Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Kinetic Langevin Diffusion for Crystalline Materials Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read KLDM generates crystalline materials by diffusing auxiliary velocities in flat space rather than running Riemannian diffusion on the hypertorus, and it matches or beats prior diffusion models on crystal structure prediction and de-novo…

desk verdict A legitimate, well-scoped adaptation of TDM to crystal coordinates with strong CSP results, but the abstract's claim overreaches: translation invariance is only mitigated at low noise, not fully accounted for. read the letter →

arxiv 2507.03602 v1 pith:BFDJSYJM submitted 2025-07-04 cs.LG

François Cornet , Federico Bergamin , Arghya Bhowmik , Juan Maria Garcia Lastra , Jes Frellsen , Mikkel N. Schmidt This is my paper · ORCID
classification cs.LG
keywords crystallinematerialsgenerationdiffusionmodelsfractionalcoordinateshypertorusLiegroupstrivializedcrystalstructurepredictionde-novo
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 introduces KLDM, a diffusion model for crystalline materials in which the fractional coordinates of atoms—points on a hypertorus—are generated by coupling them to Euclidean velocity variables and diffusing only those velocities. This generalizes Trivialized Diffusion Model to crystal unit cells, with zero initial velocities, a simplified score parameterization, and velocity fields constrained to have zero net translation. The authors claim this yields a training objective that accounts for the periodic translation symmetry of crystals while avoiding Riemannian diffusion on the torus. On crystal structure prediction and de-novo generation benchmarks, KLDM matches or improves on existing diffusion models, with the largest gains on the larger MP-20 and MPTS-52 datasets.

What carries the argument

The central object is the trivialized coupling of coordinates $\hat f$ on the Lie group $G = SO(2)^{3K}$ with velocities $\hat v$ in its Lie algebra $\mathfrak g \cong \mathbb R^{3K}$. The forward dynamics are $d\hat f_t = \hat f_t \hat v_t\,dt$ and $d\hat v_t = -\gamma \hat v_t\,dt + \sqrt{2\gamma}\,dw_t^g$, so only the Euclidean velocities receive noise and the time-reversal needs a learned score only for $\hat v_t$. The transition kernel is $p_{t|0}(f_t, v_t|f_0, v_0) = \mathrm{WN}(r_t|\mu_{r_t}, \sigma_{r_t}^2 I)\cdot \mathcal N(v_t|\mu_{v_t}, \sigma_{v_t}^2 I)$ with $r_t = \log_m(f_0^{-1} f_t)$; zero initial velocities make the Gaussian velocity score closed-form, leaving a single learnable wrapped-normal term, and projecting velocities to the mean-free subspace removes global translations.

What would settle it

A direct check is to compute the denoising score-matching target for a noisy crystal and for the same crystal shifted by a whole-cell translation at high noise; if the targets differ while the translation-invariant network produces identical scores, the mismatch the paper claims to mitigate is still present at that noise level.

Watch

Extended reading notes

Core claim

The central claim is that the key innovation lies in modeling fractional coordinates through a coupling with auxiliary Euclidean variables representing velocities, so the diffusion process is offset to a flat space. This effectively performs diffusion on the hypertorus while providing a training objective that accounts for the periodic translation symmetry of the true data distribution, in contrast to previous work that runs Riemannian diffusion directly on the torus. The paper argues that this mitigates the mismatch between the translation-invariant score architecture and the non-invariant training target at low levels of noise, and it reports competitive or better performance than prior models on both crystal structure prediction and de-novo generation tasks.

Load-bearing premise

The method's claim to respect periodic translation symmetry depends on the assumption that forcing the average atomic velocity to zero keeps noisy coordinates from drifting by a whole-cell translation during training; the paper's own 1D experiments show this holds at low noise but not reliably at high noise.

