Multiscale structure property learning of particulate metal matrix composites via graph transformers for continuous stress–strain behaviour prediction

Mechanical behaviour of particulate-reinforced metal matrix composites (PRMMCs) is highly sensitive to particle size, morphology, spacing, and volume fraction, yet existing predictive models rely heavily on image-based inputs, empirical correlations, or simplified regressors that fail to capture multiscale particle interactions. These limitations restrict the ability to generalize across heterogeneous composite systems and often produce stress–strain predictions that deviate from physically meaningful deformation trends. To address these challenges, this study proposes a Physics-Informed Graph Transformer (PIGT) that encodes literature-derived structural parameters into a particle-interaction graph and predicts full stress–strain curves without requiring SEM images. The method integrates a Graph Neural Network–based interaction layer, multiscale structural attention, and physics-guided deformation constraints, enabling the model to learn mechanistic behaviour rather than relying solely on statistical fitting. All simulations and model development were implemented using Python with PyTorch, providing an extensible environment for graph learning and transformer-based modelling. The dataset, compiled from peer-reviewed metallic composite studies, consists of standardized structural descriptors and digitized tensile curves. Experimental results demonstrate that the proposed PIGT achieves a 2.01 MAE, 7.05 RMSE, 2.95% MAPE, and 0.97 R 2 , showing an improvement of approximately 18–25% in predictive accuracy compared with leading baseline methods such as XGBoost, MLP, and deep ANN models. The generated stress–strain curves closely follow elastic–plastic transitions, strain-hardening behaviour, and failure onset, confirming strong physical consistency. Overall, the PIGT framework provides a novel image-free, physics-aware modelling paradigm for PRMMC behaviour prediction, offering improved accuracy, broader applicability, and enhanced interpretability for composite design and optimization.