that can register a mesh and colour deviations (CloudCompare works free).
Goal: produce an as-built deviation map and a corrective action Inputs: a part you fabricated, its nominal model, any 3D capture (scanner or phone photogrammetry) Time: ~75 minutes
- 1Capture the as-built part: a handheld/structured-light scan if you have one, or 30-40 phone photos run through free photogrammetry software into a mesh.
- 2Register the capture to your nominal model by best-fit or datum alignment, then generate a signed deviation map coloured by distance.
- 3Set a tolerance appropriate to the part (say plus or minus 0.5 mm on surfaces, tighter on holes) and mark every region that falls outside it.
- 4Diagnose the biggest deviation: is it random (warp, handling) or systematic (a consistent offset suggesting tool deflection or a wrong setting)?
- 5Write the corrective action you would take before making the next one - a compensation offset, a fixturing change, a slower feed - and note how you would confirm it worked.
You’ll walk away with
A coloured as-built-vs-as-designed deviation map of your part, a stated tolerance and pass/fail call, and a one-paragraph corrective action that closes the loop for the next part.
Three altitudes on the same idea
Read the band that fits you — or all three.
A digital twin gives you a defensible as-built record and catches drift before it reaches site. Scan a run of cast or milled components, compare to the nominal model, and you can prove to your engineer and client that the made thing matches the drawn one - or find, on the first article, that it does not, and correct it. The twin documents reality; structural and code judgements still rest with your qualified consultants.
This is quality control for bespoke. When a CNC-milled desk, a cast basin or a set of curved panels comes off the machine, a scan-and-compare tells you whether it hits tolerance before it goes to the client. Measure the first piece, correct the definition or machine offset, then release the batch - so a systematic error becomes one reject, not a whole order remade at your cost.
Learn to measure, not just to make. Employers value people who can verify a part, not only produce one. Practise a scan-register-deviate workflow on something you fabricated: capture it, align it to your model, read the coloured deviation, and check it against a tolerance you set. A portfolio piece that shows the made object and its as-built verification reads as rigorous and production-minded.
“We have a BIM model, so we already have a digital twin of the building.”
Do it yourself
No lab needed - reason it through.
- 1In one sentence, what turns a 3D model into a digital twin?
- 2Contrast a product twin with a process twin, with one example of each.
- 3What does registration do, and what does a deviation map show once it is done?
- 4Why does first-article inspection save more than inspecting the whole batch at the end?
- 5Give one honest failure mode of digital twins - a way they can mislead or waste effort.
The one line to carry out
Peer-reviewed journals & authoritative standards
- 01Digital twin — Wikipedia, 2026.
- 02Engineering tolerance — Wikipedia, 2026.
- 03Fourth Industrial Revolution (Industry 4.0) — Wikipedia, 2026.
- 04COMPAS — computational framework for research and collaboration in AEC — COMPAS, 2026.
- 05Gramazio Kohler Research — Digital fabrication in architecture (ETH Zurich) — ETH Zurich, 2026.
Closing the loop after the fact is good; doing it during the cut is better. Next: machines that sense the material and adapt the path in real time - adaptive fabrication.
The author
Amogh N P
Architect, interior designer, and creative polymath. Studio Matrx began in his notebooks — his vision of design made honest, useful, and open to everyone. Its Academy is written and taught in his memory, and free, forever.
More about Amogh →