Editorial extensions

If this is right

  • KLDM's coordinate model is trained with a closed-form denoising objective on flat velocity space, so no Riemannian score on the torus is needed for the coordinate target.
  • On crystal structure prediction, KLDM achieves match rates and coordinate RMSEs comparable to or better than prior diffusion models, with the largest improvements on MP-20 and MPTS-52.
  • On de-novo generation, using analog-bits or discrete diffusion for atom types gives lower energy above the hull and higher stability than the DIFFCSP baseline on MP-20.
  • The ablations show that zero initial velocities, the simplified score parameterization, and zero-net-translation velocity fields each improve validation match rate and convergence speed.
  • The paper's own analysis indicates the translation-symmetry benefit is strongest at low noise: at high noise, the Fréchet mean of noisy samples jumps among a discrete set of values rather than being preserved.

Reading between the lines

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

  • The discrete set of mean values at high noise suggests KLDM may implicitly learn a discrete choice of translation sector; exposing that choice as a latent variable could make the high-noise behavior principled rather than empirical.
  • The same trivialized velocity coupling should transfer to other periodic or Lie-group data, such as torsion-angle generation in molecules, where wrapped coordinates currently require special treatment.
  • A direct comparison against a translation-invariant noising scheme that removes the center of mass exactly would isolate how much of KLDM's gain comes from the velocity coupling itself versus from merely suppressing global translations.
  • If the high-noise mean jumps turn out to be benign, the simplified parameterization with zero initial velocities could be combined with other discrete data diffusions to extend KLDM to larger systems such as metal-organic frameworks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. KLDM applies the Trivialized Diffusion Model (TDM) to crystalline materials by coupling fractional coordinates on the hypertorus with Euclidean velocity variables. The paper proposes three design choices: zero initial velocities, a simplified score parameterization that avoids learning the velocity part of the score, and mean-free velocity fields intended to reduce the mismatch between the translation-invariant score network and the non-invariant denoising target. The training objective is derived in closed form, and the method is evaluated on CSP benchmarks (PEROV-5, MP-20, MPTS-52, CARBON-24) and on DNG on MP-20, with ablations over multiple seeds. The reported results are competitive with, and in several settings better than, DIFFCSP, EQUICSP, FLOWMM, and MATTERGEN-MP.

Significance. If the central symmetry claim were rigorously established, this would be a valuable contribution: KLDM avoids explicit Riemannian diffusion on the hypertorus, retains a simulation-free closed-form score, and transfers TDM to a practically important application. The paper's strengths include a coherent derivation of the unconstrained score in Appendix C, multi-seed ablations, evaluation on standard benchmarks, and an independent MLIP-based stability pipeline. The empirical results are competitive. However, the theoretical grounding for the handling of periodic translation symmetry is incomplete, and the training algorithm appears to use a projected noising process whose score is not derived. These gaps affect the paper's central claim and currently limit the strength of the contribution.

