MoCE: Mixture of Experts Representation for Robust Crystal Material Property Prediction

ICLR 2026 Conference Submission19887 Authors

Keywords: Material Property Prediction, Geometric Deep Learning, Graph Neural Networks

Abstract: Accurately predicting the properties of crystal materials from their atomic structure is a fundamental challenge in materials science and computational chemistry. Graph-based models for crystal property prediction face a fundamental trade-off. They typically enforce a single geometric inductive bias, such as SE(3) invariance, which excels for periodic lattices but is less suited for materials where local chemistry and molecular conformation are dominant. To resolve this, we propose Mixture of Crystal Expert (MoCE) that dynamically integrates multiple, complementary geometric representations. Our model combines three specialized experts: an SE(3)-invariant module for global periodic structures, an SO(3)-invariant module for local atomic environments, and a dynamic graph module to learn latent topological interactions beyond fixed-bond assumptions. A gating network adaptively weighs each expert's contribution, tailoring the model's focus to the specific nature of a given crystal. This multi-representation approach achieves a more flexible and powerful framework, unifying the modeling of diverse crystal systems from rigid inorganic lattices to complex molecular crystals. Extensive validation on key benchmarks demonstrates state-of-the-art performance, confirming the effectiveness of our adaptive strategy.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 19887