Title: ReciNet: Leveraging Reciprocal Space for Long-Range Modeling in Crystalline Property Prediction

Abstract:

Accurately predicting material properties based on crystal structures remains a core but difficult objective in materials science. Because crystal lattices consist of infinite, repeating atomic arrangements, effective modeling requires methods that can simultaneously capture both local and global structural information. Yet, existing approaches often struggle to adequately represent long-range interactions inherent in these periodic systems. To overcome this limitation, we utilize reciprocal space—the intrinsic framework for periodic crystals—and develop a Fourier series representation derived from fractional coordinates and reciprocal lattice vectors, employing learnable filters.

Building on this foundation, we present ReciNet, a new architecture designed for reciprocal space-based geometric modeling. This model combines geometric graph neural networks with reciprocal blocks to effectively simulate both short-range and long-range atomic interactions. Our evaluation across extensive benchmarks, including JARVIS, Materials Project, and MatBench, reveals that ReciNet delivers superior predictive accuracy across various crystal property prediction tasks. Furthermore, we investigate an extension of the model for multi-property prediction using a mixture-of-experts approach. This variant exhibits high computational efficiency and demonstrates positive transfer effects among correlated properties. These results underscore the potential of ReciNet as a robust, scalable, and precise solution for predicting crystal properties.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC