acceptodds
Under review as a conference paper at ICLR 2027

SLayerGen: a Crystal Generative Model for all Space and Layer Groups

Abstract

Crystal generative models have shown rapid progress for accelerating the discovery of bulk, periodic materials. However, many material systems such as 2D superconductors, thin film semiconductors, and catalytic surfaces are diperiodic, i.e., aperiodic along one of the lattice directions. These systems are invariant under the layer groups, which are known to influence materials properties yet not considered by existing models. In this paper, we propose SLayerGen, a generative model that produces crystals constrained to be invariant to any space or layer group. SLayerGen consists of coarse-to-fine discrete autoregressive lattice generation; transformer-based autoregressive sampling of Wyckoff positions, elements, and numbers of symmetrically unique atoms; and space or layer group equivariant diffusion of atomic coordinates. To facilitate progress in generative modeling of diperiodic materials, we assembled and filtered datasets of monolayers and bilayers, propose relevant evaluation metrics, and developed novel representations for layer group symmetries. For de novo generation of diperiodic materials, SLayerGen achieves consistent performance gains over bulk crystal generative models and is competitive when training jointly on bulk and diperiodic materials.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
space group, layer group, crystal, diperiodic, 2D material, diffusion, generative, autoregressive, equivariant

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aTQXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author/s take SGEquiDiff-style crystal generation and extends it to layer groups, so 2D materials like monolayers and bilayers come out with exact layer-group symmetry while bulk crystals still work.

Critical strengths
  • As far as I know, this is the first model that enforces exact layer-group symmetry, and the crystallographic groundwork is careful.

  • The benchmark fills a real gap for 2D generation

Critical weaknesses
  • I don't see a consistent gain from layer groups. DiffCSP++ gets a lower layer-group JSD than SLayerGen on every dataset, though the paper marks SLayerGen as best on three. SLayerGen also trails the best P1 baseline on S.U.N. structure rate, and its own P1 variant does as well or better on both S.U.N. rates where tested.

  • the method is mostly SGEquiDiff by Chang et al. (Space Group Equivariant Crystal Diffusion) adapted to layer groups. Symmetry-aware 2D generation also exists already, e.g., Xu et al. (Discovery of 2D Materials via Symmetry-Constrained Diffusion Model) and Rønne and Hammer (Atomistic Generative Diffusion for Materials Modeling, not cited), so saying prior models ignore layer groups is too strong.

Questions for authors
  • the layer group is sampled from training frequencies, so why is the layer-group JSD of SLayerGen worse than DiffCSP++ and, on two datasets, SLayerGen3D? -what can the layer-group model do that the P1 variant cannot?
3 comments

We thank the reviewer for the thoughtful comments and questions. Below we aim to address your concerns. If you find our answers helpful, we would appreciate it if you would consider raising your score.

I don't see a consistent gain from layer groups

We highlight that SLayerGen consistently does well on the SUN template rate. We will also update the text to specify that SLayerGen consistently wins on JSD only for the 18 layer groups that don’t map injectively to a space group (table 8). The reason that DiffCSP++ performs well on JSD over all 80 groups is that (1) many of them do map injectively to a space group and (2) DiffCSP++ fixes generated crystals to have combinations of space groups and Wyckoff positions in the empirical distribution; in contrast, SLayerGen must learn the Wyckoff positions, which is a more difficult learning task but enables the model to generate more diverse templates.

We also highlight results from experiments requested by reviewer hdGM, wherein we found SLayerGen is the only benchmarked model able to generalize to rare and held-out layer groups in Alex2D. Specifically, we trained SLayerGen with all crystals from 3 layer groups held out (rare layer groups 31 and 8, and the common layer group 11), and then constrained SLayerGen to generate crystals with those layer group symmetries at inference time. Impressively, we find SLayerGen generalizes to held-out layer groups as evident by nonzero SUN structure/template rates and rediscovery of a few heldout crystals. Meanwhile, DiffCSP and MatterGen achieve zero success at generating any crystals in rare layer groups even when trained on them, while by construction, DiffCSP++ cannot generate held-out groups. "In group” is the rate of generated structures with the prescribed layer group. “Rediscovered” denotes held-out structures generated by the model. Brackets give 95% CIs.

