Data-Driven Discovery of Process–Structure Relationships in Additive Manufacturing via Featurization from Kinetic Monte Carlo Simulations and Interpretable Machine Learning

Understanding process–structure linkages is essential for accelerating microstructure design in additive manufacturing (AM). Here, we develop a computational framework that integrates automated feature extraction from kinetic Monte Carlo (kMC) simulations with interpretable machine learning (ML) models to predict grain-scale morphological features from processing conditions. A dataset of 1,524 simulated three-dimensional polycrystalline microstructures was analyzed to extract descriptors including grain size, surface-to-volume ratio, sphericity, and roundness. These were correlated with process parameters such as scan velocity and heat-affected zone (HAZ) dimensions. Four ML models—Random Forest, Gradient Boosting Machine, Extreme Gradient Boosting (XGBoost), and neural networks—were implemented to benchmark predictive accuracy and interpretability. The ensemble-based models were chosen for their robustness and efficiency on structured datasets, while neural networks were included to capture nonlinear and higher-order correlations. Among them, XGBoost achieved the highest predictive performance (R² = 0.977, MAE = 6.6 pixels). Model explainability was provided using Shapley Additive Explanations (SHAP), revealing that scan velocity and HAZ parameters strongly govern grain morphology. By coupling interpretability with physics-informed learning, our framework bridges high-fidelity simulation and rapid prediction, enabling scalable microstructure design in additive manufacturing.