Physics-Informed TVAE-Based Probabilistic Framework for Uncertainty Quantification of Mechanical Properties and Failure Modes of FRP Coupons

Fiber-reinforced polymer (FRP) composites are widely used in structural applications, yet their engineering behaviour remains difficult to model reliably due to limited, unrepresentative experimental datasets. While probabilistic learning models can capture material uncertainty and complex responses, their effectiveness is constrained by data scarcity, often leading to biased or poorly generalizable predictions. To address this challenge, a physics-informed probabilistic modelling framework was proposed that integrates synthetic data generation with uncertainty-aware learning. A physics-informed tabular variational autoencoder (PI-TVAE) was employed to generate physically admissible synthetic FRP coupon data. Synthetic-to-experimental augmentation ratios ranging from 1:1 to 1:10 were systematically evaluated, with experimental data constituting 70% of the available samples. Deterministic and probabilistic performance metrics consistently identify an optimal augmentation ratio of 1:6, which was subsequently adopted for model development. The enhanced dataset was used to train a Bayesian neural network and a natural gradient boosting (NGB) model to predict FRP tensile strength, ultimate strain, and failure behaviour. The performance of the probabilistic modeling approach was also compared with that of deterministic models. Comparative results demonstrate that the NGB model trained on augmented data delivers superior prediction accuracy, calibration, and uncertainty quantification. Model interpretability was examined using Shapley Additive exPlanations, revealing fiber type as the most influential parameter governing tensile performance. To facilitate practical adoption, a web-based software application based on the NGB framework has been developed, enabling instant prediction of FRP tensile properties and associated uncertainties. The proposed framework advances reliable, interpretable, and uncertainty-informed data-driven modelling of FRP composites under limited experimental data conditions.