Machine learning models for predicting axial compressive capacity of circular CFDST columns

Abstract

Machine learning (ML), widely recognized for its advanced predictive capabilities, is progressively being employed in structural design. Concrete-filled double skin steel tubular (CFDST) members are extensively utilized due to their exceptional mechanical behavior. However, existing formulas encounter challenges in accurately calculating the axial compressive capacity of circular CFDST members, particularly when high-strength concrete and steel materials are involved, leading to complexity and low accuracy. This paper presents a data-driven framework that leverages the superior predictive performance of machine learning approaches to predict the axial compressive capacity of circular CFDST columns. Finite element models are developed through secondary development of ABAQUS using custom script, which are then validated against experimental data, confirming their effectiveness and applicability. A general database is established, comprising 226 test results and 550 finite element results obtained from the script, to facilitate the prediction of axial compressive capacity for circular CFDST columns. By incorporating the physical mechanisms of CFDST members and conducting correlation analysis, the input parameters for the machine learning models are determined. Decision Tree, Artificial Neural Network (ANN), K-nearest neighbor (KNN) and two ensemble methods including Random Forest, XGBoost, are trained and optimized using the database to provide accurate predictions. The results demonstrate that the KNN model exhibits superior performance and high accuracy, outperforming the four commonly used calculation formulas in the field, particularly for circular CFDST columns using high-strength materials. The proposed data-driven framework shows promising potential for achieving outstanding predictive performance.