Skip to main content

Machine learning for additive manufacturing

📍 Where we are: Part II · Modeling — Chapter 5 of 19. Chapter 4 built models from physics; this chapter builds them from data, and asks when each can be trusted.

The simple version

A physics model is like a cook who knows chemistry and works out a recipe from first principles. A machine-learning model is like a cook who has tasted ten thousand dishes and guesses what will taste good. The second cook is fast and often right on familiar dishes, and can be badly wrong on a new one without any warning. In AM, "dishes" are expensive, because each sample is a build, so we often have a few dozen tastes rather than ten thousand.

What this chapter covers​

We survey where machine learning (ML) is used in AM (surrogates, parameter optimisation, defect detection, physics-informed models), what the small-data problem does to each, and why reproducibility and benchmarks are the field's weakest point. Evidence is labelled lab-scale, production or contested.

Problem​

AM generates data (process settings, sensor streams, CT scans, test results) but the data are sparse where it matters: a new alloy or machine has few builds, and rare defects are rare by definition. Engineers want models that propose good parameters, flag defects as they form, and predict properties. Review articles cover the breadth of ML across the AM workflow [1][2][3], and more recent reviews take a physics-informed view [4][5].

State of the art​

Four uses​

UseIdeaEvidence level
Surrogates and property predictionLearn a fast map from parameters to density, melt-pool size or strengthLab-scale; many small studies, mostly one machine
Parameter optimisationChoose the next experiment to try, often by Bayesian optimisationLab-scale
Defect detectionClassify images or signals as normal or anomalousLab-scale; some production-oriented research
Physics-informed modelsBuild the governing equations into the loss or architectureLab-scale

Surrogates and best practice. A paper on best practices for ML aimed at process-parameter development in powder bed fusion discusses design choices that affect validity, including physics-constrained networks and data augmentation for small datasets [6].

Bayesian optimisation. A Gaussian-process model plus an acquisition rule picks each next experiment to learn the most. An autonomous system that designs, prints and tests mechanical structures in a closed loop used this approach to optimise mechanical design [7]. It is demonstrated at lab scale on small-scale printing; applying it to a metal machine, where each experiment costs far more, is the open step.

Defect detection. A trained computer-vision algorithm was applied to powder-bed images to detect and classify anomalies in an LPBF process [8]. A stronger test is detecting a process event such as a keyhole pore at the moment it forms. A machine-learning system trained with synchrotron X-ray ground truth detected keyhole pore generation in real time from thermal imaging [9], and later with sound and light sensors at sub-millisecond resolution [10]. These are lab-scale demonstrations with unusually rigorous labels; Chapter 6 on in-situ monitoring covers the sensing side.

Physics-informed models. Physics-informed neural networks (PINNs) put the heat equation, and for some versions the fluid equations, into the training loss; one study predicted temperature and melt-pool fluid dynamics this way, reducing reliance on labelled data [11]. A 2025 review (recent) organises physics-informed ML for process–structure–property modelling in AM [5]. Physics helps when data are scarce, and its limit is the same as that of Chapter 4: the physics included is only as good as its inputs.

Small data and transfer​

Transfer learning reuses a model trained on one material or machine for another. Work on deep transfer learning across materials in LPBF showed that mechanisms learned on one alloy can help on another [12]. This is lab-scale, and the data volume is the point: the target material still needs its own labelled builds, and we did not find evidence that transfer between machines is routine in production.

Four ML roles in the AM workflow (surrogate, Bayesian optimisation, defect detection, physics-informed model) between a data box on the left and a decisions box on the right, with two bars below for small machine-specific datasets and few shared benchmarks. ML in AM turns process, sensor and test data into predictions and decisions, but every role is limited by small, machine-specific datasets and by the lack of shared benchmarks. Original diagram by the authors, created with AI assistance.

Why it is hard​

  • Small, expensive, machine-specific data. A coupon build costs time and powder; sensor placement, gas flow and calibration differ between machines, so a model often does not carry over.
  • Imbalanced labels. Pores that matter are rare; ground truth requires CT or a synchrotron, which are costly [9].
  • Leakage and weak validation. If layers, tracks or slices from the same build appear in both training and test sets, accuracy is overstated. Best-practice guidance addresses validity issues of this kind [6].
  • No shared benchmark for ML. Physics modelling has blind challenges with held-back data (AM Bench [13]), which let readers compare methods. A literature-derived benchmark exists for melt-pool geometry and flaw-mode prediction [14], but we did not find a widely adopted blind-challenge counterpart for sensor-based defect detection, so comparing papers is hard; treat this as our assessment, not a verified statement about the whole field.
  • Extrapolation. A model fit within a parameter window may fail outside it with no warning.

Open questions​

  1. How much data is enough? Transfer learning aims to reduce the data needed on the target [12], but across which differences (machine, powder, alloy) it holds is not mapped.
  2. Can data-driven PSP models [15] be validated on held-out machines and laboratories rather than held-out splits of the same builds?
  3. Can uncertainty estimates be made trustworthy enough that a model can say "I do not know" outside its range?
  4. How should ML detectors be validated for use in qualification? This is a regulatory question as much as a technical one, taken up in Chapter 12.
  5. How best to combine physics and data, and how much does the physics term buy beyond a well-designed data-only model?

Industry impact​

ML's near-term industrial value lies in narrow tasks: anomaly flags on layerwise images, process-parameter screening for a new powder, and sensor-based event detection feeding the control ideas in Chapter 7. The claims of demonstrated in production for ML-driven closed-loop correction are, in our verified sources, thin; most evidence is lab-scale. The bottleneck is data infrastructure and validation, not model architecture (Chapter 13 on the digital thread).

Why it matters​

Parameter development for a new alloy is a large fraction of the cost of introducing it, and defect detection is central to qualification. ML promises to cut both, but only if its predictions are trustworthy.

In the real world​

  • Research groups pair synchrotron X-ray imaging with fast sensors to produce rare, well-labelled datasets for training detectors [9][10].
  • Open NIST benchmark data allow physics-based and data-driven methods to be tested on the same measurements [13].
  • Practical guidance on ML strategy, including small datasets, is now published specifically for powder bed fusion [6].

Key terms​

  • Surrogate model — a fast learned approximation of an expensive simulation or experiment.
  • Bayesian optimisation — choosing each next experiment using a probabilistic model of what has been learned so far.
  • Gaussian process — a probabilistic regression model that gives a prediction with an uncertainty.
  • Physics-informed neural network (PINN) — a network trained to satisfy governing equations as well as data.
  • Transfer learning — reusing a model trained on one task or machine for another.
  • Data leakage — information from test data entering training, which inflates accuracy.
  • Ground truth — the trusted measurement (for example X-ray imaging or CT) against which a model is judged.
  • Anomaly detection — flagging data that differ from normal builds.
  • Overfitting — fitting noise in a small training set so that the model fails on new data.
  • Benchmark — a shared, held-back test case that allows fair comparison of methods.

Where this leads​

A model is only useful if it sees the process. Next: In-situ monitoring and metrology.