acceptodds
Under review as a conference paper at ICLR 2027

Ghost Atoms: Flexible Crystal Generation With Voronoi Tessellations

Abstract

Generative models for materials discovery fix the number of atoms per unit cell before generation, so the model never chooses how many atoms its crystal contains and cannot add or remove them. We restore this degree of freedom by padding every cell to a fixed number of atoms with fictitious ghost atoms, placed at well-spaced interstitial sites. We place them by iterative Voronoi maximin selection, so the padded training structures contain no unphysical atomic overlaps. Training the OMatG stochastic-interpolant model on an augmented MP-20 with Voronoi-placed ghosts improves the combined rate of stable or metastable, unique, and novel structures (SUN+MSUN) by 2.6 percentage points over a matched control that places ghosts uniformly at random, and by 6.0 points over the published OMatG model. The Voronoi-trained model learns to keep ghosts out of occupied space, placing far fewer near real atoms, so that converting a ghost into a real atom happens at a vacant site. Our best run reaches a 25.27% SUN+MSUN rate on LeMat-GenBench, a new state of the art.

Primary area:
generative models
Keywords:
crystal generation, ghost atoms, Voronoi tessellation, stochastic interpolants, materials discovery

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_ZfP5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ghost atoms, which allow common generative models for crystal architectures to vary the number of atoms in a crystal unit cell during generation. The core idea of this research is Voronoi-guided ghost-atom placement, in which ghost atoms are placed in the largest empty spaces inside the unit cell, thereby avoiding overlap with real atoms. The authors claim that this design avoids non-physical behavior of atoms inside crystals and provide experiments intended to:

  1. isolate the effect of their core design choices in comparison with concurrent approaches;
  2. demonstrate state-of-the-art model performance.

The authors begin by explaining the problem, present their solution design, choose the OMatG generative methodology as the baseline model, design an iterative procedure for adding ghost atoms into the current largest empty space in the crystal, and provide a set of experiments comparing models trained on MP-20 and evaluated using SUN, MSUN [1], and SUN+MSUN in the LeMatGenBench variant [9] (with LeMatBulk [10] as the novelty reference dataset).

Critical strengths

S1

An interesting idea that addresses an important limitation.

S2

The paper is well written and easy to follow.

Critical weaknesses

The following weaknesses concern the experimental part of the work, which should provide a clear justification that the proposed methodology for constructing generative models is better than existing ones.

W1

Effect isolation

The random-ghost control usefully isolates placement under supervised coordinates, but it is not a matched MiAD-style baseline: MiAD pairs uniform/random placement with masking of the coordinate loss for ghosts, whereas the paper’s control keeps that loss unmasked. The paper states that this is not a "reimplementation of MiAD", so that is not a problem.

However, MiAD [6] proposes a particular methodology for working with ghost atoms: uniform/random placement of ghosts together with masking of the coordinate loss for those ghosts. The paper’s "random-ghost control" differs from that design mainly by not masking the coordinate loss. Therefore, a comparison with the complete MiAD-style setup on the same baseline is expected: OMatG + random ghost placement + coordinate-loss masking. Uniform placement in MiAD was proposed together with coordinate-loss masking, and MiAD demonstrates both conceptually and experimentally that this masking is important (MiAD [6], Table 5 — in the case of uniform placement, loss masking increases the metric by +35% on average). So the proposed "random-ghost control" experiment is a comparison with a weakened version of the methodology proposed in MiAD. The paper already includes MiAD in Table 1; what is missing for a fair methodology comparison is specifically this matched OMatG + random + masking ablation.

W2

Metric choice

LeMatGenBench [9] is a well-known benchmark in the field of de novo crystal generation. In my view, it is a good tool for analyzing the performance of particular generative models on the task of discovering materials that are nearly "unknown to humanity". However, its applicability to comparing generative methodologies seems questionable for the following reason. Many existing papers that propose a generative model design train and evaluate their models on the MP-20 dataset [2–8]. This dataset is well known to the community and cheap enough for training, and many previous papers compare the SUN of models trained on MP-20 using MP-20 itself (or the training set) as the novelty reference [2–8] — an entirely logical choice, since we clearly do not want a model trained on MP-20 to reproduce crystals from MP-20. (Note: MatterGen [1] is an important exception, using a larger reference set rather than MP-20 for novelty.) However, using LeMatGenBench, which employs LeMatBulk [10] (a dataset much larger than MP-20) as the novelty reference, means that we penalize models for discovering materials from LeMatBulk even when those materials are not present in MP-20. This aspect of SUN in LeMatGenBench is not an undoubted benefit. Notably, the LeMatGenBench paper itself reports that stability rates drop by factors of 2–3× and S.U.N. rates collapse even further when evaluated against LeMatBulk instead of smaller references (e.g., DiffCSP S.U.N. from 4.1% to 0.1%) [9], which confirms that the choice of reference set materially affects model rankings. Quoting the paper that proposed LeMatGenBench [9]:

"As generative models scale toward deployment in real discovery settings, evaluation against a comprehensive reference set becomes essential. MP-20 remains valuable for controlled methodological benchmarking, but LeMat-Bulk provides a more reliable basis for assessing discovery-oriented performance."

