Machine Learning Framework with Posthoc Explainability for Predicting the Strength of Mechanochemically Activated Geopolymer Paste

Abstract

Abstract Geopolymers offer a low-carbon alternative to ordinary portland cement (OPC); yet their industrial adoption is hindered by complex multivariate design spaces exceeding 10 6 possible formulations and reliance on empirical optimization. This study integrates mechanochemical activation (MCA), central composite design (CCD), ensemble machine learning (ML), and explainable artificial intelligence (XAI) to systematically optimize ground granulated blast furnace slag (GGBS)-based geopolymer pastes and derive quantitative design guidelines. Fifteen CCD experiments combined with 1,036 literature records (total n = 1,051 ) were used to train and validate seven predictive models. Among seven models evaluated, categorical boosting (CatBoost) achieved the best ensemble performance ( R 2 = 0.939 , root-mean-square error ( RMSE ) = 2.27 MPa , MAPE = 0.37 % ) and was selected for posthoc explainability analysis; support vector machine (SVM) achieved higher test R 2 (0.971) but is incompatible with TreeSHAP-based interpretation. Multiscale explainability analysis via SHapley Additive exPlanations (SHAP), partial dependence plots (PDPs), and individual conditional expectation revealed that temporal parameters dominate strength evolution, with curing time accounting for 67.3% of total model explainability and temporal-chemical interactions collectively explaining 96.4% of predictive variance. Precise optimal ranges were identified: activator molarity 10–12 M, SS/SH ratio 3.5–4.5, sodium silicate 50 – 120 kg / m 3 , curing duration ≈ 35 days (diminishing returns beyond), constraining viable formulations from > 10 6 to ∼ 10 2 combinations. MCA-enhanced 3-day strength to > 40 MPa but required optimized chemical-temporal conditions for full effectiveness. An open-source graphical interface enables real-time prediction with integrated explainability, translating black-box models into actionable design tools. This framework advances geopolymer optimization from empirical practices toward transparent, mechanistically informed, and industrially deployable protocols for sustainable construction binders.

The authors' abstract, as published at the source. Journal of Computing in Civil Engineering, 2026 · DOI ↗