Investigating how spatter evolves in metal additive manufacturing processes with machine learning

In metal additive manufacturing, laser-powder-bed fusion (LPBF) suffers from layer-to-layer instabilities; most notably molten-metal spatter and recoater streaking - that degrade surface finish and internal integrity. We investigate whether physics-informed machine-learning (PIML) can detect and predict these anomalies more efficiently than purely data-driven models. Using the Oakridge National Laboratory (ORNL) Peregrine in-situ dataset, we (i) derive physically meaningful features such as volumetric energy density, Peclet number and plume-attenuation proxies, and (ii) embed gradient penalties that enforce monotonic behaviour with respect to energy input. A lightweight PIML network attains an Root Mean Square Error (RMSE) of ≈ 3.9 × 104 spatter pixels (R2 = 0.94) while requiring 40 % less training data than an architecture-matched multilayer perceptron. SHapley Additive exPlanations (SHAP) analysis shows that the model’s attributions follow established heat-transfer mechanisms, confirming improved interpretability. These results demonstrate that even minimal physics supervision delivers data-efficient, trustworthy defect monitoring, at least in the case of neural networks tested in this work, paving the way for real-time, closed-loop LPBF control.