TL;DR
This paper introduces two novel unified models, EGNN and GemsNet, that can simultaneously generate and optimize crystalline materials by acting on both atomic positions and lattice geometry. This is a crucial step towards discovering new materials with desired properties, as it allows for the exploration of a wider range of stable crystal structures.
What Problem Does This Paper Address?
Discovering new crystalline materials with specific properties is a grand challenge for society, with applications ranging from solar energy to hydrogen storage. While machine learning has shown promise in generating organic molecules, generating crystalline materials is significantly more complex due to their inherent three-dimensional periodicity and the need to consider both atomic arrangements and the overall lattice structure. Existing generative models often struggle to simultaneously modify both the atomic positions and the crystal lattice geometry, limiting their ability to find thermodynamically stable structures. This paper aims to bridge this gap by proposing models that can jointly optimize these aspects.
Methods
The authors propose two unified models:
- Equivariant Graph Neural Network (EGNN): This model extends the existing EMPNN framework by incorporating a component that modifies atomic positions while maintaining equivariance to group actions. It uses a message-passing schema to process chemical and geometric information from edges and triplets, updating atomic features and predicting vector fields to deform the lattice.
- GemsNet: Inspired by the GemNet model for organic chemistry, GemsNet is adapted for periodic structures. It adds modules to predict actions on the lattice geometry alongside atomic positions. This model leverages triplets from the graph representation and incorporates radial and angular basis functions (RBF and CBF) to capture geometric features, allowing for simultaneous modification of atomic positions and lattice shape.
Both models are designed to be equivariant, meaning their predictions are independent of the representation of the crystal structure (e.g., orientation or cell choice). This is achieved by incorporating group actions related to Euclidean transformations and lattice transformations.
Datasets
The models were evaluated on three datasets:
- Perov-5: Contains perovskite (cubic) structures with uniform shapes but diverse chemical compositions.
- Carbon-24: Composed of carbon atoms with a wide variety of shapes.
- Mp-20: A subset of the Materials Project, considered representative of ordinary structures with a large sample of shapes and chemical compositions.
The authors used the same training, validation, and test splits as reported in a previous work ([Xie et al., 2022]).
Key Insights and Ideas
- Unified Geometry and Atomic Position Control: The core innovation is the simultaneous control over both atomic positions and lattice geometry. This is crucial because a stable crystal structure depends on the interplay between these two factors.
- Equivariance for Robustness: By enforcing equivariance to group actions, the models ensure that their predictions are independent of how the crystal is represented, making them more robust and generalizable.
- Addressing Complex Structures: GemsNet shows promise in handling more complex structures, including those with multiple elements and heavy atoms, which are notoriously difficult for existing generative models.
- Efficiency Gains: GemsNet demonstrates a significant advantage in inference time and parameter count compared to existing models, making it more practical for large-scale material discovery.
Key Results
- Reconstruction Task: Both EGNN and GemsNet achieved good reconstruction performance, restoring original crystal structures from noisy inputs. GemsNet generally outperformed EGNN, showing lower reconstruction errors for lattice parameters and atomic positions across the tested datasets.
- Denoising Task: The models demonstrated the ability to deform crystal lattices to improve their total energy, leading to more stable structures. GemsNet showed a reduction in reconstruction error for angular parameters (up to 10%), lengths (up to 40%), and atomic positions (up to 13%) on the Mp-20 dataset.
- Handling Outliers: While GemsNet performed well on most structures, it showed occasional larger errors on outlier structures, particularly those with complex compositions or unusual elements. However, it still managed to process some of these difficult cases.
Limitations and Future Work
- Outlier Performance: The paper acknowledges that while GemsNet performs well on average, it can exhibit larger errors on outlier structures, which are often more complex and challenging to analyze even with traditional methods.
- Future Work: The authors suggest that their equivariant architecture opens avenues for other applications, such as the generation of stable crystalline structures, which remains a challenging problem. Further research could focus on improving the robustness of the models to handle a wider range of complex and outlier crystal structures.
Summary updated