GEODE: Symmetry-Preserving Cartesian Diffusion for Crystal Generation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "GEODE: Symmetry-Preserving Cartesian Diffusion for Crystal Generation".
Paper summary: Mira: Thinking about this paper, "GEODE: Symmetry-Preserving Cartesian Diffusion for Crystal Generation," it seems the main implication is that explicitly treating symmetry in generative models can support both competitive structure generation and control over the resulting material properties. The authors achieved a metastable unique and novel rate of eleven point nine percent compared to seven point seven percent for the next best symmetry-aware model on unconditional generation > <ref:2610.01898#pg1,for the next best symmetry-aware model>
Kai: I think the title itself really sums it up—Symmetry-Preserving Cartesian Diffusion for Crystal Generation—it clearly tells you that this work tackles the difficulty of getting models to generate those symmetric arrangements that are so common in inorganic crystals. It moves beyond just generating random structures.
Lev: What it means for us on the hardware side is that if we want to build tools that generate materials, we can't afford to ignore these space group constraints; they have shown how this joint diffusion approach can handle those constraints while still delivering decent novelty scores >
Mira: And the fact that they show property steering capabilities later on, like controlling permittivity using classifier-free guidance with cubic templates, suggests this isn't just about generating pretty shapes anymore. It shows a path toward generating materials we know have certain functional properties.
Kai: So, for the listener driving or cooking, what does this mean? It means that when we use AI to design new materials—say, something with a specific electronic response—we can start using models like GEODE that respect the underlying atomic symmetry rules and give us a better chance at finding those useful structures >
Lev: And they admit there are still limitations. They mention that while they improved novelty by filtering by training-set frequency and orbit count, it still raises the mSUN rate to twenty-three point eight zero percent after relaxation without further training >
Mira: That’s a fair caveat; even with these improvements, getting that perfect unique structure is still tough, but this work shows the way forward for models that need to be both creative and structurally sound >
Conclusion: Kai: So, we’re wrapping up on GEODE, which is about making AI generate crystal structures that actually look like real inorganic materials by respecting their symmetry rules >
Mira: Yeah, and they did this by combining a template sampling step with a diffusion model that keeps the symmetry locked in at every single step >
Kai: The authors are focusing on this joint distribution over crystals and templates, which is how they handle the structure generation part of things >
Lev: From an engineering standpoint, it’s neat because they're not just throwing random atoms together; they’re starting with a template that already has the correct symmetry constraints built in >
Mira: And it seems like this approach gives them a really good result, achieving a novel rate of eleven point nine percent for unconditional generation >
Kai: Eleven point nine percent compared to seven point seven percent for the next best symmetry-aware model, which shows they're doing something solid here >
Lev: It’s still relatively low, but when you look at what they did—the joint diffusion in Cartesian space—it gives the coordinate noise a consistent physical scale that we empirically showed improves performance >
Mira: That scaling of the noise is key, because it means the model isn't just making up random coordinates; it’s respecting some physical reality about how atoms are positioned >
Kai: So, what does this mean for us when we talk about inverse materials design? Does this mean we can start getting structures that are both novel *and* physically plausible right out of the gate >
Lev: Well, they do show property steering capabilities too; they can use cubic templates and classifier-free guidance to control things like permittivity while keeping the material dielectric isotropic >
Mira: That’s interesting because it moves past just generating a structure to actively controlling what that structure *does* electronically >
Kai: So, the authors conclude that this two-stage training process, where they train the template generator and crystal generator independently, is a way to get both good generation and property control >
Lev: They also point out that their final result after relaxation on the MP-twenty dataset hits a rate of sixteen point four percent, which is their highest pre-relaxed entry for that data set >
Mira: So it shows that models with explicit symmetry treatment can handle both competitive generation and property control >
Kai: It’s a big step forward showing how structure and function can be guided simultaneously >
Department of Chemistry, Molecular Sciences Research Hub · Imperial College London
cond-mat.mtrl-sci
Submitted: 2026-10-01
Updated: 2026-11-18
Code: https://github.com/spglib/moyo
The gist: Most known inorganic crystals exhibit symmetric atomic arrangements, yet generative models often fail to reproduce them GEODE first samples symmetry templates…
Key concepts
- Joint Diffusion Formulation
- GEODE uses a joint diffusion model that simultaneously learns the crystal structure and its symmetry template. The distribution of crystals and templates is factored, allowing the model to generate both the physical arrangement and the required symmetry constraints together during its generation process.
- Wyckoff Constraint
- Wyckoff constraints are mathematical rules used to define symmetric atomic arrangements in crystallography. GEODE uses these constraints within a VAE component to guide the sampling of symmetry templates, ensuring that generated structures adhere to known crystallographic symmetries.
- Coordinate Noise Scaling
- The model introduces a consistent physical scale for coordinate noise during diffusion, which is empirically shown to improve generation performance. This scaling method helps the diffusion process maintain physically realistic atomic arrangements while navigating the complex space of crystal structures.