\hline \text{Layer group (train count)}&\text{Model}&\text{Group in training}&\text{In group}&\text{Valid}&\text{UN}&\text{Stable}&\text{SUN}&\text{SUN template}&\text{Rediscovered}\\\\ \hline \text{LG 8 (53)}&\text{SLayerGen}&\text{Yes}&96.5&65.0\ {\scriptstyle[62.9,\ 67.1]}&81.5&57.6\ {\scriptstyle[54.5,\ 60.6]}&41.8\ {\scriptstyle[38.8,\ 44.9]}&8.4\ {\scriptstyle[6.8,\ 10.3]}&3/29\\\\ &\text{SLayerGen}&\text{No}&96.4&58.7\ {\scriptstyle[56.5,\ 60.8]}&80.2&52.9\ {\scriptstyle[49.8,\ 56.0]}&38.0\ {\scriptstyle[35.0,\ 41.0]}&5.9\ {\scriptstyle[4.6,\ 7.5]}&3/29\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 11 (883)}&\text{SLayerGen}&\text{Yes}&99.9&71.0\ {\scriptstyle[69.0,\ 73.0]}&91.8&52.0\ {\scriptstyle[48.9,\ 55.1]}&45.6\ {\scriptstyle[42.5,\ 48.7]}&3.0\ {\scriptstyle[2.1,\ 4.3]}&5/373\\\\ &\text{SLayerGen}&\text{No}&99.8&53.4\ {\scriptstyle[51.3,\ 55.6]}&91.3&41.2\ {\scriptstyle[38.2,\ 44.3]}&34.4\ {\scriptstyle[31.5,\ 37.4]}&3.3\ {\scriptstyle[2.4,\ 4.6]}&1/373\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 31 (42)}&\text{SLayerGen}&\text{Yes}&98.3&71.2\ {\scriptstyle[69.2,\ 73.2]}&79.6&44.8\ {\scriptstyle[41.7,\ 47.9]}&31.1\ {\scriptstyle[28.3,\ 34.0]}&3.7\ {\scriptstyle[2.7,\ 5.1]}&2/17\\\\ &\text{SLayerGen}&\text{No}&99.5&79.9\ {\scriptstyle[78.1,\ 81.6]}&80.7&32.1\ {\scriptstyle[29.3,\ 35.1]}&22.5\ {\scriptstyle[20.0,\ 25.2]}&1.6\ {\scriptstyle[1.0,\ 2.6]}&1/17\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \end{array}$$ > Symmetry-aware 2D generation also exists already, e.g., Xu et al. and Rønne and Hammer, so saying prior models ignore layer groups is too strong Thank you for your comment. We will update the text to state that prior models do not enforce layer group symmetry during generation. Xu et al. uses the same framework of DiffCSP with post-hoc projection to specified space groups; they are not invariant to layer groups nor do they allow constrained generation to layer groups which map non-injectively to the layer groups. We benchmarked the space group constrained diffusion model DiffCSP++ as a direct analog to Xu et al.’s work. Thank you for sharing the Rønne and Hammer model AGeDi; we will make sure to cite it. However, we highlight that AGeDI is not symmetry constrained and instead softly conditions on layer groups using classifier free guidance; doing so achieves only ~38% accuracy in sampling the correct layer group as noted in their paper, while SLayerGen enforces layer group symmetries by construction.

the layer group is sampled from training frequencies, so why is the layer-group JSD of SLayerGen worse than DiffCSP++ and, on two datasets, SLayerGen3D?

Thank you for the question - we note that the JSD is computed only over structurally and compositionally valid crystals. We will update the text to specify where SLayerGen actually wins, which is on JSD for the 18 layer groups that don’t map injectively to a space group (table 8). The reason that DiffCSP++ performs well on JSD over all layer groups is that (1) many of them do map injectively to a space group and (2) DiffCSP++ fixes generated crystals to have combinations of space groups and Wyckoff positions in the empirical distribution; in contrast, SLayerGen must learn the Wyckoff positions, which is a more difficult learning task but enables the model to generate more diverse templates. We will promote table 8 to the main text to avoid confusion.

what can the layer-group model do that the P1 variant cannot?

The layer group model can generate realistic, high symmetry crystals and be constrained at inference time to any desired layer group. Given our result that SLayerGen can generalize to rare or unseen groups, we believe this is a strong capability given that many exotic quantum and topological properties of 2D materials are highly symmetry-dependent. In table 6, we also show that the layer group model has significantly better JSD over the layer groups than the P1 model variants, demonstrating improved learning of the distribution of layer group symmetries.

My main concern still stands. The claim of the best overall layer-group JSD was wrong, since DiffCSP++ is lower on every dataset. The response does not explain why SLayerGen trails SLayerGen3D on two datasets, and it does not address the gap to the P1 baselines on S.U.N. structure rate. On top of that, your own P1 variant beats SLayerGen on both S.U.N. rates on both datasets where it was tested. So the main benefit of layer groups is symmetry control, not higher S.U.N. rates than P1 models. I will keep my score. Thanks for the response.

Reviewer_hdGMOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors address an important problem in material generation; however, they focus on generating diperiodic materials that repeat in two directions. Their symmetries are described by layer groups rather than space groups, and no existing model accounts for them. Hence, they want to build a generative model for crystal materials that are invariant under the layer groups. Towards this goal they propose a space- and layer group-aware generative model (SLayerGen). SLayerGen combines SE(3)-invariant, coarse-to-fine sampling of discretized lattices; permutation-invariant, transformer-based autoregressive generation of occupied Wyckoff positions and atomic elements; and space- or layer-group-equivariant diffusion of atomic coordinates.

Critical strengths
  1. The problem is interesting, and the paper is well written. The problem is well motivated and enough background study is provided to understand the context.
  2. I like the idea of Coarse-to-fine discrete lattice sampling: two-level classification gives high precision with a small output space, captures multimodal lattice distributions, and enforces symmetry constraints exactly through masking. This is a really novel idea.
  3. The authors conducted extensive experiments with multiple datasets and evaluated their model.
  4. Code and Database are provided in supplimentary materials
Critical weaknesses
  1. The main weakness of the model is, authors claimed they want to develop a model that is invariant to layer/space group G, however they set p(G) to the empirical training distribution. That means p(G) follows the training distribution, so groups absent from training have zero probability. Also, Rare groups are poorly learned. Groups with few training examples get little data for the lattice, Wyckoff, and coordinate stages, so their generation quality is likely poor. An ablation/qualititive per-group results would be interesting to observe.

