Deep learning and machine learning-based prediction of capillary water absorption of hybrid fiber reinforced self-compacting concrete

Abstract

Deep auto-encoders and long short-term memory methodology (LSTM) based on deep learning as well as support vector regression (SVR) and k-nearest neighbors (kNN) based on machine learning models for the capillary water absorption prediction of self-compacting concrete (SCC) with single and binary, ternary, and quaternary fiber hybridization were developed. A macro and two types of micro steel fibers having different aspect ratios, and PVA fiber were used. One hundred and sixty-eight specimens produced from 24 mixtures were used in the prediction models. The input was the content of cement, fly ash, silica fume, fine and coarse aggregate, water, superplasticizer (SP), macro and micro steel fibers, PVA, time that the specimen was immersed in water, and splitting tensile strength. Water absorption was used as output. As per the ANOVA analysis of the experiment results, the most effective parameters were macro steel fiber and time for tensile strength and water absorption, respectively. Finally, binary hybridization of 1% macro steel fiber and PVA improved the splitting tensile strength while the use of PVA as binary, ternary, and quaternary fiber hybridization increased the water absorption of SCC specimens. The auto-encoder, LSTM, SVR, and kNN models predicted the water absorption of fiber reinforced SCC with 99.99%, 99.80%, 94.57%, and 95.50% accuracy, respectively. The performance of deep autoencoder in the estimation of water absorption of fiber-reinforced SCC was superior to the other prediction models.