Pith. sign in

REVIEW 2 major objections 3 minor 34 references

ED-CSP: Crystal Structure Prediction from Electron Diffraction

T0 review · 2 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read ED-CSP predicts a crystal's lattice and fractional atomic coordinates from a known composition and multiple unindexed electron-diffraction spot sets, reaching 57.49% MR@5 on 2,075 held-out CHILI-100K materials with CHILI-only training and…

desk verdict A new task and a strong simulated benchmark for ED-to-structure generation, but the headline registry-pretraining numbers carry an unresolved leakage risk that deserves a clean re-run. read the letter →

arxiv 2608.06448 v2 pith:DBXSPES2 submitted 2026-08-06 cs.LG cs.AI

Germain Poloudenny , Yaël Frégier , Arnaud Demortière This is my paper · ORCID
classification cs.LGcs.AI
keywords crystalstructurepredictionelectrondiffractiongenerativemodelperiodicflowmulti-viewrepresentationlearninglatticeparametersfractionalcoordinatesmatching
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that crystal structure prediction can be driven by sparse, unindexed electron-diffraction (ED) spot lists rather than by X-ray powder profiles, indexed reflections, or library lookup. For a known chemical composition and atom count, ED-CSP generates the full periodic structure — lattice parameters and fractional atomic coordinates — by encoding each detector-plane view as a relational set, pooling across views, and decoding with a periodic flow generator. On 2,075 held-out CHILI-100K materials, CHILI-only training reaches 57.49% MR@5 and beats the PXRD-conditioned PXRDGen baseline (52.92%); warm-starting from a one-million-structure precursor raises MR@5 to 66.27%. The model still reaches 53.52% MR@5 on 1,024 queries with no same-formula candidate in the retrieval library, and swapping in diffraction from a non-isomorphic same-formula donor lowers MR@5 by 22.09 percentage points. These numbers matter because electron diffraction is often the only viable signal for nanocrystalline or beam-sensitive materials, so a generative route from raw spot sets to atomic structure would let more real materials be solved automatically.

What carries the argument

The load-bearing machinery is the pairing of a permutation-invariant sparse-ED encoder with a periodic flow generator. Each detector-plane spot list is processed by a shared relational graph-attention encoder (a PE-AG-GMoE-style backbone) that consumes only two-dimensional spot coordinates and log-intensities; a mean–max pooling head merges the per-view representations so the multi-view input can be unordered. The pooled representation conditions a six-layer CSPNet-style periodic flow decoder that outputs vector fields for the lattice and periodic atomic coordinates, trained with a weighted mean-squared-error loss ($L = L_{\text{lat}} + 100L_{\text{coord}}$). Contrastive ED–structure pretraining initializes the encoder, and the registry-1M experiment warm-starts the entire model before CHILI fine-tuning. Because PXRDGen shares the same generator architecture, the comparison isolates the ED conditioning modality.

What would settle it

Run the same strict structure matcher used to build ED-CS against the registry-1M precursor and count how many of the 2,075 held-out queries (especially the 1,024 'absent-formula' ones) have a matcher-level near-duplicate in that precursor. If a sizable share do, the 66.27% and 53.52% figures would be explained substantially by near-copy retrieval; if none do, the generative interpretation survives.

Watch

Extended reading notes

Core claim

ED-CSP establishes that composition-conditioned, generative crystal structure prediction is possible from unindexed multi-view ED spot lists. Given a composition and $K$ detector-plane spot sets $S_{1:K}$, the model models $p_\theta(L,F \mid A,S_{1:K})$ and jointly predicts the lattice $L$ and fractional coordinates $F$ without Miller indices, zone axes, or crystallographic labels. The benchmark on 2,075 held-out CHILI-100K materials reports 42.12% MR@1 and 57.49% MR@5 under CHILI-only training, versus 34.02% and 52.92% for PXRDGen; full-model warm-start from the registry-1M precursor gives 51.66% MR@1 and 66.27% MR@5. Signal-use ablations show MR@1 falls from 51.66% to 17.35% when ED spots are removed, and donor-swap ED from a non-isomorphic same-formula structure lowers mean MR@5 by 22.09 percentage points. Coverage stratification separates generation from retrieval: with an exact-formula analogue available, library lookup reaches 85.82% MR@5 versus 78.69% for the registry-initialized model, while for the 1,024 queries with no same-formula candidate, ED-CSP retains 53.52% MR@5 and lookup has no candidate.

