A Surrogate Modelling Framework for Predicting the Effective Behaviour of Composite Materials

Abstract Accurate prediction of the effective properties of composite materials is essential for their design and optimisation in high-performance engineering applications. This work presents an integrated approach combining micromechanical modelling with machine learning to establish an efficient predictive framework for composite behaviour. A dataset of effective elastic moduli was generated using Representative Volume Elements (RVEs) with periodic boundary conditions via finite element analysis. Various machine learning algorithms, including Random Forest, Gradient Boosting, and Support Vector Regression, were evaluated. Random Forest demonstrates the highest accuracy in modelling complex material responses. To further enhance efficiency, a surrogate model incorporating an active learning strategy was introduced, enabling iterative improvement by focusing computational resources on high-uncertainty regions. The framework was validated against both theoretical predictions and experimental results, showing notable accuracy, especially in the prediction of transverse and shear properties. This study offers a reliable and scalable methodology that leverages the strengths of micromechanics and data-driven modelling, facilitating the accelerated design of advanced composite materials with reduced computational demand.