  2. Also, in Table-2/3/4 SUN & UN for DiffCSP, MatterGen is much better compared to SLayerGen (Structure). This shows model is not able to sample many unique and novel materials, that means not much diversity is gained. This limits practical application of the model.

  3. Few Typos Line 092: autoregressive Line 421: BJ-damped Line 1125: enantiomorphic Line 1394: tuned manually

Questions for authors

See the weaknesses

2 comments

We thank the reviewer for the thoughtful comments and questions. Below we aim to address your concerns. If you find our answers helpful, we would appreciate it if you would consider raising your score.

however, they focus on generating periodic materials that repeat in two directions

We highlight that we performed experiments for SLayerGen on both diperiodic materials (tables 1-4) and 3D bulk materials (table 5), but we agree that the primary focus is on diperiodic materials which prior literature has focused less on.

The main weakness of the model is, authors claimed they want to develop a model that is invariant to layer/space group G, however they set p(G) to the empirical training distribution. That means p(G) follows the training distribution, so groups absent from training have zero probability. Also, Rare groups are poorly learned. Groups with few training examples get little data for the lattice, Wyckoff, and coordinate stages, so their generation quality is likely poor.

Thank you for raising this concern. The empirical p(G) is only a default, and the layer group can be set freely as an input. We have conducted additional experiments on Alex2D wherein we trained SLayerGen with all crystals from 2 rare layer groups (layer groups 31, 8) and one common group (LG11) held out, and then constrained SLayerGen to generate crystals with those layer group symmetries at inference time. Impressively, we find that SLayerGen generalizes to these held-out layer groups as evident by nonzero SUN structure/template rates and rediscovery of a few crystals from the heldout set. Meanwhile, DiffCSP and MatterGen achieve zero success at generating any crystals in the rare layer groups even when trained on them, while by construction, DiffCSP++ cannot generate any held-out groups. In the table below, “In group” signifies the rate of generated structures with the prescribed layer group. “Rediscovered” denotes unseen validation and test structures generated by the model. Brackets give 95% CIs based on 10,000 resamples.

\hline \text{Layer group (train count)}&\text{Model}&\text{Group in training}&\text{In group}&\text{Valid}&\text{UN}&\text{Stable}&\text{SUN}&\text{SUN template}&\text{Rediscovered}\\\\ \hline \text{LG 8 (53)}&\text{SLayerGen}&\text{Yes}&96.5&65.0\ {\scriptstyle[62.9,\ 67.1]}&81.5&57.6\ {\scriptstyle[54.5,\ 60.6]}&41.8\ {\scriptstyle[38.8,\ 44.9]}&8.4\ {\scriptstyle[6.8,\ 10.3]}&3/29\\\\ &\text{SLayerGen}&\text{No}&96.4&58.7\ {\scriptstyle[56.5,\ 60.8]}&80.2&52.9\ {\scriptstyle[49.8,\ 56.0]}&38.0\ {\scriptstyle[35.0,\ 41.0]}&5.9\ {\scriptstyle[4.6,\ 7.5]}&3/29\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 11 (883)}&\text{SLayerGen}&\text{Yes}&99.9&71.0\ {\scriptstyle[69.0,\ 73.0]}&91.8&52.0\ {\scriptstyle[48.9,\ 55.1]}&45.6\ {\scriptstyle[42.5,\ 48.7]}&3.0\ {\scriptstyle[2.1,\ 4.3]}&5/373\\\\ &\text{SLayerGen}&\text{No}&99.8&53.4\ {\scriptstyle[51.3,\ 55.6]}&91.3&41.2\ {\scriptstyle[38.2,\ 44.3]}&34.4\ {\scriptstyle[31.5,\ 37.4]}&3.3\ {\scriptstyle[2.4,\ 4.6]}&1/373\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 31 (42)}&\text{SLayerGen}&\text{Yes}&98.3&71.2\ {\scriptstyle[69.2,\ 73.2]}&79.6&44.8\ {\scriptstyle[41.7,\ 47.9]}&31.1\ {\scriptstyle[28.3,\ 34.0]}&3.7\ {\scriptstyle[2.7,\ 5.1]}&2/17\\\\ &\text{SLayerGen}&\text{No}&99.5&79.9\ {\scriptstyle[78.1,\ 81.6]}&80.7&32.1\ {\scriptstyle[29.3,\ 35.1]}&22.5\ {\scriptstyle[20.0,\ 25.2]}&1.6\ {\scriptstyle[1.0,\ 2.6]}&1/17\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \end{array}$$

Also, in Table-2/3/4 SUN & UN for DiffCSP, MatterGen is much better compared to SLayerGen (Structure). This shows model is not able to sample many unique and novel materials, that means not much diversity is gained. This limits practical application of the model.

