acceptodds
Under review as a conference paper at ICLR 2027

FamilySpace: From Emergent Chemical Semantics to Controllable Crystal Generation

Abstract

Latent crystal generative models typically combine a variational autoencoder (VAE) for learning atom-wise representations with a flow model that transports a simple source distribution to the learned latent distribution. However, previous studies have mainly focused on the quality of generated crystals, leaving the structure and generative utility of the atom-wise latent space largely unexplored. In this work, we show that even a vanilla crystal VAE implicitly organizes atom-wise latent representations according to elemental families, with different families occupying largely separated regions of the latent space. We term this phenomenon emergent chemical semantics. This structure suggests that sampling from a family-specific latent region can bias the decoder toward atoms from the corresponding elemental family. Building on this insight, we propose FamilySpace, a framework for controlling the elemental-family composition of generated crystals. FamilySpace assigns each elemental family a distinct source component, forming a family-indexed multi-source distribution, and trains a flow model using source-target pairs matched by family. In inference, a user-specified family template selects the source component for each atom-wise representation, enabling direct control over the family composition of generated crystals. Experiments on MP-20 and MPTS-52 show that FamilySpace achieves competitive crystal-generation performance while matching prescribed elemental-family compositions. Moreover, selected family templates increase the fraction of generated crystals predicted to be stable or metastable, demonstrating that latent-space composition control can support targeted exploration of promising chemical regions without additional optimization techniques such as reinforcement learning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Controllable crystal generation, Emergent chemical semantics, Crystal variational autoencoder

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HRRbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this work, the authors observe that the latent space of a crystal generation latent diffusion model gets naturally organised into clusters of similar atoms and propose a latent diffusion model that takes advantage of this fact to generate crystals with guidance towards a specified template.

Critical strengths
  1. The paper has been properly proof-read and well organised, hence easy to follow.

  2. The work addresses an important gap in the latent diffusion literature, which is atomic-level guidance. Although the exact atom types cannot be specified, the atomic family gives controllability on the type of atoms we want to have in the final crystal.

  3. The authors have demonstrated the clustering behaviour in two different crystal datasets (MP-20 and MPTS-52) as well as the QM9 molecular dataset. This shows that this is not a dataset specific property.

  4. Table 3 shows that the model is able to adhere to the target atomic family, and tables 1 and 2 show that even within a specific family of crystals, the model shows remarkable performance with respect to stability/novelty.

Critical weaknesses
  1. In table 1, the templates are not generated arbitrarily, but takes from the training set, yet the stability and SUN metrics are lower than ADiT, on which this work is based upon. Can the authors justify this observation, or show some failure cases to highlight the reason behind this?

  2. The medoids are currently chosen as the point closest to the (roughly stated) mean of normalized posterior means, which might lie outside a cluster. How much does this affect the performance of the model? Can the authors show an ablation where they randomly select a representative from each family and use a sharp Gaussian around them as the prior distribution? Since flow matching can map any source distribution to a target distribution, does randomly selecting such representatives affect the model's performance?

  3. With the current approach, the atomic family must be known beforehand to be able to generate crystals belonging to it. A limitation towards discovery of new materials is the need for discovering novel family of crystals first, which requires domain knowledge and cannot be automated.

Questions for authors

see weaknesses

3 comments

Thank you for your valuable and constructive comments. Our responses are provided below.


Response to Weakness 1. FamilySpace uses training templates that specify the elemental families and the number of atoms from each family in a crystal, while ADiT only specifies the total number of atoms during sampling. As a result, the generated crystals may have different distributions of elemental families, which can affect the generation metrics. We analyze the generated crystals from ADiT and FamilySpace, with the results shown below.

