Machine learning-based prediction of ductility of strain-hardening fiber-reinforced cementitious composites

Abstract

The high ductility, characterized by both strain capacity and average crack spacing, of strain-hardening fiber-reinforced cement composites (SH-FRCCs) is expected to enhance the load-carrying capacity and durability of buildings and infrastructure made of SH-FRCCs. This study aimed to predict the strain capacity and average crack spacing of SH-FRCCs using four popular machine learning (ML) models: k-nearest neighbor (k-NN), decision tree (DT), random forest (RF), and adaptive boosting (ADB) models. Nine input variables, the matrix compressive strength, fiber type 1, tensile strength of fiber 1, fiber type 2, tensile strength of fiber 2, fiber index, specimen width, specimen thickness, and gauge length were considered in the ML models. Among the investigated ML models, the RF model exhibited relatively good performance in predicting the strain capacity (R2 = 0.986) and log-transform crack spacing (R2 = 0.955) of SH-FRCCs in training data. The better performance of the RF model is attributed to the model's ensemble structure, which integrates multiple decision trees, effectively reduces variance, and manages complex data structures. Fiber index is the most influential variable on both strain capacity and average crack spacing of SH-FRCCs, based on SHapley Additive Explanations (SHAP) and Partial Dependence Plots (PDP) analysis. The strain capacity of SH-FRCCs decreased with increasing specimen width and gauge length, whereas crack spacing increased with specimen width. Finally, the developed ML model was validated against experimental data, showing excellent agreement with deviations below 10 % for strain capacity and around 11 % for average crack spacing.