ReciNet: Reciprocal Space-Aware Long-Range Modeling for Crystalline Property Prediction
Researchers have developed ReciNet, a novel deep learning architecture designed to predict crystalline properties. This model effectively captures both short-range and long-range atomic interactions by leveraging reciprocal space, a natural domain for periodic crystals. Experiments on multiple benchmarks show ReciNet achieves superior predictive accuracy for various crystal property prediction tasks. An extension using a mixture-of-experts approach further enhances computational efficiency and demonstrates positive transfer learning between correlated properties. AI
IMPACT Introduces a novel deep learning architecture for materials science, potentially accelerating discovery and design of new crystalline materials.