acceptodds
Under review as a conference paper at ICLR 2027

Composable Crystals: Controllable Materials Discovery via Concept Learning

Abstract

Existing methods for de novo crystal generation largely produce complete crystal structures, either through stochastic sampling or under predefined structural constraints, providing limited mechanisms for explicitly structuring how the crystal space is explored. In this work, we introduce a concept-based compositional framework that represents crystal generation as the recombination of learned structural units. We learn a discrete vocabulary of local crystal concepts using a vector-quantized variational autoencoder, where each concept represents a recurring local atomic environment. Unlike predefined structural attributes, these concepts are automatically discovered from data, and multiple concepts can coexist and be recombined within a single crystal. Moreover, the learned concepts exhibit interpretable patterns in both local atomic environments and global crystal symmetries, and empirically transfer across crystal datasets. We further introduce a composition generator to produce novel concept combinations and refine it using high-quality samples. The resulting concept compositions guide downstream crystal generation toward structural combinations beyond those observed in the training data. Experiments on MP-20 and Alex-MP-20 show that concept composition improves the V.S.U.N. score of the base generative model by 53.2% and 51.7%, respectively, with the main gain arising from increased novelty.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
crystal generation, material discovery, compositional generation

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_z76ROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a compositional framework for crystal space exploration and generation. Generation is done through known diffusion processes. Crystal concepts are extracted through a VQ-VAE. These are then used for composition through a diffusion model with the generator being restricted to valid compositions. The main contribution is to bring together well known techniques for this task. The chosen task is not novel as people have been attempting this for at least 3 years. The validation is done on the MP20 dataset by collecting 1000 crystals in each sampling run. Then they perform another test by training on MP-20 and seeing how much of the ALEX MP 20 data they can recreate.

Critical strengths

None

Critical weaknesses

• The paper seeks to discover new crystals. But the methodology , encoding representation and generation are quite standard. I derived no new insights into how useful concepts are generated as compared to prior methods and why. • The most important part is to show new crystals can be discovered. They used the training set MP 20 crystals to evaluating stability, novelty etc using VSUN. What were the 1000 crystals found? Were they really new (not in terms of distance in structure but in material science terms?)? What were the material scientist view on these structures? These are not given, or even discussed. This is important in this context.

Questions for authors

Many aspects are poorly formulated –(eg. Derivative reference used for DFT). What is the utility of Figure 2, appears to be of little value to the understanding

3 comments
AuthorsOpenReview ↗

Rebuttal 1

Q1: The paper seeks to discover new crystals. But the methodology, encoding representation and generation are quite standard. I derived no new insights into how useful concepts are generated as compared to prior methods and why.

Discovering new crystal structures remains an open and challenging problem in materials generation. Most existing generative methods focus on learning the training data distribution, which limits their ability to explore structures that differ significantly from known materials. To overcome this limitation, we propose a novel framework that learns local structural concepts and combines them in new ways to generate previously unseen crystal structures. Our experiments show that this compositional approach achieves substantially higher novelty than existing methods. Therefore, the main contribution of our work is therefore not a new VQ-VAE, Transformer or diffusion architecture, but a new approach to crystal generation that uses learned concepts to explore structural possibilities beyond the training distribution.

As discussed in Section 3.4, existing conditioning signals, such as physical properties, space groups, and predefined structural motifs, are not directly suited to our compositional setting. In contrast, our learned concepts have three important properties: (1) locality, allowing multiple concepts to coexist within a crystal; (2) discreteness, providing identifiable and reusable structural units for explicit recombination; and (3) data-driven learning, avoiding reliance on predefined structural patterns.

We intentionally adopt standard VQ-VAE, Transformer, and diffusion architectures to isolate the contribution of compositionality. Although more advanced architectures could potentially improve generation performance, they would also introduce additional factors, making it harder to attribute improvements specifically to concept composition.

Our experiments demonstrate that introducing concept composition into the same base generative architecture substantially improves structural novelty. Moreover, our concept analyses and verification experiments in Sections 4 and 6 show that the learned concepts capture coherent local structural patterns and can be recovered from generated crystals. These results support both the usefulness of the learned concepts and their role in enabling compositional generation.

Therefore, the key novelty of our work lies in how learned structural concepts are represented, recombined, and used to guide crystal generation, rather than in the individual backbone architectures.

