Predicting joint shear in beam-column connections using convolutional neural networks

Predicting joint shear at beam-column junctions (BCJ) is essential in structural engineering to ensure the safety and reliability of systems. Current methodologies using empirical calculations may rely on simplistic assumptions and insufficiently account for the many geometric factors and material properties that influence shear in BCJ. This research introduces a novel approach using Convolutional neural networks (CNNs) to predict joint shear. The collection comprises 515 joints, categorized into 210 exterior joints and 305 interior joints, characterized by 14 fundamental factors delineating their form and material properties. The predictive performance of the CNN model has been evaluated using known engineering codes, including ACI 318-19, NZS 3101:1-2006, IS 13920:2016, and several other data-driven models in the domain. Furthermore, it has been contrasted with an ensemble regression method. The study includes a thorough sensitivity analysis using a gradient-based method to determine the relative importance of input factors in predicting shear stress. The findings demonstrate the effectiveness of CNN in identifying complex relationships among joint parameters, thereby enabling precise predictions of joint shear. This method offers a promising alternative to traditional empirical formulas and enhances the understanding of structural behavior in BCJ. This study integrates contemporary machine learning algorithms with structural engineering concepts to optimize design processes and augment the safety and reliability of built environments.