Physics-informed machine learning framework for grain size prediction in laser-processed 316L stainless steel: integration of experimental microstructure and surrogate data

Grain size is a key microstructural parameter controlling the strength, toughness, and service performance of additively manufactured (AM) alloys. However, predicting grain size remains challenging due to the combined influence of processing conditions, thermal gradients, and solidification behaviour. This work presents a physics-informed machine learning framework that integrates experimental microstructural data with physically meaningful descriptors to improve prediction accuracy and interpretability. Grain size distributions were quantified using a semi-automated ImageJ/Fiji segmentation workflow, enabling consistent characterisation of fine–coarse heterogeneity across thermally affected regions. Key metallurgical descriptors including volumetric energy density (VED = P/vht), a cooling-rate proxy (v/D), and the Hall–Petch strengthening relation were incorporated to embed physical constraints directly into model learning. A hybrid PINN/XGBoost model, trained on experimental data augmented with physically bounded surrogate samples, achieved an R² of ∼0.98 under five-fold cross-validation with low prediction error. The model also preserved essential physical trends, correctly linking rapid heat extraction to grain refinement and slower cooling to grain coarsening. By coupling metallurgical physics with data-driven inference, this study demonstrates a scalable strategy for microstructure-aware process design in AM alloys and advances the application of physics-informed machine learning in materials engineering.