Q2: Many aspects are poorly formulated –(eg. Derivative reference used for DFT). What is the utility of Figure 2, appears to be of little value to the understanding

DFT evaluation. We follow the Materials Project computational settings for our DFT calculations, with implementation details provided in Appendix C.5. The reference datasets used for the final stability and novelty evaluation are consistent with those adopted by MatterGen. We will clarify the distinction between DFT calculations and the reference datasets used for evaluation to avoid confusion.

Utility of Figure 2. Figure 2 illustrates the Transformer architectures used in the VQ-VAE, composition generator, and base generative model. Its main purpose is to clarify the inputs, outputs, and conditioning mechanisms of these components, rather than to introduce novel backbone architectures. We will revise the figure caption and accompanying text to make this purpose more explicit.

AuthorsOpenReview ↗

Rebuttal 2

Q3: The most important part is to show new crystals can be discovered. They used the training set MP 20 crystals to evaluating stability, novelty etc using VSUN. What were the 1000 crystals found? Were they really new (not in terms of distance in structure but in material science terms?)? What were the material scientist view on these structures? These are not given, or even discussed. This is important in this context.

We thank the reviewer for raising this important question.

  1. Clarification of novelty evaluation. First, we clarify that novelty in our final V.S.U.N. evaluation is not measured against the training set. Instead, we use substantially larger reference databases: the Materials Project (MP, February 7, 2023 snapshot, 154,718 structures) for MP-20, and MP + Alexandria (845,997 structures) for Alex-MP-20. These evaluations follow MatterGen.
  2. Novelty beyond structural distance. To further assess the materials science significance of the generated crystals, we conduct additional analyses on all V.S.U.N.-qualified crystals with nontrivial symmetry (space group other than P1). We evaluate novelty at three levels: chemical composition (reduced formula), structural prototype, and full crystal structure. Structural prototypes are identified using Wyckoff-based AFLOW-style labels, incorporating anonymous formulas, Pearson symbols, space groups, and Wyckoff positions.

We compare the generated crystals against two reference sets: (1) the corresponding training dataset and (2) an expanded collection comprising the original evaluation references, the latest MP release used in our analysis, the full Alexandria database, and GNoME.

Novelty levelMP-20: vs. trainingMP-20: vs. expanded referencesAlex-MP-20: vs. trainingAlex-MP-20: vs. expanded references
New reduced formula58.8%2.4%27.8%5.2%
New structural prototype27.5%11.3%15.0%13.6%
New crystal structure100%47.1%100%74.3%

These results provide three important observations:

  • Exploration beyond the training distribution. Compared with their respective training sets, 58.8% / 27.8% of the generated crystals have unseen reduced formulas, while 27.5% / 15.0% exhibit previously unseen structural prototypes on MP-20 / Alex-MP-20.
  • Novelty beyond the evaluation references. Even after comparison with the expanded databases, 47.1% / 74.3% of the generated structures remain unmatched, indicating that their novelty is not merely an artifact of the original evaluation reference.
  • Novel structural prototypes. Notably, 11.3% / 13.6% of the crystals exhibit structural prototypes absent from the expanded references, while genuinely new reduced formulas are less frequent (2.4% / 5.2%). This suggests that an important contribution of our method is discovering new structural arrangements, rather than simply generating new chemical compositions.
  1. External evidence of physical plausibility. Interestingly, 52.9% (MP-20) and 25.7% (Alex-MP-20) of the analyzed crystals, although absent from both the training set and the original novelty references, match independently reported, DFT-evaluated structures in Alexandria, GNoME, or more recent MP releases. Two of these also have experimental counterparts in ICSD.

These independent matches provide encouraging evidence that our model can generate physically meaningful crystal structures beyond its training distribution. Although they do not directly validate the remaining unmatched structures, they support the effectiveness of our generation framework in exploring chemically plausible regions of crystal space.

We acknowledge that experimental synthesis and expert assessment of the newly generated candidates remain important directions for future validation. We will include these additional analyses, along with representative CIF files of crystals exhibiting previously unreported structural prototypes, in the revised appendix.

I thank the authors for their responses, and have reflected on them. I would like to retain my original score.

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

In this work, the authors have proposed a way of breaking down the latent space of crystals into concepts, and to explore the space of crystals by composing those concepts.

Critical strengths
  1. The paper has been properly proof-read and has little to no traces of AI-generated slop.

