acceptodds
Withdrawn from ICLR 2027

Tego: Symmetry‑Preserving Local Editing for Targeted Inorganic Materials Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The inverse design of inorganic crystalline materials is generally regarded as an end-to-end, property-to-structure generation task, giving rise to a series of *ab initio* methods such as diffusion models. However, such methods overlook an inherent contradiction at the information level: properties can be viewed as a lossy compressed representation of structures, and the property-to-structure mapping follows a one-to-many non-functional relation. Reconstructing complete structures merely from properties constraints essentially attempts to recover high-dimensional full structural information from low-dimensional lossy representations, which inherently suffers from the theoretical bound of information asymmetry and constrains the structural stability and property accuracy of outputs from ab initio methods. To address these challenges, we propose the **Tiny-Edit-Good-Outcome (Tego)** framework. Tego initiates inference from an existing stable structure, thereby performing localized inverse editing to approximate target material properties, circumventing the complexity of full structure reconstruction in *ab initio* methods. Evaluations demonstrate that Tego is capable of generating undiscovered, thermodynamically stable crystalline candidates, while achieving higher property achievement rate and computational efficiency across material inverse design scenarios.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
crystal inverse design, materials informatics, large language model, local structure editing, material property optimization, inorganic crystals

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TCVtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents Tego, a framework for editing an existing stable crystal structure S0S_0 to a target structure S’S’ with a desired material property. The authors represent each crystal using Wyck-SEQ, a text representation that groups atoms into occupied Wyckoff orbits. Given this representation with the current and target property values, a large language model(LLM) outputs an ACTION that substitutes the atomic species across one entire orbit. The edited structure is then relaxed and evaluated with machine learning interatomic potentials. For the magnetic density, the authors fine-tune the LLM with LoRA, to learn which substitutions produce a given property change. The authors report that the base model without fine-tuning reaches S.U.N. rates comparable to those of de novo generators on MP20, and the fine-tuned model achieves a higher success rate than MatterGen under a magnetic density target of ≥0.2 μB/A˚3\geq 0.2\,\mu_{\mathrm{B}}/\mathrm{\mathring{A}}^{3}.

Critical strengths
  1. The method is straightforward, and each proposed edit is easy to interpret. Applying substitutions to complete symmetry orbits is a sensible way to avoid breaking the symmetry of the given structure during the discrete editing step.

  2. Beyond surrogate metrics, the authors report DFT validation of selected MP20 candidates and experimental synthesis and magnetic measurements in the property-guided editing experiments.

Critical weaknesses

Novelty and benefits of Wyck-SEQ. Wyckoff-based representations already appear in PLaID++ [1] and CrystalFormer [2]. It is unclear what Wyck-SEQ adds beyond organizing existing crystallographic information into occupied-orbit blocks; the paper itself notes that CIF can also encode crystallographic symmetry (Section 3.1). The "Wyckoff Abl." in Table 1 does not resolve this question because, if I understand correctly, it uses the same LLM to generate complete structures in Wyck-SEQ format, whereas Tego receives an existing structure in Wyck-SEQ and outputs a substitution action. A more direct test would be to finetune the same model on the same structures given either in Wyck-SEQ or in a CIF-based format with equivalent orbit information, and evaluate both under the same conditions.

Effectiveness of the learned edit policy. Moreover, the LLM only selects an occupied orbit and a replacement element. It does not directly generate coordinates or lattice parameters, which are inherited from S0S_0 and adjusted through relaxation. This distinction should be reflected in the comparisons in Table 2 and the broader conclusions. It is unclear whether an 8B LLM is more effective than simpler policies for selecting substitutions. A natural baseline is to randomly select an eligible orbit and replace all its atoms with a randomly chosen element from the same valence group in Table 4, followed by the same relaxation and evaluation.

Questionable inference cost. Table 2 reports 67.68 s per structure for MatterGen and 2.19–2.29 s for Tego, but it is unclear how the inference time of MatterGen was measured. It is also unclear whether Tego's advantage holds for batched generation or for search under matched budgets. Please specify the sampling settings, separate generation from post-processing (relaxation and property prediction), and state whether the values are single structure latencies or batch runtimes divided by the number of structures. Reporting peak GPU memory at the stated batch sizes would also help. Without these details, I cannot assess the claimed computational advantage.

Reliability of the stability evaluation. Several baseline SUN values in Table 1 are taken from the DFT-based results reproduced in MatterGen’s repository [3], whereas Tego is evaluated with MatterSim relaxation and energies. These are not evaluations under the same conditions. Because surrogate errors matter most for candidates near the 0.1 eV/atom stability threshold, a common evaluation protocol is needed to support the S.U.N. comparison.

Questions for authors
  1. On the same S0S_0's, could you compare Tego with (i) random same-valence substitution and (ii) a small model (e.g., an MLP or a small transformer with logits over {orbit, element} actions) trained on the same pairs to select substitution actions? Please use the same allowed elements, relaxation procedure, and evaluation.

  2. How does performance change when the same editing task uses a CIF-based input, with orbit information given as supplementary text, instead of Wyck-SEQ?

  3. For Table 1, which reference structure set defines the convex hull and novelty for Tego, PLaID++, and Chemeleon2? Can the main comparison be repeated under a shared protocol?

