acceptodds
Under review as a conference paper at ICLR 2027

GEODE: Symmetry-Preserving Cartesian Diffusion for Crystal Generation

Abstract

Most known inorganic crystals exhibit symmetric atomic arrangements, yet generative models often fail to reproduce them. Explicitly enforcing these symmetries has so far yielded fewer stable and novel structures than unconstrained generation. We introduce Generative Equivariant Orbit Diffusion Engine (GEODE), to our knowledge the first model to combine coordinate and lattice diffusion in Cartesian space. 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 11.9%, compared with 7.7% for the next best symmetry-aware model. We also introduce sampling time template filtering, which increases mSUN by 6% without retraining, making GEODE competitive with leading symmetry-agnostic models. Template selection also enables joint symmetry and property guidance, which we demonstrate through classifier-free guidance of permittivity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, materials discovery, crystallographic symmetry, crystal structure generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4UQ9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a method for symmetry aware crystal generation called GEODE (Generative Equivariant Orbit Diffusion Engine), building on other symmetry aware methods like SGFM (Punt et al, 2025), SGEquiDiff (Chang et al, 2025), and DynaCris (Jin et al, 2026). GEODE generates a crystal in two parts: first, a VAE is trained to generate a crystal template τ=(G,W)\tau = (G, W) consisting of the space group GG and a list of Wyckoff orbits WW. Given a template, PyXtal is used to generate an initial crystal consistent with the template. At this point, GEODE undertakes a single joint diffusion in i) the lattice parameters, ii) the atom types, and iii) the atom positions (expressed in a cartesian coordinates instead of fractional), conditioned on the template τ\tau to arrive at the final generated crystal. This diffusion preserves the symmetry template throughout to guarantee that the generated crystal is still consistent with the template. The authors then provide an evaluation of de novo generation on the MP-20 and the larger Alex-MP-20 datasets, where GEODE provides generally improved metrics over other symmetry aware methods, especially in the mSUN rate (metastable, unique, and novel structures). The authors then give several ablation studies showing the various effects of filtering templates, various diffusion settings (e.g. cartesian vs fractional diffusion), among others.

Critical strengths

-The Paper propose a new symmetry aware method for de novo crystal generation that does a good job of decoupling templates specifying the symmetry and the properties to be diffused. This allows for a simple filtering method that drastically increases quality metrics (in particular mSUN) by getting rid of improbable templates for a specific problem.

-The Paper provides comprehensive evaluations demonstrating the effectiveness of their method along with ablations to demonstrate future directions of study, with property guidance perhaps being the most interesting for real world applications.

Critical weaknesses

-While the mSUN metric was better than other symmetry aware structures, the rest of the metrics seem to be marginal improvements (at best) over existing symmetry aware methods. Furthermore, while GEODE does close the gap versus symmetry agnostic methods, I found it interesting that overall they still seem to dominate the metrics. Thus, the question of how useful GEODE is for real world chemistry remains uncertain.

-The diffusion is computationally expensive as currently laid out, requiring 1000 reverse steps per sample, with each step entailing a full pass through the network. This limits throughput, and as the authors point out, some other method like flow matching might be preferred.

Questions for authors

-While the mSUN was reported for all methods, what about the SUN rate? In general, why is the definition of "metastable" useful here? Is it the fact that these crystals are more likely to become stable after relaxation?

-From reading this, it seems like GEODE gives evidence towards doing as much joint diffusion as possible to increase expressivity. Would it ever be possible to also incorporate diffusion into the Wyckoff sites themselves (or even the spacegroups too), giving the model further flexibility?

-In Section 5.6, I didn't understand if there was anything special to GEODE that made it particularly special to property steering (apart from easy template filtering). In principle, it seems like most other symmetry aware crystal generative models could also do this fine-tuning process to learn to generate crystals with a certain material property.

