A co-training style semi-supervised artificial neural network modeling and its application in thermal conductivity prediction of polymeric composites filled with BN sheets

Abstract

Predicting the thermal conductivity of polymeric composites filled with BN sheets is helpful for fabricating thermal management material. In this study, a co-training style semi-supervised artificial neural network model (Co-ANN) was proposed to take advantage of unlabeled data to refine the prediction. The thermal conductivity of polymer matrix, the diameter, aspect ratio, and volume fraction of the BN sheets are considered as the input variables of the thermal conduction model. Two artificial neural network (ANN) learners with different architecture will label the unlabeled examples. Through estimating the labeling confidence from the mathematical influence and thermal conductive behavior, the most confidently labeled example will be used to augment the training dataset. The lower limit of the labeling confidence is introduced to reduce the data noise. After learning the augmented training information, a combination of two ANN regressors will construct the final Co-ANN thermal conduction model. Compared to other models, the newly developed Co-ANN thermal conduction model remarkably improves the thermal conductivity prediction and exhibits the best accuracy and generalization performance. The proposed method shows a vast potential in thermal conductive material design.