Physics informed ensemble learning used for interval prediction of fracture toughness of pipeline steels in hydrogen environments

Abstract

Hydrogen embrittlement leads to a loss of plasticity and a decrease in fracture toughness in pipeline steels. Fracture toughness is a critical factor in the design and safety assessment of hydrogen-carrying pipelines. However, conducting large-scale fracture toughness testing on hydrogen-charged materials is time-consuming and expensive. Analytical approaches based on micromechanics often involve numerous undetermined (or assumed) parameters. Therefore, this study presents an approach that combines physical information and ensemble learning for predicting the fracture toughness range under hydrogen environments. Initially, detailed experimental data are gathered from a series of fracture tests conducted on pipeline steels under hydrogen environments, as published in the literature. Subsequently, Given the limited availability of experimental data, a feature space was constructed, taking into account the physical information as well as the distributions of yield strength and tensile strength. Furthermore, a predictive model based on Population-Based Training (PBT) optimization and Category Boosting (CatBoost) methodology was developed. Comparative experimentation substantiated its superior predictive accuracy in comparison to Random Forest, Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). The MAE, MSE, R2, VAF, a20_index and MAPE, was 0.02139, 0.00096, 0.98741, 98.98%, 0.83 and 0.1552. Building upon these foundations, this study proposes an interval prediction method by combining Monte Carlo simulation with point prediction results to obtain fracture toughness interval prediction results with different confidence intervals (80%, 85%, 90% and 95%).