Since your work proposes a new methodology for constructing generative models, I believe comparison using SUN on MP-20 is necessary.

Also, the fact that the proposed model leads in SUN + MSUN is almost entirely due to its higher MSUN. The scale of MSUN is larger than that of SUN by design (as seen in the paper's results table, MSUN is roughly 15–85 times larger than SUN across the reported models). Therefore, leading in MSUN will almost always result in leading in SUN + MSUN, which means that introducing SUN + MSUN adds no new perspective to the model comparison and is redundant once MSUN is reported. If we remove SUN + MSUN, the proposed model is better on one metric (MSUN) and worse on another (SUN) — not better on both, and not better on one while comparable on the other. These results raise the suspicion that the model merely represents a shift along the trade-off between SUN and MSUN — which could still be a valuable result if you:

  1. justify that this shift is important; and
  2. justify that other models — or at least the main competitors — cannot make this shift.

References

[1] Zeni, C., Pinsler, R., Zügner, D., Fowler, A., Horton, M., Fu, X., Wang, Z., Shysheya, A., Crabbé, J., Ueda, S., Sordillo, R., Sun, L., Smith, J., Nguyen, B., Schulz, H., Lewis, S., Huang, C.-W., Lu, Z., Zhou, Y., Yang, H., Hao, H., Li, J., Yang, C., Li, W., Tomioka, R., & Xie, T. (2025). A generative model for inorganic materials design. Nature, 639(8055), 624–632. https://doi.org/10.1038/s41586-025-08628-5

[2] Miller, B. K., Chen, R. T. Q., Sriram, A., & Wood, B. M. (2024). FlowMM: Generating materials with Riemannian flow matching. Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235, 35664–35686. arXiv:2406.04713

[3] Sriram, A., Miller, B. K., Chen, R. T. Q., & Wood, B. M. (2024). FlowLLM: Flow matching for material generation with large language models as base distributions. Advances in Neural Information Processing Systems 37 (NeurIPS), 46025–46046. arXiv:2410.23405

[4] Joshi, C. K., Fu, X., Liao, Y.-L., Gharakhanyan, V., Miller, B. K., Sriram, A., & Ulissi, Z. W. (2025). All-atom Diffusion Transformers: Unified generative modelling of molecules and materials. Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267. arXiv:2503.03965

[5] Kazeev, N., Nong, W., Romanov, I., Zhu, R., Ustyuzhanin, A., Yamazaki, S., & Hippalgaonkar, K. (2025). Wyckoff Transformer: Generation of symmetric crystals. Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267, 29495–29526. arXiv:2503.02407

[6] Okhotin, A., Nakhodnov, M., Kazeev, N., Lazarev, M., Ustyuzhanin, A., & Vetrov, D. (2025). MiAD: Mirage Atom Diffusion for de novo crystal generation. arXiv preprint arXiv:2511.14426

[7] Yi, X., Xu, G., Zhang, Z., Liu, L., Bian, Y., Xiao, X., & Zhao, P. (2026). CrystalDiT: Simple diffusion transformers for crystal generation. Proceedings of the AAAI Conference on Artificial Intelligence. arXiv:2508.16614

[8] Hadži Veljković, T., Rosenthal, J., Lončarić, I., & van de Meent, J.-W. (2026). Crystalite: A lightweight transformer for efficient crystal modeling. arXiv preprint arXiv:2604.02270

[9] Betala, S., Gleason, S. P., Ramlaoui, A., Xu, A., Channing, G., Levy, D., Fourrier, C., Kazeev, N., Joshi, C. K., Kaba, S.-O., Therrien, F., Hernandez-Garcia, A., Mercado, R., Krishnan, N. M. A., & Duval, A. (2025). LeMat-GenBench: A unified evaluation framework for crystal generative models. arXiv preprint arXiv:2512.04562

[10] Siron, M., Djafar, I., Ramlaoui, A., du Fayette, E., Rossello, A., Fako, E., McDermott, M., Therrien, F., Barroso-Luque, L., Cipcigan, F., et al. (2025). LeMat-Bulk: Aggregating, de-duplicating, and curating quantum chemistry materials databases. arXiv preprint arXiv:2511.05178

Questions for authors

Q1 The text claims several times that

Introduction, Contributions paragraph:

We place these ghosts at well-spaced interstitial sites found by iterative Voronoi tessellation, choosing at each step the site furthest from every existing atom. Because these are the largest voids in the cell, the padded training data never contains the overlapping or nonphysically close atoms that arbitrary placement would produce. This lets fixed-cardinality generative models adjust the number of atoms during generation, and we measure the benefit by their rate of producing stable or metastable, unique, and novel structures. We compare Voronoi padding with random padding and show that the Voronoi-trained model learns to keep its ghosts out of the space occupied by real atoms.

