SHAP-Based Explainable Machine Learning Framework for Interpreting SHCC Mix Design Parameters

Artificial intelligence advancements have generated significant potential for concrete mix design optimization; however, translating machine learning predictions into practical engineering knowledge remains challenging. This investigation presents an explainable machine learning framework that addresses this limitation through application to strain-hardening cementitious composites (SHCC). Utilizing Shapley Additive exPlanations (SHAP) methodology, an analysis of 266 experimental samples incorporating various mineral admixtures and fiber types was conducted. The CatBoost algorithm demonstrated superior predictive capability (R²=0.957) while facilitating transparent interpretation of parametric relationships. SHAP analysis identified cement content as the primary determinant of compressive strength development, with water-to-binder ratio and fly ash content constituting secondary influential factors. Experimental results indicated optimal compressive strength was achieved with water-to-binder ratios not exceeding 0.3, fly ash content approximately 600 kg/m^³, and fiber volume fractions between 0.2-0.3%. Additionally, the investigation revealed that shorter fibers with larger diameters provided enhanced contribution to compressive strength development. These findings extend the fundamental understanding of SHCC behavior while establishing quantitative guidelines for sustainable mix design optimization. The methodological framework developed herein effectively transforms complex algorithmic outputs into practical engineering knowledge, providing valuable analytical capabilities for both academic researchers and industry practitioners in cementitious materials development.