Data-Driven Shear Capacity Prediction of Reinforced Concrete Deep Beams with an Uncertainty-Aware Model

Abstract

Data-driven approaches based on machine learning (ML) have become progressively popular for the shear capacity prediction of reinforced concrete (RC) deep beams because the fine-tuned ML models could mostly outperform the mechanics-driven models when approximating the hidden mapping relationship within the experimental data. However, the output of general ML models is barely a single determinate value prediction without a confidence interval, which is insufficient for engineering decision making. In such circumstances, an uncertainty-aware data-driven framework based on the artificial neural network (ANN) for predicting the shear capacity of RC deep beams is proposed. This framework can reasonably quantify both the aleatoric and epistemic uncertainties within its prediction. To validate its feasibility and compare it with existing approaches including empirical formulas and a representative ML algorithm, a comprehensive database comprising 214 experiment samples is adopted for investigation. The results indicate that this framework’s accuracy and robustness could dramatically outperform the mechanic-based prediction models, and its expectation prediction accuracy is also competitive, in contrast to the representative ML algorithm. In addition, the proposed framework could rationally estimate the uncertainties within its predictions, which is helpful and suggestive for the following decision making.