4 comments

(Part 1 of 3)

We thank the reviewer for highlighting the value of separating the symmetry template from the generation of the crystal structure, and for their questions about the evaluation and possible extensions.

W1a. Are the improvements over existing symmetry-aware methods marginal on metrics other than mSUN?

We agree that some individual metrics, such as validity, improve only slightly. For materials discovery, however, what matters is how often the same candidate is low in energy, unique and novel. Each criterion is relatively easy to satisfy in isolation. A model can generate low-energy structures by reproducing the training set, while random structures can be both unique and novel but are rarely low in energy. Meeting all three criteria simultaneously makes mSUN a more demanding measure of generation quality.

The comparison with SGFM illustrates this point. On MP-20, SGFM generates a higher fraction of metastable structures than unfiltered GEODE (45.7% versus 41.9%), yet GEODE generates substantially more candidates that are simultaneously metastable, unique and novel (11.92% versus 7.68% of all samples). GEODE also more than doubles the SUN rate (0.62% versus 0.28%) and achieves a lower mean energy above the hull (0.20 versus 0.25 eV/atom), before either template filtering or pre-relaxation.

W1b. How does GEODE compare with symmetry-agnostic models, and what does symmetry control add for materials discovery?

Several symmetry-agnostic models outperform unfiltered GEODE on MP-20, although the ranking depends on the model and metric. After pre-relaxation with a machine-learned interatomic potential, unfiltered GEODE achieves higher SUN and mSUN rates than pre-relaxed MatterGen (1.20% versus 0.24% SUN and 16.40% versus 15.00% mSUN). Crystalite achieves higher rates on both measures (1.50% SUN and 22.60% mSUN).

On Alex-MP-20, Chemeleon 2 achieves the highest mSUN rate after reinforcement learning with rewards for uniqueness, novelty and energy above the hull. This raises its mSUN rate from 11.80% to 62.40%, while its SUN rate falls from 0.56% to 0.20%. GEODE does not use reward fine-tuning, so the Chemeleon 2 base model provides a more direct comparison. Unfiltered GEODE achieves 25.78% mSUN before pre-relaxation and 29.20% afterwards. The pre-relaxed model also achieves the highest SUN rate on this dataset (1.32%).

For real-world materials discovery, we need confidence that generations are realistic suggestions. One aspect of this is that generations should reproduce the kinds of crystal symmetry seen in experiment. LeMat-GenBench's metrics do not capture this aspect of generation quality. This limitation has previously been highlighted in the literature, e.g., Jin et al. introduce the SSUN@τ\tau metric, which additionally requires a non-P1P1 space group after relaxation, in addition to being valid, unique and stable.

Crucially, only 1.20% of Alex-MP-20 training structures belong to P1P1, the space group with no symmetry beyond translations. By comparison, 40.00% of MatterGen samples and 95.52% of Chemeleon 2 base samples belong to P1P1. A search using these models would therefore systematically undersample crystals with nontrivial symmetry. For GEODE the fraction is 0.94%. GEODE also more closely reproduces the full point-group distribution, with a Jensen–Shannon distance of 0.06 from the training set, compared with 0.59 for MatterGen and 0.96 for Chemeleon 2 base.

This matters directly when a real-world application calls for a particular symmetry. In Section 5.6, for example, we search for dielectrics with an isotropic response and a target permittivity. With GEODE (or most other symmetry-aware models), we can restrict generation to cubic templates, which ensures isotropy, and use property guidance to target the desired permittivity. Every generated candidate therefore meets the symmetry requirement before we evaluate its predicted permittivity.

(Part 2 of 3)

Q1a. What are the SUN rates, and how should they be interpreted alongside mSUN?

We now report SUN alongside mSUN in all tables. GEODE achieves the highest SUN rate on both datasets, with 2.12% on MP-20 for filtered, pre-relaxed GEODE and 1.32% on Alex-MP-20 for unfiltered, pre-relaxed GEODE.

ModelValid (%) ↑\uparrowUnique (%) ↑\uparrowNovel (%) ↑\uparrowMetastable (%) ↑\uparrowSUN (%) ↑\uparrowmSUN (%) ↑\uparrowEhullE_{\mathrm{hull}} (eV/atom) ↓\downarrowRMSD (Å) ↓\downarrow
MP-20
Symmetry-agnostic
Crystalite*97.20‾\underline{97.20}95.8053.2051.60‾\underline{51.60}1.5022.600.09‾\underline{0.09}0.13‾\underline{0.13}
Chemeleon 2 RL95.2088.1071.6039.800.0421.200.160.42
MCFlow*97.20‾\underline{97.20}96.30‾\underline{96.30}52.2049.300.7018.900.100.17
MatterGen*95.7095.1070.5033.400.2415.000.180.39
DiffCSP95.7094.8066.2029.800.128.500.280.59
Symmetry-aware
SGFM95.2093.9042.6045.700.287.680.250.44
DiffCSP++95.3095.1062.0026.400.164.960.410.70
SGEquiDiff91.9091.3058.2022.700.082.920.530.90
CrystalFormer69.7069.1030.8028.000.002.560.670.65
WyFormer93.7093.3066.2016.500.082.400.490.77
SymmCD73.4073.0047.0018.600.082.360.880.87
GEODE95.8094.9055.8041.900.6211.920.200.42
GEODE*95.6094.9054.3043.701.2016.400.140.15
GEODE-filtered*96.4095.3074.00‾\underline{74.00}38.802.12‾\underline{2.12}23.80‾\underline{23.80}0.160.20
Alex-MP-20
Chemeleon 2 RL97.40‾\underline{97.40}93.5093.70‾\underline{93.70}67.400.2062.40‾\underline{62.40}0.100.27
Chemeleon 194.0093.9059.8054.400.9628.000.110.31
CrystaLLM-π\pi93.4093.2023.9071.30‾\underline{71.30}0.3212.480.090.19
GEODE95.0094.90‾\underline{94.90}54.1057.701.0625.780.100.30
GEODE*95.0094.90‾\underline{94.90}52.5061.801.32‾\underline{1.32}29.200.08‾\underline{0.08}0.11‾\underline{0.11}
GEODE-filtered*94.8094.6064.2061.701.2036.760.090.12

Table 1 of the paper with the SUN rate added, 2,500 structures per model. ∗^{*}Relaxed before evaluation. Best value per dataset underlined.

An important detail of LeMat-GenBench is that mSUN excludes stable structures. SUN counts unique and novel structures with Ehull≤0E_{\mathrm{hull}}\leq 0, while mSUN counts those with 0<Ehull≤0.10<E_{\mathrm{hull}}\leq 0.1 eV/atom. In practice, when searching for new materials, the metric of interest would be the combination of both stable and low-energy metastable candidates. On MP-20, filtered, pre-relaxed GEODE gives the highest combined rate (25.92%), followed by pre-relaxed Crystalite (24.10%).

Q1b. Why include metastable structures? Are they expected to become stable after relaxation?

A structure can remain above the hull after relaxation and still be experimentally accessible. Relaxation can bring a structure to a local energy minimum even when another phase, or combination of phases, has a lower energy. Sun et al. found that 50.5% of 29,902 experimentally observed inorganic phases lay above the calculated 0 K hull (Sun et al., Sci. Adv. 2, e1600225, 2016).

There are several reasons for this. Entropy can stabilise a phase at finite temperature, and kinetic barriers can allow a metastable phase to persist (e.g. diamond is metastable above the ground state of graphite). The DFT energies used to train the machine-learned interatomic potentials also carry uncertainties that affect estimated hull distances. Including candidates up to 0.1 eV/atom above the hull therefore lets us retain plausible materials that a strict ground-state criterion would exclude.

(Part 3 of 3)

W2. Does the cost of 1,000 reverse steps limit throughput?

With the same 1,000-step sampler, the optimised implementation now generates 2,500 MP-20 structures in 2.16 h on one Quadro RTX 6000, down from 17.4 h. The corresponding reported times are 6.06 h for MatterGen and 3.5 min for Crystalite. The EquiformerV3 denoiser now accounts for most of GEODE's generation time. Our symmetry-conserving framework is agnostic to the choice of denoiser. Replacing it with Crystalite's would bring the total generation time down significantly. We describe the implementation changes in the general response.

Q2. Could jointly diffusing the Wyckoff sites or even the space group give GEODE more flexibility?

This is an interesting suggestion that we are currently exploring. Allowing the Wyckoff assignments or space group to change during denoising would let GEODE revise its initial symmetry template as the atom types, coordinates and lattice evolve. The difficulty is that changing a symmetry assignment also changes which atomic coordinates can move independently. For example, lowering the space group can split one Wyckoff orbit into several, allowing previously equivalent atoms to move independently. If the symmetry instead increases, those atoms must be brought into positions that are exactly related by the higher symmetry, and their atom types must agree. A joint diffusion process would therefore need to update the discrete symmetry assignment, the atom types and the continuous geometry consistently at each transition. For our subsequent work, we are exploring whether these symmetry changes can be incorporated directly into GEODE's coordinate and lattice diffusion. Whether the added flexibility improves generation quality remains an open question.

DynaCrys explores a simpler version of this problem (arXiv:2608.07401, 2026). It jointly diffuses only the atom types, space group and Wyckoff occupations. The coordinates are generated in a second stage with all three held fixed. This makes the symmetry transitions considerably easier to handle because the model can change the occupied Wyckoff orbits without having to map the free coordinates of the old orbits onto those of the new ones.

Q3. Is property steering specific to GEODE?

We agree that other symmetry-aware models can also be fine-tuned for property-guided generation. CrystalFormer-RL, for example, has been used to target dielectric and band-gap properties (Cao and Wang, arXiv:2504.02367, 2025).

The purpose of Section 5.6 is to show how symmetry control and property guidance can be used together. Restricting GEODE to cubic templates requires no additional training and ensures that the dielectric tensor is isotropic. Property guidance then steers the magnitude of the predicted permittivity towards the target value. We will clarify this distinction in Section 5.6.

Reviewer_4UQ9OpenReview ↗

Response to the Authors

I thank the authors for their thorough responses to my questions. I also think the performance improvements detailed are significant to the work as the overall pipeline is much more usable at this point. I also think it would be interesting to see how using Crystalite's decoder further improves throughput. In general, I am very optimistic about symmetry-aware approaches to this problem, and feel like they are ultimately the right way to solve this. Thus, I have increased my score as I think this would be a fine paper to accept. I think the only remaining reservation I have is that it's still a bit unclear to me how much these additional architecture changes/joint diffusion really help over the existing symmetry aware approaches.

Reviewer_vpa4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GEODE, a symmetry-preserving Cartesian diffusion framework for crystal generation. The framework first samples space groups and Wyckoff positions using a template VAE, then jointly generates lattice parameters, atomic coordinates, and atom types under the corresponding symmetry constraints. A Wyckoff-aware denoising objective incorporates permutation remapping and reachable periodic copies. Experiments on MP-20 and Alex-MP-20 demonstrate strong generation performance and reliable symmetry preservation, with further improvements through inexpensive template filtering. The framework also enables property-guided generation under a fixed symmetry class.

Critical strengths
  1. The paper addresses the useful problem of combining Cartesian coordinate diffusion with lattice generation under space-group constraints. Using Cartesian noise gives coordinate perturbations a consistent physical scale across different cells. The controlled comparison with fractional-coordinate diffusion supports this choice, showing improvements in generation quality primarily through increased metastability.

