Innovative uncertainty-aware probabilistic framework for quantification of fiber-reinforced cementitious matrix-concrete bond

• Developed a data-augmented probabilistic framework to predict FRCM–concrete bond behaviour with quantified uncertainty. • Integrated synthetic data generation, deterministic/probabilistic models, uncertainty propagation, and SHAP analysis. • NGB model trained on synthetic data showed superior calibration and accuracy (testing MAPE ≈7.61%). • Operationalized validated model in DAB-FRCM tool for risk-informed design and reduced experimental burden. Fiber-reinforced cementitious matrix (FRCM) composites provide durable and sustainable alternatives for strengthening concrete structures, but their effectiveness depends on a complex FRCM-concrete bond whose variability defies deterministic design. This study introduced a data-augmented probabilistic framework that integrates synthetic data generation (conditional tabular generative adversarial networks and tabular variational autoencoders, deterministic and probabilistic (Bayesian neural networks, natural gradient boosting (NGB)), uncertainty propagation, and SHAP explainability to predict bond stress response and failure modes with quantified uncertainty. Models were trained on a dataset of 559 literature tests and synthetic augmentation from the 70% training subset; evaluation under train-on-experimental and train-on-synthetic schemes shows that NGB trained on synthetic data achieved superior calibration and accuracy (testing MAPE ≈7.61%). Tabular variational autoencoders produced higher-fidelity augmentations than a conditional tabular generative adversarial network. SHAP analysis identified reinforcement configuration and fiber/matrix properties as dominant drivers. The study also operationalizes the validated model in DAB-FRCM, an online software for probabilistic bond design, reducing experimental burden and enabling risk-informed retrofit decisions.