Soft computing-based prediction models for compressive strength of concrete

The complexity of concrete's composition makes it difficult to predict its compressive strength, which is a highly valuable and desired characteristic. Traditional methods for prediction are expensive and time-consuming, resulting in limited data availability. However, modern soft-computing models have emerged as a reliable solution for accurately forecasting concrete's compressive strength. The research proposes a novel Deep Neural Network (DNN), Multivariate Adaptive Regression Splines (MARS) and Extreme Learning Machine (ELM) based machine learning (ML) models for forecasting the compressive strength of concrete added with various proportions of fly ash and silica fume. For this purpose, a dataset of 144 trials, having 8 input parameters is taken from the literature. The performance of the models is confirmed using various statistical parameters. Rank Analysis reveals that DNN is the best-performing model (Rank =52, RTR2=0.983 and RTs2=0.954), closely followed by MARS (Rank =38, RTR2=0.974 and RTs2=0.956); while ELM lags behind the other two counterparts. The results are further confirmed using an error matrix, external validation and AIC criteria. The visual interpretation is provided using the Taylor diagram. MARS has the edge over the other two models in terms of providing a user-friendly solution.