Machine Learning for Performance Prediction and Optimization of Polymer Composites: Unveiling the Dominant Role of Thermally Conductive Pathways.

Polymer composites with multifiller systems have gained significant attention due to their potential to enhance both thermal conductivity (TC) and mechanical properties. However, traditional methods for predicting composite properties often struggle to account for the complex effects of filler distribution and thermally conductive pathways formation, particularly at varying filler volume fractions. In this study, we developed a novel dense sphere-packing algorithm to construct geometric models of multifiller composites, generating a database of 1024 representative volume element (RVE) models using COMSOL. Random Forest Regression (RFR) and Convolutional Neural Network (CNN) models were trained to predict TC and Young's modulus (E). Recognizing the limitation of traditional machine learning algorithms in capturing thermally conductive pathways, we defined a series of descriptors related to it, which improved prediction accuracy. Furthermore, a transformer-based generative model, along with an auxiliary algorithm, was employed to generate RVE structures with thermally conductive pathways at low total filler volume fractions (vol). We found that at low vol value, the presence of thermally conductive pathways markedly enhanced TC, while their influence on E remains negligible. This combined strategy of traditional machine learning, deep learning, and generative modeling offers an efficient and accurate framework for predicting and designing polymer composites, offering valuable insights for the development of next-generation composite materials that are optimized for both thermal and mechanical performance.