A multi-fidelity physics-informed machine learning framework for probabilistic low-cycle fatigue life prediction of shot-peened materials

Inherent randomness in the shot peening process induces significant dispersion in surface integrity (SI) parameters. Due to the limitation of high-fidelity (HF) fatigue test data, severe challenge exists in constructing high-precision probabilistic low-cycle fatigue (LCF) life prediction for turbine disks processed by shot peening. In this study, a multi-fidelity physics-informed machine learning (MF-PIML) framework integrating physical mechanisms with multi-source data is proposed. First, through LCF tests and SI characterization across four typical surface conditions, the governing mechanisms of surface roughness and residual stress distribution on the scatter of life are revealed. Accordingly, SI correction terms are introduced to modify the classical Smith-Watson-Topper (SWT) model, establishing a physical baseline with predictions falling within the scatter band of 2.8. Subsequently, a mean–variance dual-output neural network based on transfer learning is constructed. Abundant low-fidelity (LF) finite element simulation data are utilized for pre-training to capture underlying physical mapping laws, followed by fine-tuning with sparse HF experimental data. Concurrently, physical constraints and negative log-likelihood terms are embedded into the loss function to ensure physical consistency and enable uncertainty quantification. Validation results demonstrate that the MF-PIML model successfully narrows the life prediction accuracy to within the scatter band of 1.8. Furthermore, SHapley Additive exPlanations (SHAP) analysis is employed to quantify the contributions of key SI parameters, significantly enhancing model transparency and interpretability. This study provides an efficient and reliable predictive tool for the design and process optimization of high-performance turbine disks.