Results, Reproducing the Tessellation paragraph:

Where ghosts are placed matters because their coordinate loss is not masked: they are supervised targets for the coordinate field rather than inert padding, and a model trained on physically implausible targets learns to reproduce them. We see this directly in where each model places its ghosts relative to real atoms.

Overlapping and physically implausible atom configurations are unintended behavior — but why is this important for ghost atoms? Especially when they are erased from the unit cell anyway, and the final structure will not contain overlaps or physically implausible configurations.

Q2 Results, Reproducing the Tessellation paragraph:

We also verify that the Voronoi-trained model reproduces the geometry rather than memorizing positions by taking the generated structures across all 5 seeds, stripping the model’s ghosts, and re-running the Voronoi maximin procedure of Section 3.5 on the remaining real-atom cell. Then, we compare the ghost–real distance distribution the model produced against the one the algorithm would have produced on the same real atoms. Over 11,593 structures containing 121,489 ghosts, the model’s median V1/3V^{1/3}-normalized ghost–real distance is 0.357, against 0.359 for the re-tessellation — a difference of 0.6%. The agreement is stable across all five seeds (Table 4 in Appendix G).

Why is a 0.6% difference enough to support such a claim?

Q3 The text (Appendix I) states:

3000 epochs; best-validation checkpoint (epoch 1963)

Does this mean the checkpoint was selected based on the best loss on the validation set?

3 comments

We thank the reviewer for the careful feedback and address their concerns below. All new numbers come from our own runs (N = 2500 structures, 5 seeds).

W1 (MiAD-style baseline).
We agree this was a missing comparison, so we trained the appropriate MiAD baseline. The ghost-coordinate loss is masked as in MiAD, for both random and Voronoi ghosts, with the position loss averaged either over all atoms or over real atoms only (3000 epochs). MiAD uses the real-atom average. SUN+mSUN (%):

Ghost placementMaskedMasked, real-atom average (= MiAD loss)Supervised (standard loss)
Random23.89 ± 0.9524.21 ± 0.4922.80 ± 0.92
Voronoi23.83 ± 0.2425.66 ± 0.5325.0 ± 0.3

The MiAD baseline on our model (random ghosts, MiAD loss) gives 24.21; our submitted model gives 25.0. With Voronoi ghosts, the supervised and MiAD losses agree within error bars (25.0 vs 25.66), so Voronoi ghosts do not require modifying the loss. Random ghosts do, and with the standard loss they give 22.80. With the same loss, Voronoi placement matches or beats random placement in every column. On DiffCSP the effect is the same. With DiffCSP's architecture and loss unchanged, Voronoi ghosts give 23.31 against 21.10 for random ghosts, and random ghosts reach 23.32 only with MiAD's modified loss.

W2 (LeMat-GenBench, SUN on MP-20). With MP-20 as the novelty reference:

ModelSUNSUN+mSUN
Voronoi2.4235.44 ± 0.39
Random (supervised loss)2.6834.59 ± 1.03
OMatG, no ghosts (tuned schedule)1.4929.94 ± 0.83

Ghosts raise SUN, not only mSUN. On the OMatG framework we measure 2.42 and 2.68 vs 1.49 SUN, and 5.35 (Voronoi) and 6.11 (random, MiAD loss) vs 4.76 on DiffCSP. So the gain is not a trade of SUN for mSUN.

The large stability gap in the submission came from comparing against a different OMatG model (encoder-decoder interpolant, ODE sampling) than the one our model is built on (linear interpolant, SDE sampling). Run through our pipeline, the matching OMatG checkpoint is 7.75–8.14% stable, against 7.3% for our model. On DiffCSP, ghosts do not lower stability (about 14–15% with and without). Ranking by SUN+mSUN and using LeMat-Bulk as the reference are choices made by the benchmark. We will report both references in the revision.

Q1 (why does ghost placement matter if ghosts are removed?). During generation, a ghost can turn into a real atom wherever it currently sits. So the ghost positions during generation are part of the structure being built. The model's predictions for the real atoms also depend on the ghost positions at every step. In practice, Voronoi placement does better than random placement with the same loss (W1).

Our analysis of late ghost→real switches did not show that late insertions hurt stability. We will therefore reword the sentence about keeping ghosts out of occupied space as a description of what the model does, not as the reason it works.

Q2 (is a 0.6% difference enough?). We agree it is not, so we tested it directly. For each generated crystal we remove the model's ghosts, rerun Voronoi maximin placement on the real atoms, and match the model's ghosts to the new sites with a periodic-distance Hungarian assignment. On a scale where 0 is chance and 1 is agreement within the placement jitter (Voronoi training data scores 1, random training data scores 0):

ScoreMean distance to re-tessellated siteWithin 0.5 Å
Our model (Voronoi)0.700.91 Å40%
Random-ghost model0.001.45 Å7%
Chance / jitter floor0 / 11.43 / 0.70 Å7% / 54%

(5 seeds, about 122,000 ghosts per model.)

The model puts most of its ghosts in the Voronoi cavities of new structures it has never seen, while the random-ghost model stays at chance. That is not memorization, and we will remove that word.