Terminology
Summary
Most known inorganic crystals exhibit symmetric atomic arrangements, yet generative models often fail to reproduce them GEODE first samples symmetry templates, then jointly generates the lattice, atomic coordinates and atom types while preserving the specified symmetry with a novel Wyckoff-constrained loss Cartesian diffusion gives coordinate noise a consistent physical scale that we empirically demonstrate improves performance Unconditional generation achieves a metastable, unique and novel (mSUN) rate of 11. p. 09%, compared to 7. p. 07% for the next best symmetry-aware model
Model Architecture and Methodology
GEODE is introduced as the first model to combine coordinate and lattice diffusion in Cartesian space GEODE first samples a symmetry template, then jointly generates the lattice, atomic coordinates and atom types while preserving the specified symmetry The joint distribution over crystals and templates factorises as p(C, τ) = pθ(C τ) pψ(τ), where z is the VAE latent variable The diffusion model is conditioned only on the template, and both its forward and reverse processes preserve the template’s symmetry at every step
Training Signal Construction
The training signal is built through a two-stage process involving a Wyckoff Template VAE and a diffusion model The template VAE uses an E(3)-equivariant message-passing encoder to parameterise a Gaussian posterior over the latent vector z The training objective for the VAE is LVAE = λspg LCE + λwyc LNLL + λKL DKL(qϕ(z C) ∥ N (0, I)) The diffusion model is trained to predict a scaled target σts⋆ using a wrapped-normal score target modified to respect Wyckoff constraints Permutation remapping is used to match atoms to reference sites within the same Wyckoff orbit using minimum-cost linear assignment
Performance Gains and Ablations
The gains come from several key innovations, including the joint diffusion formulation and template filtering GEODE achieves the highest metastable, unique and novel (mSUN) rate among symmetry-aware models, achieving 11.9% compared with 7. p. 07% for the next competitor Sampling time template filtering increases mSUN by approximately 6% without retraining, making GEODE competitive with leading symmetry-agnostic models at negligible additional cost Replacing Cartesian with fractional coordinate diffusion lowers mSUN from 11. p. 19% to 9.3%, consistently scoring below GEODE across sampling runs Filtering by training-set frequency and orbit count raises mSUN to 23. p. 880% after relaxation, without further training
Property Steering Capabilities
Joint symmetry and property guidance are demonstrated by combining cubic templates with classifier-free guidance to control permittivity while ensuring dielectric isotropy The fine-tuning of the MP-20 diffusion model on the AnisoNet dataset allows for control over predicted permittivity through feature-wise linear modulation (FiLM) layers At guidance strength γg = 5, targets of 2, 5, 10 and 15 give median predicted permittivities of 2.1, 6.2, 10.5 and 12.2 respectively
Conclusion on Competitive Advantage
GEODE achieves an mSUN rate of 16.40% after relaxation on MP-20, which is the highest pre-relaxed MP-20 entry in Table 1 The two stages are trained independently, so the template generator receives no feedback on the quality of the resulting crystals Future work could investigate training the template generator using the outcomes of crystal generation GEODE shows that models with explicit symmetry treatment can support both competitive generation and control over the properties of generated crystals
AI Use Statement
We used large language model assistants throughout this work The authors verified all outputs of these assistants, including code, results and references The research questions, the model design and the interpretation of the results are the authors’ own, and the authors are responsible for the contents of this paper
Ethics Statement
This work introduces a generative model for inorganic crystal structures The training data are the publicly available MP-20 and Alex-MP-20 sets, and the generated structures are candidate materials whose stability and properties are predicted, not measured; any use would require experimental validation
Reproducibility Statement
The model and both diffusion processes are specified in Section 4, and the sampling procedure is given in full as Algorithm 1 in Appendix B Training and evaluation details are collected in Appendices G and H Code is provided as supplementary material and will be released publicly upon publication
References
Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. p. 9 Structured denoising diffusion models in discrete state-spaces Luis Barros-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M. p. 9 Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, Matt Uyttendaele, C. p. 9 Lawrence Zitnick, and Zachary W. p. 9 Ulissi. The Open Materials 2024 (OMat24) inorganic materials dataset and models Ilyes Batatia, Dávid Péter Kovács, Gregor N. p. 9 C. Simm, Christoph Ortner, and Gábor Csányi. p. 9 MACE: Higher order equivariant message passing neural networks for fast and accurate force fields Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. p. 9 Elena, Dávid P. p. 9 Kovács, Janosh Riebesell, Xavier R. p. 9 Advincula, Mark Asta, Matthew Avaylon, William J. p. 9 Baldwin et al. p. 9 A foundation model for atomistic materials chemistry Siddharth Betala, Samuel P. p. 9 Gleason et al. p. 9 LeMatGenBench: A unified evaluation framework for crystal generative models