Leakage-Controlled and Applicability-Domain-Aware Machine Learning Optimization of Fly Ash-GGBS Geopolymer Concrete
Fly ash-ground granulated blast-furnace slag (GGBS) geopolymer concrete is a promising low-clinker binder technology, yet its compressive strength is governed by coupled precursor chemistry, activator dosage, liquid-solid balance, aggregate proportioning, curing history, and testing age. These couplings make purely empirical mixture design expensive and make conventional random-split machine-learning validation prone to optimistic conclusions when repeated or compositionally related mixtures occur in compiled databases. This study presents a leakage-controlled, physics-guided, explainable, and applicability-domain-aware workflow for compressive-strength prediction and multi-objective mixture screening of fly ash-GGBS geopolymer concrete. A restarted database containing 548 records, 13 raw input variables, and compressive strength was audited and expanded using 33 physics-guided descriptors representing alkali activation, calcium–silicate balance, liquid-solid proportioning, curing intensity, age transformations, and nonlinear interaction terms. Exact duplicate records were removed before modeling, and a composition-level grouped split was used to evaluate generalization to unseen mixture families, giving 444 training rows, 98 held-out test rows, 124 training groups, 32 test groups, and zero train–test group overlap. Raw-feature SVR, XGBoost, and CatBoost models were compared with physics-guided XGBoost and CatBoost variants. Raw-feature CatBoost achieved the strongest held-out global accuracy, with R² = 0.903, RMSE = 5.915 MPa, and MAE = 4.207 MPa, while the physics-guided CatBoost surrogate was retained for mechanism-aware diagnosis and downstream constrained optimization. Regional analysis showed that physics-guided descriptors reduced low-strength-tail RMSE by 10.19% but worsened high-strength-tail RMSE by 16.04%, revealing that global accuracy masked uncertainty in the performance region most relevant to optimization. SHAP analysis linked these tail behaviors to curing-age interactions, Ca/Si balance, alkali–aluminate balance, chemistry-reactivity descriptors, and high-strength extrapolation effects. An NSGA-II optimization using compiled material cost and CO₂ factors then produced 260 Pareto candidates spanning 31.67-92.75 MPa predicted strength, 0.191-0.277 USD/kg binder material cost, 0.349-0.444 kg CO₂/kg binder material emissions, and 15.94-382.43 curing-severity units. Crucially, all optimized candidates were outside the 99% nearest-neighbor applicability-domain threshold. The results demonstrate that high apparent predictive accuracy is insufficient for responsible data-driven geopolymer mixture design; optimization outputs should be filtered through leakage-aware validation, regional reliability diagnosis, explainable feature attribution, and applicability-domain screening before being advanced to laboratory validation.