Abstract:One of the greatest challenges facing our society is the discovery of new innovative crystal materials with specific properties. Recently, the problem of generating crystal materials has received increasing attention, however, it remains unclear to what extent, or in what way, we can develop generative models that consider both the periodicity and equivalence geometric of crystal structures. To alleviate this issue, we propose two unified models that act at the same time on crystal lattice and atomic positions using periodic equivariant architectures. Our models are capable to learn any arbitrary crystal lattice deformation by lowering the total energy to reach thermodynamic stability. Code and data are available at <a class="link-external link-https" href="https://github.com/aklipf/GemsNet" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The paper primarily aims to address the following issues: ### Research Background and Objectives - **Challenges in Discovering New Materials**: A major challenge faced by current society is how to discover new crystalline materials with specific properties. These materials have wide applications in modern technology, such as metal alloys or semiconductors. - **Difficulties in Generating Crystalline Materials**: Although Machine Learning (ML) methods are widely used in high-throughput screening to find new materials, there are still many challenges in generating crystalline materials. The periodicity of crystal structures makes it difficult to train generative models. ### Solution - **Proposing Two Unified Models**: To address the above issues, the authors propose two unified models that can simultaneously handle crystal lattices and atomic positions, utilizing a periodic equivariant architecture. These models aim to overcome the limitations of existing methods that can only modify atomic positions or lattice geometry, achieving optimization of the entire crystal structure. ### Specific Contributions - **Equivariant Graph Neural Network (EGNN)**: This is a model extended from the Equivariant Message Passing Neural Network (EMPNN), which can operate on lattices and change atomic positions. Compared to EMPNN, it enhances the ability to manipulate atomic positions while maintaining equivariant properties. - **GemsNet**: This is a geometry-based architecture derived from GemNet, capable of modifying both atomic positions and lattice geometry while utilizing all geometric information. This is achieved by fully leveraging the geometric information of triplets in the Equivariant Graph Neural Network (EGNN). ### Experimental Validation - **Datasets**: The paper uses three datasets for experimental validation, including Perov-5, Carbon-24, and Mp-20, covering different types of crystalline materials, reflecting diversity in chemical composition and shape. - **Task Settings**: The experiments include two main tasks—**reconstruction task** (or lattice reconstruction task) and **denoising task** (or geometric optimization task). The former evaluates the model's ability to construct the correct lattice from scratch; the latter assesses whether the model can achieve thermodynamic stability by improving the structure's total energy, i.e., lowering the energy level to make the structure more stable. ### Results Analysis - **Performance Comparison**: The paper demonstrates the superior performance of GemsNet in various tasks, especially in the denoising task, where the reconstruction errors for both lattice parameters and atomic positions are significantly lower than other models. - **Qualitative Analysis**: Examples are provided to show how GemsNet effectively restores noisy structures to near-original stable crystal structures, proving the model's effectiveness and robustness. In summary, this paper provides effective solutions for the generation of crystalline materials by proposing two new unified models—EGNN and GemsNet, achieving breakthrough progress in simultaneously handling lattice geometry and atomic positions.

Unified Model for Crystalline Material Generation

Data-Driven Score-Based Models for Generating Stable Structures with Adaptive Crystal Cells

Physics guided deep learning for generative design of crystal materials with symmetry constraints

Vector Field Oriented Diffusion Model for Crystal Material Generation

Generative Design of Crystal Structures by Point Cloud Representations and Diffusion Model

Scalable Diffusion for Materials Generation

Predicting crystal growth via a unified kinetic three-dimensional partition model

WyCryst: Wyckoff Inorganic Crystal Generator Framework

Self-Supervised Generative Models for Crystal Structures

Unleashing the power of novel conditional generative approaches for new materials discovery

3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning

Generative Hierarchical Materials Search

Crystal Structure Generation with Autoregressive Large Language Modeling

Crystal Nucleation in Al-Ni Alloys: an Unsupervised Chemical and Topological Learning Approach

Statistically conditioned polycrystal generation using denoising diffusion models

CrysText: A Generative AI Approach for Text-Conditioned Crystal Structure Generation using LLM

Efficient Probabilistic Modeling of Crystallization at Mesoscopic Scale

Generative Inverse Design of Crystal Structures via Diffusion Models with Transformers

CrysToGraph: A Comprehensive Predictive Model for Crystal Materials Properties and the Benchmark

Crystal Structure Generation Based On Material Properties