Machine learning and nonlinear finite element analysis of fiber‐reinforced polymer‐confined concrete‐steel double‐skin tubular columns under axial compression

Abstract

AbstractFiber‐reinforced polymer (FRP)‐confined double‐skin tubular columns (DSTCs) are an innovative type of hybrid columns that consist of an outer tube made of FRP, an inner circular steel tube, and a concrete core sandwiched between them. Available literature focuses on hollow DSTCs with limited research on DSTCs made with inner steel tubes filled with concrete. Overall, DSTCs have many applications, highlighting the importance of studying the effects of concrete filling and strength on the composite system. To address this gap, finite element models (FEMs) and both traditional and innovative machine learning (ML) techniques were used to develop accurate models for predicting load‐bearing capacity and confined ultimate strain under axial loads. A comprehensive database of 60 experimental tests and 45 FEMs simulations of columns was analyzed, with five parameters selected as input variables for ML‐based models. New techniques like gradient boosting (GB), random forest (RF), convolutional neural networks, and long short‐term memory are compared with established algorithms like multiple linear regression, support vector regression (SVR), and empirical mode decomposition (EMD)‐SVR. Regression error characteristics curve, Shapley Additive Explanation analysis, and statistical metrics are used to assess the performance of these models using a database containing 105 FEMs test results that cover a range of input variables. While EMD‐SVR and GB perform well for confined ultimate strain, the suggested EMD‐SVR, GB, and RF models show superior predictive accuracy for confined ultimate load. To be more precise, for confined ultimate load prediction, EMD‐SVR, GB, and RF obtain values of 0.99, 0.989, and 0.960, respectively. The values for GB and EMD‐SVR at confined ultimate strain are 0.690 and 0.99, respectively. However, design engineers are limited by the “black‐box” nature of ML. In order to solve this, the study presents an open‐source GUI based on GB, which gives engineers the ability to precisely estimate confined ultimate load and strain under various test conditions, enabling them to make well‐informed decisions about mix proportion.