Average atomic fraction of each elemental family per crystal

Elemental familyMP-20 training setADiTFamilySpace
Alkali6.69%6.38%6.69%
Alkaline earth5.20%5.57%5.21%
Boron group6.80%6.24%6.96%
Carbon group6.71%6.44%6.70%
Pnictogens5.01%5.07%4.85%
Chalcogens20.79%23.51%20.66%
Halogen7.85%7.43%7.76%
Noble gas0.05%0.04%0.11%
Transition metal28.69%26.55%28.79%
Rare earth12.21%12.76%12.27%

Distribution of the number of distinct elemental families per crystal

Number of familiesMP-20 training setADiTFamilySpace
15.83%5.11%6.03%
235.13%31.95%35.74%
349.22%48.36%48.71%
≥49.82%14.58%9.52%

For the first table, we calculate the fraction of atoms belonging to each family within each crystal, assigning zero if the family is absent, and then average the fractions across all crystals with equal weight. The second table reports the number of distinct elemental families in each crystal. Different elements from the same family are counted as one family. These results show that ADiT and FamilySpace generate different distributions of elemental families, which may contribute to the differences in their generation metrics.

As shown in Table 1 of the revised manuscript, FamilySpace achieves a trade-off between stability and novelty. On MP-20, FamilySpace achieves a substantially higher UN rate than ADiT (69.53% vs. 30.81%), while its stability rate is lower (6.26% vs. 11.03%). Although SUN is slightly lower (1.38% vs. 1.68%), MSUN improves from 13.39% to 15.31%. These results indicate broader exploration of novel structures, but not necessarily higher stability.


Response to Weakness 2. Thank you for this suggestion. We have conducted additional ablation experiments to examine how the selection of source centers affects generation performance. The results are provided in Appendix B.4 (pages 20–21, Table 11 of the revised manuscript).

Source centersUN (%)S (%)SUN (%)MS (%)MSUN (%)
Our medoid69.536.261.3842.9915.31
Random representative 170.635.861.0142.2714.14
Random representative 270.435.771.1343.7015.32
Random representative 372.405.571.0740.9314.73
Evenly spaced centers75.004.950.9438.3414.55

The results show that our medoid achieves the highest stability and SUN rates. Random representatives achieve comparable results in some metrics, while evenly spaced centers improve UN but reduce stability and metastability. Overall, our medoid provides a balanced choice across the evaluated metrics. Please refer to the revised manuscript for further details.


Response to Weakness 3. The elemental families are general categories based on the periodic table. By specifying predefined elemental families, FamilySpace can generate crystals containing only elements from the selected families, allowing users to control the types of materials being generated. We agree that domain knowledge can help guide more targeted exploration, as demonstrated in our application-oriented exploration experiments.

In LLM-driven automated crystal discovery, such as autonomous laboratories or agents research, the domain knowledge of LLMs may help select suitable family templates and make better use of the control capabilities of FamilySpace.

For general exploration, we can check whether the combinations of elemental families and atom counts satisfy basic chemical requirements, such as charge neutrality. In fact, other diffusion and flow matching crystal generation methods also require specifying the number of atoms during generation. FamilySpace provides additional control over the elemental families and their ratios.

I thank the authors for the additional experiments. Table 11 clearly shows the advantage of using medoids over random representatives or evenly spaced centers, but the elemental family distribution table does not really justify the dip in stability compared the ADiT. There might be novel compositions, but it can't be inferred from the elemental distributions, more so because the distribution closely follows that of the MP-20 training set. A little more investigation into what drives the novelty would be good, but overall, I maintain my score.

Thank you very much for taking the time to review our additional experiments. We appreciate your comments and suggestions, and we will further investigate.

Reviewer_cFdhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the latent representations learned by VAEs for crystal generation. The authors observe that atom-wise latent representations naturally exhibit clustering patterns corresponding to elemental families, which they refer to as "emergent chemical semantics." Building on this observation, they propose FamilySpace, a framework that introduces a family-indexed multi-source distribution and family-matched source-target coupling into latent flow matching to enable controllable crystal generation. Experiments on MP-20 and MPTS-52 demonstrate that the proposed method achieves competitive crystal generation performance, accurate control over elemental-family compositions, and improved stability metrics under selected family templates.

Critical strengths
  1. High research quality and comprehensive empirical analysis. The paper is well organized, technically detailed, and supported by extensive experiments. I appreciate the in-depth investigation of the VAE latent space across different architectures and datasets, including both qualitative visualizations and quantitative nearest-neighbor analyses.

  2. Potential value for materials discovery. The ability to control elemental-family compositions and explore chemically meaningful regions may be useful for practical applications in crystal and materials generation.

