Evaluation of Kolmogorov-Arnold Networks in predicting bearing capacity of hollow circular and hollow square CFST columns

Abstract

In recent years, with the explosion of the industrial revolution 4.0, terms such as artificial intelligence (AI) have become familiar and increasingly widely applied in the engineering field. This study focuses on the study and evaluation of AI models to predict the axial strength of concrete-filled steel tube columns (CFST). In particular, this study introduces and highlights a new AI model, Kolmogorov-Arnold Networks (KAN), and compares its performance with the previously existing AI model, support vector regression (SVR), along with Eurocode 4. A large dataset consisting of two types of CFST columns (hollow circular and hollow square CFST columns) with different concrete strengths was created using ABAQUS software. The AI models were evaluated based on important statistical indices such as MAPE, MAE, RMSE, and correlation coefficient R. The analysis results showed that the KAN model was the most effective AI model when compared with other models. The R indices were always greater than 0.9 and the MAPE, MAE, RMSE indices were the lowest among the compared models. At the same time, the predicted data from the KAN model showed the highest similarity with the actual data in predicting the axial strength of four types of CFST columns. Therefore, the KAN model can be considered as a powerful and accurate tool in predicting the compressive strength of CFST columns.