Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK, 38000 Grenoble, France.
Emmanuel Jehanno, Romain Menegaux, Julien Mairal, Sergei Grudinin · 2025
Read Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. by Emmanuel Jehanno, Romain Menegaux, Julien Mairal, Sergei Grudinin (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-2509-00832.txt.
Introduction
Emmanuel Jehanno, with three co-authors, tackles crystal structure prediction. The target is non-polymeric molecules, packed as crystals. Predicting such structures is a prerequisite for designing materials and drugs. They focus on the molecular assembly problem. A set of identical rigid molecules must be packed into a crystal.
Recent methods have grown more sophisticated in technique. Yet the learning objective beneath them remains ill-posed. Many structures can score alike under that objective. The team proposes a better loss function for the task. It captures key geometric properties of molecules. It also ensures the order of molecules does not change the score. A simple regression model then beats prior approaches. It wins on the COD-Cluster17 benchmark, including against flow matching methods.
Catalog record
| CREATOR | Emmanuel Jehanno, Romain Menegaux, Julien Mairal, Sergei Grudinin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 90 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-00832
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. Emmanuel Jehanno, Romain Menegaux, Julien Mairal, Sergei Grudinin Abstract Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties. Yet, it is still an open challenge in materials design and drug discovery. Despite recent advances in computational materials science, accurately predicting three-dimensional non-polymeric crystal structures remains elusive. In this work, we focus on the molecular assembly problem, where a set 𝒮 \mathcal{S} of identical rigid molecules is packed to form a crystalline structure. Such a simplified formulation provides a useful approximation to the actual problem. However, while recent state-of-the-art methods have increasingly adopted sophisticated techniques, the underlying learning objective remains ill-posed. We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over 𝒮 \mathcal{S} . Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD). Our code is available at https://github.com/EmmanuelJhno/SinkFast 1 Introduction Generative modeling and deep learning have enabled rapid progress in the understanding and design of materials, molecules, and drugs. On the one hand, for material property prediction , advances in graph neural networks and transformers have significantly improved the understanding of molecular structures ( Joshi et al. 2023 ; Lin et al. 2023 ; Choudhary and DeCost 2021 ) , linking their three-dimensional (3D) geometry to physical and chemical properties. Particular attention has been paid to SE(3)-equivariant representations, which present higher expressivity by preserving geometric symmetries ( Schütt et al. 2021 ) . These methods have been adapted to crystalline structures, with their inherent challenges of infinite periodicity and rich symmetry patterns ( Yan et al. 2024a ) . Yan et al. 2022 ; Yan et al. 2024a ; Ito et al. 2025 yield state-of-the-art performance in property prediction of crystalline structures thanks to physically grounded methods, reflecting the need to integrate physics knowledge in models. On the other hand, for material design , generative models such as diffusion models ( Song et al. 2021 ) and flow matching methods ( Liu et al. 2023 ) have greatly enhanced the capacity to generate valid and diverse molecular and material structures ( Watson et al. 2023 ) . This work aims to combine these two aspects for the task of molecular assembly prediction , where a finite set of identical rigid molecules is packed into a crystalline structure. A fundamental step in designing a material with specific properties is to know its crystallization pattern. As represented in Figure 1 , a crystal is conventionally described by a unit cell, the smallest volume that contains all the structural and symmetry information necessary to generate the whole crystal by translation. This three-dimensional infinitely periodic shape largely determines the physical and chemical properties of the resulting material. This shape can be predicted either by regression ( Liang et al. 2020 ; Cao et al. 2024 ) or by flow matching/diffusion methods that allow for probabilistic answers ( Merchant et al. 2023 ; Xie et al. 2022 ; Luo et al. 2025 ; Pakornchote et al. 2024 ; Jiao et al. 2023 ) . Most of the previous methods model atoms in the unit cell individually. While such an approach works well ( Miller et al. 2024 ) for simple crystals of atomic point clouds from the Materials Project ( Jain et al. 2013 ) , the performance degrades on more complex molecular materials with symmetries other than translations. These contain internal point-group symmetries within the unit cell. An asymmetric unit (ASU) is defined as an elementary pattern of the unit cell, irreducible under the symmetry group transformations. A unit cell can be composed of multiple ASUs and an example is shown in Figure 1 -left . As this basic structure maintains a fixed internal structure, generating the crystal by directly predicting the ASU position, orientation, and symmetry operations in the world frame significantly reduces the dimensionality of the task, compared to moving each atom individually. In this setting, the goal of the molecular assembly prediction problem can be formulated as follows: given an elementary structure – an ASU –, predict its local crystalline structure, or in other words, how it packs in space. Related work Historically, the problem of computational material design has been extensively studied through the lens of Crystal Structure Prediction (CSP) challenge. This was first tackled through iterative process involving expensive spatial optimization ( Martínez et al. 2009 ) and energy assessment of predicted structures with first-principles calculations based on the density functional theory (DFT) ( Pickard and Needs 2011 ; Kresse and Furthmüller 1996 ) . However, these methods are slow, scale poorly with the number of atoms in the unit cell and thus may not be adapted to infinite materials. More recently, generative models have emerged as promising candidates for this task, especially for simple inorganic crystals ( Xie et al. 2022 ; Jiao et al. 2023 ; Levy et al. 2025 ; Nam et al. 2025 ) . However, they have not yet been widely adapted for complex organic materials. Nonetheless, very recent deep learning-based approaches study the molecular assembly prediction task by atom-wise ( Liu et al. 2024a ) and rigid-body ( Guo et al. 2025 ) flow matching. We detail the SE(3) flow matching setup in Appendix B.6 . Despite integrating sophisticated techniques, such as diffusion and flow matching, some essential building blocks for non-polymeric crystal structure prediction are still missing in these methods. A thorough related works section is provided in Appendix A . Contributions In this work we show the need of integration of domain-specific physics knowledge in the training scheme of models and the challenges that constitute the task of material generation. Our contributions can be summarised as follows: 1. Physics grounded loss. We show that a domain-specific rigid-body, model agnostic loss, grounded in physical principles, leads to improved prediction of crystalline structures. 2. Permutation-invariant loss. We propose an effective differentiable soft matching objective that is invariant to global geometric transformations and to the order permutation of repeated molecular units. 3. Remaining challenges. While the proposed domain-driven learning objective enables us to outperform prior approaches with a simple regression model, we also witness the challenges that remain to be tackled to reach real-world applicability. Figure 1 : A crystalline material at three different scales. From left to right: (a) The asymmetric subunit (ASU). (b) The unit cell with mirror images of the ASU. (c) The unit cell is repeated periodically in all three directions. Illustrations correspond to the COD-4316210 crystal structure from Crystallographic Open Database ( Gražulis et al. 2009 ) . 2 Problem setting Problem formulation A non-polymeric crystal is a solid material in which molecules are arranged in a highly ordered pattern repeating in the three spatial dimensions (3D). The asymmetric units ASU that constitute it are molecules that are identical objects in 3D. The unit cell is then defined by a finite number of symmetry operators applied to