Load-bearing premise

The 66.27% registry-1M result assumes that the earlier one-million-structure precursor, assembled before the final CHILI-exclusion filtering, contains no structural near-duplicates of the 2,075 held-out test materials; the authors' protection is a post-hoc sensitivity analysis on a subset rather than a guaranteed exclusion.

Editorial extensions

If this is right

  • Electron diffraction becomes a usable conditioning signal for full structure generation, not just for classifying spot patterns or retrieving library entries.
  • Scaling the training corpus helps: switching from CHILI-only initialization to a one-million-structure warm start raises MR@1 by about 9.5 points and MR@5 by about 8.8 points.
  • Generation and exact-formula lookup can be combined: using retrieval when a same-formula analogue exists and ED-CSP otherwise yields 69.88% MR@5, a paired gain of 3.61 points over ED-CSP alone.
  • Predicted candidates remain improvable by energy-based post-processing: relaxing and re-ranking with a learned interatomic potential raises top-1 recovery by about 13 points, with the relaxed-pool ceiling at 61.69% MR@5 pointing to room for ED-aware refinement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct experimental-transfer test is the natural next step: fine-tuning on even a few hundred measured 3D-ED structures with matched detector conditions would show whether the simulated spot statistics survive contact with real detector calibration, background, and missing reflections.
  • The view-count trend (Top-5 encoder retrieval rising from 7.81% at 10 views to 13.38% at 20) suggests the 10-view limit used in generator training under-samples the available orientation information, so larger view sets may yield further gains.
  • The composition-only reference at 50.94% MR@5 implies that a sizable share of recoveries may come from learned chemical priors; a paired, identical-layout comparison against composition-only generation would calibrate how much ED itself contributes on the full split.
  • The same-formula donor-swap result indicates the model has learned some orientation-dependent geometry; an extension would be to vary simulated thickness and see whether the predictions track dynamical scattering effects or only kinematic spot geometry.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper introduces ED-CSP, a generative model that predicts lattice parameters and fractional atomic coordinates from a known composition, atom count, and multiple detector-plane electron diffraction (ED) spot lists. The authors also construct ED-CS, a 4.85-million-structure dataset with simulated multi-view ED, deduplicated across seven materials repositories and filtered to exclude CHILI-100K overlaps. On 2,075 held-out CHILI-100K materials, CHILI-only ED-CSP achieves 57.49% MR@5 versus 52.92% for PXRDGen, and warm-starting from a one-million-structure precursor raises MR@5 to 66.27%. On 1,024 compositions absent from the train/validation retrieval library, the registry-initialized model achieves 53.52% MR@5, which the paper presents as evidence of true generative capability beyond exact-formula retrieval. The evaluation also includes input interventions, coverage-stratified retrieval controls, and comparisons to indexed-reflection reconstruction.

Significance. The evaluation design is a clear strength: it uses a fixed external benchmark, the training objective (flow MSE on lattice and coordinates) is separate from the evaluation metric (StructureMatcher MR@5), and the paper reports bootstrap confidence intervals, paired significance tests, input interventions, and coverage-stratified retrieval controls. The public release of code, frozen protocols, provenance certificates, and a registry checkpoint supports reproducibility. If the registry-1M overlap concern is resolved, the contribution is significant: ED-CSP is, to my knowledge, the first generative crystal structure prediction method from unindexed sparse multi-view ED spot lists, and ED-CS is a large, carefully documented simulated dataset that can serve as a benchmark for future work.

