Machine learning based prediction models for spilt tensile strength of fiber reinforced recycled aggregate concrete

Abstract

The demand for concrete production has led to a significant annual requirement for raw materials, resulting in a substantial amount of waste concrete. In response, recycled aggregate concrete has emerged as a promising solution. However, it faces challenges due to the vulnerability of the hardened mortar attached to natural aggregates, leading to susceptibility to cracking and reduced strength. This study focuses on predicting the split tensile strength of fiber reinforced recycled aggregate concrete using five prediction models, including two deep neural network models DNN1 and DNN2, one optimizable Gaussian process regression (OGPR), and two genetic programming based GEP1 and GEP2 models. The models exhibited high accuracy in predicting spilt tensile strength with robust R2, RMSE, and MAE values. DNN2 has the highest R2 value of 0.94 and GEP1 has the lowest R2 value of 0.76. DNN2 model R2 was 3.3% and 13.5% higher than OGPR and GEP2. Similarly, DNN2 and GEP2 model performed 9.3% and 9.21% better than DNN1 and GEP1 respectively in terms of R2. DNN2 model performed 20.32% and 31.5% better than OGPR and GEP2 in terms of MAE. Similarly, GEP2 and DNN2 MAE were 13.1% and 31.5% better than GEP1 and DNN1. Sensitivity analysis using the relevance factor and permutation feature importance revealed that the most significant positive factors are cement, natural coarse aggregates, density of recycle aggregates, and superplasticizer while recycle aggregate concrete, max size, and water content of recycle aggregates and water content have the most negative effect on STS values. The proposed ML methods, especially DNN2 and OGPR can be effectively utilized in practical projects, saving time and cost for eco-friendly fiber reinforced recycled aggregate concrete mixes. However, it is required to study more input variables and utilize hybrid models to further enhance the accuracy and reliability of the models.