Creep-Fatigue Life Prediction of 316H Stainless Steel through Physics-Informed Data-Driven Models

Abstract

316H stainless steel is a critical material for fourth-generation nuclear reactors, yet it is prone to creep-fatigue failure under high-temperature and high-pressure conditions. This study evaluates physics-driven models (including time fraction model, ductile exhaustion model, modified strain energy density exhaustion model, and plastic strain energy model) and data-driven models (including support vector regression, random forests, generalized regression neural networks, and backpropagation neural networks) for predicting the creep-fatigue life of 316H base metal and welded joints. On the basis of data-driven models, physical information from the creep-fatigue damage is further integrated to embed the physics-informed input features and the physics-informed loss function, thereby constructing physics-informed data-driven models  to predict creep-fatigue life. Results demonstrate that physics-informed data-driven models significantly outperform conventional approaches, with the physics-informed generalized regression neural network achieving the highest accuracy (R² = 0.9277). This work provides a robust framework for enhancing life prediction in high-temperature structural applications.