Interpretable and physics-informed machine learning for predicting the compressive strength of geopolymer concrete

Geopolymer concrete (GPC) synthesized from fly ash (FA), ground granulated blast-furnace slag (GGBS), and, in part of the mixes, glass powder (GP) offers a low-carbon alternative to ordinary Portland cement concrete, yet its compressive strength (CS) depends on numerous nonlinearly interacting mix and activator variables poorly captured by conventional empirical equations. This study compiles a multi-source database of 713 ambient-temperature FA-GGBS GPC specimens (with and without GP) from 34 studies, including the authors’ own experiments, and reduces it via deduplication to 251 unique mix designs. Seven machine learning algorithms (CatBoost, MLP, LightGBM, XGBoost, Gaussian Process Regression (GPR), Extra Trees, and Stacking) were benchmarked under ten-fold cross-validation and leave-one-study-out validation. (CatBoost achieved the highest mean accuracy (R ²  = 0.8291 ± 0.0647, RMSE = 7.93 MPa), while GPR intervals covered 93.2% of the held-out mixes at the nominal 95% level and conformal intervals for CatBoost reached the nominal coverage at every level. Under leave-one-study-out validation the pooled R ² of CatBoost was 0.11, so the random-split accuracy applies to mixes from studies represented in the database. A physics-informed neural network enforcing the constraint dCS/dWB < 0 reduced the share of held-out mixes with a predicted increase of CS with water-to-binder ratio (WB) from 45% to 7–21% without loss of accuracy (R ²  = 0.81–0.82). SHAP identified the SS/SH-WB interaction as the leading local predictor, while Sobol sensitivity analysis ranked the GGBS share of the binder first on four surrogate models when only feasible binder compositions were sampled (S  = 0.40–0.48); the SS/SH ratio led only under independent sampling. The proposed CatBoost model, with SHAP and Sobol interpretability layers, offers rapid mix-design support for FA-GGBS GPC, with and without GP, within the range of the compiled data.