Q3 (checkpoint selection). Yes, it is the checkpoint with the lowest total validation loss.

I thank the authors for the additional experiments and clarifications.

W1. I appreciate the matched MiAD-style ablation. A remaining concern is training budget. The original MiAD work trains for 8000 epochs and reports that, at 3000 epochs, performance is still substantially below the peak (MiAD, Fig. 2 / Fig. 3). While the number of epochs required for a MiAD-style objective on the OMatG backbone and diffusion process may differ, a single matched run at 3000 epochs does not yet establish which placement-loss combination is preferable at peak performance. A more conclusive comparison would report performance over training for a sufficiently long run, identify the peak for each configuration, and compare MiAD-style random placement against Voronoi placement at those peaks.

I would also ask the authors to explicitly document the reproduced MiAD-style configuration and how it differs from the original MiAD setup (e.g., number of epochs, batch size, learning rate, learning-rate schedule, and related optimization choices). In particular, MiAD reports that learning-rate schedulers hindered further gains in their experiments, whereas the Ghosts-OMatG pipeline uses a learning-rate scheduler. Such differences do not confound the within-pipeline Voronoi vs. random comparison under a shared training recipe, but they do matter for interpreting how closely the matched baseline follows the original MiAD protocol, and they should be stated clearly in the revision, since incomplete matching of MiAD’s training protocol may underestimate the MiAD-style baseline relative to its peak performance under a more faithful setup.

W2. Thank you for the additional MP-20-reference results. To clarify: I did not claim that ghosts improve only mSUN. My concern was specifically that leadership on SUN+mSUN can be driven largely by mSUN because mSUN is much larger in scale than SUN (roughly 15–85× across the reported models), so the sum can add little new information once both components are reported. I also note that SUN+mSUN is used in LeMat-GenBench, which I had overlooked; given that, and given that the additional results support gains beyond mSUN alone, I no longer view the use of this aggregate as a substantive issue for the paper’s main claims.

Q1. Understood; thank you for the clarification.

Q2. Thank you — the Hungarian-matching analysis addresses my concern.

Q3. Confirmed. As a suggestion for the revision: consider selecting checkpoints by a generation-quality metric (e.g., SUN+mSUN, if that is the primary target) rather than by validation loss. For diffusion/flow models, validation loss is often weakly correlated with sample-quality metrics [1–4]. Metric-based checkpointing would also put the Voronoi and MiAD-style runs on a more equal footing: long training, periodic evaluation every K checkpoints, selection of each run’s best checkpoint by the primary metric, and comparison at peak performance.

[1] Ho et al., DDPM (2020).
[2] Nichol & Dhariwal, Improved DDPM (2021).
[3] When Are Two Scores Better Than One? (2026).
[4] Bigger Isn’t Always Memorizing (2025).

Thank you for the follow-up and the careful reading.

W1. We agree that the OMatG MiAD-style runs use our training recipe rather than MiAD's: AdamW (learning rate 1.97e-4, weight decay 3.1e-4), cosine annealing over 2000 epochs, batch size 256, best checkpoint by validation loss. All OMatG variants share this recipe, and we will state it clearly in the revision. To test MiAD under its own protocol, our DiffCSP runs use MiAD's released code and its MP-20 configuration (8000 epochs, Adam with learning rate 1e-3, no learning-rate scheduler, batch size 256; we only use 20 instead of 25 atom slots). There, random ghosts with the MiAD loss give 23.32 ± 0.52 and Voronoi ghosts with the standard loss give 23.31 ± 1.23 (random ghosts with the standard loss: 21.10 ± 1.34). So when the MiAD baseline is trained as MiAD recommends, it matches but does not exceed Voronoi placement, and Voronoi placement needs no change to the loss. We agree that validation loss is an imperfect selection criterion. In the revision we will select checkpoints by SUN+mSUN for all runs, evaluating the saved checkpoints of the existing OMatG and DiffCSP runs over training, and compare each configuration at its best checkpoint. We will also train both the MiAD-style and Voronoi configurations on OMatG with MiAD's recipe (8000 epochs, Adam with learning rate 1e-3, no scheduler) and compare them the same way.

W2. Thank you for clarifying, we agree. Since mSUN is one to two orders of magnitude larger than SUN for current models, a ranking by the sum is mostly a ranking by mSUN, and SUN carries little weight in it. We still see raising mSUN as useful in its own right, and Table 1 reports SUN and mSUN separately; in the revision we will discuss both rather than only the sum.

Q3. Thank you for the suggestion, we will adopt it (see W1).

Reviewer_P9E2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the fixed atom-count limitation in existing crystal generative models by proposing a flexible generation strategy based on Voronoi-guided ghost atoms. The key idea is to pad crystal structures with fictitious ghost atoms placed at geometrically plausible interstitial sites using iterative Voronoi maximin selection. By treating ghost atoms as an additional atomic species, the model can dynamically determine the number of real atoms during generation without substantially modifying the underlying generative architecture.