Erik Bitzek, Pekka Koskinen, Franz Gähler, Michael Moseler, and Peter Gumbsch. p. 0 Structural relaxation made simple Cyprien Bone et al. p. 7 Discovery and recovery of crystalline materials with property-conditioned transformers Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. p. 0 Space group constrained crystal generation Rees Chang et al. p. 0 Space group equivariant crystal diffusion Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. p. 9 Analog bits: Generating discrete data using diffusion models with self-conditioning Stefano Curtarolo et al. p. 9 The high-throughput highway to computational materials design Daniel W. p. 0 Davies, Keith T. Butler, Adam J. Jackson et al. SMACT: Semiconducting materials by analogy and chemical theory Bowen Deng, Peichen Zhong et al. p. 9 CHGNet as a pretrained universal neural network potential for chargeinformed atomistic modelling Filip Ekström Kelvinius et al. p. 1 WyckoffDiff: A generative diffusion model for crystal symmetry Scott Fredericks, Kevin Parrish, Dean Sayre, and Qiang Zhu. p. 0 PyXtal: A Python library for crystal structure generation and symmetry analysis Tin Hadži Veljkovic et al. p. 7 Crystalite: A lightweight transformer for efficient crystal modeling Theo Hahn (ed. p. 0). International Tables for Crystallography, Volume A: Space-Group Symmetry Jonathan Ho and Tim Salimans. p. 0 Classifier-free diffusion guidance Jonathan Ho, Ajay Jain, and Pieter Abbeel. p. 0 Denoising diffusion probabilistic models Anubhav Jain, Shyue Ping Ong et al. p. 9 Commentary: The Materials Project: A materials genome approach to accelerating materials innovation Rui Jiao, Wenbing Huang, Peijia Lin et al. p. 0 Crystal structure prediction by joint equivariant diffusion Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. p. 2 Space group constrained crystal generation Zhuotao Jin et al. p. 0 DynaCrys: Crystal generation with dynamic space-group diffusion Nikita Kazeev et al. p. 9 Wyckoff transformer: Generation of symmetric crystals Harold W. p. 0 Kuhn. The Hungarian method for the assignment problem Daniel Levy, Siba Smarak Panigrahi et al. p. 9 SymmCD: Symmetry-preserving crystal generation with diffusion models Yi-Lun Liao, Alexander J. p. 0 Hoffman, Sabrina C. Shen et al. EquiformerV3: Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers Yuchen Lou and Alex M. p. 0 p. 0 Ganose. Discovery of highly anisotropic dielectric crystals with equivariant graph neural networks Anubhav Jain, Shyue Ping Ong et al. p. 9 Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis Shunye Ping Ong, William Davidson Richards et al. p. 9 Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis Hyunsoo Park and Aron Walsh. p. 6 Guiding generative models to uncover diverse and novel crystals via reinforcement learning Hyunsoo Park, Zhenzhu Li, and Aron Walsh. p. 6 Has generative artificial intelligence solved inverse materials design? p. 9 Matter Hyunsoo Park, Anthony Onwuli, and Aron Walsh. p. 6 Exploration of crystal chemical space using text-guided generative artificial intelligence Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen et al. EquiformerV3: Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers Yuchen Lou and Alex M. Ganose.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Better crystal generation by combining coordinate and lattice diffusion in Cartesian space, which GEODE achieves, allows for a
consistent physical scale
of noise across different cell dimensions compared to models that use fractional coordinates wherethe physical scale and shape of coordinate perturbations depend on the shape and size of the cell.
-
Improved structure uniqueness and novelty through template selection filtering, as GEODE's sampling time template filtering increases mSUN by approximately 6% without retraining, making it competitive with leading symmetry-agnostic models.
-
Enhanced property control through joint symmetry and property guidance, GEODE demonstrates this by combining cubic templates with classifier-free guidance to control permittivity while ensuring dielectric isotropy.
-
More robust training targets via Wyckoff-constrained loss, the introduction of a
Wyckoff-constrained loss with permutation remapping that restricts coordinate targets to reachable sites and accounts for equivalent atom assignments
reduces noise in the coordinate diffusion process. -
Increased generation quality by utilizing template selection, as filtering by training set frequency and orbit count raises mSUN to 23.80% after relaxation without further training, demonstrating
template selection can adjust the balance between novelty and metastability at sampling time.
-
Property steering capabilities are improved by fine-tuning on property labels using FiLM layers during generation, allowing for control over predicted permittivity with guidance strengths of 2, 5, 10 and 15.
Sources
- Structured Denoising Diffusion Models in Discrete State-Spaces
- LeMat-GenBench: A Unified Evaluation Framework for Crystal Generative Models
- Discovery and recovery of crystalline materials with property-conditioned transformers
- Crystalite: A Lightweight Transformer for Efficient Crystal Modeling
- DynaCrys: Crystal Generation with Dynamic Space-Group Diffusion
- EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers
- Space Group Conditional Flow Matching
- Orb-v3: atomistic simulation at scale
- Multi-property directed generative design of inorganic materials through Wyckoff-augmented transfer learning
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion