Compressive strength prediction of fly ash/slag-based geopolymer concrete using EBA-optimised chemistry-informed interpretable deep learning model

Abstract

Geopolymer concrete (GPC) is a sustainable alternative to conventional Portland cement concrete, utilising industrial by-products like fly ash (FA) and ground-granulated blast-furnace slag (GGBS). However, optimising GPC's compressive strength (CS) often requires costly and time-consuming experimental trials. This study develops a deep learning (DL) model based on convolutional neural networks (CNN) to predict the CS of FA/GGBS-based GPC. The model integrates key mix parameters such as material proportions, curing conditions, and the chemical composition of FA/GGBS binders, making it chemistry-informed. The CNN architecture includes two convolution layers, global max-pooling, and two fully connected layers, with 11 input variables and a single output for CS prediction. To optimise model accuracy, the enhanced bat algorithm (EBA) is designed for metaparameter tuning. The model is trained and tested on a comprehensive dataset comprising experimental data extracted from published literature. The results demonstrate that the EBA-optimised CNN outperforms traditional learning models, including support vector machine (SVM), extreme gradient boosting (XGBoost), and artificial neural networks (ANN), with higher performance in terms of R 2 , MAE, and RMSE. The model achieved R 2 values of 0.997 for training and 0.978 for testing. Additionally, the Shapley additive explanations (SHAP) method was used to interpret the model, identifying the Na 2 O to binder ratio and curing age as the most influential factors on CS. This study highlights the potential of DL techniques, particularly chemistry-informed CNN with metaparameter optimisation, for accurately predicting the strength of GPC, providing a cost-effective solution for mix design and performance evaluation. • A chemistry-informed interpretable model for geopolymer concrete strength prediction. • Enhanced bat algorithm is developed to optimise the model metaparameters. • Hybrid model outperforms other learning methods in terms of various evaluation metrics. • SHAP is used to analyse the contributions of different influence factors.