Generative adversarial network approach for predicting tensile behavior and failure pattern of fiber-reinforced cementitious matrices

Abstract

Fiber-reinforced cementitious matrices (FRCM) are a sustainable solution for rehabilitating aging civil infrastructure. Yet, there is a lack of consistent models for predicting the tensile strength, ultimate strain, and failure pattern of FRCM coupons, posing hurdles against effective design and wider applications. The present study resolves this gap by coining a novel machine learning (ML) framework based on conditional tabular generative adversarial network (CTGAN) to estimate the tensile strength, ultimate strain, and failure patterns of FRCM coupons. Firstly, an extensive dataset of FRCM coupons considering tensile strength, ultimate strain, and failure patterns was collected from relevant publications. CTGAN was then employed to generate synthetic data, thus alleviating the problem of limited experimental data. A training subset encompassing 70 % of the collected data was used for synthetic data generation using CTGAN. The augmented dataset was used to develop ML models to prognosticate the tensile behavior of FRCM coupons. Results show that the synthetic dataset offers credibility enabling the development of ML models with higher prediction accuracy in estimating the tensile behavior of FRCM coupons compared to models trained with real datasets. Among the developed models trained with synthetic data, eXtreme gradient boosting showed the highest prediction accuracy, achieving testing R 2 and MAE values of 0.9690 and 84.50 MPa, respectively, for the tensile strength of FRCM coupons. SHAP feature importance analysis identified fiber density, width of FRCM coupons, thickness of fabric, and length of FRCM coupons as the most influential parameters affecting tensile strength and ultimate strain, conforming to domain knowledge in the open literature. • Novel conditional tabular generative adversarial network generated reliable synthetic data on FRCM tensile behavior. • Machine learning models trained on synthetic data yield superior accuracy to models trained on limited experimental data. • eXtreme gradient boosting was the most accurate in predicting tensile behavior and failure patterns. • Accurate estimation of FRCM’s coupon tensile behavior enables accurate FRCM design in structural rehabilitation.

Topics & Concepts