  4. For Table 2, were all 247 MP20 outputs evaluated with DFT, or only a screened subset? How are invalid actions, failed relaxations, and unevaluated candidates counted in the denominators? The earlier MatterGen preprint [4] reports the 47 hit result after MLFF filtering and subsampling before DFT property evaluation. The published paper reports up to 18 qualifying SUN structures with a budget of 180 DFT property calculations [5]. How are the screening and validation budgets matched between Tego and the cited MatterGen experiment?

  5. Does changing the target value on the same seed lead to appropriate changes in the proposed action and final magnetic density, including for targets below the current value? Section 4.3 retains pairs with high or strongly increased magnetic density, and when multiple candidates are generated, the selection rule favors the highest value rather than the one closest to the target. How do the authors distinguish following the requested target from simply increasing magnetic density? In addition, Table 5 specifies an absolute sum of magnetic moments, and CHGNet predicts unsigned site moments. How does this quantity relate to the net magnetization when local moments are antiparallel (e.g., in Gd–Fe compounds), and how was this checked for the reported candidates?

Other comments

  • Citations and related work : please use \citep for parenthetical citations (e.g., "identified Merchant et al. (2023)") and merge duplicate bibliography entries (e.g., Gebauer et al. 2019a/b). P-G-SchNet was introduced by Xie et al. (2022), not Gebauer et al. (2019b). DiffCSP++ constrains both the lattice and the Wyckoff positions, using templates from known structures. DiffCSP defines CSP as predicting the structure for a given composition, and the pipeline. The manuscript describes the design process as differentiable, but it does not show an end-to-end gradient path from the property objective through the discrete action selection.
  • Internal consistency problems : the relaxation settings differ between Table 3 (0.01 eV/Å, 100 steps) and Appendix A.3 (0.05 eV/Å, 300 steps) for dataset construction. And Section 4.3 says one Wyckoff block is selected at random per structure, whereas Appendix A.2 and Table 3 say all blocks are used.

References

[1] A. Xu et al., "PLaID++: A Preference Aligned Language Model for Targeted Inorganic Materials Design," ICML 2026.

[2] Z. Cao, X. Luo, J. Lv, L. Wang, "Space group informed transformer for crystalline materials generation," Science Bulletin 70(21):3522–3533, 2025.

[3] MatterGen repository, https://github.com/microsoft/mattergen (README benchmark table and MODEL_CARD.md).

[4] C. Zeni et al., “MatterGen: a generative model for inorganic materials design,” arXiv:2312.03687v2, 2024.

[5] C. Zeni et al., "A generative model for inorganic materials design," Nature 639:624–632, 2025.

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

This paper uses an LLM to suggest element substitutions on Wyckoff groups, then relaxes and evaluates the edited structures. The idea is easy to follow, but I’m not convinced that an LLM is needed here.

Critical strengths

See Questions

Critical weaknesses

See Questions

Questions for authors

This paper uses an LLM to suggest element substitutions on Wyckoff groups, then relaxes and evaluates the edited structures. The idea is easy to follow, but I’m not convinced that an LLM is needed here.

  1. My main concern is the small search space. Each step only chooses a Wyckoff group and a replacement element. With the same-valence restriction used to build the training data, why not enumerate the possible substitutions? Pls compare with exhaustive search, random substitution, and simple chemical rules, using the same starting structures and evaluation budget. Otherwise, it’s hard to tell what the LLM adds.

  2. The comparison with de novo models isn’t on equal terms. Tego starts from low-energy structures for the SUN test and already-magnetic structures for the targeted-design test. That’s a strong prior.

  3. I also don’t think the training-data energy filter is well justified. Replacing an element changes the composition, so directly subtracting the original and edited total energies per atom doesn’t tell us whether the substitution is favorable. Elemental reference energies or chemical potentials need to be considered.

  4. The property-control claim seems broader than the experiments support. The main test is to increase magnetic density beyond one threshold, and candidate selection simply favors the highest value. Does the model actually use the requested target, or has it mostly learned to suggest substitutions that increase magnetism?

Reviewer_g7nMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on the problem of crystal structure design with target properties, in other words, inverse design of crystalline materials with target properties. Most existing works on this topic use generative models that generate the structure conditioned on a property value. The paper argues that a property value is a compressed summary that many different structures share, and states that this constrains the performance of such methods. The paper proposes Tego, which instead starts from a known stable crystal and changes it a little every step.

Critical strengths

The paper proposes a novel approach towards inverse design of crystalline materials. Experiments demonstrate some unique characteristics of the proposed method, compared to existing generative modeling-based inverse design methods.

Critical weaknesses

I am not fully convinced of the primary argument of the paper, that the limitation of existing crystal structure design methods (essentially, conditional generative models) comes from the fact that the mapping from property to structure is a one-to-many mapping. Typically, one such method works by using a generative model conditioned on a target property to generate the structure. This process can simply be repeated to obtain multiple structures that, ideally, all have the target property, thanks to the inherent randomness of common learning-based generative models.

It is also unclear to me how the edit-based framework of the proposed method can effectively tackle the inherent challenge of learning this one-to-many mapping from property to structure. It might make more sense if the proposed method's unique capability/functionality that is made possible thanks to this edit-based framework could be explored.

Questions for authors

Could the authors reiterate how their proposed method can more effectively tackle the claimed challenges in inverse design of crystalline materials?

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

Related papers

Discussion (0)

Sign in to comment.