Scaling law-informed machine learning for predicting thermal and electrical properties of polymers: A physics-based approach

Abstract

Polymer materials hold great promise for a wide range of applications due to their unique structures and properties. Using machine learning to predict their properties is a highly promising approach. However, the limited number of polymer databases hinders the application of machine learning methods. In this work, we choose to introduce physical prior knowledge at the output layer of the model by fitting the scaling law relationships satisfied by polymer to predict their properties. This approach of introducing physical prior knowledge at the output layer does not conflict with other methods of incorporating prior knowledge, such as molecular encoding and model pre-training, indicating its good portability and potential for application in more polymer property prediction models. We design the model structure from a perspective that aligns with physicochemical intuition. We first validate the model’s effectiveness and portability on a small dataset containing 1,070 data points and finally use this method to successfully predict the electrical conductivity of polymer electrolyte materials with an R² value of approximately 0.8. This modeling approach is expected to open up new avenues for polymer property prediction.