Enhanced data-driven shear strength prediction for RC deep beams: analyzing key influencing factors and model performance

Abstract

Accurately predicting the shear strength of reinforced concrete (RC) deep beams remains a challenging task due to the complex interplay of influencing factors. This study advances the field by integrating sophisticated machine learning (ML) techniques. A comprehensive dataset of 386 beams was compiled, covering a diverse range of geometries, materials, loading conditions, and reinforcements. Utilizing a random forest (RF) algorithm, a robust ML model was developed that significantly outperforms existing models, including those by Feng et al., in both accuracy and consistency. This model achieved a nearly perfect mean prediction (1.03) and the lowest coefficient of variation (19.4 %) for predicted versus target values. It identifies key parameters (beam width, effective depth, shear span-depth ratio, load and support plate widths, and material properties of concrete and steel) as critical factors influencing shear strength. In addition, an innovative nonlinear model based on insights from the ML model and fundamental mechanical principles was proposed. This nonlinear model, refined using data from 1650 deep beams, surpasses traditional models (e.g., those specified by ACI-318 and Eurocode-2) demonstrating superior predictive accuracy. The study illustrates the successful integration of traditional engineering principles with advanced ML techniques, highlighting the substantial potential of interdisciplinary approaches. These findings offer valuable insights into the complex relationships among material properties, reinforcement types, and shear strength, advancing our understanding and predictive capabilities for RC deep beams.