Contrastive loss with dynamic margin for rubber-recipe physical-property prediction

Abstract

Predicting the physical properties of rubber presents difficulties because of the intricate and nonlinear relationships between its constituent elements. The laborious nature of rubber manufacture results in restricted data sets, hence exacerbating this difficulty. To achieve successful training of a neural network with a smaller amount of data, it is crucial to use an appropriate weight initialization approach and carefully construct the network architecture to avoid overfitting. This study presents a novel approach to representation learning that uses a contrastive loss with dynamic margin loss function. The approach is specifically designed to capture semantic similarities in rubber material quantities and physical-property recipes. In addition, a novel neural network architecture is proposed that uses latent vectors for assigning initial weights. The proposed approach, validated using data from a rubber manufacturing company, showed a significant improvement of up to 30% in predictive accuracy for physical properties with a restricted amount of data.