Predicting Freeze–Thaw Resistance in Geopolymer Concrete: Interpretable ML Model with Practical Deployment

Geopolymer concrete offers a sustainable alternative to portland cement, reducing CO2 emissions by up to 80%; however, its durability in freeze–thaw (F-T) cycles in cold climates remains a critical challenge. Although prior studies have characterized geopolymer strength, few address F-T degradation, and none leveraged machine learning (ML) for predictive modeling, representing a significant research gap. This study develops a novel interpretable ML framework to predict compressive strength degradation in geopolymer concrete subjected to F-T cycles. A curated dataset of 212 samples was compiled from peer-reviewed experimental studies involving varying base materials (fly ash, ferrochrome slag), alkali activator types and dosages (NaOH, KOH, Na2SiO3), additives, curing durations, and F-T cycle regimes. Five boosting algorithms were optimized using grid search and 10-fold cross-validation. Gradient boosting outperformed others ML models with coefficient of determination R2 values of 0.984, and the lowest mean absolute error (MAE) (1.30 MPa) and mean absolute percentage error (MAPE) (2.66%) on test data. Model transparency was enhanced by employing Shapley additive explanations and partial dependence plots. Results revealed that base material content, NaOH dosage, and F-T cycle count were the most influential variables, whereas other features showed limited standalone effects. For practical deployment, a user-friendly graphical user interface (GUI) was developed, marking the first integration of an interpretable ML model into a predictive tool for geopolymer F-T durability. The GUI generates three outputs—initial strength, post-F-T strength, and strength loss—allowing users to assess both performance and deterioration. This study advances sustainable material design by providing a validated, accessible tool to support mix optimization and resilience assessment.