Efficient thermal and mechanical property prediction of highly filled bimodal particle-reinforced polymer composites using a modified embedded-element method and machine learning

Efficient prediction of effective properties in highly filled bimodal particle-reinforced polymer composites (PRCs) is challenging because conventional finite-element (FE) simulations require complex particle-matrix conforming meshes at high filler contents. In this work, a modified embedded-element FE (EE-FE) formulation is developed to overcome this limitation by removing overlapping matrix regions and introducing local thermal-conductivity corrections, enabling accurate property prediction without conforming particle-matrix meshes. Random bimodal particle structures are generated using the discrete element method, and an FE database is constructed by systematically varying filler volume fraction, particle-size composition, and constituent properties. The proposed framework is validated against conforming-mesh FE models, with prediction discrepancies below 4% for the benchmark configurations. Parametric analysis shows that filler volume fraction strongly affects all three effective properties and produces the largest observed variation in thermal conductivity and elastic modulus within the investigated parameter ranges, while the constituent thermal-expansion properties also strongly affect CTE. Machine-learning surrogate models based on support vector regression (SVR), Gaussian process regression (GPR), and feedforward neural networks (FNN) are further developed to approximate the FE response surfaces. GPR achieves the highest in-domain prediction accuracy for thermal conductivity, elastic modulus, and CTE, with R 2 values of 0.998, 0.998, and 0.995, respectively. For unseen constituent-property combinations within the investigated parameter ranges, FNN gives the best thermal-conductivity prediction (R 2 = 0.94), whereas GPR gives the best results for elastic modulus (R 2 = 0.91) and CTE (R 2 = 0.88). The framework supports rapid prescreening of filler loading, particle-size composition, and constituent combinations, thereby reducing the number of candidate formulations requiring detailed simulations or experimental validation.