Graph Contrastive Learning for Materials
This work addresses the need for efficient material property prediction in materials science by reducing reliance on costly labeled data, though it is incremental as it builds on existing contrastive learning and graph neural network techniques.
The paper tackles the problem of training graph neural networks for material property prediction without large labeled datasets by introducing CrystalCLR, a contrastive learning framework that uses material-specific transformations and a novel loss function. The result is that CrystalCLR learns representations competitive with engineered fingerprinting methods, improves prediction performance via finetuning, and produces clusters by compound class.
Recent work has shown the potential of graph neural networks to efficiently predict material properties, enabling high-throughput screening of materials. Training these models, however, often requires large quantities of labelled data, obtained via costly methods such as ab initio calculations or experimental evaluation. By leveraging a series of material-specific transformations, we introduce CrystalCLR, a framework for constrastive learning of representations with crystal graph neural networks. With the addition of a novel loss function, our framework is able to learn representations competitive with engineered fingerprinting methods. We also demonstrate that via model finetuning, contrastive pretraining can improve the performance of graph neural networks for prediction of material properties and significantly outperform traditional ML models that use engineered fingerprints. Lastly, we observe that CrystalCLR produces material representations that form clusters by compound class.