A machine learning approach to predict demand-to- capacity ratio for reinforced concrete jacketing of columns in seismic-deficient buildings

Abstract

Existing reinforced concrete (RC) buildings risk seismic damage because they were not constructed in compliance with seismic design standards and may have irregular mass distribution and construction defects. Typically, columns in these buildings are designed to withstand only gravity loads, making them vulnerable to damage or collapse during earthquakes. Retrofitting these columns using an RC jacket system is a standard way to enhance seismic resilience. However, conventional parametric modeling for RC jacketed structures using physics-based (finite element) modeling can be time-consuming and non-intuitive. To address this challenge, the present study proposes a novel data-driven machine-learning approach to predict RC jacketed columns' demand-to-capacity ratio (DCR), aiming for a reasonably accurate design with reduced computational time. Various design parameters related to RC column jacketing are considered when predicting the DCR. The datasets generated in post-processing are used to train Graphical Neural Network (GNN) and Gaussian Mixture Model (GMM). The dataset encompasses parameterization of design variables, including retrofit location, concrete compressive strength, cross-sectional dimensions, jacket thickness, longitudinal and transverse reinforcement areas, yielding reinforcement strength, and slenderness ratio. Subsequently, both models are fitted and evaluated against a test dataset to identify the optimal performer, using a multiple scorer performance index as the model evaluation metric. The analysis indicates that the GMM model emerges as the most suitable regressor for DCR estimation, exhibiting lower residual error than the GNN model.