Physics-Guided Explainable Surrogate Framework for Topology-Aware Local Strain Prediction in Defect-Rich Additively Manufactured Materials

Abstract

Predicting defect-driven strain localization in additively manufactured materials remains challenging because experimental datasets are often limited, expensive to generate, and characterized by complex defect topologies. Conventional machine learning models can provide accurate predictions but often lack physical interpretability and uncertainty quantification. To address these limitations, this study proposes a physics-guided and explainable surrogate learning framework for topology-aware damage evolution and localized strain prediction. The framework integrates physics-guided defect descriptors, surrogate-assisted data augmentation, Gaussian Process Regression (GPR), and SHapley Additive exPlanations-based explainable artificial intelligence (AI) to enable accurate, interpretable, and uncertainty-aware predictions under small-data conditions. Physically meaningful descriptors, including topology damage, defect concentration, inter-defect distance, heterogeneity, and damage entropy, were formulated to capture nonlinear defect interactions governing localized deformation. The proposed GPR model achieved an R² of 0.895, with a mean absolute error (MAE) of 0.017 and root mean square error (RMSE) of 0.020, outperforming the benchmark XGBoost model (R² = 0.794, MAE = 0.031, RMSE = 0.041). Explainability analysis identified heterogeneity and topology damage as the dominant factors influencing strain localization. The proposed framework provides a computationally efficient, physically interpretable, and uncertainty-aware approach for defect-sensitive strain prediction, supporting reliable small-data modelling and future digital twin applications in additive manufacturing.