Generative adversarial network-enhanced machine learning models for high-precision prediction of rectangular concrete-filled steel tube strength

Abstract

Concrete-filled steel tubes (CFST) are widely recognized for their superior mechanical properties, including increased strength, ductility, and seismic resistance, making them popular in construction. However, accurately predicting the ultimate load (Nu) of CFST remains challenging due to the complex interactions between steel and concrete, and the varying parameters such as column dimensions, steel yield strength, and concrete compressive strength. Existing models and standards often lack precision, mainly when working with limited datasets. This study applies an advanced approach to improve Nu prediction for Rectangular CFST (RCFST) by combining Generative Adversarial Networks (GAN)-augmented data with machine learning and deep learning models. Four models (Gradient Boosting Regressor, Random Forest, Convolutional Neural Network, Residual Network) were initially trained on the original dataset. Subsequently, a GAN was utilized to generate synthetic data, expanding the dataset and improving model performance. The Random Forest model achieved the highest accuracy, with an R² of 0.9989, the root mean square error (RMSE) of 90.1, and the mean absolute percentage error (MAPE) of 1.3 %. A lightweight version of the Random Forest model was also developed to reduce computational complexity while maintaining an R² of 0.9979. Compared to three major standards (EN 1994, ACI, DBJ) and 18 machine learning models, the proposed models outperformed across key metrics including R², RMSE, and MAPE, demonstrating their effectiveness in predicting RCFST strength. Finally, a user-friendly graphical user interface (GUI) was developed, enabling direct engineering applications.