Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Open Materials Generation with Stochastic Interpolants

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read OMatG applies stochastic interpolants to inorganic crystal generation and claims state-of-the-art results over diffusion and flow-matching baselines.

desk verdict A careful, well-verified framework for crystal generation whose genuine contribution—interpolation choice matters—is undercut by an overstated SOTA claim and missing error bars. read the letter →

arxiv 2502.02582 v2 pith:YKNCFFTS submitted 2025-02-04 cs.LG cond-mat.mtrl-sci

Philipp Hoellmer , Thomas Egg , Maya M. Martirossyan , Eric Fuemmeler , Zeren Shui , Amit Gupta , Pawan Prakash , Adrian Roitberg
show 6 more authors
Mingjie Liu George Karypis Mark Transtrum Richard G. Hennig Ellad B. Tadmor Stefano Martiniani
This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sci
keywords stochasticinterpolantsgenerativemodelingcrystalstructurepredictiondenovomaterialsgenerationequivariantgraphneuralnetworksperiodicboundaryconditionsdiscreteflowmatching
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

This paper introduces OMatG, a generative framework for inorganic crystals built on stochastic interpolants, and argues that the extra design freedom of interpolants is what lets it beat existing diffusion and flow-matching models. The central idea is to bridge a random base structure and a target crystal through a tunable path $x_t=\alpha(t)x_0+\beta(t)x_1+\gamma(t)z$, with periodic boundary conditions handled by interpolating along the shortest geodesic in unwrapped coordinates and wrapping back, and atomic species handled by discrete flow matching. The paper claims that this framework achieves a new state of the art on crystal structure prediction (CSP) and de novo generation (DNG), outperforming both DiffCSP and FlowMM on CSP and DNG, and MatterGen on DNG, across the four benchmark datasets considered. If true, the practical payoff is that materials discovery pipelines can treat the generative process itself as tunable rather than fixed to one diffusion or flow scheme.

What carries the argument

The load-bearing object is the stochastic interpolant $x_t = \alpha(t)x_0 + \beta(t)x_1 + \gamma(t)z$, with loss functions that learn the velocity $b_\theta(t,x)$ and, for SDE sampling, the denoiser $z_\theta(t,x)$. The interpolation coefficients are flexible except for endpoint constraints, and the paper tests linear, trigonometric, encoder-decoder, and score-based-diffusion (VP/VE) interpolants, with ODE or SDE sampling and a tunable latent-noise amplitude $\gamma(t)$. For periodic fractional coordinates, the path is defined by unwrapping $x_1$ to its closest periodic image, interpolating in Euclidean space along the shortest geodesic, and wrapping the result back onto the torus; atomic species are generated by discrete flow matching through a continuous-time Markov chain. This mechanism is what lets the same code reproduce conditional flow matching (linear interpolant, ODE, $\gamma=0$) and score-based diffusion (VP/VE interpolants with Gaussian base) while also exploring new in-between schemes.

What would settle it

On a one-dimensional torus, take a uniform base $\rho_0$, a target density whose exact form is known, and a nonlinear or stochastic interpolant with $\gamma(t) > 0$; numerically integrate the learned ODE/SDE and compare the empirical endpoint density with the exact wrapped target marginal. A mismatch would mean the wrapped process is not sampling the target distribution, so the training objective itself is unsound.

Watch

Extended reading notes

Core claim

OMatG is described as the first implementation of stochastic interpolants for inorganic crystal generation, and as such it unifies conditional flow matching and score-based diffusion as special cases inside a single, tunable framework. The paper's central empirical claim is that by optimizing the interpolation functions for the fractional coordinates and lattice vectors separately, and by coupling the continuous flow with discrete flow matching for atomic species, OMatG reaches a new state of the art: higher structural match rates than DiffCSP and FlowMM on perov-5, MP-20, and MPTS-52, the first CSP baseline on Alex-MP-20, and higher S.U.N. (stable, unique, novel) rates than MatterGen-MP in de novo generation. The best OMatG models also generate structures with lower average energy above the convex hull and closer match to reference coordination-number and crystal-system distributions. The authors interpret the results as evidence that the choice of interpolant, latent noise, sampling scheme, and per-degree-of-freedom tuning is a decisive axis of model design for materials generation.

Load-bearing premise

The load-bearing assumption is that interpolating fractional coordinates in unwrapped Euclidean space along the shortest periodic image and then wrapping back onto the torus gives the correct marginal distribution on the torus for nonlinear and stochastic interpolants with $\gamma(t)>0$; the paper uses this construction for all interpolants but proves it only for the linear geodesic case.

Editorial extensions

If this is right

  • If the SOTA claim holds, any new generative model for crystals should be compared against OMatG under the same refined metrics, including validity filters and the coordination-number benchmark, before claiming improvement.
  • The reported CSP gains on MP-20 and MPTS-52 mean that specified-composition structure prediction can be run at generative-model speed with better starting candidates for downstream DFT relaxation.
  • The DNG improvements in stability and S.U.N. rate mean that a larger fraction of generated structures should survive DFT screening, cutting the cost of de novo materials discovery pipelines.
  • Because OMatG's linear/ODE setting reproduces FlowMM and its SBD interpolants reproduce diffusion models, the framework gives a single code base for fair comparisons and ablations across previously separate methods.

