A unified theory and physics-guided machine learning framework for thermomechanical fatigue life prediction of nickel-based superalloys

Accurate thermomechanical fatigue (TMF) life prediction is critical for the safe service of high-temperature components. To address the limited accuracy and applicability of conventional models under complex thermo-mechanical coupling, a TMF database covering polycrystalline, single-crystal, and directionally solidified Ni-based superalloys was established. A unified TMF life model (UTLF) was developed by incorporating strain, temperature, phase relation, and dwell time into a single characterization framework. Compared with the classical Manson–Coffin model, the UTLF showed lower scatter, with most predictions falling within the five-fold error band. Furthermore, a UTLF-constrained LSTM-PINN framework was proposed by embedding the life-correlation model into network training as a prior constraint. Among the compared models, including SVM, RF, LSTM, and PINN, the proposed model achieved the best overall performance, with most predicted lives located within the three-fold error band. SHAP analysis further showed that the dominant variables identified by the model were consistent with the key controlling factors in TMF life evolution. The proposed method provides a practical approach for TMF life assessment of Ni-based superalloys under complex loading conditions.