PREDICTING COMPRESSIVE STRENGTH OF ULTRA-HIGH-PERFORMANCE CONCRETE USING SUPPORT VECTOR MACHINE REGRESSION: MODEL DEVELOPMENT, FEATURE ANALYSIS, AND VALIDATION

Abstract

Ultra-High-Performance Concrete (UHPC) is a complex composite material whose mechanical properties depend on intricate interactions between multiple constituents and curing conditions. Traditional mix design relies on extensive experimentation, which is time-consuming and costly. This study develops and validates a Support Vector Machine (SVM) regression model to predict the compressive strength of UHPC using a dataset of 810 mixtures with 13 input features. The model was configured as a ν-SVM with a Radial Basis Function (RBF) kernel and automatic hyperparameter tuning. Performance evaluation on a held-out test set yielded an R² of 0.908, RMSE of 12.205 MPa, and MAPE of 8.969%, indicating high predictive accuracy and generalization. Feature importance analysis, supplemented by k-fold validation, identified silica fume, fiber content, and curing age as the most influential factors—findings consistent with contemporary material science and recent machine learning literature. The study demonstrates that SVM regression, coupled with robust feature selection and validation, provides a reliable, interpretable, and efficient tool for UHPC mix optimization, reducing the need for physical trials and accelerating the development of high-performance concrete formulations.