多尺度条件下，传统纯数据驱动方法在疲劳裂纹扩展预测中存在物理约束不足、模型可解释性有限及预测精度不佳的问题。为了准确预测多尺度疲劳裂纹扩展行为，并提升模型的物理一致性、可解释性和泛化能力，本工作选用304奥氏体不锈钢为研究对象，基于人工神经网络(ANN)进行含物理信息的特征扩展构建了特征扩展神经网络(FENN)模型，并在FENN模型基础上，进一步耦合多尺度疲劳裂纹扩展模型构建了物理信息神经网络(PINN)模型。结果表明，与轻量梯度提升机(LGBM)、极限梯度提升(XGBoost)和岭回归(RR)算法相比，ANN算法更适用于预测多尺度疲劳裂纹扩展行为。相较于ANN模型，引入物理信息特征的FENN模型的训练集和测试集精度分别提升了7.23%和18.75%，耦合了多尺度疲劳裂纹扩展模型的PINN模型预测精度更高。PINN模型能够有效捕捉实际疲劳裂纹扩展行为在不同尺度下的复杂规律，具有良好的误差可控性、鲁棒性和泛化能力。此外，利用夏普利加性解释(SHAP)方法分析了PINN决策过程，并量化了每个特征对预测结果的贡献。

Although machine learning methods have demonstrated promising potential in fatigue crack propagation prediction, they are currently limited by insufficient incorporation of physical constraints, low model interpretability, and suboptimal predictive accuracy. These limitations restrict the reliability of data-driven models on complex engineering structures, in which fatigue crack propagation is governed by nonlinear interactions across multiple length scales and loading conditions. When explicit physical information is lacking, models may output unstable predictions and cannot be generalized by extrapolating beyond the training data domain. To accurately predict multiscale fatigue crack propagation behavior while enhancing the physical consistency, interpretability, and generalization capability, the present authors investigated the multi-scale fatigue crack propagation behavior of 304 austenitic stainless steel samples using an artificial neural network (ANN). As purely data-driven models cannot properly represent complex crack propagation across different scales, the authors incorporated physics-informed features into the ANN framework, creating a feature-extended neural network (FENN) that incorporates prior physical knowledge related to fatigue crack propagation. Next, a multi-scale fatigue crack propagation mechanism was coupled into the PENN model, establishing a physics-informed neural network (PINN) model that can explicitly integrate physical constraints into the learning process. The ANNs more accurately predicted multi-scale fatigue crack propagation than the traditional light-gradient boosting machine, extreme gradient boosting, and ridge regression machine learning methods. The neural-network-based models well-fitted the nonlinear data of fatigue crack growth data under multi-scale conditions. Expanded with physics-informed features, the FENN model improved the prediction accuracy by 7.23% and 18.75% on the training and testing datasets, respectively, relative to the baseline ANN, confirming the effectiveness of physical feature embedding. Incorporating the physical information into the learning framework significantly enhanced both the training performance and generalization capability. The prediction accuracy was further improved in the PINN model, which integrates multi-scale physical mechanisms to capture the complex patterns of fatigue crack propagation across different scales. The PINN model is robust with good error controllability and high generalization capability. This performance improvement demonstrates the advantage of combining data-driven learning with explicit physical constraints into multiscale fatigue crack propagation prediction. To improve the transparency and physical interpretability of the model, the decision-making process of the PINN model was interpreted through the SHapley Additive exPlanations (SHAP) method, which quantitatively evaluates the contribution of each input feature to the model’s prediction. The SHAP-based analysis determines the relative importance of different input variables, further supporting the physical rationality and reliability of the proposed PINN model.