major comments (2)
  1. [Datasets and ED Simulation; Results, Registry Scaling and Signal Use] The registry-1M warm-start pool is not certified disjoint from the CHILI-100K test queries. The paper states in Datasets and ED Simulation that the completed one-million-structure precursor predates the final ED-CS filtering, and the headline 66.27% MR@5 and the 53.52% absent-formula MR@5 are both produced by the registry-initialized model. The reported post-hoc sensitivity analysis excludes 888 queries with a strict StructureMatcher near-duplicate and is reported only for the full split; it does not cover the 1,024-query absent-formula subset that is the basis of the generative-capability claim, and it does not state against which pool the near-duplicate matches were found. To support the scaling and out-of-library claims, the authors should either certify the registry pool disjoint from the 2,075 test structures, or report a near-duplicate analysis for both the absent-formula subset and the full split against the registry-1M pool, or retrain the registry model on a certified-disjoint pool and report the revised numbers.
  2. [Results, Registry Scaling and Signal Use] The sensitivity analysis is not described precisely enough to be reproducible. The sentence 'excluding 888 queries with a strict StructureMatcher near-duplicate' does not specify the reference pool (registry-1M, ED-CS, or the CHILI train/validation library), the direction of matching, or how many of the 1,024 absent-formula queries are among the 888. Since the reported 38.84%/53.92% figures are the only leakage control for the registry-initialized model, this ambiguity is load-bearing for the paper's main scaling result.
minor comments (3)
  1. [Results, Registry Scaling and Signal Use] The sentence beginning 'WeretainthereferenceCHILI-100Ksplitforcomparability' appears to have a typo; it should read 'We retain' or 'We retrain' the reference CHILI-100K split, and the intended meaning should be clarified.
  2. [Table 2] In the 'Zero ED spots' row, only MR@1 is reported, while MR@5 is a dash. Clarifying why MR@5 is not reported for this ablation would make the table easier to interpret.
  3. [Datasets and ED Simulation] The throughput-oriented online policy that extends a case to at most 100 views uses a 'break-even threshold' that is described only in the supplementary material; a one-sentence summary in the main text would help readers assess the simulation cost and dataset construction choices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation is against an external CHILI-100K held-out split, the training objective and evaluation metric are distinct, and no fitted parameter or self-citation is load-bearing.

full rationale

ED-CSP's claimed derivation is self-contained. The model is trained to minimize a weighted flow MSE on lattice parameters and fractional coordinates (L = L_lat + 100 L_coord) conditioned on composition and ED spot lists, while evaluation uses StructureMatcher MR@5 on a fixed external CHILI-100K held-out split. These are distinct, so prediction performance is not equivalent to the training objective by construction. The CHILI-only run uses 16,611 training structures and evaluates on 2,075 held-out queries; no test material enters that training run. The ED-CS corpus is explicitly filtered to exclude CHILI-100K matches under a strict StructureMatcher certificate, and the registry-1M precursor is reported separately with the caveat that it predates the final filtering. That caveat is an acknowledged data-provenance limitation rather than a circular step: the paper's post-hoc sensitivity analysis on the full split shows the ranking persists after excluding near-duplicates, and the core comparison is against external baselines (PXRDGen, deCIFer, XRDSol) under the same split. The input-intervention experiments (zero ED spots, donor ED, corruption) test signal use rather than assuming it. No self-citation chain is used to justify the method's validity; architecture references are to external prior work. Therefore no circular step is exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical ML benchmark, so the ledger lists configuration choices and domain assumptions rather than fitted physical constants. No invented physical entities are introduced.

free parameters (3)
  • ED simulation parameters = 300 keV, 20 nm, 2.00 Å^-1 cutoff, 10 random orientations
    Chosen to match EDiffCrystals/py4DSTEM benchmarks; these set the input distribution and are not optimized on the held-out set.
  • Number of ED views and spots per view = K=10, 16 strongest spots
    Configuration choice; the paper's own diagnostic shows retrieval improves at K=20, so K=10 is a capacity/input choice, not an optimized constant.
  • Coordinate loss weight = 100
    Training uses L = L_lat + 100*L_coord; the weight is hand-set and no tuning curve is reported.
assumptions (5)
  • domain assumption Simulated py4DSTEM ED patterns (300 keV, 20 nm thickness, 2.00 Å^-1 cutoff) are a sufficient proxy for the conditioning signal.
    All training and evaluation use simulated ED; experimental transfer is explicitly deferred as the principal limitation.
  • domain assumption Known chemical composition and atom count are available at inference.
    The problem formulation fixes atom types and count, a strong input not available in all experimental workflows.
  • domain assumption pymatgen StructureMatcher equivalence is the correct notion of structural match.
    Every MR number depends on this matcher's strictness and tolerance settings, which are not specified numerically in the main text.
  • domain assumption The CHILI-100K split provides correct ground-truth structures.
    The held-out labels are taken from an experimental database; no independent verification is performed in this paper.
  • domain assumption Flow-matching MSE on lattice and coordinates is a sufficient training proxy for match-rate evaluation.
    The generator is trained with L = L_lat + 100*L_coord, while evaluation rewards exact structural matches under StructureMatcher.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ED-CSP: Crystal Structure Prediction from Electron Diffraction." pith.science (2026). https://pith.science/paper/DBXSPES2

@misc{pith2026260806448,
  author       = {Pith},
  title        = {Pith review of: ED-CSP: Crystal Structure Prediction from Electron Diffraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBXSPES2}},
  note         = {Machine review of arXiv:2608.06448}
}
read the original abstract

