Probabilistic prediction method for shear strength capacity of RC deep beams based on the fusion of multiple machine learning models

Abstract

Reinforced concrete deep beams (RCDBs) exhibit complex shear mechanisms due to low shear span-to-depth ratios, resulting in a lack of universally accepted computational models. This study develops a probabilistic prediction method that fuses multiple machine learning (ML) models to identify non-linear relationships between design parameters and shear capacity, providing a novel approach to predicting the shear strength of RCDBs. A comprehensive database of 1577 experimental RCDB samples was collected from the literature. Various ML models, including deep learning approaches, were applied to predict shear capacity, with hyperparameter optimization to enhance model performance. To increase reliability, a fusion model was created by assigning weights to individual models based on their predictive capabilities. Additionally, a method was introduced to estimate the 95 % confidence interval for shear capacity. Results indicated that overfitting occurred in the default ML models; however, hyperparameter optimization significantly improved prediction accuracy and reduced the overfitting. The fusion model surpassed individual models in predictive accuracy and robustness, with 96.32 % of the experimental shear capacities falling within the established confidence interval.