Critical weaknesses
  1. Limited novelty of the main observation. The paper presents the clustering of elemental families in the VAE latent space as a major finding. However, I do not consider this observation particularly surprising from a machine learning perspective. VAEs are designed to learn representations that preserve information relevant to reconstruction. Since the reconstruction objective explicitly includes atomic species, it is reasonable to expect the latent representations to encode elemental identities and reflect similarities among related elements.

  2. Limited methodological novelty. The proposed family-matched latent flow is a relatively straightforward modification of an existing VAE-based latent flow framework by simply replacing the Gaussian latent with a mixture of Gaussian. This is what existing class-conditional generative modeling already doing.

  3. Limited contribution to the broader machine learning community. Overall, I appreciate the quality of the experiments and the potential practical utility of this work for crystal generation. However, the main contributions appear to be the empirical characterization of a domain-specific latent representation and a relatively standard adaptation of latent generative modeling to incorporate chemical-family information. I believe this work may be more suitable for a specialized venue in computational materials science than for a general machine learning conference such as ICLR.

Questions for authors

I do not have more questions.

1 comment

Thank you for your valuable comments and for recognizing the strengths of our work. We would like to clarify the following points.


Response to Weakness 1. We agree that VAEs are expected to learn information useful for reconstruction. However, learning to reconstruct atomic species does not necessarily imply that their latent representations will form coherent clusters according to elemental families.

The VAE objective only includes reconstruction and KL regularization, without any family label supervision or constraints on distances between latent representations. Therefore, even if atomic identities can be reconstructed accurately, elements from the same family are not guaranteed to occupy neighboring regions in the latent space.

To further investigate this, we have added an analysis of crystal systems, space groups, and structural prototypes (Figure 5, page 18 of the revised manuscript). The results show that not all properties relevant to reconstruction exhibit similar clustering patterns. For example, the same-label 1-NN accuracy is 99.7% for elemental families, compared with 49.5% for crystal systems at the atom level. At the crystal level, the corresponding accuracies are 45.3% for crystal systems, 30.7% for space groups, and 50.8% for structural prototypes.

These findings demonstrate that the strong family organization is not simply a guaranteed consequence of VAE reconstruction. More importantly, our contribution goes beyond observing this phenomenon: we show how the learned organization can be directly exploited for controllable crystal generation.


Response to Weakness 2. FamilySpace is not simply replacing a Gaussian prior with a Gaussian mixture for conventional class-conditional generation.

First, the Gaussian prior used for VAE training remains unchanged. We only modify the source distribution of the subsequent latent flow.

Second, rather than using arbitrary Gaussian components, we construct family-specific source distributions based on the emergent organization of the VAE posterior. Importantly, we introduce family-matched source–target coupling, which pairs each atomic source representation with a target representation from the same elemental family.

Third, unlike conventional class-conditioned generation, family labels are not provided as additional inputs to the flow network. Instead, the family of each atomic representation is specified through source selection. A single shared flow can therefore generate crystals with different family combinations and ratios without retraining or additional conditioning modules.

Thus, our contribution lies in connecting the emergent latent organization with source construction and family-matched flow training, rather than proposing Gaussian mixtures themselves. Please see Sections 4.2–4.3 (pages 5–6) for details.

The effectiveness of this design is also supported by our experiments. FamilySpace achieves 99.992% family-set accuracy and 99.968% family-ratio accuracy (Table 6, page 10). We have further added comparisons with existing composition-conditioned generation methods, including MatterGen, Chemeleon, and OMatG (Tables 5–6, pages 9–10 of the revised manuscript), demonstrating more accurate control and higher MSUN rates in the evaluated settings.


Response to Weakness 3. The official ICLR 2027 Call for Papers explicitly includes "applications to physical sciences (physics, chemistry, biology, etc.)" among its subject areas. It also welcomes research on generative models and the interpretation of learned representations, all of which are central to our work.

Moreover, ICLR has previously accepted numerous papers on crystal generation, including CDVAE (ICLR 2022 [1]), UniMat (ICLR 2024 [2]), DiffCSP++ (ICLR 2024 [3]), SymmCD (ICLR 2025 [4]), TGDMat (ICLR 2025 [5]), CrysBFN (ICLR 2025 [6]), RG-VFM (ICLR 2026 [7]), LLEMA (ICLR 2026 [8]), etc.

