A defect and multi-physics informed neural network for fatigue life prediction in friction stir welded additive manufacturing AlSi7Mg

Additively manufactured aluminium alloys joined by friction stir welding exhibit great potential for large-scale structural applications, yet their fatigue life prediction remains challenging due to the complex influence of internal defects and process-induced microstructural heterogeneity. To address this issue, a defect-based, physics-informed neural network framework was proposed for fatigue life prediction of additively manufactured AlSi7Mg friction stir welded joints. A multi-route physics-informed modeling strategy is introduced, incorporating physical prior knowledge through feature engineering, loss function design, and adaptive feature weighting. Four data augmentation methods are employed to enhance data diversity and model robustness, and the results show that among these approaches, Precise Physics Law (PPL) yields the best overall performance. The proposed Defect and Multi-Physics Coupled Informed Neural Network (DM-PINN) dynamically adjusts the weights of physical loss terms based on defect and multi-physics features, achieving a better balance between data-driven learning and physical consistency. Experimental results demonstrate that the DM-PINN model outperforms both conventional artificial neural networks and standard physics-informed neural networks in terms of prediction accuracy, generalization capability, and physical interpretability. The model successfully predicts conditional fatigue limits under different processing conditions with errors below 15%, providing a reliable tool for fatigue life assessment and process optimization of defect-sensitive additive manufacturing components.