Predicting the thermal conductivity of polymer composites with one-dimensional oriented fillers using the combination of deep learning and ensemble learning

Abstract

• Deep learning (CNN) and ensemble learning (RFR) was successfully used to accurately predict the thermal conductivity of polymer composites with one-dimensional oriented fillers. • Introduced a new descriptor, "Orientation degree (Od)," to quantitatively describe the spatial distribution of one-dimensional fillers, significantly enhancing prediction accuracy. • The mathematical expressions between the different descriptors and the thermal conductivity of the composite were obtained by using the symbolic regression algorithm. • The proposed method reduces the need for extensive retraining of models, offering a practical solution for material property prediction in advanced materials design and energy efficiency applications. Polymer composites with one-dimensional (1D) oriented fillers, recognized for their high thermal conductivity (TC), are extensively utilized in cooling electronic components. However, the prediction of the TC of composites with 1D oriented fillers poses a challenge due to the significant impact of filler orientation on composite TC. In this paper, we use a strategy that combines deep learning and ensemble learning to efficiently and quickly predict the TC of composites with 1D oriented fillers. First, as a control, we used convolutional neural network (CNN) model to predict the TC of 1D carbon fiber-epoxy composite, and the R-squared (R 2 ) on the test set reached 0.924. However, for composites consist of different matrices and fillers, the CNN model needs to be retrained, which greatly wastes computing resources. Therefore, we define a descriptor ‘Orientation degree ( O d )’ to quantitatively describe the spatial distribution of the 1D fillers. CNN model was used to predict this structural parameter, the accuracy R 2 can reach 0.950. Using O d as a feature, random forest regression (RFR) was used to predict the TC, and the accuracy R 2 reached 0.954, which was higher than that of CNN control group. We further successfully extended this strategy to composites consist of different 1D fillers and matrices, and only one CNN model and one RFR model needed to be trained to achieve fast and accurate TC prediction. This strategy provides valuable insights and guidance for machine learning-based material property prediction.

Topics & Concepts