An effective design method of high-strength steel columns with limited datasets using physics-guided conditional tabular GAN

Abstract

High-strength steel structures received increasing interest but are lack of suitable design methods. Most studies revise the existing design codes, which, however, is still inaccurate for failure load predictions. Some researchers explore machine learning (ML) but require extensive high-quality datasets. This is challenging for high-strength steel structures, where the collection of either experimental data or numerical simulation data is time-consuming and expensive. To address it, this study develops an effective ML-based design method which only needs very limited datasets. In particular, Physics-guided Conditional Tabular Generative Adversarial Network (Pg-CTGAN) is proposed to be integrated with suitable existing design code for similar structural members and further generate sufficient synthetic datasets for subsequent ML training. Two benchmark tests were designed to assess the quality of synthetic data and its effects on the predictive performance of the ML models. To evaluate the performance of Pg-CTGAN, two case studies were conducted, including the prediction of ultimate compressive strength of high-strength steel bolted built-up angle section columns and that of high-strength steel bolted built-up channel section columns. The results are then compared with the current design codes, showing that the proposed method can improve the prediction accuracy of various ML models in both case studies, and outperforms the current design specifications. It can be demonstrated that, the proposed method can provide accurate failure load predictions and has the great potential for other structural components.