Stochastic Multiscale Modeling of Electrical Conductivity of Carbon Nanotube Polymer Nanocomposites: An Interpretable Machine Learning Approach

This study introduces an interpretable machine learning (ML) framework for efficiently predicting the electrical conductivity of carbon nanotube (CNT)/polymer nanocomposites. A stochastic multiscale numerical model based on representative volume element (RVE) is employed to generate a representative dataset. This dataset is used to train three ML models, including random forest, XGBoost, and artificial neural networks (ANN). The dataset includes six input features: CNT length, aspect ratio, intrinsic CNT conductivity, number of CNT conduction channels, energy barrier height, and volume fraction, with the electrical conductivity of the nanocomposites as the output feature. The findings highlight the exceptional accuracy of the ANN model in predicting electrical conductivity at significantly lower computational costs. Furthermore, the use of Shapley additive explanations (SHAP) enhances the interpretability of these ML models, identifying the volume fraction, energy barrier height, and intrinsic CNT conductivity as the most influential factors affecting conductivity. This approach sets the stage for rapid and efficient modeling of CNT/polymer nanocomposites facilitating the design of materials with tailored electrical properties for diverse applications.