major comments (3)
  1. [Abstract, Section 3.2, Appendix D, Eq. (18)] The abstract states that the training objective 'accounts for the periodic translation symmetry of the true data distribution,' but this is not established by the analysis in Appendix D. The DSM target in Eq. (18) depends on rt = logm(f0^-1 ft), i.e., on the full relative displacement between the clean and noisy coordinates, not only on the torus-valued Fréchet mean of ft. Appendix D only checks whether the Fréchet mean of the noisy sample equals that of f0; it does not measure the variance of the score target over translations of f0 for fixed (ft, vt). Moreover, the appendix itself shows that at high noise the Fréchet mean is not preserved and jumps among values separated by 2π/n. The argument that this 'turns a continuous problem into a discrete one' is not quantified and does not follow from the mean-preservation experiment. Please either provide a direct analysis or diagnostic of the DSM target under the mean-free constraint, or revise the abstract and contributions to state that the mismatch is mitigated empirically at low noise.
  2. [Section 3.2, Eqs. (16)-(18), Algorithms 1-2] The actual training distribution differs from the transition kernel used to derive the training target. In Algorithm 1, velocities are sampled from a projected Normal, rt is sampled with mean-free noise, and ft is centered; Section 3.2 additionally states that the score in Eq. (17) is projected. The density of this projected/centered process is not the wrapped normal of Eq. (16), and no correction term (e.g., a Jacobian factor or a modified score formula for the projected wrapped normal) is provided. Consequently, the DSM loss may be minimizing against a target that is not the score of the distribution actually used to noise the training examples. Since this objective is the basis for the paper's central contribution, the authors should either derive the score for the projected process or explicitly justify the projection as an approximation with a quantitative assessment of its effect.
  3. [Section 5.3, Table 2] The DNG comparison is not controlled. DIFFCSP* and MATTERGEN-MP* were trained on the re-optimized, filtered MP-20 dataset, whereas KLDM was trained on the original, unfiltered MP-20. The text nonetheless concludes that KLDM 'outperforms DIFFCSP' on stability-related metrics. Because dataset preprocessing can materially affect stability and energy-above-hull metrics, this conclusion should either be supported by retraining the baselines on the same data split or be presented only as a suggestive, not fully controlled, comparison.
minor comments (5)
  1. [Section 5.1] The text says KLDM is evaluated across '4 datasets' but lists only PEROV-5, MP-20, and MPTS-52; CARBON-24 is missing from the list.
  2. [Table 2] The caption contains a contradictory sentence: KLDM 'outperforms DIFFCSP in terms of RMSD, energy above the hull, and stability, while being slightly subpar on RMSD.' The second occurrence should presumably read 'S.U.N.' as in the body text.
  3. [Conclusion] The conclusion refers to 'MPTS-20', but the dataset used in the paper is MP-20; please correct the name.
  4. [Algorithms 3 and 4] In Algorithm 4, the line constructing the score uses 'vt' without a prior definition in that pseudocode block; it should refer to the predicted velocity vpred_n (and similar clarity would help in Algorithm 3).
  5. [Section 3.1, Eq. (15)] The wrapping function is defined inconsistently across the paper: Eq. (15) uses w(·) = atan2(sin, cos), while earlier text uses w(·) = · − floor(·) for wrapping in Eq. (10). Please unify the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KLDM's derivation is self-contained and its central claims are empirical, not definitional.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The forward process (Eq. 12), transition kernel (Eq. 16), and denoising score-matching target (Eqs. 17-19) are obtained by direct algebraic manipulation of the TDM framework of Zhu et al. (2024), an external reference, not a self-citation. The zero-net-translation velocity constraint is introduced as a modeling choice, and its benefit is supported by ablations (Fig. 1, Appendix E) rather than by assuming the desired conclusion. The paper explicitly acknowledges the limitation in Appendix D: 'the mismatch may not be fully mitigated across all noise levels' and 'At higher levels of noise, we note that the Frechet mean jumps between n different values.' This is an honest empirical caveat, not a circular step. The final CSP and DNG metrics are computed against external benchmarks (StructureMatcher, MatterSim via MATTERGEN's pipeline) and are not quantities fitted into the derivation. No load-bearing self-citation or definitional equivalence makes any 'prediction' equal to an input by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The derivation of the score target rests on the closed-form transition kernel of TDM (prior work) and on the Abelian structure of the torus. The domain-specific improvements, zero initial velocity, simplified score, and mean-free velocities, are empirical and are supported by ablations and a limited 1D analysis. No new physical entities are introduced.