  2. The authors have proposed a way of exploring the chemical space by combining existing concepts in known crystal structures.

  3. Concepts are learned from the dataset itself, hence all are valid existing concepts and do not require any prior effort in defining or categorizing them.

  4. The authors have demonstrated how concepts capture a local view as well as represent some global properties.

  5. Table 3 shows that this approach is able to achieve a higher novelty rate than most existing models.

Critical weaknesses
  1. In Wu et al., 2024 [1], the concepts are classified into different orthogonal categories like color, shape, etc. The concepts learned by the VQ-VAE model in this work are not guaranteed to be orthogonal and might have redundant or overlapping concepts. Can the authors provide an ablation on the effect of the codebook size on the model's performance? If possible, can they also look into how the stability and novelty are affected when the concepts are clustered and one embedding from each cluster is used as the codebook?

  2. Since the concepts are learned from the dataset, the searchable space becomes dependent on the concepts present in that dataset. Section 5.2 claims 61% reconstruction of Alex-MP-20 crystals with the concepts learned from MP-20 dataset. Although this shows that concepts can overlap between datasets, this also shows that the concepts learns from one dataset are not enough to reconstruct another dataset to a significant extent. How much of the crystals from a non-overlapping dataset like Perov-5 or GNoME can be reconstructed? Can the authors also provide the reconstruction metrics for Alex-MP-20 trained on the same dataset, for comparison?

  3. In tables 3 and 4, there are two variants, base model and w/composition, but the variants are not clearly differentiated. The "condition incorporation" paragraph (line 261) talks about the w/composition variant but it is not labeled as that. Referring to fig. 2(d) would also help with clarity.

  4. The positional encoding scheme enforces an artificial ordering among the atoms. Have the authors tried other schemes like sinusoidal embedding of the x, y, z coordinates?

  5. The representation of an atom contains the lattice parameters which are global properties of the crystal. In the w/composition variant, the generated atoms will have different lattice parameters. How are such inconsistencies resolved? Are there any noticeable variation in performance when the lattice parameters are not decoded from the transformer embeddings and instead by a pool-and-decode operation?

  6. Table 3 shows the Stability of the base model as the highest 48.4 (marked in bold) whereas the Stability achieved by ADiT is even higher 69.5, and the same thing goes for Novelty where the w/comopsition variant has been marked as highest 90.0, whereas the Novelty of CDVAE is even higher at 96.5.

[1] Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty. Advances in Neural Information Processing Systems, 37:86004–86047, 2024.

Questions for authors

see weaknesses

  1. In defining a local environment, up to K neighbors are retained (12 in this case). How are the neighbors selected if there are more than K?

  2. Can the authors demonstrate a workflow of their model on some practical use cases? If some target properties (like band gap, crystal system, formation energy, etc.), or partial elemental constitution are known, how can a user generate the concepts and decode them into a crystal structure?

5 comments
AuthorsOpenReview ↗

Rebuttal 1

Q1: In Wu et al., 2024 [1], the concepts are classified into different orthogonal categories like color, shape, etc. The concepts learned by the VQ-VAE model in this work are not guaranteed to be orthogonal and might have redundant or overlapping concepts. Can the authors provide an ablation on the effect of the codebook size on the model's performance? If possible, can they also look into how the stability and novelty are affected when the concepts are clustered and one embedding from each cluster is used as the codebook?

We thank the reviewer for this valuable suggestion. Unlike Wu et al. (2024), where concepts are factorized into predefined attribute categories (e.g., color and shape), our concepts represent prototypes of local chemical environments. Similar local environments may naturally have similar embeddings, and therefore orthogonality is not an explicit design objective. Nevertheless, as the concepts are learned in an unsupervised manner, potential redundancy or overlap among them is a valid concern. We investigate this issue through the following ablation.

Experimental setup. Following the reviewer's suggestion, we apply K-means clustering to the original 10,000 codebook embeddings, reducing them to (K' \in {5000, 2000, 1000, 500}) clusters. Each cluster is represented by an existing codebook embedding closest to its centroid. We then replace each original code with the representative embedding of its corresponding cluster, while keeping all model parameters frozen. We evaluate both reconstruction accuracy and generation quality on MP-20.