We respectfully disagree with the characterization that SLayerGen “is not able to sample many unique and novel materials” given that SLayerGen attains the highest SUN and UN template rates on most datasets we benchmarked. While true that other models like MatterGen yield higher SUN and UN structure rates, we found that most of these are P1 materials, which cannot exhibit many symmetry-dependent quantum and topological properties. We quantified the degree to which P1 structures from MatterGen and DiffCSP inflate their SUN structure rates by computing the SSUN rate from Kazeev et al (2503.02407), which measures the rate of non-P1 SUN generated structures. As expected, we found that MatterGen and DiffCSP contain a much larger share of P1 crystals in their generated SUN structures: $$ \begin{array}{llccccc} \hline \text{Dataset}&\text{Model}&\text{SUN}&\text{SSUN (0.1 Å)}&\text{SSUN (0.01 Å)}&\text{SSUN/SUN (0.1 Å)}&\text{SSUN/SUN (0.01 Å)}\\ \hline \text{2DMatPedia}&\text{MatterGen}&29.0&8.0\ {\scriptstyle[6.5,\ 9.8]}&4.7\ {\scriptstyle[3.6,\ 6.2]}&0.28&0.16\\ &\text{DiffCSP}&11.1&4.5\ {\scriptstyle[3.4,\ 6.0]}&1.8\ {\scriptstyle[1.1,\ 2.8]}&0.41&0.16\\ &\text{DiffCSP++}&17.7&14.7\ {\scriptstyle[12.6,\ 17.0]}&10.9\ {\scriptstyle[9.1,\ 13.0]}&0.83&0.62\\ &\text{SLayerGen}&24.8&20.6\ {\scriptstyle[18.2,\ 23.2]}&15.7\ {\scriptstyle[13.6,\ 18.1]}&0.83&0.63\\ \hline \text{C2DB}&\text{MatterGen}&45.2&23.2\ {\scriptstyle[20.7,\ 25.9]}&8.7\ {\scriptstyle[7.1,\ 10.6]}&0.51&0.19\\ &\text{DiffCSP}&35.5&20.4\ {\scriptstyle[18.0,\ 23.0]}&8.8\ {\scriptstyle[7.2,\ 10.7]}&0.57&0.25\\ &\text{DiffCSP++}&33.6&29.0\ {\scriptstyle[26.3,\ 31.9]}&24.9\ {\scriptstyle[22.3,\ 27.7]}&0.86&0.74\\ &\text{SLayerGen3D}&24.7&19.0\ {\scriptstyle[16.7,\ 21.5]}&16.3\ {\scriptstyle[14.1,\ 18.7]}&0.77&0.66\\ &\text{SLayerGen}&33.0&24.8\ {\scriptstyle[22.2,\ 27.6]}&19.4\ {\scriptstyle[17.1,\ 22.0]}&0.75&0.59\\ \hline \text{BiDB}&\text{MatterGen}&51.2&27.2\ {\scriptstyle[24.5,\ 30.0]}&7.3\ {\scriptstyle[5.8,\ 9.1]}&0.53&0.14\\ &\text{DiffCSP}&63.6&33.2\ {\scriptstyle[30.4,\ 36.2]}&8.3\ {\scriptstyle[6.7,\ 10.2]}&0.52&0.13\\ &\text{DiffCSP++}&53.1&45.3\ {\scriptstyle[42.2,\ 48.4]}&39.5\ {\scriptstyle[36.5,\ 42.6]}&0.85&0.74\\ &\text{SLayerGen}&36.2&29.4\ {\scriptstyle[26.7,\ 32.3]}&25.5\ {\scriptstyle[22.9,\ 28.3]}&0.81&0.70\\ \hline \text{Alex2D}&\text{MatterGen}&49.9&24.7\ {\scriptstyle[22.1,\ 27.5]}&9.0\ {\scriptstyle[7.4,\ 10.9]}&0.49&0.18\\ &\text{DiffCSP}&45.1&26.2\ {\scriptstyle[23.6,\ 29.0]}&13.4\ {\scriptstyle[11.4,\ 15.7]}&0.58&0.30\\ &\text{DiffCSP++}&37.1&30.8\ {\scriptstyle[28.0,\ 33.7]}&26.2\ {\scriptstyle[23.6,\ 29.0]}&0.83&0.71\\ &\text{SLayerGen}&34.2&31.0\ {\scriptstyle[28.2,\ 33.9]}&25.6\ {\scriptstyle[23.0,\ 28.4]}&0.91&0.75\\ \hline \end{array}

Few Typos Line 092: autoregressive Line 421: BJ-damped Line 1125: enantiomorphic Line 1394: tuned manually

Thank you for pointing these out. We will fix them in the text.

Reviewer_vEk6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SLayerGen is the first from-scratch generative diffusion model for layer groups (the diperiodic symmetry groups of free-standing 2D materials), extending the space-group-equivariant SGEquiDiff to the periodic-in-xy / aperiodic-in-z setting via anisotropic, center-of-mass-free-only-along-z noise, layer-group + Wyckoff embeddings, and asymmetric-unit (ASU) enumeration. Along the way the authors correct two ASU definitions in the International Tables (layer groups 7 and 46) and a hexagonal non-orthogonality issue in fractional-coordinate distances, and curate four 2D datasets (2DMatPedia, C2DB, BiDB, Alex2D). The modeling contribution is genuine and the crystallographic corrections independently check out, but the empirical "SOTA" framing is overstated: SLayerGen loses the standard S.U.N.-structure rate on all four datasets and the headline is re-anchored on the template-level metric (the chemistry-agnostic U.N.-template rate follows SymmCD, Levy et al. 2025; the paper adds the stability filter to form the S.U.N.-template rate it reports as headline), with no DFT verification, no error bars on baselines, and no comparison against existing 2D-specific generators.

Critical strengths
  • the equivariance and the crystallography are correct and non-trivial. independent check: CoM-free-along-z preserves layer-group equivariance including the R_z=−1 flip case (the headline novelty over sgequidiff, currently only asserted — promote it to a proposition), the hex non-orthogonality fix is right, and the lg7/lg46 ASU corrections reproduce decisively (lg46 original is provably broken). these are real, checkable contributions beyond the ML result.

