Explainable hybrid stacked ensemble of LSTM and DNN with SHAP-driven insights for predicting geopolymer concrete compressive strength for sustainable constructions

Abstract

As the world look forward to reduce carbon emissions, it leads to the adoption of sustainable construction materials, with geopolymer concrete (GPC) emerging as a green alternative to traditional portland cement concrete. GPC, produced using industrial waste materials like fly ash and slag, offers superior chemical resistance, durability, and strength, significantly reducing environmental impact. Analyzing a dataset of 820 geopolymer mixes, critical parameters were identified and evaluated. Traditional models often tend to identify the linear trends rather than non-linear relationships present in the data leading to inaccurate predictions. This study aims to study the linear and non-linear trends simultaneously for best accurate performance. This study predicts the compressive strength of fly ash-slag GPC by the utilization of advanced machine learning models, namely, Random Forest (RFR), Gradient Boosting (GBR), Long Short-Term Memory (LSTMs) and Deep Neural Networks (DNN). A novel heterogeneous ensemble stacking framework integrating base models namely, RFR, GBR, and LSTMs through a DNN meta-learner is proposed to simultaneously capture complementary non-linear relationships among geopolymer concrete mix parameters. The study also explores the effect of hyperparameter optimization of the proposed model using different optimization measures for GPC predictions. Selecting the best optimal parameters can enhance the performance of predictions. Hybrid Techniques like Optuna, Bayesian optimization combined Hyperband, Bayesian optimization combined with random search is utilized. Among the optimization models, Optuna outperformed both Bayesian methods, achieving the lowest errors with a Mean Squared Error value of 3.16 MPa, Root Mean Square Error of 1.77 MPa, Mean Absolute Error of 1.07 MPa and the highest R² score of 0.9890 during testing, making the most effective approach for a balance in trade-off between prediction accuracy and minimize computational cost. To enhance model interpretability, Shapley Additive exPlanations (SHAP) analysis was employed, revealing slag content, NaOH molarity, and curing temperature as the most influential factors affecting compressive strength. These insights enable optimized mix designs, highlighting robust prediction accuracy and the potential to guide sustainable, high-performance concrete mixtures.