Physics-informed cross-scale prediction of cooling rates and grain morphology in additive manufacturing

Accurate microstructural prediction is critical for enhancing process control and mechanical performance in additive manufacturing (AM). This study proposes a hybrid deep learning–physics framework that couples a CNN-Transformer model with a cellular automaton (CA) simulator to predict spatial grain morphology in laser wire-feed additive manufacturing (LWAM). The CNN-Transformer network leverages local and global feature extraction to predict high-fidelity 10 × 20 cooling rate matrices from process parameters, trained on simulation data rigorously calibrated via XGBoost-enhanced finite element analysis. These predicted thermal fields serve as drivers for CA simulations to reconstruct grain structures, capturing spatial grain distribution and growth orientation. Experimental validation using 10CrNi3MoV high-strength steel shows that the framework accurately reproduces grain features, with a quantitative relative error of 9.5% in the grain aspect-ratio distributions compared to EBSD results. This physics-informed approach significantly reduces computational costs while maintaining high predictive accuracy, offering a robust tool for microstructural control in large-scale metal AM.