Beyond its application to crystal generation, our work investigates the internal organization of learned representations and how this organization can be used to design controllable generative models. We believe these contributions are relevant to the broader ICLR community.

[1] Crystal diffusion variational autoencoder for periodic material generation. ICLR 2022.

[2] Scalable Diffusion for Materials Generation. ICLR 2024.

[3] Space Group Constrained Crystal Generation. ICLR 2024.

[4] SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models. ICLR 2025.

[5] Periodic Materials Generation using Text-Guided Joint Diffusion Model. ICLR 2025.

[6] A Periodic Bayesian Flow for Material Generation. ICLR 2025.

[7] Riemannian Variational Flow Matching for Material and Protein Design. ICLR 2026.

[8] Accelerating Materials Design via LLM-Guided Evolutionary Search. ICLR 2026.

Reviewer_CeGdOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors analyze emergent semantic properties of crystal VAE latents, showing empirically that the latent space naturally organizes by elemental family. They leverage this insight to enable controllable generation via elemental family-conditioned flow matching.

Critical strengths
  • The approach is simple and clear
  • The insight that crystal VAE models naturally organize the latent space by elemental family is interesting
Critical weaknesses
  • No comparisons to Crystal Structure Prediction models or composition-conditioned (e.g. via classifier free guidance) latent diffusion models, which achieve the same goal of composition-conditioned generation through existing techniques
Questions for authors
  • Did you explore whether the VAE latent spaces also naturally organize other semantic properties like structural properties (prototype structures, space groups, etc.)? It would be interesting to see what semantic properties are and are not automatically captured by VAE training
3 comments

Thank you for your valuable and insightful comments. Our responses are provided below.

Response to Weaknesses. We have added comparisons with composition-conditioned generation methods (MatterGen and Chemeleon) and a CSP model (OMatG). We also include Chemeleon2 as an additional latent diffusion baseline (reinforcement learning) in Table 1 (page 7) of the revised manuscript. The new comparisons are presented in Tables 5 and 6 (pages 9–10) of the revised manuscript, with the results summarized below.

Table 4: Comparison with other composition-conditioned generative models.

MethodValid (%)U (%)N (%)UN (%)S (%)SUN (%)MS (%)MSUN (%)
RE–Halogen (1:3)
MatterGen90.0923.1147.7921.1111.512.1786.2613.00
Chemeleon84.7729.5267.0227.9413.263.7771.4415.16
OMatG (CSP)1008.1434.727.129.120.7781.875.04
FamilySpace91.2264.5392.5863.007.874.9770.3641.50
Alkali–Halogen (1:1)
MatterGen73.8240.2551.9536.9317.444.4883.6829.18
Chemeleon82.3943.9158.1941.4614.202.3779.6328.29
OMatG (CSP)99.9122.8840.4420.6517.320.8994.8117.52
FamilySpace99.8547.7458.4045.2513.560.7090.2336.12

Table 5: Comparison of composition control accuracy.

MethodFamily-set accuracy (%) ↑Family-ratio accuracy (%) ↑
RE–Halogen (1:3)
MatterGen77.0266.91
Chemeleon92.7981.90
FamilySpace99.6699.66
Alkali–Halogen (1:1)
MatterGen97.8968.88
Chemeleon98.9981.34
FamilySpace100.0099.92

FamilySpace achieves the highest MSUN rates on both templates and more accurate family-composition control than MatterGen and Chemeleon. These results demonstrate the effectiveness of FamilySpace for targeted crystal generation. Please refer to Tables 5–6 (pages 9–10) of the revised manuscript for further details.

Response to Questions. We have added an analysis of crystal systems, space groups, and structural prototypes in the VAE latent space (Figure 5, page 18 of the revised manuscript). The results show that elemental families form clear clusters, while structural properties show much weaker clustering. For example, the same-label 1-NN accuracy is 99.7% for elemental families, compared with 45.3% for crystal systems, 30.7% for space groups, and 50.8% for structural prototypes at the crystal level. These findings suggest that the VAE captures elemental-family information more clearly than structural properties.

Thank you to the authors for the added experiments, which are interesting and raise confidence in the results. I will raise my score accordingly.

Thank you very much for your positive feedback and for taking the time to review our additional experiments. We greatly appreciate your consideration and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.