Analyzing the mechanical performance of fly ash-based geopolymer concrete with different machine learning techniques

Fly ash (FA)-based geopolymer concrete is considered as an alternative system with potentially lower environmental impact than Portland cement mixes. However, the prediction accuracy of compressive strength still needs to be improved. This study demonstrated the feasibility of predicting the 28-day strength of geopolymer concrete through mix proportions and pre-curing conditions by using three machine learning algorithms (back-propagation neural network (BPNN), support vector machine (SVM) and extreme learning machine (ELM)) and provided a comparison of their differences, highlighting variations in prediction accuracy. As a part of the evaluation of model performance and error analysis, the prediction accuracy differences of these three models in training, validation and testing sets were discussed, and the influence weight of each input parameter on results was analyzed by permutation feature importance (PFI) index. Results showed that all models revealed good prediction performance for the overall database. BPNN model had the largest number of instances where the error percentage was within +/- 20%. SVM model showed the highest generalization capability and most stable prediction accuracy among all three. Out of different variables investigated, SiO2 content in FA had the highest influence on strength, followed by Al2O3 content and activator content/concentration. These outcomes can enable reductions in experimental time, labor, materials and costs; and facilitate the adoption of alternative binders in the concrete industry.