Interpretable Machine Learning Method for Modelling Fatigue Short Crack Growth Behaviour

Interpretable machine learning (ML) has become a popular tool in the field of science and engineering. This research proposed a domain knowledge combined with ML method to increase interpretability while ensuring the accuracy of ML models and verifies the generality of the ML approach in fatigue crack growth (FCG) modelling. LZ50 steel single edge notch tension (SENT) specimens were tested for short crack (SC) growth rate and microstructure characterization under various R-controls. Based on the test results, the SC growth process was divided into 3 stages: microstructural short crack (0–145 μm), physical short crack (145–1000 μm), and long crack (1000 μm–fracture). Following the analysis of 8 semi-empirical FSCG rate equations with different driving forces, 6 impact variables that may affect the FCG rate characteristics were identified. Random forest and Pearson correlation analysis were used to investigate the influence of each feature on the FCG rate and the relationships among the features. The main influential features for the short crack symbolic regression (SCSR) model were found to be |ΔK–ΔK_at|, Δγ_xy, |a–a_t|, and e^α(1−R). After considering these 4 input features, the predicted FSCG rate equation generated by the SR model has a concise mathematical structure. Finally, an elastic net multiple linear regression method was proposed to determine the parameters of the predicted equation, while retaining the physical characteristics of each parameter. The SCSR model for SC demonstrated good prediction performance on various metallic materials.