Equivariant Hypergraph Diffusion for Crystal Structure Prediction

Yang Liu, Chuan Zhou, Shuai Zhang, Peng Zhang, Xixun Lin, Shirui Pan · 2025

Read Equivariant Hypergraph Diffusion for Crystal Structure Prediction by Yang Liu, Chuan Zhou, Shuai Zhang, Peng Zhang, Xixun Lin, Shirui Pan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-18850.txt.

Introduction

Yang Liu and co-authors predict crystal structures with a generative model that thinks in groups of atoms, not pairs, in a 2025 paper. Knowing a material's crystal structure unlocks its properties, but searching the possibilities is brutally hard. Diffusion models that dream up structures have shown promise, yet they describe crystals as graphs of pairwise bonds. Real chemistry runs on higher-order patterns: rings, cages, and clusters no pair list captures.

Their model replaces the graph with a hypergraph, whose edges can join many atoms at once. The diffusion process respects the symmetries crystals demand, including shifts that repeat by the lattice. The authors argue the richer representation describes complex structures more faithfully. The work extends the generative wave in materials science from molecules to the periodic world of solids.

Catalog record

CREATORYang Liu, Chuan Zhou, Shuai Zhang, Peng Zhang, Xixun Lin, Shirui Pan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-18850

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Equivariant Hypergraph Diffusion for Crystal Structure Prediction Yang Liu, Chuan Zhou, Shuai Zhang, Peng Zhang, Xixun Lin, Shirui Pan Abstract Crystal Structure Prediction (CSP) remains a fundamental challenge with significant implications for the development of new materials and the advancement of various scientific disciplines. Recent developments have shown that generative models, particularly diffusion models, hold great promise for CSP. However, traditional graph-based representations, where atomic bonds are modeled as pairwise graph edges, fail to fully capture the intricate high-order interactions essential for accurately representing crystal structures. In this work, we propose a novel approach that utilizes hypergraphs to represent crystal structures, providing a more expressive abstraction for modeling multi-way atomic interactions. By adopting hypergraphs, we can effectively capture complex high-order relationships and symmetries, such as permutation and periodic translation invariance, which are crucial for characterizing crystal structures. In this work, we propose the E quivariant H ypergraph Diff usion Model ( EH-Diff ), a generative model designed to take advantage of the symmetry-preserving properties of hypergraphs. EH-Diff exploits these features to offer an efficient and accurate method for predicting crystal structures with a strong theoretical justification to preserve invariance properties. Empirically, we conduct extensive experiments on four benchmark datasets, and the results demonstrate that EH-Diff outperforms state-of-the-art CSP methods with only one sample. 1 Introduction Crystal Structure Prediction (CSP), a long-standing challenge in the physical sciences since the early 2000s ( Desiraju 2002 ; Woodley & Catlow 2008 ) , aims to determine the stable 3D structure of a compound based solely on its chemical composition. Given the fundamental role of crystals in a wide range of materials ( Steiger 2005 ; Cartwright & Mackay 2012 ) , accurately predicting their spatial configuration is crucial ( Wooster 1953 ; Sato 2016 ) for understanding the physical and chemical properties that govern their applications in various academic and industrial domains, including drug development ( Zhang et al. 2023 ; Bannan et al. 2025 ) , battery technologies ( Xie & Grossman 2018 ; Yin et al. 2024 ) , and catalysis ( Bao et al. 2023 ) . Figure 1: Illustration of Crystal Structure Modeling. (a) Simple Graph and (b) Hypergraph. Interference on the hypergraph demonstrates effects involving the atom, with pairwise (first-order) interference evident in (a), while higher-order (non-pairwise) interference is observed in (b). Conventional CSP methods mainly use density functional theory (DFT) ( Kohn & Sham 1965 ) to iteratively evaluate the system energies. Techniques like random search ( Pickard & Needs 2011 ) and Bayesian optimization ( Yamashita et al. 2018 ) explore the energy landscape to identify stable local or global minima ( Oganov et al. 2019 ) . Predicting crystal properties remains challenging because of the inherent complexity of crystalline materials. Crystalline structures, characterized by periodic repetition of a unit cell ( Yan et al. 2022 ) in a 3D lattice, feature highly ordered atomic arrangements. The physical properties of these materials, including electronic, mechanical, and thermal characteristics, depend on their atomic configurations. Figure 2: Crystal Structure Illustration. (1) crystal structure in a Cartesian coordinate system; (2) crystal structure in a fractional coordinate system. (2) \to (3) orthogonal transformation of the