Comparative evaluation of machine learning and deep learning approaches for compressive strength prediction of geopolymer concrete

Abstract Geopolymer concrete has emerged as a promising sustainable alternative to ordinary Portland cement concrete due to its lower environmental impact and favorable engineering performance. However, accurately predicting its compressive strength remains challenging because it is governed by complex interactions among binder composition, activator characteristics, curing conditions, and other mixture parameters. Therefore, reliable data-driven predictive models are needed to support efficient design and performance estimation. For this the study develops a data-driven framework for predicting the compressive strength of geopolymer concrete using both machine learning (ML) and deep learning (DL). A unified database of 594 geopolymer mix designs was compiled from published literature, covering key mix parameters (binder constituents, activator contents, aggregates, water/superplasticizer, fiber contents, and curing age). The dataset was cleaned and divided into 80% training and 20% testing for unbiased evaluation. Multiple supervised ML models (SVM, Decision Tree, Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost) were trained and assessed using MAE, MSE, R 2 , adjusted R 2 , and MAPE, then compared to identify the most reliable predictor. Among ML approaches, CatBoost achieved the best overall performance (MAE = 14.740, MSE = 759.484, R 2 = 0.968), indicating high accuracy and strong generalization. On the other hand, Decision Tree showed the weakest results among this set, reflected by higher MAE (21.842) and lowest R² (0.912). DL benchmarks were also implemented using ANN Shallow, ANN Deep, and TabNet; the TabNet model emerged as the strongest DL performer, achieving MAE = 18.27, MSE = 765.80, and R 2 = 0.968, effectively rivaling the best-performing ML approach. Visually, TabNet demonstrated the tightest clustering around the 45° reference line, whereas the ANN Deep model showed the weakest fit and the highest overall error (MAE = 30.15, R 2 = 0.9106). To enhance interpretability, SHAP was applied to quantify feature contributions and identify the governing variables affecting compressive strength; the results indicate that prediction is dominated by binder and curing related variables, with fibers and minor constituents providing secondary refinements. The study provides a practical predictive tool for strength estimation and supports more efficient geopolymer mix optimization.