  2. The paper demonstrates a useful combination of symmetry constraints and property guidance by generating cubic crystals with target dielectric constants. This provides a concrete example of how explicit symmetry control can support property-oriented crystal design.

Critical weaknesses
  1. The paper leaves an important theoretical connection underexplored. When an input crystal has a given space-group symmetry, an appropriately equivariant network constrains its outputs to respect that symmetry, regardless of whether it has been trained. Atoms in the same Wyckoff orbit should receive identical type probabilities. Coordinate updates should transform consistently under space-group operations and remain within the directions permitted by site symmetry. The second-order lattice tensor should satisfy the point-group constraints. The paper discusses the last property through Neumann’s principle, but the first two warrant a similar treatment. Actually, under exact equivariance to the target space group and an exactly symmetric input, the group projection introduced in Section 4.3 would leave the predicted coordinate field unchanged, serving only to correct numerical deviations in an otherwise exact implementation. Developing this connection would explain how symmetry preservation across types, coordinates, and lattices follows from a common principle, and what the network learns within these constraints. Connecting these results would provide an elegant, unified account of how crystallographic constraints emerge through equivariance and deepen the paper’s scientific contribution. The reviewer would raise the score if the authors develop this unified perspective in depth and clarify the conditions under which it holds.
Questions for authors
  1. What motivated the choice to diffuse cell lengths and angles? Have the authors explored alternatives based on a symmetric lattice matrix or its logarithm, as in related approaches such as MatterGen and DiffCSP++? Since the lattice head already predicts a second-order symmetric tensor, it would be interesting to understand whether these representations offer advantages or introduce practical difficulties.

  2. Have the authors explored other ways of coupling Cartesian coordinate diffusion with lattice diffusion? Appendix E.1 explains the motivation for directly defining the coordinate corruption kernel and using alternating coordinate and lattice updates during sampling. It would be interesting to hear whether alternative coupling schemes were considered, and what practical observations led to the current formulation.

3 comments

(Part 1 of 2)

We thank the reviewer for highlighting the connection between equivariance and the symmetry constraints on atom types, coordinates and the lattice. We agree that explaining this connection would clarify both what the architecture guarantees and what the denoiser learns.

W1. Under what conditions does equivariance preserve the symmetry of the input crystal, and when is the group projection needed?

The reviewer is correct that these constraints follow from a common principle. If a symmetry operation leaves the input crystal unchanged and the denoiser is equivariant to that operation, then applying the operation to the predicted output must also leave it unchanged. This holds for any choice of network weights, including before training. For a periodic crystal, the operation includes the induced permutation of equivalent atoms, and the input graph must preserve that symmetry.

The consequences depend on what the network predicts. Consider a space-group operation with rotational part RR that maps atom ii to atom jj. The two atoms must receive the same type probabilities, pj=pip_j=p_i, and their predicted displacements must satisfy Δrj=RΔri\Delta\mathbf r_j=R\Delta\mathbf r_i. If the operation fixes atom ii, its displacement must therefore lie in a direction left unchanged by the operation. Applying this condition to every operation that fixes the site gives the allowed Wyckoff directions. The same argument requires the predicted lattice tensor TT to satisfy T=RTRTT=RTR^\mathsf{T} for every operation in the point group, as in Neumann's principle. Within these constraints, the denoiser learns which atom types are likely, how far and in which allowed direction an atom should move, and how the lattice should deform.

The coordinate projection plays different roles during corruption and denoising. During corruption, an unconstrained noise draw generally breaks the template symmetry. Projecting the noise gives displacements that follow the allowed site directions and transform consistently across each orbit. During denoising, as the reviewer notes, projecting the predicted coordinate field would leave it unchanged in exact arithmetic if the input were exactly symmetric and the denoiser were equivariant to every operation in the template.