The authors implement this strategy within the existing OMatG stochastic-interpolant framework and further investigate the influence of inference-time noise schedules. Experiments on MP-20 using LeMat-GenBench show improvements over random ghost placement and existing crystal-generation baselines, achieving 25.02% SUN+MSUN on average and 25.27% in the best run. Additional analyses of ghost–real atom distances show that the model learns to place ghost atoms in unoccupied regions, supporting the effectiveness of the proposed geometry-aware padding strategy.

Critical strengths
  1. Flexible and potentially plug-and-play design. The proposed ghost atom strategy is simple and flexible, allowing existing fixed-cardinality crystal generative models to support variable atom counts with minimal architectural modifications. This makes the approach practically appealing and potentially applicable to a broader range of generative frameworks beyond OMatG.

  2. Promising empirical improvements with well-controlled experiments. The proposed method achieves improved generation performance over existing baselines, including a 2.6 percentage-point improvement in SUN+MSUN over the matched random-ghost control. The authors also provide controlled experiments that isolate the effect of Voronoi-based placement, along with geometric analyses showing that the generated ghost atoms avoid overlapping with real atoms. These results provide meaningful evidence for the effectiveness of the proposed strategy.

Critical weaknesses
  1. Limited validation across different generative backbones. One of the main advantages claimed by the authors is that the proposed ghost atom strategy can be easily integrated into existing fixed-cardinality crystal generative models. However, the experiments use only OMatG. Since the method is presented as a general, flexible training strategy, evaluating it on additional architectures, such as diffusion-based or flow-matching models, would provide stronger evidence of its generalizability and practical applicability.

  2. Limited methodological novelty and insufficient differentiation from prior work. The proposed approach primarily combines existing techniques, including ghost-based variable-cardinality generation, Voronoi tessellation, and the OMatG framework.

  3. Insufficient direct evaluation of flexible atom-count generation. A central motivation of the paper is to enable generative models to determine the number of atoms during generation rather than relying on a predefined atom-count distribution. However, the experimental evaluation mainly focuses on SUN+MSUN and the geometric placement of ghost atoms. A more detailed analysis of the generated atom-count distributions, as well as generation quality across different atom counts, would help establish whether the proposed method effectively addresses the original fixed-cardinality limitation.

Questions for authors
  1. The proposed method is presented as a flexible and potentially plug-and-play training strategy, but the experiments are limited to OMatG. Could the authors clarify whether the approach can be effectively integrated into other crystal generative architectures?

  2. Could the authors further clarify the methodological novelty of the proposed approach beyond combining existing ghost atom techniques with Voronoi-based placement?

  3. Since enabling flexible atom-count generation is a central motivation of the paper, could the authors provide more direct evidence that the model learns to determine appropriate atom counts based on crystal composition and geometry, rather than simply reproducing the atom-count distribution in the training data?

3 comments

We thank the reviewer for the feedback and suggestions. All new numbers below come from our own runs (N = 2500 structures, 5 seeds).

W1 / Q1 (other models). We added ghost atoms to DiffCSP without changing its architecture or loss. The only addition is an extra atom type for ghosts, which MiAD's codebase already provides, and the ghosts enter through the training data. SUN+mSUN goes from 19.93 ± 0.88 without ghosts to 23.31 ± 1.23 with Voronoi ghosts, against 21.10 ± 1.34 with random ghosts. Stability stays at about 14–15% in all three. Random ghosts reach the Voronoi level only with MiAD's modified loss, which masks the ghost coordinates (23.32 ± 0.52). Voronoi ghosts need no such change, which is what makes them plug-and-play.

W2 / Q2 (novelty). We do not claim a new architecture. The contribution is a way to build training data with a variable number of atoms (ghost atoms at Voronoi maximin sites) that existing fixed-count models can use as is, as the DiffCSP results show. We combine existing methods from different areas to address a key limitation of generative models in materials science. The aim is to let the model choose the atom count together with the composition and geometry, which matters for guided generation (see our reply to YWU5, Q1).

W3 / Q3 (does the model choose sensible atom counts?). Our model chooses N itself, and its atom-count distribution matches MP-20 closely:

Atoms1–45–89–1213–1617–20Mean
MP-20 train19.6%23.1%24.2%16.8%16.3%10.42
Ours (Voronoi)20.1%22.1%23.4%18.6%15.8%10.28
Random ghosts21.3%21.3%23.9%18.2%15.3%10.17
No ghosts20.1%23.8%24.4%16.1%15.6%10.26

We also checked whether the atom count fits the composition the model generates. About a third of generated crystals have a composition that also appears in MP-20; for example, a composition that MP-20 contains only as 5- and 20-atom cells. For 87.8 ± 1.1% of these crystals, the generated cell has one of the atom counts that composition has in MP-20. If N were drawn at random from the training distribution of atom counts (keeping only counts the formula allows), this would happen 41% of the time. The crystals are also better than those of the no-ghost model, which samples N from the training data, at every atom count. SUN+mSUN (%):

Atoms1–45–89–1213–1617–20
Ours (Voronoi)7.1523.3927.2434.2435.90
Random ghosts5.5124.1825.2828.1732.14
No ghosts4.8321.7026.4034.0530.74

Choosing N matters most in guided generation, where the training distribution of atom counts no longer applies (see our reply to YWU5, Q1). Testing this directly needs a guided-generation setup, which is beyond the scope of this paper.

