Robust crystal property prediction from unrelaxed structures via expert-guided coordinate refinement

Graph neural networks are efficient surrogates for materials property prediction, but their reliance on computationally expensive relaxed geometries creates a major bottleneck in high-throughput screening. Here, we present RUGE-Net, an equivariant graph neural network that predicts properties directly from unrelaxed crystal structures. RUGE-Net integrates a Mixture-of-Experts (MoE) architecture, an iterative coordinate refinement module that progressively adjusts atomic positions to mitigate structural noise, and a perturbation-augmented training strategy. On the JARVIS database, RUGE-Net achieves competitive accuracy across multiple thermodynamic properties including formation and total energies, and maintains stable prediction performance under significant coordinate perturbations. We demonstrate that coordinate refinement and perturbation training are mutually essential to prevent catastrophic error amplification. Furthermore, the MoE gating network spontaneously develops interpretable element-specific specialization, routing covalent and ionic elements through distinct experts without explicit chemical supervision. By preserving near-perfect material ranking fidelity (Spearman ρ > 0.99) on noisy inputs, our approach enables rapid and reliable computational materials discovery directly from unrelaxed candidate structures.