Predicting shear strength of FRP-reinforced concrete beams using novel synthetic data driven deep learning

Abstract

Machine learning algorithms have emerged as a powerful technique to predict the engineering properties of composite materials and structures where traditional statistical methods have resulted in poor accuracy and high uncertainty. However, the lack of reliable and comprehensive experimental data hinders developing high-throughput computational models. To mitigate such a limitation, this study deploys the state-of-the-art tabular generative adversarial network (TGAN) to generate synthetic data for training generalized ML models. ‘Train on Synthesized - Test on Real’ approach was used to pioneer a novel framework for predicting the shear capacity of FRP-reinforced concrete beams. Accordingly, the models were trained using 8816 synthesized design data and tested using 304 real experimental data. The TGAN approach exhibited tremendous potential in generating credible data for training robust machine learning models, achievingR2of 0.96 when tested on the entire experimental dataset. Furthermore, a Bayesian optimization was performed on the extensive synthesized data to propose a new predictive shear strength equation. Results demonstrate that the proposed design model attained superior accuracy and vastly outperformed both existing design code provisions and empirical and theoretical models in the literature. The TGAN technique could transcend the lack of available experimental datasets in engineering problems via synthetizing numerous plausible data points to enhance the prediction accuracy and generalization ability of machine learning models.