A Deep Learning-Informed Design Scheme for Shear Friction at Concrete-to-Concrete Interface: Recommendations for Inclusion in AASHTO LRFD Guidelines

Abstract

Recent advancements in construction technology have led to high-strength concrete and steel. However, these developments have depreciated the accuracy of the design equations in current provisions, which were based on normal-grade materials. To fill such a research gap, this study presents a novel deep learning-based computation scheme that can replace the current design provisions by virtue of its superior accuracy and reliability. The proposed approach exploits Neural Additive Models (NAMs) in which geometric and material properties associated with a normalweight concrete-to-concrete shear interface are inputted to individual neural network blocks. The outputs of the individual blocks are linearly combined to produce the prediction for interfacial shear strength. This model provides a way to identify and quantify the individual contributions of the input parameters, thus enhancing the interpretability of the model predictions for shear strength at the normalweight concrete-to-concrete interface. The deep learning-informed design (LID) scheme improves the prediction accuracy of the shear strength equation in the existing AASHTO LRFD Bridge Design Specifications by over 32%.