Reliable forecasting non-linear triaxial mechanical response of recycled aggregate concrete by knowledge-enhanced, modified, explainable and replicable machine learning algorithms

Abstract

The constitutive modelling is the only method to describe the triaxial stress–strain behavior of cement-based materials while its theoretical deduction, modelling parameters determination and numerical calibrations made it difficult to be further applied. To overcome the limitation of constitutive modelling, a comprehensive machine learning (ML) approach, including Artificial Neural Network (ANN), Gaussian Process (GP), Gradient Boosting (GB) and Optimized Gaussian Process (OGP) was firstly proposed to predict triaxial mechanical behavior of recycled aggregate concrete (RAC). The data augment technology was employed to increase the training data size from 249 to 580, effectively improving the generalization performance. The performance statistics of the aforementioned ML models were compared and validated by R², MAE, RMSE, and Taylor diagram, showing that the OGP had the best study ability and prediction accuracy. The 99 % prediction results generated by the OGP model concentrated within the ± 10 % confidence interval (R² = 0.991, MAE = 1.04, RMSE = 0.122). Furthermore, to address the black box nature of ML models, the shapley additive explanation and partial dependence analysis were employed to elucidate the underlying arithmetic mechanism. Finally, the best OGP model was compared with previous constitutive method and further utilized to validate its applicability. Unlike classical constitutive modeling, which requires specialized expertise, the proposed ML approach, available as open source at https://doi.org/10.13140/RG.2.2.15784.89608, offered an accessible and effective solution for predicting triaxial behavior with experimental data.