  • honest on-target metric + reproducibility. the layer-group JSD win is the metric actualy tied to what they built, it's a clean result, and the four curated diperiodic datasets are a useful artifact on their own. orb-v3 being eval-only (not a training reward) is the right call and avoids the reward-hacking trap most crystal papers fall into.

Critical weaknesses

Coverage assessment

Coverage of the bulk generative-model literature is good and the parent method is properly positioned. The gap is on the 2D-specific side: the paper's "first/SOTA for layer groups" claim is evaluated only against bulk generators (DiffCSP, DiffCSP++, MatterGen, SGEquiDiff), while the existing 2D-specific generators it cites are dismissed in prose and never run.

Missing citations / comparisons

  • Benchmark (or explicitly justify not benchmarking) against Dong et al. 2412.18414 and at least one C2DB-native generator. "First generative model for layer groups" is defensible only if the distinction from space-group-constrained 2D generators is made crisp and the overlap is measured, not asserted.

  • the headline leads with the metric where they win while the standard one regresses. advertised wins are S.U.N.-template (best 3/4, alex2d only with a TG,W=2.0 tempering trick), but S.U.N.-structure loses on all four datasets. to be fair the template metric is not their invention — U.N.-template is from symmcd (Levy 2025), they only add the stability filter — so this isnt metric-cooking, it's emphasis: they foreground the winning metric and underplay the structure-level loss. and chemistry-agnostic templates mechanicaly favor a wyckoff-tuple sampler, so it should sit next to the structure numbers, not replace them. report both, dont crown the template rate.

  • weak statistical and physical backing around the win. no error bars on any baseline (only slayergen, only one dataset), so every close gap is unsupported; no DFT spot-check of a single generated structure; pre-relaxation U.N.-structure also drops vs baselines (bidb 50.3 vs 96.8, −46pp) and that's buried; and "sota for layer groups" is measured only against bulk generators while the 2d-specific ones they cite (Dong 2412.18414; Lyngby & Thygesen 10.1038/s41524-022-00923-3) are never run. plus free-standing stability labels ignore the substrate, which they concede, and heavy filtering (22-50% kept) couples training selection to the eval cutoff.

Questions for authors

can you address weaknesses above?

  • Re-anchor the headline on what is actually controlled/won. State plainly: SLayerGen wins layer-group JSD and S.U.N.-template (metric adopted from SymmCD), but trails on S.U.N.-structure across all four datasets. Present the template metric alongside — not instead of — the structure metric, and address directly why the structure-level regression is acceptable for a 2D-discovery tool.
  • with ≥3 seeds and error bars on every baseline, and paired tests, which of the template-metric and JSD wins survive, and does slayergen still beat its own 3d ablation?
  • DFT-relax a random 30-50 of your S.U.N. hits (structure and template) and the baselines': what's the DFT-confirmed rate, and the orb-v3→dft agreement on generated 2d structures (not held-out labels)?
  • Quantify the P1 claim: report the P1 fraction of MatterGen outputs and a controlled test, and reconcile with Table 6 (SLayerGen-P1 raising structure-SUN) — or remove the "inflated" framing.
  • Promote the equivariance argument from "proof sketch" to a Proposition: state the hypothesis R Σ Rᵀ = Σ (and that in-plane isotropy σ_x=σ_y is required), and add the 2-line P_z-equivariance proof for the z-flip (R_z=−1) case. This is your headline novelty and currently only asserted.
  • Confirm in text that the 18 non-injective layer groups (e.g. LG 4 vs LG 11, both → space group 6) receive distinct embeddings; the experiments suggest yes but the construction doesn't say so.
3 comments

We thank the reviewer for the thoughtful comments and questions. Below we aim to address your concerns. If you find our answers helpful, we would appreciate it if you would consider raising your score.

empirical "SOTA" framing is overstated

We will qualify the writing to specify that we get SOTA performance on the template-level metric, which we believe is still a strong result. Given that many exotic quantum and topological properties of 2D materials are highly symmetry-dependent, enabling generation of realistic structures with targeted symmetries has potential to enable new functional materials.

Beyond existing performance metrics, we highlight that to the best of our knowledge, SLayerGen is the only model capable of constraining generation to prescribed layer groups at inference time. We added new experiments leveraging this capability to demonstrate transferability to rare and completely unseen layer groups. In the table below, “In group” signifies the rate of generated structures with the prescribed layer group. “Rediscovered” denotes unseen validation and test structures generated by the model. Brackets give 95% CIs.