Recovering a periodic 3D crystal structure from sparse, unindexed electron diffraction (ED) observations is a challenging generative inverse problem. Existing ED-based learning methods mainly predict crystallographic labels, reconstruct structures from indexed reflections, or retrieve candidates from finite structure libraries. Here, we introduce ED-CSP, a machine learning framework that predicts crystal structures from chemical composition, atom count, and multiple detector-plane ED spot sets. ED-CSP combines a relational set encoder, permutation-invariant multi-view aggregation, and a periodic flow generator to jointly predict lattice parameters and fractional atomic coordinates. To train the model, we construct ED-CS, a dataset of 4.85 million simulated multi-view ED crystal structures, deduplicated across seven materials repositories and filtered to exclude CHILI-100K overlaps. On 2,075 held-out CHILI-100K materials, ED-CSP trained only on CHILI achieves a structural match rate of 57.49% MR@5, outperforming PXRDGen (52.92%), a state-of-the-art crystal structure prediction model conditioned on powder X-ray diffraction. Scaling training data further improves performance: initializing from a one-million-structure precursor raises MR@5 to 66.27%. On 1,024 compositions absent from the training retrieval library, the model still achieves 53.52% MR@5, demonstrating true generative capability beyond exact-formula retrieval. Replacing target ED observations with diffraction from non-isomorphic structures of identical composition decreases MR@5 by 22.09 percentage points, confirming that predictions depend on the input diffraction patterns rather than composition alone. ED-CSP and ED-CS establish a benchmark for generative crystal structure prediction from sparse ED observations and provide a foundation for future transfer to experimental data.

Figures

Figures reproduced from arXiv: 2608.06448 by the authors.

Figure 1
Figure 1. Diffraction observation regimes. ED-CSP con￾ditions on discrete sparse ED views from multiple orienta￾tions, whereas PXRD aggregates randomly oriented crystal￾lites into a one-dimensional radial profile. Contributions. This paper makes the following focused contributions: • A formulation of composition-conditioned CSP from unindexed sparse multi-view ED, instantiated by ED-CSP. • Controlled ED-input interventions an… view at source ↗
Figure 2
Figure 2. ED-CSP pretraining, inference, and optional post-processing. Contrastive pretraining initializes the ED encoder before joint generator training; panel (b) shows inference with the resulting model. The optional uMLIP branch receives generated candidates but neither ED observations nor ground truth; structure matching is evaluation only. transferred ED encoder remains trainable and the structure encoder is discarded. … view at source ↗
Figure 3
Figure 3. Effect of pretraining scale and exact-formula retrieval availability. (a) Sequential ED-to-structure stages: each nested-pool checkpoint resumes its converged predecessor, so the points are not independent fits. (b) Registry-initialized ED-CSP recovery with and without an exact-formula train/validation candidate; exact-formula lookup has no candidate in the absent stratum. Error bars are query-bootstrap 95% CIs. We … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Train-time Kmodel ablation. Single-seed check￾points; error bars are exact binomial 95% CIs over 1,024 queries and exclude training-run variation. Post-Generation Relaxation We test whether a target-free interatomic potential can sta￾bilize and rank a frozen five-candi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    , title =

    Jain, Anubhav and Ong, Shyue Ping and Hautier, Geoffroy and Chen, Wei and Richards, William Davidson and Dacek, Stephen and Cholia, Shreyas and Gunter, Dan and Skinner, David and Ceder, Gerbrand and Persson, Kristin A. , title =. APL Materials , volume =. 2013 , doi =

    2013

  2. [2]

    The Journal of Chemical Physics , volume =

    Batatia, Ilyes and others , title =. The Journal of Chemical Physics , volume =. 2025 , doi =

    work page 2025

  3. [3]

    and Dzamba, Misko and Gao, Meng and Rizvi, Ammar and Zitnick, C

    Barroso-Luque, Luis and Shuaibi, Muhammed and Fu, Xiang and Wood, Brandon M. and Dzamba, Misko and Gao, Meng and Rizvi, Ammar and Zitnick, C. Lawrence and Ulissi, Zachary W. , title =. arXiv preprint arXiv:2410.12771 , year =

    arXiv

  4. [4]

    and Barroso-Luque, Luis and Levine, Daniel S

    Fu, Xiang and Wood, Brandon M. and Barroso-Luque, Luis and Levine, Daniel S. and Gao, Meng and Dzamba, Misko and Zitnick, C. Lawrence , title =. arXiv preprint arXiv:2502.12147 , year =

    arXiv

  5. [5]

