Metaheuristic-Enhanced Machine Learning for Accurate Shear Strength Assessment of RC Deep Beams

Abstract

Extensive testing required by traditional structural engineering methods can be time-consuming and costly due to the complexity of the procedures involved. This work presents a novel machine learning approach for predicting the shear strength of reinforced concrete (RC) deep beams. It employs a Gradient Boosting (GB) algorithm, optimized using metaheuristic techniques, specifically the Golden Jackal Optimization (GJO) and Honey Badger Algorithm (HBA). To develop this approach, a comprehensive dataset of 314 experimentally tested RC deep beams with web openings was compiled from peer-reviewed literature. The dataset includes key features governing the shear strength. The GB model's hyperparameters were fine-tuned using GJO and HBA, with the GJO-optimized model (GB_06) showing superior performance. It achieved a coefficient of determination (R²) of 0.9664 and a root mean squared error (RMSE) of 70.258 kN on the test dataset. Feature importance analysis using SHAP values identified the shear span-to-depth ratio, horizontal web reinforcement ratio, and vertical web reinforcement ratio as the key factors influencing shear strength. The proposed model offers significant improvements in accuracy and reliability, providing structural engineers with an efficient tool for design optimization and safety assessment of RC deep beams.