Thank you to the authors for the detailed response and additional experiments.

W1: The new DiffCSP experiments provide useful evidence for the generalizability and plug-and-play nature of the proposed method. While integrating the approach into more generative backbones would make the claim more convincing, I understand the time constraints of the rebuttal period and consider the additional evidence sufficient to substantially address my concern.

W2: Thank you for the clarification.

W3: The additional atom-count analysis is helpful. However, since conventional fixed-cardinality models can also learn dependencies between atom count and composition through p(A∣N)p(A \mid N), could the authors report the same composition–atom-count consistency metric for the no-ghost baseline? This would help clarify whether the observed 87.8% consistency represents an advantage of the proposed ghost-based formulation over conventional generation.

Thank you for the follow-up.

W1. We agree that integrating ghost atoms into more backbones would strengthen the generality claim, and will add them to MatterGen and FlowMM in the revision (both use a fixed atom count, like DiffCSP).

W3. The no-ghost baseline scores higher on this metric: 96.5 ± 0.4% (chance 42%), against 87.8 ± 1.1% for our model. As you note, a fixed-count model learns p(A | N) and chooses a composition that fits the N it drew. It also stays closer to the training compositions: 58.5% of its crystals have a composition in MP-20, against 35.1% for ours. So in unconditional generation this metric does not show an advantage for ghost atoms. What it shows is that our model, which has to choose N itself, recovers consistent atom counts well above chance.

The difference we expect is in guided generation. There, a fixed-count model still has to draw N before generating, from a training distribution that no longer matches the steered compositions, while the ghost model chooses N jointly with the composition (see the perovskite example in our reply to Reviewer YWU5, Q1). Testing this needs a guided-generation setup, which we leave to future work; we will state this limitation explicitly in the revision.

Reviewer_YWU5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies crystal generation which is the problem of learning the distribution of crystals described by a Lattice L∈R3×3L \in \mathbb{R}^{3 \times 3} and atoms with coordinates X∈[0,1)N×3X \in [0, 1)^{N \times 3} and atom types. The authors argue that there is a bottleneck in prior works that the number of atoms per unit cell needs to be predetermined. Then authors then propose to fix the number of atoms for all unit cells by padding ghost atoms up to a maximum cardinality. However, there is an issue with naively placing ghost atoms uniformly/randomly into the cell. These ghost cells will land very close to real atoms. The network is trained to predict the coordinates of both real and ghost atoms, random ghost placement trains the network with impossible 3D geometries. The users propose to place ghost cells using Voronoi diagrams, ensuring that ghost atoms won't be too close to a real atom or a different ghost atom.

Experiments show that the proposed method performs worse on metrics Stable, Metastable, SUN but better results on MSUN. The explanation is that this approach helps with generating unique and novel crystals (note: SUN means Stable, Unique, and Novel, MS in MSUN means Metastable).

Critical strengths

None except the ones mentioned in summary already.

Critical weaknesses
  1. Lack of comprehensive experiments. For example, in the referenced paper Crystalite [1], there are also MPTS-52 Alex-MP-20 datasets. In the same referenced paper, additional metrics like MR, RMSE, Struct. Val., Comp. Val., Unique, Novel (See tables 1, 2, and 3 in the referenced paper).

  2. No statics of distribution of the generated crystals is shown. It is important to show that the generated molecules still have similar cardinality distribution of the number of atoms to the datasets as well as volume.

[1] Crystalite: A Lightweight Transformer for Efficient Crystal Modeling

Questions for authors
  1. It might be a personal opinion, but I do not fully get why having a fixed number of atoms is bad. In introduction, lines 40-45, you state This matters because composition and atom count are not independent of structure. The thermodynamically preferred number of atoms in a unit cell depends on the lattice and species the model is simultaneously generating, so fixing NN a priori prevents the model from exploring structures whose stability requires a different count." However, I do not think this is valid. Given a fixed count, the model can just explore stable structures for this count. Also, from your reported numbers, your approach does perform worse than the baselines in terms of stability.

  2. What's the cause in degrading the stability e.g. from ~12% to 7.3%? Are late-unmasked ghost atoms distorting the lattice or is the model just producing unstable chemistry?

4 comments

(Response 1/2)

We thank the reviewer for the thoughtful feedback and constructive criticism. All new numbers below come from our own runs (N = 2500 structures, 5 seeds).

W1 (more experiments). MP-20 is the standard benchmark for de novo crystal generation; DiffCSP, MatterGen, MiAD, OMatG and Crystalite all report on it. Training on MPTS-52 and Alex-MP-20 would mean retraining our model, its controls and the baselines on each dataset. Alex-MP-20 has about 600,000 structures, against about 45,000 in MP-20, and training on MP-20 alone already takes several days per model, so this was not possible within the rebuttal period.

