CINAS-PINN: Causal inference-based neural architecture search in physics-informed neural networks for fatigue life prediction with welding strain energy

Welded joints are critical components in engineering structures, yet accurate fatigue life prediction remains challenging due to multiaxial loading complexity and material nonlinearity. Conventional physics-based models often fail to capture intricate load–material interactions, while data-driven approaches demand extensive datasets and lack physical interpretability. To address these limitations, this study introduces CINAS-PINN, a causal inference-based neural architecture search integrated with physics-informed neural networks for welded joint fatigue life prediction. By constructing equivalent tensors, multiaxial load paths are converted into scalar strain energy densities, aiming to capture the physical characteristics of multiaxial loading and provide input support for neural networks. We integrate causal inference with neural architecture search (NAS) in physics-informed neural networks (PINNs). Based on this, we implemented the PINN structure and optimized the model parameters, addressing the challenge of accurate fatigue life prediction. To overcome issues related to poor model interpretability and low accuracy, we employed a causal graph-constrained architecture, enabling the model to focus on key physical factors. Additionally, a dynamic loss function, adjusted through Granger causality analysis, prioritizes key physical constraints during training, improving model efficiency and physical consistency. Case studies on AISI316L, GH4169, and TC4 alloys demonstrate that CINAS-PINN achieves superior accuracy, reducing prediction errors by more than 30% compared with benchmark methods. The proposed framework offers enhanced physical consistency, robustness, and generalization for fatigue life prediction under complex service conditions.