Synergizing machine learning and experimental analysis to predict post‐heating compressive strength in waste concrete

Abstract In the current study, the impact of utilizing granite and marble construction waste powders as replacements (1%–9%) for cement on concrete compressive strength was investigated. In the second stage of the experimental program, combined mixtures were designed to evaluate their response to high temperatures using various machine learning (ML) techniques. Models employing water cycle algorithm (WCA) and genetic algorithms (GA) were developed based on 288 experimental results, featuring input variables such as temperature, exposure time, waste powder type, and cement replacement ratio, with residual compressive strength (RCS) as the sole output. Artificial neural networks (ANN), fuzzy logic (FL), and multiple linear regression (MLR) models were also developed for comparison. Optimal performance, with a 22% increase in compressive strength at 28 days, was observed by replacing 9% of cement with waste granite powder (WGP). At high temperatures, the best performance occurred with 9% WGP + 5% waste marble powder (WMP), resulting in a 59.6% increase in RCS value after exposure to 800°C for 2 h. The predictive WCA model outperformed GA and MLR, closely aligning with ANN and FL models, with a mean absolute error of 3.96 kg/cm2. Additionally, nonlinear prediction equations of RCS with high regression values were successfully developed using WCA and GA. Furthermore, sensitivity analyses were conducted using the weights of the hidden layers of the idealized neural networks and revealed that the RCS value exhibits high sensitivity to temperature variations. Exposure time had the second‐highest impact on RCS value, followed by the WGP ratio, and then the WMP ratio.