GEODE's EquiformerV3 denoiser is SE(3)SE(3)-equivariant. Its equivariance covers translations and proper rotations, including the operations in a screw axis, but does not guarantee the required transformation under a mirror, glide or inversion. For templates containing these operations, GEODE projects the predicted coordinate updates and assigns the same atom type to all sites in each Wyckoff orbit. The coordinate projection also corrects numerical drift during sampling. When we omitted it, we observed deviations from the template symmetry within about 50 steps. We will add the common equivariance argument to the appendix and clarify which constraints follow from the denoiser's equivariance and which GEODE enforces explicitly.

In future work, we could use an E(3)E(3)-equivariant denoiser such as MACE. Given an exactly symmetric input, its predicted coordinate updates would already respect the full space-group symmetry, so projecting these updates would serve only to correct numerical drift. Equivalent sites would also receive identical type probabilities, and sampling one atom type per orbit would preserve that symmetry in the sampled structure.

(Part 2 of 2)

Q1. Why diffuse cell lengths and angles? Did you consider a symmetric lattice matrix or its logarithm?

We chose cell lengths and angles because the constraints for each crystal system can be imposed directly by tying lengths or angles together and fixing angles where required. A symmetric lattice matrix, as in MatterGen, or a matrix-logarithm representation, as in DiffCSP++, can also preserve lattice symmetry. To understand the practical differences, we compared the noisy cells produced by the three forward processes on 200 MP-20 test structures (see figure below).

The main difficulty was the occasional highly compressed cell. GEODE builds a neighbour graph from all periodic neighbours within 7 Å, so a single compressed cell can make a graph too large to process. The mean neighbour count for MatterGen was manageable, but individual noisy cells in our comparison produced more than 10,000 neighbours for a single atom. Figure 1b shows these rare cases more clearly than the mean in Figure 1a. Our length-and-angle process kept the maximum neighbour count lower over the noise draws we tested. This was the practical reason for our choice.

Figure: https://postimg.cc/876GcCjW

Number of neighbours within 7 Å of each atom as a function of the lattice diffusion timestep for GEODE, MatterGen and DiffCSP++. Panel (a) shows the mean and standard deviation across atoms in 200 MP-20 test structures. Panel (b) shows the largest neighbour count for any atom on a logarithmic scale, highlighting highly compressed cells that are difficult to detect from the mean alone.

Q2. Did you consider other ways of coupling Cartesian coordinate diffusion to lattice diffusion?

We considered both orders for the coordinate and lattice updates. The practical issue is that changing the cell also changes the Cartesian positions of atoms held at fixed fractional coordinates. A Cartesian displacement predicted in the original cell must therefore be transformed if it is applied after the cell changes.

In the forward process, GEODE first deforms the clean cell and then adds coordinate noise in the resulting cell, rt=LtL0−1r0+σtϵr.\mathbf r_t=\mathbf L_t\mathbf L_0^{-1}\mathbf r_0 +\sigma_t\boldsymbol\epsilon_r. During sampling, the denoiser predicts both updates in a single evaluation at (Lt,rt)(\mathbf L_t,\mathbf r_t). Its predicted Cartesian displacement is therefore expressed in the frame of the current cell. The simplest consistent implementation applies that displacement first, then updates the lattice while holding fractional coordinates fixed. This separation of internal atomic displacements from homogeneous cell deformation is also used in models of crystal relaxation (Tadmor et al., Phys. Rev. B 59, 235, 1999).

Alternatively, applying the lattice update first gives exactly the same result if the displacement is transformed with the cell. For a lattice deformation F=I+εF=\mathbf I+\boldsymbol\varepsilon, this follows directly from F(r+Δr)=Fr+FΔrF(\mathbf r+\Delta\mathbf r)=F\mathbf r+F\Delta\mathbf r. Applying the original displacement after the deformation would omit the term εΔr\boldsymbol\varepsilon\Delta\mathbf r. Thus, the two consistent update orders are equivalent.