#codes10,0005,0002,0001,000500
Recon. (%)83.065.642.028.715.2
Valid & Stable (%)45.740.437.735.931.4
Valid & Novel (%)82.883.884.585.384.5
V.S.U.N. (%)36.031.931.429.927.5

Results and discussion. As the codebook is reduced from 10,000 to 500 entries, reconstruction accuracy drops substantially (83.0% → 15.2%), accompanied by decreases in the valid-and-stable rate (45.7% → 31.4%) and V.S.U.N. (36.0% → 27.5%). These results suggest that even geometrically similar codebook embeddings can encode distinctions that are important for accurate reconstruction and stable crystal generation. Thus, simply merging similar embeddings does not preserve the information required by the original model.

The valid-and-novel rate remains comparable or increases slightly under codebook compression, reaching 85.3% with 1,000 codes. However, this does not translate into better overall generation quality, as stability and V.S.U.N. consistently deteriorate. This observation suggests a trade-off between preserving fine-grained structural information and exploring novel structures.

Overall, the original 10,000-code vocabulary achieves the best reconstruction accuracy, stability, and V.S.U.N. among the evaluated settings. We will include this ablation in the revised appendix.

Q2: In tables 3 and 4, there are two variants, base model and w/composition, but the variants are not clearly differentiated. The "condition incorporation" paragraph (line 261) talks about the w/composition variant but it is not labeled as that. Referring to fig. 2(d) would also help with clarity.

We thank the reviewer for this helpful suggestion. We agree that the distinction between the two variants should be made clearer. We will revise the Condition Incorporation paragraph in Section 3.3 to explicitly clarify that the base model is an unconditional generative model, whereas the w/ composition variant additionally incorporates the generated concept compositions as conditions, as illustrated in Fig. 2(d). We will also clarify this distinction in the captions of Tables 3 and 4.

AuthorsOpenReview ↗

Rebuttal 2

Q3: Since the concepts are learned from the dataset, the searchable space becomes dependent on the concepts present in that dataset. Section 5.2 claims 61% reconstruction of Alex-MP-20 crystals with the concepts learned from MP-20 dataset. Although this shows that concepts can overlap between datasets, this also shows that the concepts learns from one dataset are not enough to reconstruct another dataset to a significant extent. How much of the crystals from a non-overlapping dataset like Perov-5 or GNoME can be reconstructed? Can the authors also provide the reconstruction metrics for Alex-MP-20 trained on the same dataset, for comparison?

We thank the reviewer for this suggestion. To further investigate the transferability of learned concepts, we conduct additional experiments on both in-distribution and out-of-distribution datasets.

  1. Reconstruction on Alex-MP-20 using its own learned concepts. We retrain the VQ-VAE on Alex-MP-20 to extract a new concept vocabulary. The resulting reconstruction rate on the Alex-MP-20 validation set reaches 94.4%. This confirms that the proposed VQ-VAE can effectively learn dataset-specific structural patterns.
  2. Cross-dataset reconstruction on Perov-5 and GNoME. We further evaluate the concept vocabularies learned from MP-20 and Alex-MP-20 on two additional datasets, Perov-5 and GNoME. For GNoME, we filter out crystals containing more than 20 atoms per unit cell or containing Tc, Pm, or elements with atomic numbers of 84 or higher, following the constraints of Alex-MP-20.
Reconstruction Ratio (%)Perov-5 #18,928GNoME #198,655
MP-20 concepts53.356.5
Alex-MP-20 concepts57.981.6

The concepts learned from both datasets retain meaningful reconstruction capability on previously unseen crystal datasets. In particular, the Alex-MP-20 concepts achieve 81.6% reconstruction on GNoME, suggesting substantial transferability of the learned structural vocabulary. The Perov-5 reconstruction was lower, because a significant portion of the materials are not thermodynamically stable as stated in CDVAE, far from the stable / meta-stable distribution in MP-20 / Alex-MP-20.

These results clarify that our learned concepts are transferable but not universally exhaustive. A vocabulary learned from one dataset can capture structural patterns in other datasets, while expanding the diversity of training structures can improve its coverage. We will incorporate these additional results and discuss this limitation in the revised manuscript.

Q4: The positional encoding scheme enforces an artificial ordering among the atoms. Have the authors tried other schemes like sinusoidal embedding of the x, y, z coordinates?

