Strength Prediction of Sustainable UHPC Via VAE-Augmented Data and Explainable Boosting Models

This study develops an interpretable data-centric machine learning paradigm to predict the compressive strength of ultra-high-performance concrete (UHPC) with recycled concrete powder (RCP). Given limited experimental data, an experimental dataset of 80 samples was augmented with 300 synthesized concrete mix designs to improve the model. We evaluated the performance of two machine learning methods using extreme learning machine (ELM) and CatBoost, employing three different data strategies: experimental only (M1), synthetic only (M2), and combined (M3). Our results showed that CatBoost in the combined strategy (M3) outperformed the competition with an R2 of 0.968 and the best error metrics in the test phase. To demonstrate the feature contributions, we used SHAP (SHapley Additive exPlanations) and found that the most important features explaining compressive strength included cement, steel fibers, and water. Our proposed method demonstrates that combining generative data with explainable machine learning significantly improves both predictive performance and transparency, suggesting potential for designing and optimizing concrete mix recipes for sustainable UHPC.