The distinction from fractional-coordinate diffusion lies in the corruption kernel and the training loss. GEODE defines the noise scale and measures displacement errors in ångströms, so a given displacement has the same physical meaning across cells. Our MP-20 ablation tests this choice of coordinates: switching to fractional-coordinate diffusion reduced mSUN from 11.92% to 9.28%.

The reviewer appreciates the clarification and will keep the current positive score. Given that preserving the space group symmetry is central to this work, the reviewer still finds an E(3)-equivariant denoiser a more natural choice. It would respect reflections directly, and it remains unclear whether the extra freedom of the SE(3)-equivariant model is useful for this task.

Reviewer_SUFNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GEODE, a two-stage model for generating new inorganic crystal structures while keeping their symmetry. First, a VAE samples a "symmetry template": a space group plus a list of Wyckoff orbits, which are the sets of symmetry-equivalent atom sites. Then a diffusion model generates the cell, atom positions and atom types together, keeping the template's symmetry at every step. The authors proposes a "Wyckoff-constrained" loss that ignores target sites an atom can't reach along its allowed directions, and it re-pairs equivalent atoms with their nearest sites. On the MP20 dataset, GEODE reaches the best mSUN compared with Symmetry-agnostic and Symmetry-aware methods. GEODE generated structures broadly match the training set's point-group distribution. Only 1.7% fall in the trivial space group P1, versus 82% for MatterGen.

Critical strengths
  1. On the MP20 dataset, the proposed method reaches the best mSUN compared with Symmetry-agnostic and Symmetry-aware methods.
  2. On the MP20 dataset, structures generated by the proposed method matchs the training set's point-group distribution. Only 1.7% fall in the trivial space group P1, versus 82% for MatterGen.
  3. The proposed template filtering strategy can improve mSUN without retraining.
Critical weaknesses
  1. On a large dataset Alex-MP-20, the mSUN of the proposed method is significantly worse compared to a symmetry-agnostic baseline, i.e., GEODE reaches an mSUN rate of 29.20%, compared with 62.4% for Chemeleon 2. On the other hand, the authors do not report point-group distribution comparison (Figure 3) on the Alex-MP-20 dataset. This raises a concern that whether if the main argument, i.e., existing methods like MatterGen / Chemeleon 2 have difficulty learning the target space group while the proposed GEODE can, holds only for small datasets.

  2. The proposed method is very inefficient, i.e., 17 hours for generating 2500 structures compared to 6.06 h for MatterGen and 3.5 min for Crystalite.

Questions for authors

Following on weakness 1. For a larger dataset like Alex-MP-20, how much space group gain does GEODE obtain by sacrifising the quality of the generated structures significantly compared to existing symmetry-agnostic methods? Can the authors generate Figure 3 for Alex-MP-20 with MatterGen (MatterGen was trained on Alex-MP-20 originally), Chemeleon 2 and GEODE?

3 comments

We thank the reviewer for the questions on Alex-MP-20 and on the generation cost, which we answer below.

W1+Q1. How well does GEODE reproduce the symmetry distribution on Alex-MP-20, and how does this compare with mSUN?

On Alex-MP-20, GEODE closely reproduces the training-set point-group distribution, with a Jensen–Shannon distance of 0.06 before pre-relaxation. After pre-relaxation, the distance increases to 0.28, compared with 0.59 for MatterGen, 0.96 for Chemeleon 2 base and 0.97 for Chemeleon 2 RL. The symmetry breaking we observe during pre-relaxation is due to the SE(3)SE(3)-equivariant EquiformerV3 relaxation model.

The 62.40% mSUN rate cited by the reviewer is for Chemeleon 2 after RL fine-tuning with rewards for uniqueness, novelty and energy above the hull. GEODE does not use this additional optimisation, so Chemeleon 2 base provides a more direct comparison. Its mSUN rate is 11.80%, compared with 25.78% for GEODE before pre-relaxation and 29.20% afterwards. MatterGen, Chemeleon 1 and CrystaLLM-π\pi also do not apply RL fine-tuning. Against them, GEODE reaches an mSUN rate in a comparable range and the highest SUN rate of all models evaluated. RL fine-tuning could also be applied to GEODE, although we have not tested its effect.

