Crystal twins: self-supervised learning for crystalline material property prediction

Abstract Machine learning (ML) models have been widely successful in the prediction of material properties. However, large labeled datasets required for training accurate ML are elusive and computationally expensive to generate. Recent advances Self-Supervised Learning (SSL) frameworks capable on unlabeled data mitigate this problem demonstrate superior performance computer vision natural language processing. Drawing inspiration from developments SSL, we introduce Crystal Twins (CT): a generic SSL method crystalline materials property that can leverage datasets. CT adapts twin Graph Neural Network (GNN) learns representations by forcing graph latent embeddings augmented instances obtained same system be similar. We implement Barlow SimSiam CT. By sharing pre-trained weights when fine-tuning GNN downstream tasks, significantly improve 14 challenging benchmarks.