- 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 to a target structure 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 .
Critical strengths
-
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.
-
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 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
-
On the same '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.
-
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?
-
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?
-
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?
-
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.