An interpretable machine learning technique to predict size effect and fracture behavior of concrete

The size effect phenomenon, stemming from the inherent fracture characteristics of materials, is notably widespread in concrete. Traditional approaches to investigate this effect and concrete fracture behaviors are typically laborious. To address these issues, this study employs a machine learning (ML) methodology (e.g., light gradient boosting machine (LGBoost) and categorical boosting (CatBoost)). The findings indicate that these models deliver high accuracy in predicting the nominal flexural strength of concrete, reaching R² values of 0.933 and 0.929, respectively, for testing the models. Additionally, SHapley additive exPlanations (SHAP) analysis was used to differentiate the influence of material and geometric parameters on the predictions and determine the input parameters influences. These ML models offer a generalized framework for predicting the nominal flexural strength and fracture toughness across various materials and sizes, highlighting their utility in advancing the understanding of concrete behaviors under different mechanical stresses.