Application of Gaussian processes and transfer learning to prediction and analysis of polymer properties

Machine learning has shown superior performance in polymer properties prediction research. However, due to the complex experimental process or simulation process of polymers, it is not practical to obtain a large amount of data for fitting machine learning hyperparameters. This study aims to construct an effective performance prediction model using a small amount of data on polymers and the interaction between properties. In this paper we apply Gaussian Process Regression (GPR) and transfer learning for the establishment of polymer properties prediction models with few samples. Experiments show that when the algorithm is used to predict important properties such as diffusion coefficient and mechanical properties of polymers, the predicted value of the established model is close to the real value of the simulation, which verifies the validity of the model. Transfer learning can build a Gaussian model that describes the relationship between two performances, which not only improves the prediction accuracy of the prediction model, but also visualizes the relationship between performances and provides guidance for further experiments or simulations.

Automated summary

The study constructs an effective performance prediction model with a small amount of data by applying GPR and transfer learning for polymer properties prediction. The experiments validate the model's accuracy in predicting important polymer properties and provide guidance for further experiments or simulations. Transfer learning enhances the prediction accuracy and visualizes the relationship between properties.