We report crystal structure prediction results, match rate and RMSE, in Appendix F (Table 3: “CSP match rate and reconstruction error”). Over all generated structures, the match rate is 69.1% and the RMSE is 0.082 Å.

For this rebuttal we added:

  • a second model, DiffCSP, with and without ghosts
  • a MiAD-style ablation with the ghost-coordinate loss masked
  • SUN computed with MP-20 as the novelty reference

For every model we report validity, uniqueness, novelty, stability, metastability, SUN and mSUN.

Validity, uniqueness and novelty (%):

ModelStructural validityCompositional validityUniquenessNovelty
Ours (Voronoi)99.95 ± 0.0596.32 ± 0.3499.23 ± 0.3067.46 ± 1.02
Voronoi ghosts, MiAD loss99.94 ± 0.0596.51 ± 0.4899.14 ± 0.2064.94 ± 0.63
Random ghosts99.97 ± 0.0596.58 ± 0.4799.15 ± 0.1962.06 ± 0.63
Random ghosts, MiAD loss99.93 ± 0.0396.78 ± 0.5399.00 ± 0.2656.94 ± 0.56
No ghosts (tuned schedule)99.98 ± 0.0497.30 ± 0.3398.85 ± 0.3061.93 ± 1.22
DiffCSP, Voronoi ghosts100.00 ± 0.0096.40 ± 0.3497.71 ± 0.2946.07 ± 1.06
DiffCSP, random ghosts99.99 ± 0.0296.46 ± 0.3298.07 ± 0.2141.46 ± 1.64
DiffCSP, random ghosts, MiAD loss99.98 ± 0.0297.50 ± 0.2197.80 ± 0.2843.36 ± 0.81
DiffCSP, no ghosts99.18 ± 0.2197.27 ± 0.1998.08 ± 0.3141.89 ± 0.80

Stability, metastability and SUN (%):

ModelStableMetastableSUNSUN+mSUN
Ours (Voronoi)7.30 ± 0.6146.96 ± 0.680.68 ± 0.1625.02 ± 0.31
Voronoi ghosts, MiAD loss8.05 ± 0.7250.16 ± 1.460.83 ± 0.1225.66 ± 0.53
Random ghosts8.65 ± 0.1448.73 ± 0.560.74 ± 0.1322.80 ± 0.92
Random ghosts, MiAD loss10.45 ± 0.5054.11 ± 0.660.84 ± 0.1824.21 ± 0.49
No ghosts (tuned schedule)8.14 ± 0.9550.20 ± 0.660.57 ± 0.1122.85 ± 0.72
DiffCSP, Voronoi ghosts14.84 ± 0.8058.82 ± 1.461.46 ± 0.1423.31 ± 1.23
DiffCSP, random ghosts14.93 ± 0.7561.15 ± 0.971.11 ± 0.1921.10 ± 1.34
DiffCSP, random ghosts, MiAD loss14.09 ± 0.2761.29 ± 1.171.54 ± 0.1823.32 ± 0.52
DiffCSP, no ghosts14.35 ± 0.5059.74 ± 0.701.11 ± 0.2419.93 ± 0.88

Structural validity is LeMat-GenBench's minimum-distance and physical-plausibility check. Compositional validity is its charge-neutrality check. All other rates are over valid structures, with novelty measured against LeMat-Bulk. "Random ghosts" is the random-ghost control trained for the full 3000 epochs used for our model.

(Response 2/2)

W2 (statistics of generated crystals). The number of real atoms per crystal generated by our model matches MP-20 closely:

Atoms1–45–89–1213–1617–20Mean
MP-20 train19.6%23.1%24.2%16.8%16.3%10.42
Ours (Voronoi)20.1%22.1%23.4%18.6%15.8%10.28
Random ghosts21.3%21.3%23.9%18.2%15.3%10.17
No ghosts20.1%23.8%24.4%16.1%15.6%10.26

The no-ghost model draws N from the training data, so it matches by construction. The ghost models choose N themselves and still match.

Volume per atom is 20.57 ų for our model, against 20.08 ų in MP-20, 20.90 ų for random ghosts and 20.69 ų for no ghosts. Our model is the closest to the training distribution (Wasserstein-1 distance 0.52, against 0.85 for random ghosts and 0.64 for no ghosts).

Q1 (why is a fixed atom count a problem?). We agree this needs a clearer argument. For unconditional generation a fixed count works, since the model can look for stable structures at that count. The problem arises with guided generation, where the right count depends on the composition and geometry being generated.

Take ABX₃ perovskites. The stable tilt pattern, and so the number of atoms in the primitive cell (5, 10 or 20), depends on the species. If we steer toward a target band gap, the composition changes and so does the correct count. A fixed-count model has to pick N first. It can guess, or sample N from the training data, which no longer fits once generation is steered. With ghost atoms, the model picks N together with the species and lattice. We will rewrite the sentences in lines 40–45 to make this argument. On stability, see Q2.

