Accurate prediction of fatigue crack growth rate based on physics-informed transfer learning

Predicting fatigue crack growth (FCG) rates, particularly based on limited FCG data, remains a persistent challenge in engineering. Traditional models, such as the Paris law, often suffer from limited accuracy because of their simplified assumptions and semi-empirical nature. To address this issue, this study develops a physics-informed transfer learning model for FCG-rate prediction. More specifically, a pretraining-finetuning framework is introduced to adaptively select high quality pretraining data through similarity identification, thereby alleviating the scarcity of available FCG data. Meanwhile, a temporal data-driven predictor is further improved by embedding physical constraints into both the feature and output spaces, enabling it to provide continuous predictions with non-decreasing FCG trajectory. In six evaluation tasks that emulate early-warning scenarios, most predictions from the proposed model fall within the ± 5% error band. The results demonstrate that the proposed model not only furnishes good accuracy across diverse FCG scenarios but also maintains strong physical consistency, highlighting its potential for precisely characterizing complex FCG processes.