free parameters (4)
  • Kinetic Langevin drift gamma = 1
    Chosen by hand following Zhu et al. (2024); controls how strongly velocities relax in Eqs. (12) and (13).
  • Diffusion time horizon T = 2
    Chosen by hand; coordinates diffuse on [0,2) while lattice and atom types diffuse on [0,1).
  • Initial velocity distribution p(v0) = delta at 0
    Design choice; zero initial velocity enables the simplified score parameterization in Eq. (19) and improves validation match rate in ablations.
  • Atom-type loss weights lambda_a = 20 (one-hot), 1 (analog-bits), 0.33 (discrete)
    Hand-selected to balance loss magnitudes in DNG training; affects final DNG metrics.
assumptions (4)
  • standard math Transition kernel Eq. (16) for the left-trivialized kinetic Langevin process on SO(2)^(3K) is taken as given from Zhu et al. (2024).
    The paper does not re-derive this kernel; it is the mathematical foundation of the training objective.
  • standard math The hypertorus T^(3K) is an Abelian Lie group, so the wrapped normal transition has a closed form.
    Used in Appendix A and Eq. (16) to justify simulation-free training.
  • domain assumption A periodic-translation-invariant score network combined with zero-mean velocity fields yields a sufficiently translation-invariant learned distribution.
    Invariance of the network is architectural; the mean-free constraint only approximately preserves the frame, as shown in Appendix D.
  • domain assumption The Frechet mean of the clean sample is approximately preserved at moderate noise under zero-net-translation velocities.
    Supported only by 1D synthetic experiments in Appendix D, not by a general proof for 3D crystals.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kinetic Langevin Diffusion for Crystalline Materials Generation." pith.science (2026). https://pith.science/paper/BFDJSYJM

@misc{pith2026250703602,
  author       = {Pith},
  title        = {Pith review of: Kinetic Langevin Diffusion for Crystalline Materials Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BFDJSYJM}},
  note         = {Machine review of arXiv:2507.03602}
}
read the original abstract

Generative modeling of crystalline materials using diffusion models presents a series of challenges: the data distribution is characterized by inherent symmetries and involves multiple modalities, with some defined on specific manifolds. Notably, the treatment of fractional coordinates representing atomic positions in the unit cell requires careful consideration, as they lie on a hypertorus. In this work, we introduce Kinetic Langevin Diffusion for Materials (KLDM), a novel diffusion model for crystalline materials generation, where the key innovation resides in the modeling of the coordinates. Instead of resorting to Riemannian diffusion on the hypertorus directly, we generalize Trivialized Diffusion Model (TDM) to account for the symmetries inherent to crystals. By coupling coordinates with auxiliary Euclidean variables representing velocities, the diffusion process is now offset to a flat space. This allows us to effectively perform diffusion on the hypertorus while providing a training objective that accounts for the periodic translation symmetry of the true data distribution. We evaluate KLDM on both Crystal Structure Prediction (CSP) and De-novo Generation (DNG) tasks, demonstrating its competitive performance with current state-of-the-art models.

Figures

Figures reproduced from arXiv: 2507.03602 by the authors.

Figure 1
Figure 1. Ablation study results. We report the mean and standard error across 6 seeds of the validation Match Rate (MR). (Left) Impact of the initial velocity distribution variance. A variance of 0 corresponds to the zero-initial velocity case (i.e., a delta distribution). (Center) Effect of the score parameterization on convergence. (Right) Impact of enforcing zero-net translation in velocity fields under zero-initial veloc… view at source ↗
Figure 2
Figure 2. Effect of sampling from the transition kernel in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Effect of sampling from the transition kernel in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation of velocity fields with zero net translation for non-zero initial velocities. ( [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Discovery and recovery of crystalline materials with property-conditioned transformers

    cond-mat.mtrl-sci 2025-11 conditional novelty 6.0 of 10

    Conditioning the attention layers of a crystal-writing transformer on continuous property values enables XRD-based structure recovery and targeted generation of photovoltaic candidates.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    for training iterations do x0 = {(f0, l0, a0)}B i=1 ∼ q(x), t ∼ U(t), v0 ∼ p0(v) ▷ Bindicates the batch size

    For DNG task we require also an atom type transition kernel pt|0(at|a0). for training iterations do x0 = {(f0, l0, a0)}B i=1 ∼ q(x), t ∼ U(t), v0 ∼ p0(v) ▷ Bindicates the batch size. if task is DNG then (vt, ft, lt, at), (targetv, targetl, targeta) = training targets(f , v, l, a, t) outv, outl, outa = sθ(t, ft, vt, lt, at) ▷ The network takes t, ft, vt, l...

    work page 2023

  2. [2]

    (2024) while using a classic Euler–Maruyama step for lattice parameters lt and atom type at in Algorithm

