Optimized Generative Modeling and Interpretable Machine Learning for Predicting the Compressive Strength of High-Performance Concrete

This study provides a comprehensive analysis of ultrahigh-performance concrete (UHPC) compressive strength, focusing on data augmentation, predictive modeling, and model interpretability. The research utilized 808 experimental data points with 16 features to form the original UHPC data set. An optimized Gaussian copula generative adversarial network (Opt-CopulaGAN), enhanced via Bayesian hyperparameter tuning, was employed to augment the data set. Postaugmentation, the inception-fully connected network (Incep-FC Net) and machine learning models were used for prediction, resulting in a significant R2 value increase from 0.9035 to 0.9545. Partial dependence plots (PDP) and Shapley additive explanations (SHAP) analysis were applied to understand model behavior, revealing that high silica fume and fly ash concentrations improved UHPC strength, while nano-silica had a minor impact. SHAP also highlighted the negative effects of low-dosage silica powder and fine aggregates during extended curing. Elevated temperatures and humidity positively influenced UHPC compressive strength, especially during the initial curing phase.