Hybrid generative-probabilistic machine learning approach for predicting residual tensile strength and elastic modulus of FRP coupons under environmental exposure

The continuous deterioration of concrete and masonry structures has increased the demand for rehabilitation and strengthening of aging civil assets. Fiber-reinforced polymer (FRP) composites have been utilized to enhance the structural capacity of deteriorated structures. While these composites can be used in various environments, their performance is often compromised after exposure to harsh conditions, leading to a reduction in the structural capacity of the strengthened elements. Therefore, understanding the post-exposure residual performance of FRP composites, specifically residual tensile strength and elastic modulus is crucial for developing effective rehabilitation strategies. However, predicting the post-exposure performance of FRP coupons is inherently complex due to the involvement of multiple environmental and material parameters, as well as limited experimental data. This challenge necessitates the implementation of innovative modelling approaches to enhance both prediction accuracy and reliability and potentially mitigate the limitations of available small datasets. To address these challenges, this study employed credible synthetic data generated by advanced generative models, specifically a conditional tabular generative adversarial network (CTGAN) and a tabular variational autoencoder (TVAE), for machine learning model development. Moreover, probabilistic neural networks (PNNs) and natural gradient boosting (NGB) models were developed to predict the residual tensile strength and elastic modulus of FRP coupons and quantify the prediction uncertainty. The results indicated that the TVAE synthesizer demonstrated greater capability for generating credible synthetic data for predictive model development than the CTGAN synthesizer. Furthermore, the NGB model exhibited marginally better performance than the PNN model in predicting the residual tensile strength and elastic modulus of the FRP coupons, achieving testing coefficient of determination values of 0.9635 and 0.9914, respectively. In addition, the Shapley additive exPlanations feature importance analysis was employed to evaluate the influence of input parameters on the predictions, providing valuable insights into the performance evaluation of FRP composites under exposure conditions.