    We then focus on sampling, presenting both a sampling scheme that uses an exponential integrator for simulating the reverse SDE of the dynamics of the velocities vt and fractional coordinates ft as proposed by Zhu et al. (2024) while using a classic Euler–Maruyama step for lattice parameters lt and atom type at in Algorithm

    work page 2024

  3. [3]

    In algorithm Algorithm 4, instead, we present how we sample from our model using the predictor-corrector steps as proposed in (Song et al., 2021). Algorithm 1 training targets(f , v, l, a, t): Routine for sampling ft, vt, lt, at from the transition kernels and the corresponding target scores Require: task (either CSP or DNG), timestep t, scheduler α(t) an...

    work page 2021

  4. [4]

    ( Left) Initial velocities are sampled from a N (0, 0.1 · I), i.e

    0k 1k 2k 3k Epoch 0.2 0.3 0.4 Match Rate [%] ∑ k vk ⁄= 0 ∑ k vk = 0 0k 1k 2k 3k Epoch ∑ k vk ⁄= 0 ∑ k vk = 0 0k 1k 2k 3k Epoch ∑ k vk ⁄= 0 ∑ k vk = 0 Figure 4: Ablation of velocity fields with zero net translation for non-zero initial velocities. ( Left) Initial velocities are sampled from a N (0, 0.1 · I), i.e. σ2 v0 = 0.1. (Center) Initial velocities ar...

    work page 2023

  5. [8]

    Mattersim: A deep learning atomistic model across elements, temperatures and pressures

    Yang, H., Hu, C., Zhou, Y ., Liu, X., Shi, Y ., Li, J., Li, G., Chen, Z., Chen, S., Zeni, C., et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967, 2024a. Yang, S., Cho, K., Merchant, A., Abbeel, P., Schuurmans, D., Mordatch, I., and Cubuk, E. D. Scalable diffusion for materials gen...

    arXiv

  6. [9]

    doi: 10.1038/s41586-025-08628-5

    ISSN 1476-4687. doi: 10.1038/s41586-025-08628-5. Zhu, Y ., Chen, T., Kong, L., Theodorou, E. A., and Tao, M. Trivialized momentum facilitates diffusion generative modeling on lie groups. arXiv preprint arXiv:2405.16381,

    arXiv

  7. [10]

    +∞X k=−∞ exp − (rt − µrt + 2πk)2 2σ2rt # = C ·

    13 Kinetic Langevin Diffusion for Crystalline Materials Generation Organization of the Supplementary Material Appendix A We provide a short informal introduction to Lie groups and manifolds. We provide intuition behind key equations Eqs. (12) and (13), and the TDM and KLDM models more broadly. Appendix B We discuss why data on a torus can be represented e...

    work page 2024

  8. [11]

    Effect of the zero-net translation velocity field Mismatch between translation-invariant score parameterization and non-invariant training target As noted by Lin et al

    D. Effect of the zero-net translation velocity field Mismatch between translation-invariant score parameterization and non-invariant training target As noted by Lin et al. (2024), there is a potential mismatch between the translation-invariant parameterization of the score and the non-invariant training target. For instance, a noisy point cloud and its pe...

    work page 2024

Show all 21 references
  1. [14]

    COV-R (recall) and COV-P (precision) are obtained by comparing the distances between generated and test fingerprints

    and Magpie compositional fingerprints (Ward et al., 2016). COV-R (recall) and COV-P (precision) are obtained by comparing the distances between generated and test fingerprints. Finally, the property statistics are obtained by comparing distributions of properties, computed on ...

    2016

  2. [18]

    vn = exp(dt)vn−1 + 2(exp(2dt) − 1)out(n−1) v + p exp(2dt) − 1ϵv ▷ Update on v fn = w(fn−1 − vndt) ▷ Update on f ## Update step for l Sample ϵl ∼ N(0, I) ln = ln−1 − (f (t) − g2(t)s(out(n−1) l ))dt + √ dtϵl ▷ EM step for l if task is DNG then ## Update step for a Sample ϵa ∼ N(...

