CrysGNN: Distilling Pre-trained Knowledge to Enhance Property Prediction for Crystalline Materials

In recent years, graph neural network (GNN) based approaches have emerged as a powerful technique to encode complex topological structure of crystal materials in an enriched repre- sentation space. These models are often supervised nature and using the property-specific training data, learn relation- ship between different properties like formation energy, bandgap, bulk modulus, etc. Most these methods require huge amount property-tagged data train system which may not be available for prop- erties. However, there is availability with its chemical composition structural bonds. To leverage untapped this paper presents CrysGNN, new pre-trained GNN framework crystalline materials, captures both node level information graphs unla- belled material data. Further, we extract distilled knowledge from CrysGNN inject into state art erty predictors enhance their property prediction accuracy. We conduct extensive experiments show that model, all SOTA algo- rithms able outperform own vanilla version good margins. also observe distillation process provides significant improvement over conventional ap- proach finetuning model. will release model along large dataset 800K crys- tal carefully curated; so can plugged any existing upcoming