Interpretable Deep Learning With Sobol Sensitivity and Interaction Analysis for Splitting Tensile Strength of Sustainable Recycled Aggregate Concrete

Abstract

The variability of Recycled Concrete Aggregates (RCAs) poses a significant challenge in predicting the mechanical properties of Recycled Aggregate Concrete (RAC), particularly its splitting tensile strength (STS). Accurate, interpretable predictive models are essential to support mix-design decisions and promote the use of sustainable construction materials. This study develops a Deep Neural Network (DNN) framework to predict the STS of RAC and to interpret the model-inferred parameter influence using global sensitivity analysis. Four models were developed: two feedforward neural networks (FNNs) and two long short-term memory (LSTM) networks, trained on an experimental dataset comprising 257 data points. Bayesian optimization was employed for hyperparameter tuning, while stratified cross-validation ensured model robustness. The results indicate that the stacked LSTM (sLSTM) architecture achieved superior generalization performance (R ${}^{2} =0.946$ , MSE = 0.060) within the analyzed dataset, outperforming the other models. To address the limited interpretability of deep learning, Sobol global sensitivity analysis and partial dependence plots (PDPs) were integrated to quantify the influence of model parameters and their interactions. The analysis identified water content as the most influential variable within the model, followed by cement content and recycled aggregate density. PDP interaction analysis highlighted regions of the input space associated with higher predicted STS values, characterized by low water content and specific fiber types, particularly steel and carbon fibers, under the conditions represented in the dataset. Overall, the proposed framework provides a highly accurate and interpretable computational tool for the analyzed dataset, enhancing understanding of the factors associated with STS predictions in sustainable recycled aggregate concrete.