ModelmSUN (%)SUN (%)P1P1 (%)JS distance
Alex-MP-20 training set0.840.001.200.00 (reference)
Chemeleon 2 RL62.400.2099.480.97
Chemeleon 2 base11.800.5695.520.96
MatterGen34.480.8040.000.59
Chemeleon 128.000.9650.240.64
CrystaLLM-π\pi12.480.32N/AN/A
GEODE25.781.060.940.06
GEODE*29.201.325.010.28
GEODE-filtered*36.761.204.800.36

JS denotes the Jensen–Shannon distance between the generated and training-set point-group distributions. A distance of zero indicates identical distributions. An asterisk marks GEODE structures pre-relaxed with EquiformerV3.

RL fine-tuning raises Chemeleon 2's mSUN rate from 11.80% to 62.40%, while its stable fraction falls from 2.60% to 0.36% and its SUN rate falls from 0.56% to 0.20%. Fine-tuning also changes the chemistry of the generated structures. For Chemeleon 2 RL, 65% of structures fall in chemical systems absent from the LeMat-Bulk reference, including alloys with four to seven elements. The corresponding fraction is 3–7% for the other models.

Point-group distribution on Alex-MP-20

The figure below compares the Alex-MP-20 training-set point-group distribution (purple) with those generated by (a) GEODE, (b) MatterGen, (c) Chemeleon 2 RL and (d) Chemeleon 2 base (green). Across all point groups, the JS distance from the training set is 0.06 for GEODE before pre-relaxation and 0.28 after pre-relaxation, compared with 0.59 for MatterGen and 0.96–0.97 for the two Chemeleon 2 models.

Figure: https://postimg.cc/K1MbjYRR

Generation cost

As detailed in the general response, GEODE now generates 2,500 MP-20 structures in 2.16 h, compared with 6.06 h for MatterGen.

Dear Reviewer, thank you again for your helpful feedback. We appreciate that our response was posted close to the discussion deadline. With the discussion period ending tomorrow, we wanted to highlight the additional comparison of symmetry distributions on Alex-MP-20 and GEODE’s substantially reduced generation time. If you have any remaining questions, we would be happy to address them before the deadline.

I appreciate the authors additional experiments and responses. The results resolved my concerns about the efficiency and the effectiveness of the method on large datasets. Adding RL to GEODE might be helpful for further improvement, but the current status looks good to be accepted. The optimization for the generation cost should be included in to the manuscript. I will raise my score to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Optimized generation speed

We thank all reviewers. This response addresses the generation cost, which Reviewers 1 and 3 raised.

Generation cost

We have reduced the time needed to generate 2,500 MP-20 structures from 17.4 h to 2.16 h on one Quadro RTX 6000. We used the same trained model and sampler throughout. We made three changes:

  1. denoiser and memory optimisations, including a fused attention kernel;
  2. graph construction on the GPU;
  3. evaluation of the denoiser on the asymmetric unit (ASU) instead of the conventional cell.
ImplementationTimeSpeedup
Original implementation17.40 h1.00×
Denoiser and memory optimisations11.60 h1.50×
Graph construction on the GPU14.50 h1.20×
(1) and (2) together6.63 h2.60×
+ denoiser on the ASU2.16 h8.10×
MatterGen6.06 h–

The optimised implementation gives the same outputs up to single-precision rounding. For 2,500 structures, the reported generation times are 2.16 h for GEODE, 6.06 h for MatterGen and 3.5 min for Crystalite. GEODE is now faster than MatterGen, although Crystalite remains substantially faster than both. The EquiformerV3 denoiser now accounts for most of the generation time. In future work, we will explore whether a less expensive denoiser, such as the one used in Crystalite, can reduce this time further.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.