A novel transfer learning framework for predicting the elevated-temperature fatigue properties of aero-bearing steels

• A gated recurrent unit-based transfer learning framework enables accurate elevated-temperature fatigue life prediction using limited data. • The framework demonstrates robust generalization across diverse aero-bearing steels and manufacturing processes. • Experiments show temperature-driven hardness loss and carbide-controlled crack-initiation evolution governing cross-temperature transferability. • Prediction head optimization and a residual transfer strategy are employed to mitigate accuracy loss from temperature-driven domain shifts. Fatigue life prediction is essential for evaluating the service reliability of bearing steels, yet conventional physics-based models are costly, time-consuming, and often limited under complex conditions. After comparing Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) networks, we develop a novel transfer learning framework, TL-GRU, that leverages room-temperature S–N data of M50 steel to predict elevated-temperature behavior at 100 °C, 200 °C, and 300 °C. Under small-sample conditions, TL-GRU substantially outperforms Base-GRU, delivering accurate and robust S–N predictions. TL-GRU is further validated on GCr15 and CSS-42 L steels, maintaining high accuracy despite specimen and process variability. Experimental and fracture-mechanics analyses further reveal that elevated temperature reduces matrix hardness and activates coarse Mo-rich M 2 C carbides as crack-initiation sites, clarifying the defect evolution responsible for temperature-induced domain shift and degraded cross-temperature transferability. To address this discrepancy, prediction-head optimization and a residual transfer framework are introduced, effectively compensating temperature-dependent deviation and mitigating performance loss at 300 °C. Overall, TL-GRU enables accurate fatigue life prediction under limited data and elevated temperatures, and exhibits strong cross-material and cross-process adaptability, offering a powerful and broadly applicable approach for fatigue life prediction of aero-bearing steels.