Generative data augmentation, explainable machine learning, and uncertainty analysis for compressive strength prediction of sustainable UHPC

Abstract

Recent interest in ultra-high-performance concrete (UHPC) has been on the use of construction and demolition waste (CDW) due to the growing need to use sustainable construction materials. Recycled concrete powder (RCP), a product of CDW, is a promising way to reduce cement use, lower CO2 emissions, and make green construction more cost-effective. This research establishes a data-based machine learning framework to predict the compressive strength of UHPC incorporating RCP. To build simplified models, a Variational Autoencoder (VAE) was employed to produce a synthetic dataset. This resulted in the development of three models: M1 (experimental data only), M2 (synthetic data only), and M3 (merged dataset). M3 served as the basis for training three different machine learning models: Extreme Learning Machine (ELM), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost). CatBoost was selected from the three because it performed best in the initial class and, across all cross-validated systems, was selected as the default for the group. CatBoost_M3 outperformed the others, with an R2 of 0.95 and the lowest mean and root-mean-square errors. Analysis of the mean and root-mean errors identified CatBoost_M3 and its performance as optimal. A SHapley Additive exPlanations (SHAP) analysis was conducted to meet the explainable AI requirement; the dominant explanatory variables for the predicted values were cement, steel fibers, and water content. At a conceptual level, performance predictions for UHPC and the explanatory AI serve as complementary support for reducing experimental work and enabling the rapid adoption of eco-efficient recycled UHPC. A novel machine learning framework is proposed to predict the compressive strength of RCP-based UHPC. Synthetic data were generated using a Variational Autoencoder (VAE) to enhance model robustness. The combined dataset strategy yielded the highest predictive performance across all models. CatBoost_M3 achieved superior accuracy, with an R2 of 0.968 and the lowest MAE and RMSE. SHAP analysis provided model explainability, identifying cement, steel fibers, and water as key predictors. A graphical user interface (GUI) has been developed and deployed, enabling both single-input and batch comma-separated values (CSV) predictions. A novel machine learning framework is proposed to predict the compressive strength of RCP-based UHPC. Synthetic data were generated using a Variational Autoencoder (VAE) to enhance model robustness. The combined dataset strategy yielded the highest predictive performance across all models. CatBoost_M3 achieved superior accuracy, with an R2 of 0.968 and the lowest MAE and RMSE. SHAP analysis provided model explainability, identifying cement, steel fibers, and water as key predictors. A graphical user interface (GUI) has been developed and deployed, enabling both single-input and batch comma-separated values (CSV) predictions.