Machine learning (ML) models have shown great promise in predicting material properties, but they often require large labeled datasets, which are expensive and time-consuming to generate. This paper introduces Crystal Twins (CT), a self-supervised learning (SSL) method that leverages large unlabeled datasets for crystalline material property prediction. CT employs a twin Graph Neural Network (GNN) architecture, learning representations by enforcing similarity between embeddings of augmented instances from the same crystalline system. Using Barlow Twins and SimSiam frameworks, CT significantly improves GNN performance on 14 material property prediction benchmarks.

Rishikesh Magar, Yuyang Wang, Amir Barati Farimani