    2023

  3. [19]

    Experimental details I.1

    vn = vpred n + δoutv + √ 2δϵv ▷ Update on v fn = w(f pred n − vndt) ▷ Update on f ## Update step for l Sample ϵl ∼ N(0, I) ln = ln−1 − (f (t) − g2(t)s(outl))dt + √ dtϵl ▷ EM step for l if task is DNG then ## Update step for a Sample ϵa ∼ N(0, I) an = an−1 − (f (t) − g2(t)s(out...

    2021

  4. [20]

    For this dataset, the CSP@1 task is ill-defined due to its one-to-many nature, and we only report the obtained values for completeness

    Error bars for @1 represent the standard deviation over the mean at sampling time across 20 different seeds. For this dataset, the CSP@1 task is ill-defined due to its one-to-many nature, and we only report the obtained values for completeness. CARBON -24 M ODEL MR [%] ↑ RMSE ...

    1988

  5. [21]

    The scaling is needed to make the different loss terms have similar magnitudes

    In the case of analog-bits, we used λa = 1, while when using discrete diffusion for the atom types, we scale the losses using λa = 0.33. The scaling is needed to make the different loss terms have similar magnitudes. I.3. Parameters for the KLDM forward process We kept the dri...

    2024

  6. [22]

    for the diffusion process. For lattice parameters and atom types, we rely on the standard Euclidean diffusion model where the drift and diffusion coefficients are defined by a linear schedule on the interval [0, 1). We trained all the networks using AdamW with the default PyTo...

    2023

  7. [2011]

    Piloting structure-based drug design via modality-specific optimal schedule

    Qiu, K., Song, Y ., Fan, Z., Liu, P., Zhang, Z., Zheng, M., Zhou, H., and Ma, W.-Y . Piloting structure-based drug design via modality-specific optimal schedule. arXiv preprint arXiv:2505.07286,

    arXiv

  8. [2013]

    H., Zhu, R., Dai, H., Aggarwal, S., Wei, N., Chen, A

    Petersen, M. H., Zhu, R., Dai, H., Aggarwal, S., Wei, N., Chen, A. P., Bhowmik, A., Lastra, J. M. G., and Hip- palgaonkar, K. Dis-CSP: Disordered crystal structure predictions. In AI for Accelerated Materials Design - ICLR 2025,

    2025

  9. [2020]

    M., and Bengio, Y

    Hoffmann, J., Maestrati, L., Sawada, Y ., Tang, J., Sellier, J. M., and Bengio, Y . Data-driven approach to encod- ing and decoding 3-d crystal structures. arXiv preprint arXiv:1909.00949,

    1909 arXiv

  10. [2022]

    and Aspuru-Guzik, A

    Flam-Shepherd, D. and Aspuru-Guzik, A. Language models can generate molecules, materials, and protein binding sites directly in three dimensions as xyz, cif, and pdb files. arXiv preprint arXiv:2305.05708,

    arXiv

  11. [2023]

    S., Kaba, S.-O., Zhu, Q., Galkin, M., Miret, S., and Ravanbakhsh, S

    Levy, D., Panigrahi, S. S., Kaba, S.-O., Zhu, Q., Galkin, M., Miret, S., and Ravanbakhsh, S. SymmCD: Symmetry- preserving crystal generation with diffusion models. InAI for Accelerated Materials Design - NeurIPS 2024,

    2024

  12. [2024]

    Crys- talgan: learning to discover crystallographic structures with generative adversarial networks

    Nouira, A., Sokolovska, N., and Crivello, J.-C. Crys- talgan: learning to discover crystallographic structures with generative adversarial networks. arXiv preprint arXiv:1810.11203,

    arXiv

  13. [2025]

    Pickard, C. J. Airss data for carbon at 10gpa and the c+ n+ h+ o system at 1gpa. Materials Cloud Archive, doi: 10.24435/materialscloud:2020. 0026/v1,

    2020 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.