    Crystallography Open Database (

    Gra. Crystallography Open Database (. Nucleic Acids Research , volume =. 2012 , doi =

    work page 2012

  6. [6]

    Schmidt, Jonathan and Cerqueira, Tiago F. T. and Romero, Aldo H. and Loew, Antoine and J. Improving Machine-Learning Models in Materials Science through Large Datasets , journal =. 2024 , doi =

    work page 2024

  7. [7]

    and Nelson, Lance J

    Curtarolo, Stefano and Setyawan, Wahyu and Wang, Shidong and Xue, Junkai and Yang, Kesong and Taylor, Richard H. and Nelson, Lance J. and Hart, Gus L. W. and Sanvito, Stefano and Buongiorno-Nardelli, Marco and Mingo, Natalio and Levy, Ohad , title =. Computational Materials Science , volume =. 2012 , doi =

    work page 2012

  8. [8]

    and Aykol, Muratahan and Cheon, Gowoon and Cubuk, Ekin Dogus , title =

    Merchant, Amil and Batzner, Simon and Schoenholz, Samuel S. and Aykol, Muratahan and Cheon, Gowoon and Cubuk, Ekin Dogus , title =. Nature , volume =. 2023 , doi =

    work page 2023

Show all 34 references
  1. [9]

    and Kirklin, Scott and Aykol, Muratahan and Meredig, Bryce and Wolverton, Christopher , title =

    Saal, James E. and Kirklin, Scott and Aykol, Muratahan and Meredig, Bryce and Wolverton, Christopher , title =. JOM , volume =. 2013 , doi =

    2013

  2. [10]

    and Reid, Andrew C

    Choudhary, Kamal and Garrity, Kevin F. and Reid, Andrew C. E. and DeCost, Brian and Biacchi, Adam J. and others , title =. npj Computational Materials , volume =. 2020 , doi =

    2020

  3. [11]

    and Kolb, Ute and Palatinus, Luk

    Gemmi, Mauro and Mugnaioli, Enrico and Gorelik, Tatiana E. and Kolb, Ute and Palatinus, Luk. ACS Central Science , volume =. 2019 , doi =

    2019

  4. [12]

    and Oliver, Allen G

    Unge, Johan and Nannenga, Brent L. and Oliver, Allen G. and Gonen, Tamir , title =. Acta Crystallographica Section C , volume =. 2025 , doi =

    2025

  5. [13]

    , title =

    Krivovichev, Sergey V. , title =. Angewandte Chemie International Edition , volume =. 2014 , doi =

    2014

  6. [14]

    Journal of Applied Crystallography , volume =

    Palatinus, Luk. Journal of Applied Crystallography , volume =. 2007 , doi =

    2007

  7. [15]

    Advanced Science , volume =

    Lai, Qingsi and Xu, Fanjie and Yao, Lin and Gao, Zhifeng and Liu, Siyuan and Wang, Hongshuai and Lu, Shuqi and He, Di and Wang, Liwei and Zhang, Linfeng and Wang, Cheng and Ke, Guolin , title =. Advanced Science , volume =. 2025 , doi =

    2025

  8. [16]

    Nature Communications , volume =

    Li, Qi and Jiao, Rui and Wu, Liming and Zhu, Tiannian and Huang, Wenbing and Jin, Shifeng and Liu, Yang and Weng, Hongming and Chen, Xiaolong , title =. Nature Communications , volume =. 2025 , doi =

    2025

  9. [17]

    Transactions on Machine Learning Research , year =

    Johansen, Frederik Lizak and Friis-Jensen, Ulrik and Dam, Erik Bj. Transactions on Machine Learning Research , year =

  10. [18]

    and Valsecchi, Michele and Billinge, Simon J

    Guo, Gabe and Saidi, Tristan Luca and Terban, Maxwell W. and Valsecchi, Michele and Billinge, Simon J. L. and Lipson, Hod , title =. Nature Materials , volume =. 2025 , doi =

    2025

  11. [19]

    Nature Communications , volume =

    Yu, Dongfang and Zhu, Zhewen and Leng, Fucheng and Zhu, Yizhou , title =. Nature Communications , volume =. 2026 , doi =

    2026

  12. [20]

    arXiv preprint arXiv:2603.00965 , year =

    Li, Qi and Guo, Mingyu and Jiao, Rui and Gao, Jing and Xu, Fanjie and Xue, Haonan and Zhang, Weixiong and Huang, Wenbing and Yan, Junchi and Zhang, Linfeng and Wang, Cheng and Yan, Zhuang and Ke, Guolin and E, Weinan and Tang, Zhiyong and Jin, Shifeng and Yao, Lin , title =. a...

