An image-based deep learning framework for predicting thermal conductivity of polymer composites with rotational data augmentation

초록

Predicting the thermal conductivity of polymer-based composites with complex microstructures is critical for enhancing thermal flow in applications like electronics and electric vehicles. Traditional methods such as theoretical models and simulations often lack generalizability and are computationally expensive. This study proposes an efficient deep learning framework that predicts thermal conductivity from scanning electron microscopy (SEM) images. The model employs a multi-input convolutional neural network that combines 2D structural images, region-specific data, and latent features. A simulation dataset was created using random sequential adsorption to generate statistical volume elements (SVEs), which were analyzed through finite element thermal simulations. The resulting 3D models were converted into 2D images aligned along the heat-transfer direction and augmented via rotation. Regions of interest were annotated using bounding boxes and labeled as filler, inter-filler, or matrix. Global features were compressed into latent vectors using an autoencoder. The model was trained on 1,728 image-label pairs and achieved high prediction accuracy (R2 = 0.9315, RMSE = 0.0664). These results confirm the model's effectiveness in capturing the complex relationship between composite microstructures and their thermal conductivity.

키워드