Q2 (why does stability drop from ~12% to 7.3%?). The ~12% in our table is LeMat-GenBench's entry for a different OMatG model, which uses the encoder-decoder interpolant with ODE sampling. Our ghost model is built on OMatG's linear interpolant with SDE sampling, the OMatG variant with the highest S.U.N. rate in the OMatG paper (22.5% vs 18.8% for the encoder-decoder model, which is more stable but less novel). Run through our pipeline, that checkpoint is 7.75% stable with our sampling schedule and 8.14 ± 0.95% with a schedule tuned for it, against 7.30 ± 0.61% for our model. So adding ghosts to that backbone costs us little stability.

We also removed all generated crystals with a late ghost→real switch, and stability did not go up. On DiffCSP, ghosts do not lower stability. We measured 14.35% stability without ghosts, 14.84% with Voronoi ghosts. We will replace the baseline numbers with our own.

Thanks for the response. I have some remaining concerns:

  1. You agree that "for unconditional generation a fixed count works" and argue variable cardinality is specifically needed for guided generation. Yet every experiment in the paper is unconditional generation. The paper provides no experimental validation for the only regime where you now claim variable atom counts are necessary.
  2. Against the fair "No ghosts (tuned schedule)" baseline, Voronoi ghosts still reduce raw stability (7.30% vs. 8.14%) and metastability (46.96% vs. 50.20%), gaining +2.17% SUN+mSUN purely via novelty (67.46% vs. 61.93%). Moreover, MPTS-52 has ~40k structures (comparable to MP-20) with up to 52 atoms per cell, making it feasible and important for testing how iterative 3×3×33 \times 3 \times 3 Voronoi padding scales beyond N=20N=20.

I will maintain my score for now.

Thank you for the follow-up.

Unconditional vs. guided generation. We phrased this poorly in Q1. What we meant is that a fixed count does not prevent unconditional generation, not that nothing is gained by choosing N. Our claim in the paper is about unconditional generation, and that is what the experiments test: ghost atoms raise SUN+mSUN on both backbones (OMatG: 25.02 ± 0.31 vs 22.85 ± 0.72 for the tuned no-ghost baseline; DiffCSP: 23.31 ± 1.23 vs 19.93 ± 0.88). Guided generation is a further reason to want a variable count, but we have not tested it, and in the revision we will present it as motivation and future work rather than as a result.

Stability vs. novelty. The no-ghost model is more often metastable (50.20 ± 0.66% vs 46.96 ± 0.68%); the stability difference (8.14 ± 0.95% vs 7.30 ± 0.61%) is within the seed-to-seed spread. But more of its structures are already known: 58.5% of its crystals have a composition in MP-20, against 35.1% for ours (see our reply to Reviewer P9E2, W3). So the trade-off you point out is real. Our model gives up some metastability, and in return more of what it generates is new. This trade-off is common in de novo generation: a model that fits the training distribution more closely tends to be more stable but less novel (see Crystalite, Fig. 4, which shows this over training). For discovery, the relevant quantity is the overlap of the two, structures that are both stable and new, and that is what SUN and mSUN measure. With MP-20 train as the novelty reference, SUN is 2.42 ± 0.26% for our model against 1.49 ± 0.32%, and mSUN is 33.02 ± 0.34% against 28.45 ± 0.69%. On DiffCSP, ghosts do not reduce stability (14.84% vs 14.35%) and raise SUN from 1.11% to 1.46%.

MPTS-52. We agree this is an important test of whether the approach scales. A ghosted OMatG run on MP-20 takes about 3.5 GPU-days, and a 52-slot run together with its matched no-ghost baseline would have taken most of the rebuttal period, so we prioritized the second backbone and the MiAD-style ablations. To see how the approach scales beyond N=20N = 20, we measured the two costs that ghosts add.

The first, dataset preparation, is a one-time cost: padding the MP-20 training set took 2 hours on 16 CPU cores, and on a random sample of 2,000 MPTS-52 training structures, padding to 52 slots takes about 8× longer per structure, or roughly 16 hours for the full training set on the same cores. Ghosts make up 65% of the slots (48% on MP-20) and still keep away from the atoms: the median minimum ghost–atom distance is 1.38 Å, against 0.45 Å for the same number of randomly placed ghosts (1.67 Å vs 0.68 Å on MP-20).

The second, training cost, depends on the architecture. OMatG's network passes messages between all pairs of atoms present in each crystal, so ghosts increase the number of edges by 2.9× on MP-20 and 5.5× on MPTS-52, and a training step of the network on a batch of 256 takes 652 ms at 52 slots against 109 ms at 20 (one A100). Transformers such as Crystalite already pad every crystal to the dataset maximum (52 on MPTS-52) and attend over all slots, so for them ghosts would occupy slots that are currently padding, at no extra cost to the network.

In the revision we will train OMatG with and without ghosts on MPTS-52 and report validity, stability, metastability, novelty, SUN and mSUN.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.