  13. [21]

    arXiv preprint arXiv:2606.14003 , year =

    Segal, Nofit and Li, Mingda and Miller, Benjamin Kurt and G. arXiv preprint arXiv:2606.14003 , year =

  14. [22]

    and Anker, Andy S

    Friis-Jensen, Ulrik and Johansen, Frederik L. and Anker, Andy S. and Dam, Erik B. and Jensen, Kirsten M. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2024 , doi =

    2024

  15. [23]

    and Persson, Kristin A

    Ong, Shyue Ping and Richards, William Davidson and Jain, Anubhav and Hautier, Geoffroy and Kocher, Michael and Cholia, Shreyas and Gunter, Dan and Chevrier, Vincent L. and Persson, Kristin A. and Ceder, Gerbrand , title =. Computational Materials Science , volume =. 2013 , doi =

    2013

  16. [24]

    and Zeltmann, Steven E

    Savitzky, Benjamin H. and Zeltmann, Steven E. and Hughes, Lauren A. and Brown, Hamish G. and Zhao, Shiteng and Pelz, Philipp M. and Pekin, Thomas C. and Barnard, Edward S. and Donohue, Jennifer and Rangel DaCosta, Luis and Kennedy, Ellis and Xie, Yujun and Janish, Matthew T. a...

    2021

  17. [25]

    and Rakowski, Alexander and Ribet, Stephanie M

    Gleason, Samuel P. and Rakowski, Alexander and Ribet, Stephanie M. and Zeltmann, Steven E. and Savitzky, Benjamin H. and Henderson, Matthew and Ciston, Jim and Ophus, Colin , title =. Physical Review Materials , volume =. 2024 , doi =

    2024

  18. [26]

    Nathani, Anvesh and McCray, Arthur R. C. and Liu, Yingtao and Ding, Hanping and Kazempoor, Pejman and Xu, Shuozhi and Ophus, Colin and Ghamarian, Iman , title =. npj Computational Materials , volume =. 2026 , doi =

    2026

  19. [27]

    Advanced Functional Materials , volume =

    Peng, Qianwen and Han, Xiaocang and Wang, Zijian and Hong, Yanhui and Meng, Fanqi and Zhang, Zezhong and Zhang, Jin and Zhao, Xiaoxu , title =. Advanced Functional Materials , volume =. 2026 , doi =

    2026

  20. [28]

    Journal of the American Chemical Society , volume =

    Melgalvis, D. Journal of the American Chemical Society , volume =. 2026 , doi =

    2026

  21. [29]

    Malik, S. A. and Doherty, T. A. S. and Colmey, B. and Roberts, S. J. and Gal, Y. and Midgley, P. A. , title =. Nature Communications , volume =. 2026 , doi =

    2026

  22. [30]

    Klar, P. B. and Krysiak, Y. and Xu, H. and Steciuk, G. and Cho, J. and Zou, X. and Palatinus, L. , title =. Nature Chemistry , volume =. 2023 , doi =

    2023

  23. [31]

    Advances in Neural Information Processing Systems , volume =

    Jiao, Rui and Huang, Wenbing and Lin, Peijia and Han, Jiaqi and Chen, Pin and Lu, Yutong and Liu, Yang , title =. Advances in Neural Information Processing Systems , volume =. 2023 , doi =

    2023

  24. [32]

    Atomistic Language Models Understand and Generate Materials , journal =

    Edamadaka, Sathya and Ramesh, Krithik and Li, Ju and G. Atomistic Language Models Understand and Generate Materials , journal =

  25. [33]

    arXiv preprint arXiv:2504.06231 , year =

    Rhodes, Benjamin and Vandenhaute, Sander and. arXiv preprint arXiv:2504.06231 , year =

    arXiv

  26. [34]

    and Ceder, Gerbrand , title =

    Deng, Bowen and Zhong, Peichen and Jun, KyuJung and Riebesell, Janosh and Han, Kevin and Bartel, Christopher J. and Ceder, Gerbrand , title =. Nature Machine Intelligence , volume =. 2023 , doi =

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.