Transfer learning for materials informatics using crystal graph convolutional neural network

TL;DR

This paper introduces a transfer learning approach using crystal graph convolutional neural networks to improve property predictions in materials informatics, especially when data is limited.

Contribution

It proposes a pretrained crystal graph CNN model that enhances prediction accuracy for various material properties with small datasets.

Findings

TL-CGCNN improves prediction accuracy for multiple properties.

Prediction accuracy increases with larger pretrained datasets.

TL-CGCNN outperforms other regression methods on small datasets.

Abstract

For successful applications of machine learning in materials informatics, it is necessary to overcome the inaccuracy of predictions ascribed to insufficient amount of data. In this study, we propose a transfer learning using a crystal graph convolutional neural network (TL-CGCNN). Herein, TL-CGCNN is pretrained with big data such as formation energies for crystal structures, and then used for predicting target properties with relatively small data. We confirm that TL-CGCNN can improve predictions of various properties such as bulk moduli, dielectric constants, and quasiparticle band gaps, which are computationally demanding, to construct big data for materials. Moreover, we quantitatively observe that the prediction of properties in target models via TL-CGCNN becomes more accurate with an increase in size of training dataset in pretrained models. Finally, we confirm that TL-CGCNN is superior to other regression methods in the predictions of target properties, which suffer from small amount of data. Therefore, we conclude that TL-CGCNN is promising along with compiling big data for materials that are easy to accumulate and relevant to the target properties.