Comparison of deep LSTM and machine learning models for predicting compressive strength of fly ash/slag-based geopolymer concrete

In the production of geopolymer concrete (GPC), using ground granulated blast furnace slag (GGBFS) and fly ash (FA) can reduce the carbon dioxide footprint and decrease the amount of waste materials released into the environment. Finding the compressive strength (fc) of GPC through experiments is time-consuming and costly; thus, applying artificial intelligence models can expedite this process. This study aims to compare the performance of deep Long Short-Term Memory (LSTM) and the machine learning (ML)-based algorithms in predicting the f_c of FA/GGBFS-based GPC. Artificial neural networks (ANN), Bootstrap aggregating (Bagging), Least-Squares Boosting (LSBoost) and K-Nearest-Neighbours (kNN) were used for ML-based algorithms. For this goal, data were collected from the previous studies in the literature. The selected input characteristic variables included the chemical composition and quantities of FA and GGBFS, fine and coarse aggregates, sodium hydroxide molarity, alkaline activators, superplasticizer dosage, and curing temperature. Based on sensitivity analysis, the most influential parameter in the f_c of FA/GGBFS-based GPC was the fine aggregate content. Performance metrics, error percentage distribution, and Taylor diagrams indicate that the highest accuracy was achieved by LSTM, which had an R-squared value of 0.98. This was followed by ANN, LSBoost, Bagging, and kNN. Notably, LSBoost and ANN also demonstrated strong performance, with R-squared values of 0.94 and 0.95, respectively. Also, Bagging showed acceptable ability for f_c estimation of FA/GGBFS-based GPC due to having an R-squared value of 0.88, but kNN had very poor performance.