AI-powered interpretable models for the abrasion resistance of steel fiber-reinforced concrete in hydraulic conditions

Significant financial losses have arisen from the failure of concrete in hydraulic structures due to excessive levels of abrasion. Despite significant advancements in this field, the construction of durable hydraulic concrete structures capable of withstanding abrasion damage remains a formidable challenge. This is particularly pertinent when endeavoring to achieve a balance between economic incentives and environmental sustainability. This study utilizes variables such as hydraulic conditions, curing age, and concrete mixture proportions to develop predictive models for the attrition depth of concrete, employing machine learning approaches including gene expression programming (GEP) and multi-expression programming (MEP). Supplementary metrics utilized to assess the developed models encompassed statistical tests, R² values, Taylor's diagram, and the comparative analysis of test and forecasted components. With R 2 =0.930, MAE=0.188, MAPE=19.40%, and RMSE = 0.248, the MEP somewhat outperformed the GEP in terms of model fit and prediction accuracy. The study on SHapley Additive exPlanations (SHAP) revealed that abrasion depth exhibited a positive correlation with testing time while demonstrating an indirect (negative) relationship with both age and cement content. Consequently, various alternative mix proportions and hydraulic conditions may be examined utilizing MEP and GEP models to generate and assess diverse concrete compositions.