An Interpretable Hybrid Machine Learning and Deep Learning Framework for Predicting Compressive Strength of Recycled Powder Mortar

The accurate prediction of compressive strength (CS) is a fundamental requirement for structural safety, yet conventional empirical models remain limited in their ability to capture the complex nonlinear response of cementitious materials. Hybrid machine learning (ML) and deep learning (DL) models are investigated for predicting the CS of recycled powder mortar (RPM) using 10-fold cross-validation. Material type (classified as kind), particle size, water/binder (W/B), and mass replacement ratio (MRR) are used as input variables, with CS as the target response. Model performance was quantified using RMSE, R², MAE, and MAPE, and analysis of variance (ANOVA) was conducted to examine the statistical significance of the input variables. Based on 10-fold cross-validation, the SVR–PSO model exhibited the highest average predictive performance, attaining R² = 0.950 and RMSE = 2.042 MPa, thereby demonstrating superior accuracy and reliability relative to the other evaluated models. The interpretable hybrid ML–DL framework achieved high accuracy and robust generalization across folds, with error and stability analyses confirming unbiased predictions. MRR is the dominant factor controlling CS, with suggested optimal ranges; this influence is consistently supported by ALE, PDP, SHAP, and ICE analyses. Feature importance and PDP demonstrate Kind has a secondary yet meaningful effect on CS, reflecting differences in composition, morphology, and reactivity, with its subtler role in ALE and ICE analyses arising from its categorical and indirect mechanistic influence on hydration and pozzolanic activity. The framework offers a robust, interpretable tool for predicting CS, supporting data-driven mix design and sustainable mortar optimization.