Physics-informed and explainable machine learning framework for performance prediction and design of Ti (C, N)-based cermets

Abstract

Ti (C, N)-based cermets exhibit an exceptional combination of ceramic hardness and metallic toughness, making them essential for high-performance applications. However, their complex multi-component nature and lengthy experimental cycles hinder rapid development. This study presents a physics-informed and explainable machine learning (ML) framework for accurate and interpretable performance prediction. A comprehensive dataset was compiled from literature (1980–2024) and in-house experiments. Dimensionality reduction was applied to extract key features while minimizing data noise. To ensure physical consistency, domain knowledge was embedded via constraints such as mass conservation, monotonic trends, and property trade-offs. Explainable artificial intelligence (XAI) tools such as SHapley Additive exPlanations (SHAP) and adopts Local Interpretable Model-Agnostic Explanations (LIME) were employed to identify globally influential features and validate local predictions. Experimental validation confirmed the framework's predictive accuracy. The proposed approach enables high-throughput, physically consistent, and interpretable modelling, offering a generalizable strategy for the intelligent design and optimization of Ti (C, N)-based cermets and other complex materials.