A Novel Machine-Learning Based Framework for Calibrating Micromechanical Fracture Model of Ultra-Low Cycle Fatigue in Steel Structures
Mingming Yu,Xu Xie,Zhiyuan Fang,James B. P. Lim
DOI: https://doi.org/10.1016/j.engfracmech.2024.110200
IF: 5.4
2024-01-01
Engineering Fracture Mechanics
Abstract:Calibrating micromechanical fracture model of ultra-low cycle fatigue (ULCF) life in steel structures typically involves laborious and time-consuming tests of structural steels. In response, this paper presents a novel framework for calibrating the micromechanical fracture model using the predicted ULCF life of structural steels from machine learning (ML) model. A database comprising 307 experimental ULCF life results was constructed for training the ML models, considering variations in steel grades (ranging from 235 to 690 MPa), structural dimensions, and loading conditions. In total, six trained ML models were compared in terms of predictive performance. From the comparison, it was shown that the developed eXtreme Gradient Boosting (XGBoost) model demonstrated excellent predictive accuracy and exhibited the highest R2 and the lowest MAPE outperforming other ML models in predictions. The developed XGBoost model was further interpreted using SHAP (Shapley Additive exPlanations) method to identify the main influencing variables. Then, the generalization of the developed XGBoost model was validated in low carbon steels in other data source, showing acceptable predictive accuracy. Additionally, the parameters of the cyclic void growth model (CVGM) were calibrated for Q355qC and AP50 steels, aligning with the corresponding values reported in the literature. The application of the calibrated parameters in predicting ULCF life of steel bridge piers was verified by comparing with experimental results.