Transfer learning-based prediction of high-temperature fatigue performance in Fe-based structural alloys with limited data

High-temperature fatigue life serves as a crucial performance metric for assessing the structural integrity and service safety of structural materials under elevated-temperature conditions. However, conventional assessment approaches rely on high-temperature fatigue testing, which is typically time-consuming, costly, and experimentally challenging. Consequently, the scarcity of reliable high-temperature fatigue data poses a major obstacle to accurate fatigue life prediction and limits the effectiveness of conventional machine-learning models. In this study, a transfer-learning framework is proposed to address this data limitation by leveraging data-rich room-temperature fatigue datasets. An optimal feature subset is used with a Gradient Boosting Decision Tree model, and limited high-temperature samples are progressively incorporated through incremental retraining to enable effective knowledge transfer across temperature domains. The results demonstrate significantly improved predictive accuracy and confirm the consistency of the dominant features across room and high temperatures. Overall, the proposed framework offers a practical strategy for materials design and high-temperature fatigue-life assessment in engineering applications.