acceptodds
Under review as a conference paper at ICLR 2027

A Unified Framework for Joint Crystal Generation and Property Prediction

Abstract

Crystal generation and property prediction are two fundamental tasks in the discovery of novel materials. Recently, AI-based frameworks have shown substantial promise for both tasks; however, they have largely evolved as separate paradigms, since each requires distinct representations and inductive biases. Predictive models typically rely on local, neighborhood-based multigraph representations, whereas generative models often operate on global three-dimensional crystal representations. In this work, we investigate whether these complementary representations can be unified to jointly support crystal generation and property prediction. To this end, we propose CrysUniNet, a unified framework that learns a shared embedding space by integrating local and global representations of crystal structures, enabling a single latent representation to support both generative and predictive tasks. We first train a VAE with a unified encoder and a multi-task decoder that maps crystals into this latent space, after which a latent diffusion model learns to generate new latent embeddings. Once trained, CrysUniNet can generate novel crystal materials via the latent diffusion model, while the VAE's property-prediction head enables property prediction for any target material. Empirically, CrysUniNet performs strongly on both tasks, generating more SUN crystals than the strongest SUN baseline while also predicting the properties of existing structures more accurately.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ML4Materials, Latent Diffusion Models, Crystal Materials, Material Generation, AI4Science, Crystal Property Prediction

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_KZm9OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
3(+2)
Read review
Paper summary

CrysUniNet combines local and global crystal features and uses a multi-task VAE and latent diffusion to support both crystal generation and property prediction.

Critical strengths
  1. The paper provides a unified framework for crystal generation and property prediction, achieving promising results on both tasks.
  2. Merging graph and 3D information.
Critical weaknesses

The methodological novelty is limited. The proposed framework mainly combines existing techniques, such as VAE, multi-task learning, and latent diffusion, without introducing substantial architectural innovations.

Questions for authors
  1. Why does the paper mainly report validity and stability metrics? Additional diversity and distribution matching metrics would provide a more comprehensive evaluation.
  2. Why is crystal generation only evaluated on MP-20? More datasets are needed to demonstrate the generalization ability of the model.
  3. The baseline comparisons seem limited. Why are recent models such as CrystalLLM not included?
  4. In Figure 2, removing the property prediction objective improves SUN and MSUN. How does this support the claim that multi-task learning benefits both generation and prediction?
Reviewer_JQP9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce CrysUniNet, a latent flow matching model with the joint purpose of generating stable, unique, and novel crystal structures as well as fitting properties of materials. Their work applies a standard two step training procedure for latent flow matching models.  First they learn a latent representation with a VAE trained with usual reconstruction + regularization alongside an auxiliary property prediction loss based on the prediction of the formation energy. They then train a standard flow matching model on the latent samples to decrease latent distribution mismatch. Their VAE concatenates atomic embedding vectors from an exactly invariant ComFormer graph encoder and a standard Transformer architecture trained with data augmentation. With the trained model they demonstrate 6.54 % SUN and 31.59 % MSUN metrics alongside strong demonstrated fine-tuning of many properties of the JARVIS DFT dataset.

Critical strengths

The authors demonstrate strong performance on several cross-domain benchmarks involving both regression of properties and generation of SUN crystalline materials. Their application of a joint augmented + exactly invariant latent space seems to lead to a more expressive representation capable of better describing materials. Figure 2 reveals an interesting trade-off between stability and generative diversity. An apparent decrease in novelty and/or uniqueness among (meta)stable structures arises upon introduction of invariant graph in addition to the transformer features despite increasing (meta)stability, highlighting a useful quantitative result. Their property objective ablations further support the value of joint training. The authors also present a clear and concise exposition of their proposed method, which is certainly a strength in the modern research landscape.

Critical weaknesses

The literature review and empirical comparisons are seemingly incomplete. Firstly, there is a noticeable lack of discussion on Chemeleon2, published in Nature Machine Intelligence, which also uses a VAE-based latent diffusion framework and demonstrates strong metastable-material generation after reinforcement learning. Chemeleon also applies the same latent diffusion model that the proposed work does, showing a lack of novelty in the proposed work. A critical ablation of their joint latent space could be performed by comparing embeddings from Chemeleon and CrysUniNet to further show the importance of the ComFormer + Transformer representation. Furthermore, more credit must be assigned to existing work on transformer-based generators such as Crystalite which serve as the SOTA for SUN performance as far as the reviewer knows.

The authors should acknowledge recent crystal generators (CSP and DNG alike) such as Crystalite, OMatG, and Kinetic Langevin Diffusion, alongside symmetry-aware approaches including SymmCD, SGEquiDiff, Space Group Conditional Flow Matching and MaskGXT. These works provide context for assessing the contribution of their geometric representation and generation framework. The discussion of property-guided generation should also include OMatGRPO which improves MSUN yield and demonstrates how single-element repackings and insufficient reference-hull coverage can inflate apparent discovery performance. These findings motivate additional diagnostics of the generated chemical and structural distributions.

The authors should evaluate their pretrained generator on LeMat-GenBench under a matched sampling budget of k=2500 structures as opposed to their 10,000.  The benchmark forces a standard thermodynamic stability definition and proper novelty assessment allowing the authors to more effectively place themselves relative to other generators.  We also recommend the authors sample at least 10 independent k=2500 batches to permit reporting mean performance and error bars, while retaining comparisons at the same batch size rather than padding predictions up to 10000.   This requires additional inference and benchmark evaluation, but no retraining. Matching the benchmark’s relaxation, novelty, and stability protocols would make comparisons substantially more informative than juxtaposing percentages from incompatible pipelines. If time permits, comparisons against MP-20-only models do not isolate the contribution of the proposed representation. A matched-data ablation is needed, this means training the VAE and latent flow model on only MP20 is needed for a proper comparison.

If the claimed improvements survive matched data, evaluation protocols, and uncertainity estimates I will consider raising my score.

Questions for authors
  • How does the reconstruction loss account for physically equivalent representations of the same crystal, including rotations, periodic translations, atom permutations, unimodular changes of lattice basis, and supercells? The stated coordinate and lattice losses appear to penalize such equivalent representations as they do not look invariant as presented to such symmetries. Do the authors choose to canonicalize/align structures before computing the loss? How is orientation recovered from a latent representation that is intended to be rotation invariant/equivariant? This discussion is lacking from the paper.
Reviewer_P25HOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new generative model based on a VAE to build a global latent vector encoding crystals with a diffusion model to generate crystalline structures. A property classifier is also proposed.

Critical strengths
  • Strong on SUN related metrics
Critical weaknesses
  • Property prediction seems close to SOTA and wasn't compared with known works like orbV3 or MACE.

  • SOTA performance might not be true if the MP20+QM9 checkpoint was used for ADiT

  • SUN is not computed with DFT and relies on an MLIP approach which is less precise.

  • Decoder architecture is not well explained (VAE decoder)

  • Benchmark didn't include MatterGen

  • Conditional generation is not tested against other approaches so what is the point of property prediction?

  • Limited to small systems of up to 20 atoms.

Questions for authors
  • You train on MP20 like ADiT, but haven't you included ADiT trained on MP20+QM9?

  • In table 1, metastable materials are ~82% but SUN is ~7%, does it mean that most structures are not novel or unique?

  • Why haven't you tested MatterGen in table 1? It's a well known generative model.

  • ADiT exists in small, medium and large versions, which one have you used?

  • Haven't you tested conditional generation against generative models with classifier-free guidance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.