We thank the reviewer for this suggestion. While coordinate-based positional encoding is a reasonable alternative, it is not directly applicable across all three components of our framework:

  1. VQ-VAE. The encoder processes each local atomic environment independently, without modeling an ordering across atoms in the crystal. The decoder takes concept embeddings as input and reconstructs atomic coordinates, which are therefore unavailable as input to its positional encoding.
  2. Composition generator. This module generates concept embeddings rather than atomic structures and thus has no access to atomic coordinates.
  3. Base generative model. During diffusion sampling, atomic coordinates are initialized from Gaussian noise. Therefore, coordinate-based positional embeddings would initially be derived from noisy coordinates rather than meaningful structural positions.

We did try sinusoidal embedding of x,y,z at the beginning of the project, but this approach did not provide good results. Instead, we adopt a consistent atom-ordering convention based on atomic species and coordinates. The resulting index-based positional embeddings provide a shared positional reference across the three components: the VQ-VAE decoder reconstructs atoms in this order, the composition generator learns concept sequences following the same convention, and the composition-based generative model uses the corresponding concept ordering for conditioning.

Our objective is therefore not to encode precise spatial positions through positional embeddings, but to maintain consistent atom indexing across concept extraction, composition, and crystal generation.

AuthorsOpenReview ↗

Rebuttal 3

Q5: The representation of an atom contains the lattice parameters which are global properties of the crystal. In the w/composition variant, the generated atoms will have different lattice parameters. How are such inconsistencies resolved? Are there any noticeable variation in performance when the lattice parameters are not decoded from the transformer embeddings and instead by a pool-and-decode operation?

We thank the reviewer for raising this concern. Although lattice parameters are included in the input atom representations, a learned concept does not correspond to a fixed lattice. In fact, each used concept is shared by an average of 45 different training crystals, and 89% of them appear in at least 10 different crystals. This suggests that the learned concepts capture recurring local structural patterns rather than being tied to a specific crystal lattice.

Lattice prediction ablation. Following the reviewer's suggestion, we compare two approaches for predicting lattice parameters of validation crystals from the learned concepts: (1) mean pooling of concept embeddings followed by an MLP, and (2) jointly processing all concepts using our original Transformer decoder.

Prediction errorLengthAngleVolume
mean-pooled concepts + MLP11.9%10.6°17.3%
concepts + Transformer decoder4.5%4.5°5.4%

The Transformer decoder substantially outperforms mean pooling followed by an MLP across all three metrics. These results indicate that interactions among the concepts are important for accurately predicting the global lattice, and that simple aggregation of concept embeddings is insufficient to capture these relationships.

Lattice consistency during composition. Importantly, concept composition does not directly combine independently predicted lattice parameters. Instead, the composed concepts are jointly processed to predict a single lattice for the entire crystal. The composition-conditioned generative model follows the same principle, using all conditioning concepts to guide the generation of a globally consistent crystal structure. Therefore, combining concepts does not inherently introduce conflicting lattice parameters at the atom level.

Q6: Table 3 shows the Stability of the base model as the highest 48.4 (marked in bold) whereas the Stability achieved by ADiT is even higher 69.5, and the same thing goes for Novelty where the w/ composition variant has been marked as highest 90.0, whereas the Novelty of CDVAE is even higher at 96.5.

We thank the reviewer for pointing out these formatting errors. We will correct the bold highlighting in Table 3 to accurately reflect the best-performing methods for each metric.

Q7: In defining a local environment, up to K neighbors are retained (12 in this case). How are the neighbors selected if there are more than K?

We thank the reviewer for this question. Following the default setting of MinimumDistanceNN in pymatgen, we consider atoms within a distance of (1.1$d_{\min})asneighbors,where() as neighbors, where (d_{\min}$) is the distance from the central atom to its nearest neighbor. If more than (K=12) neighbors satisfy this criterion, we retain the 12 closest ones.

In practice, such truncation is rarely needed. The median number of neighbors is 4 in MP-20 and 6 in Alex-MP-20, and only 1.6% and 1.9% of atoms, respectively, have more than 12 neighbors. Therefore, truncation affects only a small fraction of local environments, while preserving the nearest neighbors of each central atom. We will clarify this selection procedure in the revised manuscript.

