Interpretable Deep Learning Framework for Predicting Compressive Strength of Steel Fiber Reinforced Geopolymer Concrete

Quynh-Anh Thi Bui^1,*, Son Hoang Trinh¹, Maryam Sayadi², Reza Khanali³
1 Research group on Industry 4.0 in Transportation (I4T Group), University of Transport Technology, Trieu Khuc, Thanh Liet, Hanoi, Vietnam
2 Division of Water Resources Engineering, Lund University, Lund, Sweden
3 Department of Water Resource, University of Tabriz, Tabriz, Iran
* Corresponding Author: Quynh-Anh Thi Bui. Email: email

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.081794

Received 09 March 2026; Accepted 27 April 2026; Published online 19 May 2026

Download PDF

Abstract

Geopolymer concrete has attracted increasing attention as a lower-carbon alternative to ordinary Portland cement concrete because it can utilize aluminosilicate-rich industrial by-products while still achieving satisfactory mechanical performance. However, the 28-day compressive strength of steel fiber-reinforced geopolymer concrete (SFGPC) is governed by multiple interacting mixture variables, which makes reliable prediction difficult, especially for medium-sized experimental datasets. This study developed an interpretable deep-learning framework to predict the 28-day compressive strength (CS28) of SFGPC using an original experimental dataset of 189 mixtures produced under a consistent laboratory protocol in Vietnam. The dataset covered nine mixture variables, including activator chemistry (NaOH, Na₂SiO₃, and Na₂O), binder constituents (Fly Ash (FA), Ground Granulated Blast-Furnace Slag (GGBS), and silica fume (SF), aggregate contents, and steel fiber dosage. Three tabular models, namely Tabular Deep Polynomial Transformer (TabDPT), Tabular Mixer (TabM), and Extreme Gradient Boosting (XGBoost), were trained and evaluated. Among them, TabDPT achieved the best performance on the independent test set, with R² = 0.978 and RMSE = 2.214 MPa, while also showing more stable behavior under repeated 5-fold cross-validation than XGBoost. SHapley Additive exPlanations indicated that binder composition and activator chemistry were the dominant variables in the trained model, with GGBS and Na₂O showing the strongest influence, whereas steel fiber dosage had a comparatively smaller contribution to CS28 within the investigated domain. Totally, the proposed framework can support preliminary mixture screening and strength-oriented design of SFGPC within the investigated material and curing domain.