A Bayesian physics-informed neural network model for probabilistic fatigue life assessment considering the size effect in additively manufactured materials

Abstract

This study investigates the pronounced size effect in high-cycle fatigue of additively manufactured metallic specimens and proposes a probabilistic fatigue life assessment approach that combines a size-effect physical model with a Bayesian neural network (BNN). The model takes two physical-model parameters and the applied stress level as inputs, with these two parameters jointly capturing the influences of specimen size, defect characteristics, and Vickers hardness. By normalizing these inputs, the BNN is able to learn fatigue response patterns in a concise yet comprehensive manner. Statistical fatigue test data are used to construct training, validation, and test sets, followed by a systematic hyperparameter search to determine the optimal model configuration for the dataset in this study. Cross-validation is then performed on three different materials with a total of eight specimen sizes. The results show that the proposed Bayesian Physics-Informed Neural Network model delivers reliable fatigue life predictions across multi-material and multi-size conditions, demonstrating strong generalization capability.