\hline \text{Layer group (train count)}&\text{Model}&\text{Group in training}&\text{In group}&\text{Valid}&\text{UN}&\text{Stable}&\text{SUN}&\text{SUN template}&\text{Rediscovered}\\\\ \hline \text{LG 8 (53)}&\text{SLayerGen}&\text{Yes}&96.5&65.0\ {\scriptstyle[62.9,\ 67.1]}&81.5&57.6\ {\scriptstyle[54.5,\ 60.6]}&41.8\ {\scriptstyle[38.8,\ 44.9]}&8.4\ {\scriptstyle[6.8,\ 10.3]}&3/29\\\\ &\text{SLayerGen}&\text{No}&96.4&58.7\ {\scriptstyle[56.5,\ 60.8]}&80.2&52.9\ {\scriptstyle[49.8,\ 56.0]}&38.0\ {\scriptstyle[35.0,\ 41.0]}&5.9\ {\scriptstyle[4.6,\ 7.5]}&3/29\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 11 (883)}&\text{SLayerGen}&\text{Yes}&99.9&71.0\ {\scriptstyle[69.0,\ 73.0]}&91.8&52.0\ {\scriptstyle[48.9,\ 55.1]}&45.6\ {\scriptstyle[42.5,\ 48.7]}&3.0\ {\scriptstyle[2.1,\ 4.3]}&5/373\\\\ &\text{SLayerGen}&\text{No}&99.8&53.4\ {\scriptstyle[51.3,\ 55.6]}&91.3&41.2\ {\scriptstyle[38.2,\ 44.3]}&34.4\ {\scriptstyle[31.5,\ 37.4]}&3.3\ {\scriptstyle[2.4,\ 4.6]}&1/373\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \text{LG 31 (42)}&\text{SLayerGen}&\text{Yes}&98.3&71.2\ {\scriptstyle[69.2,\ 73.2]}&79.6&44.8\ {\scriptstyle[41.7,\ 47.9]}&31.1\ {\scriptstyle[28.3,\ 34.0]}&3.7\ {\scriptstyle[2.7,\ 5.1]}&2/17\\\\ &\text{SLayerGen}&\text{No}&99.5&79.9\ {\scriptstyle[78.1,\ 81.6]}&80.7&32.1\ {\scriptstyle[29.3,\ 35.1]}&22.5\ {\scriptstyle[20.0,\ 25.2]}&1.6\ {\scriptstyle[1.0,\ 2.6]}&1/17\\\\ &\text{MatterGen}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP}&\text{Yes}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ &\text{DiffCSP++}&\text{No}&0&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}&\text{–}\\\\ \hline \end{array}$$ > no DFT We conducted DFT calculations and found good agreement with Orb-v3; we drew 30 structures predicted stable by Orb-v3 uniformly randomly from SLayerGen's SUN sets on Alex2D (15 SUN structures from SLayerGen and 15 SUN templates from SLayerGen T=2). Each was relaxed with VASP using Materials Project-compatible settings. DFT confirms 18 of the 30 structures as stable at the 200 meV/atom threshold. However, 28 of the 30 structures fall below 250 meV/atom. Geometries are largely unchanged: 24 of 30 structures move by less than 0.1 A under DFT relaxation, and 28 of 30 retain their layer group and template. > measured only against bulk generators while 2d ones are never run We stress that we did benchmark against models used by prior papers focused on 2D materials generation. Specifically, Lyngby and Thygesen [1] trained CDVAE on 2D materials, SCIGEN [2] trained DiffCSP on 2D lattices, and Xu et al [3] trained a space group constrained generative model on 2D materials. We benchmarked CDVAE, DiffCSP, and DiffCSP++ as direct analogues of these works. We will state this clearly in the text to avoid confusion. We highlight that none of these prior models are invariant to layer groups, nor can they constrain generated crystals to specified layer groups; instead, they train bulk generative models on 2D materials. [1] P Lyngby and Thygesen KS 10.1038/s41524-022-00923-3. [2] R Okabe et al. https://doi.org/10.1038/s41563-025-02355-y [3] S Xu et al. J Phys Chem C 10 April 2025; 129 (14): 6794–6802

no error bars

We computed 95% bootstrapped confidence intervals with 10,000 resamples. Due to space constraints, we only report JSD and SUN rates, finding SLayerGen (T=2.0) still has the highest SUN template rate on every dataset, and SLayerGen maintains JSD wins compared to SLayerGen3D on C2DB and BiDB while effectively tying on 2DMatPedia:

\hline \text{Dataset}&\text{Model}&\text{UN temp}\uparrow&\text{JSD}_G\downarrow&\text{SUN temp}\uparrow&\text{SUN struct}\uparrow\\\\ \hline \text{2DMatPedia}&\text{DiffCSP}&9.4\ {\scriptstyle[7.7,\ 11.3]}&0.762&2.4\ {\scriptstyle[1.5,\ 3.4]}&10.6\ {\scriptstyle[8.7,\ 12.5]}\\\\ &\text{DiffCSP++}&6.3\ {\scriptstyle[4.8,\ 7.9]}&0.180&3.9\ {\scriptstyle[2.8,\ 5.2]}&19.9\ {\scriptstyle[17.4,\ 22.4]}\\\\ &\text{MatterGen}&9.8\ {\scriptstyle[8.0,\ 11.7]}&0.701&4.2\ {\scriptstyle[3.0,\ 5.5]}&30.7\ {\scriptstyle[27.9,\ 33.6]}\\\\ &\text{SLayerGen3D}&27.3\ {\scriptstyle\pm 1.8}&0.208\ {\scriptstyle\pm 0.011}&4.7\ {\scriptstyle\pm 0.9}&21.3\ {\scriptstyle\pm 1.3}\\\\ &\text{SLayerGen}&30.3\ {\scriptstyle\pm 0.9}&0.200\ {\scriptstyle\pm 0.008}&5.8\ {\scriptstyle\pm 0.4}&24.5\ {\scriptstyle\pm 0.9}\\\\ &\text{SLayerGen T=2.0}&48.0\ {\scriptstyle[44.8,\ 51.2]}&0.301\ {\scriptstyle[0.281,\ 0.320]}&7.0\ {\scriptstyle[5.4,\ 8.6]}&21.5\ {\scriptstyle[19.0,\ 24.0]}\\\\ \hline \text{C2DB}&\text{DiffCSP}&13.0\ {\scriptstyle[10.9,\ 15.1]}&0.617&4.3\ {\scriptstyle[3.1,\ 5.6]}&35.5\ {\scriptstyle[32.5,\ 38.4]}\\\\ &\text{DiffCSP++}&5.0\ {\scriptstyle[3.7,\ 6.4]}&0.138&4.4\ {\scriptstyle[3.2,\ 5.8]}&35.8\ {\scriptstyle[32.9,\ 38.8]}\\\\ &\text{MatterGen}&10.1\ {\scriptstyle[8.3,\ 12.0]}&0.521&5.0\ {\scriptstyle[3.7,\ 6.4]}&45.2\ {\scriptstyle[42.1,\ 48.3]}\\\\ &\text{SGEquiDiff}&17.5\ {\scriptstyle[15.2,\ 19.9]}&0.284&3.0\ {\scriptstyle[2.0,\ 4.1]}&23.9\ {\scriptstyle[21.3,\ 26.6]}\\\\ &\text{SLayerGen3D}&17.1\ {\scriptstyle[14.8,\ 19.4]}&0.301\ {\scriptstyle[0.287,\ 0.315]}&5.2\ {\scriptstyle[3.9,\ 6.6]}&29.4\ {\scriptstyle[26.6,\ 32.2]}\\\\ &\text{SLayerGen}&19.9\ {\scriptstyle[17.4,\ 22.4]}&0.160\ {\scriptstyle[0.141,\ 0.180]}&5.6\ {\scriptstyle[4.2,\ 7.1]}&33.0\ {\scriptstyle[30.1,\ 36.0]}\\\\ &\text{SLayerGen T=2.0}&38.4\ {\scriptstyle[35.4,\ 41.4]}&0.239&6.8\ {\scriptstyle[5.3,\ 8.4]}&30.0\ {\scriptstyle[27.2,\ 32.8]}\\\\ \hline \text{BiDB}&\text{DiffCSP}&10.6\ {\scriptstyle[8.7,\ 12.6]}&0.597&4.7\ {\scriptstyle[3.4,\ 6.0]}&63.6\ {\scriptstyle[60.6,\ 66.5]}\\\\ &\text{DiffCSP++}&2.2\ {\scriptstyle[1.3,\ 3.2]}&0.121&5.5\ {\scriptstyle[4.1,\ 6.9]}&56.8\ {\scriptstyle[53.7,\ 59.9]}\\\\ &\text{MatterGen}&8.9\ {\scriptstyle[7.2,\ 10.7]}&0.411&5.0\ {\scriptstyle[3.7,\ 6.4]}&51.2\ {\scriptstyle[48.2,\ 54.3]}\\\\ &\text{SLayerGen3D}&15.6\ {\scriptstyle[13.4,\ 17.9]}&0.182&9.0\ {\scriptstyle[7.2,\ 10.8]}&41.6\ {\scriptstyle[38.5,\ 44.6]}\\\\ &\text{SLayerGen}&10.9\ {\scriptstyle[9.0,\ 12.9]}&0.158\ {\scriptstyle[0.139,\ 0.177]}&7.6\ {\scriptstyle[6.0,\ 9.3]}&36.2\ {\scriptstyle[33.3,\ 39.2]}\\\\ &\text{SLayerGen T=2.0}&22.5\ {\scriptstyle[19.9,\ 25.1]}&0.266\ {\scriptstyle[0.245,\ 0.286]}&11.3\ {\scriptstyle[9.4,\ 13.3]}&37.2\ {\scriptstyle[34.2,\ 40.2]}\\\\ \hline \text{Alex2D}&\text{DiffCSP}&7.3\ {\scriptstyle[5.7,\ 8.9]}&0.614&2.5\ {\scriptstyle[1.6,\ 3.5]}&45.1\ {\scriptstyle[42.0,\ 48.4]}\\\\ &\text{DiffCSP++}&1.8\ {\scriptstyle[1.0,\ 2.7]}&0.108&1.3\ {\scriptstyle[0.7,\ 2.1]}&37.1\ {\scriptstyle[34.1,\ 40.1]}\\\\ &\text{MatterGen}&6.9\ {\scriptstyle[5.4,\ 8.5]}&0.451&3.9\ {\scriptstyle[2.7,\ 5.1]}&49.9\ {\scriptstyle[46.8,\ 53.0]}\\\\ &\text{SLayerGen3D}&7.5\ {\scriptstyle[5.9,\ 9.2]}&0.124&2.0\ {\scriptstyle[1.2,\ 2.9]}&35.9\ {\scriptstyle[33.0,\ 38.9]}\\\\ &\text{SLayerGen}&7.5\ {\scriptstyle[5.9,\ 9.2]}&0.149&2.1\ {\scriptstyle[1.3,\ 3.0]}&34.2\ {\scriptstyle[31.3,\ 37.2]}\\\\ &\text{SLayerGen T=2.0}&20.0\ {\scriptstyle[17.6,\ 22.5]}&0.212&4.4\ {\scriptstyle[3.2,\ 5.7]}&32.3\ {\scriptstyle[29.4,\ 35.2]}\\\\ &\text{SLayerGen +MP20}&7.4\ {\scriptstyle[5.8,\ 9.1]}&0.131&2.2\ {\scriptstyle[1.4,\ 3.1]}&36.9\ {\scriptstyle[33.9,\ 39.9]}\\\\ \hline \end{array}$$