lattice vectors; (2) \to (4) periodic translation of the fractional coordinates. Recent advances in AI4Material research have utilized graph-based deep learning models ( Xie & Grossman 2018 ; Xie et al. 2021 ) , with a particular focus on diffusion models ( Lin et al. 2024 ; Jiao et al. 2024 ) . These models employ graph-based diffusion techniques to generate multiple atoms while preserving the invariant properties within the models. However, traditional graph-based approaches, including graph neural networks and graph transformers, are restricted to pairwise relationships ( Kim et al. 2022 ) , where atoms are represented as nodes and atomic bonds as edges. This limitation restricts their ability to model higher-order interactions, which are crucial to capturing the structural complexity of crystal materials. Crystal structures often exhibit complex interactions, such as shared lattice points and multi-atomic coordination ( Voet & Rich 1970 ; Grimmer et al. 1974 ) , which are difficult to represent comprehensively using only pairwise edges. To overcome the limitation above, we employ hypergraphs , a robust graph data structure capable of capturing high-order relationships, to model the structure of crystal materials. Hypergraphs offer significant advantages over traditional graph representations by naturally encoding multi-atom interactions through hyperedges that link more than two atoms. In this framework, the model can offer a more expressive means of capturing the intricate coordination and structural dependencies within crystal structures. Hypergraphs represent atomic arrangements, enabling analysis of crystal stability, coordination geometry, and bond distributions. Their ability to incorporate symmetries and periodicity makes them ideal for modeling the fundamental properties and behaviors of crystalline materials. However, several challenges exist in the introduction of hypergraphs into CSP with generative models. First, a major challenge is effectively constructing hypergraph-based material representations. The next challenge involves developing diffusion models that operate on hypergraph data. Unlike traditional graph-based methods, hypergraph representations must capture complex multi-atomic interactions while preserving key structural invariants, such as permutation and periodic translation symmetries inherent in crystalline materials. Moreover, designing a hypergraph-based diffusion model is critical, as it introduces additional complexities. The model must simultaneously ensure that these symmetries are maintained throughout both the forward and reverse diffusion processes. This necessitates the integration of equivariant architectures capable of handling the high-order interactions encoded in hypergraphs, thereby ensuring accurate and symmetry-consistent predictions of crystal structures. In this work, we address the CSP problem by using a hypergraph-based modeling approach and we propose the Equivariant Hypergraph Diffusion Model, a novel framework designed to capture high-order interactions while preserving intrinsic symmetries and invariance properties of crystal structures (defined in Section 2.3 ). It simultaneously generates the lattice vectors and fractional coordinates of all atoms by employing a proposed denoising hypergraph model with periodic translation (defined in Section 3 ). Through an iterative process of expansion and refinement, our framework captures complex structural dependencies and trains a denoising diffusion model specifically for CSP, while preserving invariance properties (theoretical justification provided in Section 3.2 ). Empirically, we conduct comprehensive experiments across four real-world datasets and results position EH-Diff as a reliable approach for modeling complex crystalline materials, outperforming simple graph-based diffusion models. Our contributions are summarized as follows: • We formalize the Crystal Structure Prediction (CSP) problem by introducing a hypergraph modeling framework to describe high-order atomic interactions. To the best of our knowledge, this is the first work to represent crystal structures beyond traditional pairwise relationships, offering a novel perspective on capturing complex atomic coordination. • We propose the E quivariant H ypergraph Diff usion Model ( EH-Diff ), a novel framework that integrates hypergraph-based representation learning with a diffusion process. EH-Diff effectively captures high-order atomic interactions while explicitly preserving key symmetries, including permutation and periodic translation invariance, which are fundamental properties of crystal structures. • We evaluate the proposed framework through extensive experiments on four benchmark datasets, demonstrating its superior performance compared to state-of-the-art generative models with only one sample. 2 Problem Setup and Analysis In this section, we formally define the problem and provide a concise theoretical analysis of the problem. Some basic mathematical definitions are given in

Read the complete text