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

The paper proposes treating de novo crystal generation as recombining learned local building blocks. The proposed method trains a VQ-VAE on MP-20 and learns a vocabulary of 10,000 discrete "concepts." Each concept is one recurring local atomic environment. Its Transformer encoder sees each atom's neighborhood, while its decoder rebuilds the whole crystal. A diffusion model samples new combinations of these concepts. It is fine-tuned on samples whose decoded crystals pass validity, stability, uniqueness and novelty (V.S.U.N.) checks against the training set only. The sampled combinations then steer a diffusion Transformer that generates the crystals. The authors show the concepts are interpretable. Environments assigned to the same concept share geometry, and chemically related elements fall into the same concept. The authors conduct experiments on MP20 and Alex-MP-20 datasets. On MP20, the proposed method outperforms 10 baseline methods. On the Alex-MP-20 dataset, the proposed method outperforms MatterGen.

Critical strengths
  1. The authors demonstrate that the concepts learned from the dataset are interpretable, locally and globally. The concept is also transferable to another dataset.
  2. The proposed method outperforms the baseline methods on the MP20 dataset.
Critical weaknesses
  1. The authors do not conduct an ablation experiment to isolate the influence of the concepts and the refinement on the improvement. As shown in Table 5, on the MP20 dataset, refinement contributes to most of the improvement, i.e., baseline SUN: 27.8, w/ concept w/o refinement: 28.3, w/ concept w/ refinement: 36.4.

  2. The proposed method sacrifices stability of the generated structures.

Questions for authors

Following on weakness 1, one question to answer is that, if we simply add refinement to a baseline method like MatterSim, does that change the conclusion?

1 comment

Q1: The authors do not conduct an ablation experiment to isolate the influence of the concepts and the refinement on the improvement. As shown in Table 5, on the MP20 dataset, refinement contributes to most of the improvement, i.e., baseline SUN: 27.8, w/ concept w/o refinement: 28.3, w/ concept w/ refinement: 36.4. Following on weakness 1, one question to answer is that, if we simply add refinement to a baseline method like MatterSim, does that change the conclusion?

We would like to clarify the distinct roles of concept composition and refinement in our framework.

  1. Refinement improves the overall V.S.U.N. score, as expected by design. As explained in Section 3.2 (Validating and Refining the Generator), refinement is explicitly designed to improve V.S.U.N. by using the training set as a proxy for the reference set in the final evaluation. Importantly, this process relies only on training data, without accessing any external evaluation references.
  2. Concept composition plays a key role in improving novelty. While refinement contributes substantially to the overall V.S.U.N. score, the results in Table 5 indicate that concept composition is important for promoting novel structural exploration. Therefore, the two components serve complementary purposes: concept composition facilitates the exploration of novel structures, while refinement improves the quality of the generated compositions.
  3. We have already evaluated refinement on a baseline model, DiffCSP, as reported in Table 5. Although refinement improves its overall V.S.U.N. score (25.6→34.4), the improvement in novelty is limited (83.6→84.3). In contrast, our compositional framework achieves a substantially greater novelty improvement (78.0→88.1). This comparison further supports our conclusion that concept composition, rather than refinement alone, is crucial for effective exploration of novel crystal structures.

Q2: The proposed method sacrifices stability of the generated structures.

We acknowledge that concept composition introduces a trade-off between stability and novelty.

  • On MP-20, composition without refinement reduces stability by approximately 5 percentage points, which is further narrowed to only 1.5 percentage points after refinement (Table 5). Meanwhile, novelty improves by 12.0 percentage points, resulting in a substantial improvement in V.S.U.N. from 23.3% to 35.7%.
  • On Alex-MP-20, the stability reduction is more pronounced (69.0% → 59.1%).This is because the extracted codes from MP-20 cannot precisely cover all structural patterns from Alex-MP-20. However, this reduction is accompanied by a much larger improvement in novelty (61.8% → 85.9%), leading to an increase in V.S.U.N. from 28.8% to 43.7%. Moreover, our method outperforms MatterGen on the joint V.S.U.N. metric (43.7% vs. 37.4%), despite its lower individual stability.

These results suggest that concept composition enables broader exploration of novel crystal structures, at the cost of some stability. Importantly, state-of-the-art crystal generation methods are evaluated using the joint V.S.U.N. metric rather than individual metrics in isolation, as the critical objective is to generate crystals that simultaneously satisfy validity, stability, uniqueness, and novelty. From this perspective, our method achieves a substantial improvement in V.S.U.N. on both datasets, demonstrating its stronger ability to balance these competing objectives and generate novel yet physically plausible crystal structures.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.