Strength prediction of circular and square CFDST columns using generalised machine learning models
TL;DR
Three ML algorithms, including Categorical Boosting (CatBoost), Multilayer Perceptron (MLP) and Kolmogorov‐Arnold Networks (KAN), were tested and the results show that the KAN model provides the most accurate predictions of CFDST strengths, with an R2 of 0.9903, compared with 0.9084 for Eurocode 4.
Abstract
Concrete‐filled double skin tubular (CFDST) members exhibit excellent structural performance owing to composite action between the steel tube and infilled concrete. However, the varied cross‐sections and complex composite action make it difficult for regression‐based empirical equations derived for a specific cross‐section to be generally applicable to all geometries. Although finite element simulations can capture cross‐sectional differences with high accuracy, their substantial computational demand makes them impractical for design and generalisation. In contrast, machine learning (ML), through leveraging its strong capability in capturing complex nonlinear relationships and data heterogeneity, offers a promising alternative for the efficient and accurate prediction of the strength of CFDST columns. In this study, a database of 1,650 CFDST columns, covering both circular and square sections, was compiled. Three ML algorithms, including Categorical Boosting (CatBoost), Multilayer Perceptron (MLP) and Kolmogorov‐Arnold Networks (KAN), were tested. The results show that among these ML models, the KAN model provides the most accurate predictions of CFDST strengths, with an R2 of 0.9903, compared with 0.9084 for Eurocode 4.