Quantify the P1 claim

Per 1000 structures, we observe that MatterGen generates much fewer non-P1 structures compared to SLayerGen at a 0.01 Ang tolerance:

\hline \text{Dataset}&\text{MatterGen (symprec=0.1)}&\text{SLayerGen (symprec=0.1)}&\text{MatterGen (symprec=0.01)}&\text{SLayerGen (symprec=0.01)}\\\\ \hline \text{2DMatPedia}&85&230&7&230\\\\ \text{C2DB}&263&267&12&265\\\\ \text{BiDB}&333&331&45&331\\\\ \text{Alex2D}&275&289&6&289\\\\ \hline \end{array}$$ To quantify the claim that P1 structures inflate SUN structure rate, we compute the SSUN rate from Kazeev et al (2503.02407), which measures the rate of non-P1 SUN generated structures. As expected, we find that the non-symmetry-constrained baselines (MatterGen and DiffCSP) contain a much larger share of P1 crystals in their generated SUN structures: $$\begin{array}{llccccc} \hline \text{Dataset}&\text{Model}&\text{SUN}&\text{SSUN (0.1 Å)}&\text{SSUN (0.01 Å)}&\text{SSUN/SUN (0.1 Å)}&\text{SSUN/SUN (0.01 Å)}\\\\ \hline \text{2DMatPedia}&\text{MatterGen}&29.0&8.0\ {\scriptstyle[6.5,\ 9.8]}&4.7\ {\scriptstyle[3.6,\ 6.2]}&0.28&0.16\\\\ &\text{DiffCSP}&11.1&4.5\ {\scriptstyle[3.4,\ 6.0]}&1.8\ {\scriptstyle[1.1,\ 2.8]}&0.41&0.16\\\\ &\text{DiffCSP++}&17.7&14.7\ {\scriptstyle[12.6,\ 17.0]}&10.9\ {\scriptstyle[9.1,\ 13.0]}&0.83&0.62\\\\ &\text{SLayerGen}&24.8&20.6\ {\scriptstyle[18.2,\ 23.2]}&15.7\ {\scriptstyle[13.6,\ 18.1]}&0.83&0.63\\\\ \hline \text{C2DB}&\text{MatterGen}&45.2&23.2\ {\scriptstyle[20.7,\ 25.9]}&8.7\ {\scriptstyle[7.1,\ 10.6]}&0.51&0.19\\\\ &\text{DiffCSP}&35.5&20.4\ {\scriptstyle[18.0,\ 23.0]}&8.8\ {\scriptstyle[7.2,\ 10.7]}&0.57&0.25\\\\ &\text{DiffCSP++}&33.6&29.0\ {\scriptstyle[26.3,\ 31.9]}&24.9\ {\scriptstyle[22.3,\ 27.7]}&0.86&0.74\\\\ &\text{SLayerGen3D}&24.7&19.0\ {\scriptstyle[16.7,\ 21.5]}&16.3\ {\scriptstyle[14.1,\ 18.7]}&0.77&0.66\\\\ &\text{SLayerGen}&33.0&24.8\ {\scriptstyle[22.2,\ 27.6]}&19.4\ {\scriptstyle[17.1,\ 22.0]}&0.75&0.59\\\\ \hline \text{BiDB}&\text{MatterGen}&51.2&27.2\ {\scriptstyle[24.5,\ 30.0]}&7.3\ {\scriptstyle[5.8,\ 9.1]}&0.53&0.14\\\\ &\text{DiffCSP}&63.6&33.2\ {\scriptstyle[30.4,\ 36.2]}&8.3\ {\scriptstyle[6.7,\ 10.2]}&0.52&0.13\\\\ &\text{DiffCSP++}&53.1&45.3\ {\scriptstyle[42.2,\ 48.4]}&39.5\ {\scriptstyle[36.5,\ 42.6]}&0.85&0.74\\\\ &\text{SLayerGen}&36.2&29.4\ {\scriptstyle[26.7,\ 32.3]}&25.5\ {\scriptstyle[22.9,\ 28.3]}&0.81&0.70\\\\ \hline \text{Alex2D}&\text{MatterGen}&49.9&24.7\ {\scriptstyle[22.1,\ 27.5]}&9.0\ {\scriptstyle[7.4,\ 10.9]}&0.49&0.18\\\\ &\text{DiffCSP}&45.1&26.2\ {\scriptstyle[23.6,\ 29.0]}&13.4\ {\scriptstyle[11.4,\ 15.7]}&0.58&0.30\\\\ &\text{DiffCSP++}&37.1&30.8\ {\scriptstyle[28.0,\ 33.7]}&26.2\ {\scriptstyle[23.6,\ 29.0]}&0.83&0.71\\\\ &\text{SLayerGen}&34.2&31.0\ {\scriptstyle[28.2,\ 33.9]}&25.6\ {\scriptstyle[23.0,\ 28.4]}&0.91&0.75\\\\ \hline \end{array}$$ > Promote the equivariance argument to Proposition We agree. > Confirm that the 18 non-injective layer groups receive distinct embeddings We confirm we have unit tested all layer groups receive distinct embeddings and will include it in the text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.