Abstract
The discovery of novel materials depends on identifying stable crystal configurations, which are defined by periodic arrangements of atoms. Traditional machine learning models often fail to account for the inherent symmetries of these structures, such as lattice "repaving" and Euclidean transformations. We introduce the Equivariant Message Passing Neural Network (EMPNN), a framework designed to handle these symmetries explicitly. By integrating the modular group $SL_3(\mathbb{Z})$ and the Euclidean group $Euc(3)$, EMPNN learns to deform and relax crystal lattices in an unsupervised manner. Our results demonstrate that EMPNN significantly improves lattice accuracy and successfully identifies lower-energy configurations across multiple datasets.
Introduction
Material discovery is fundamentally a search for stable crystal structures—configurations where the system's energy is minimized. Crystals are characterized by their periodicity, represented mathematically by a unit cell that tiles 3D space. However, a single physical crystal can be described by infinitely many different unit cells (lattices). Standard neural networks lack equivariance to these choices, leading to inconsistent predictions for identical materials. This paper addresses the "repaving" problem by developing a model that understands the fundamental geometry of periodic structures.
Methods: Equivariant Message Passing
The core of our approach is the Equivariant Message Passing Neural Network (EMPNN). Unlike standard Graph Neural Networks (GNNs), EMPNN simultaneously updates atomic positions and the global lattice tensor.
Symmetry Integration
The model is designed to be equivariant to the product of three critical groups: permutations ($S_n$), Euclidean transformations ($Euc(3)$), and the modular group ($SL_3(\mathbb{Z})$). The latter accounts for the different ways a lattice can be defined without changing the underlying physical structure.
Model Architecture
The architecture consists of stacked Message Passing Neural Network (MPNN) and Periodic Equivariant Graph Neural Network (PEGNN) layers. These layers use vector fields and gradients of geometric invariants (e.g., interatomic distances and bond angles) to guide infinitesimal deformations of the lattice.
Experimental Results
We evaluated EMPNN on three datasets: Carbon-24, Mp-20 (diverse materials), and Perov-5 (highly symmetric perovskites). In a denoising task where stable crystals were randomly deformed, EMPNN demonstrated superior performance. On the Carbon-24 dataset, EMPNN improved lattice length predictions by nearly 4x compared to invariant feed-forward baselines. Furthermore, for the Perov-5 dataset, the model successfully moved 93.8% of noisy structures toward lower energy states, as validated by high-fidelity Density Functional Theory (DFT) calculations.
Conclusion
EMPNN represents a significant step forward in applying deep learning to crystallography by explicitly modeling the $SL_3(\mathbb{Z})$ symmetry. While current limitations include the handling of chiral structures and the need for integrated element prediction, the model's ability to perform unsupervised relaxation offers a fast, geometrically-grounded alternative to expensive physics-based simulations.
Summary updated