Evaluation of the compactness of granular mixes for roller compacted concrete design using SPO-ANN

Granular compactness is a critical parameter in the mix design of roller-compacted concrete (RCC), although existing empirical and theoretical methods often require extensive experimental characterisation and perform poorly for mixtures containing fillers and large aggregates. This study proposes a data-driven model to accurately predict granular compactness, specifically for RCC applications. A comprehensive laboratory database of 163 granular mixtures, with maximum aggregate sizes ranging from 15 to 63 mm, was developed, and compactness values were experimentally measured using a specially designed testing system. A hybrid predictive model combining an artificial neural network with the Stochastic Paint Optimizer was then trained using key geometrical descriptors, including intermediate diameters, curvature and Hazen coefficients and filler rate. The model achieved strong predictive performance (training R2=0.9081; testing R2=0.9017) with minimal error (MSE=2.718E-05; RMSE=0.0052). Validation using 24 additional mixtures and sensitivity analysis confirmed both accuracy and interpretability. A user-friendly graphical interface was also developed to facilitate practical RCC design applications.