Reading between the lines

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

  • The periodic-boundary construction, based on the same unwrap-interpolate-wrap recipe, could carry over to other periodic or compact manifolds such as angle variables or SO(3) orientations, although the paper does not develop that generalization.
  • If the match-rate/RMSE trade-off observed on perov-5 is generic, then match rate alone may over-reward models that find the right chemical environment but miss exact symmetric sites; a combined metric penalizing both errors would sharpen future comparisons.
  • The framework's arbitrary base distributions suggest a direct way to inject prior knowledge, such as LLM-suggested compositions or machine-learned-potential relaxed motifs, as the base $\rho_0$, which the paper only begins to explore with LLM-generated structures.
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

3 major / 5 minor

Summary. The paper introduces OMatG, a generative model for inorganic crystal structures built on stochastic interpolants (SI), jointly modeling lattice vectors, fractional coordinates, and atomic species (via discrete flow matching). It extends SI to periodic boundary conditions by an unwrap–interpolate–wrap procedure, and evaluates the method on crystal structure prediction (CSP) on perov-5, MP-20, MPTS-52, and Alex-MP-20, and on de novo generation (DNG) on MP-20, claiming a new state of the art against DiffCSP, FlowMM, and MatterGen-MP. The paper also releases code, recomputes baseline metrics with SMACT 3.0, and adds a new average-coordination-number metric and a MatterSim/DFT stability pipeline.

Significance. The paper makes a valuable contribution to the materials-generation literature by demonstrating the practical flexibility of the stochastic-interpolants framework, with an open-source implementation and careful attention to evaluation (recomputed baselines, updated SMACT rules, DFT relaxation for stability metrics, and a new coordination-number metric). If the technical concerns below are resolved, the work would constitute a strong benchmark contribution. However, the headline state-of-the-art claim is currently not robustly supported: it rests on single-run, best-of-hyperparameter-optimization results with no seed variance, and no single OMatG configuration dominates across metrics. The torus extension of SI is also asserted without proof, leaving the soundness of the training objective unresolved for nonlinear and stochastic interpolants.

major comments (3)
  1. [Section I.B, Table III, Appendix C.3] The central claim that OMatG outperforms DiffCSP, FlowMM, and MatterGen in DNG is not established as stated, because no single OMatG configuration dominates across the reported metrics: MatterGen-MP achieves a markedly lower RMSD (0.1038 Å versus the best OMatG value of 0.4187 Å in Table III), DiffCSP achieves higher structural validity (99.91% versus the best OMatG 99.04% in Table II), and the best OMatG S.U.N. rate (22.48%) exceeds MatterGen-MP (20.30%) by only 2.18 points without any seed variance. Since each configuration is selected from 27–32 Bayesian HPO trials per interpolant (Appendix C.3) and reported once, the observed margin could be favorable noise. Please report multiple seeds with standard errors for the headline metrics, or revise the claim to a per-metric best-configuration statement.
  2. [Section III.B.1, Appendix B.5, Eq. (2)] The unwrap–interpolate–wrap procedure for periodic boundary conditions is asserted to extend stochastic interpolants to the torus, including the latent term γ(t)z, but no proof is given that the wrapped stochastic process x_t = wrap(α(t)x0 + β(t)x1' + γ(t)z) satisfies the SI velocity or Fokker–Planck equations on the torus. For the linear interpolant with γ=0 this reduces to the FlowMM geodesic, but for nonlinear interpolants (trig, enc-dec, VP/VE SBD) and for SDE sampling with γ>0 the path crosses the boundary and the time derivative of the wrapped path is not simply the wrapped derivative; hence the conditional velocity in Eq. (2) may not be the correct one for the target density on the torus. Please provide a proof or direct numerical validation (e.g., exact density recovery on a known toroidal target); without this, the training objective is not shown to be sound.
  3. [Appendix B.7, Table VII] Velocity annealing, which rescales the learned velocity as b_theta(t,x) → (1+st)b_theta(t,x), is applied with large coefficients (e.g., s=14 for positions in Table VII) and is acknowledged to lack theoretical justification. Because this ad-hoc modification changes the generative ODE/SDE dynamics, the reported SOTA results cannot be attributed solely to the SI framework, and the rescaling may invalidate the marginal density guaranteed by the SI construction. Please provide an ablation with s=0 or a demonstration that the annealed flow still samples the target distribution, or treat velocity annealing as a core component of the method and provide a theoretical or systematic empirical justification.
minor comments (5)
  1. [Table I] There is a formatting error in the DiffCSP row of the MPTS-52 columns: the text reads "15.79 / 14.290.1533/ 0.1489" without a separator between the match rate and RMSE; this should be corrected.
  2. [Section III.B.1 and Appendix B.5] The manuscript repeatedly refers to the "four-dimensional torus" for fractional coordinates in three dimensions; please clarify whether a three-dimensional torus is intended or explain the extra dimension.
  3. [Table III] The column header "⟨E⟩/N(↓) RMSD" is ambiguous; the energy above hull is reported in eV/atom and should be labeled clearly, with the RMSD unit (Å) separated.
  4. [Section V.C and Appendix C.3] The manuscript does not state explicitly how the single OMatG configuration shown in Tables II and III is chosen (e.g., best validation evalDNG, or best S.U.N. on a hold-out set), nor whether any test-set information was used in the selection; please make this selection rule explicit.
  5. [Figure 6 caption] The phrase "score based diffusion" should be hyphenated as "score-based diffusion".

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: OMatG's derivation is self-contained and externally benchmarked; the SOTA claim's per-metric HPO selection is a robustness concern, not a circular reduction.

full rationale

The central derivation chain is not circular. The SI losses in Eqs. (2)-(4) are taken from Albergo et al. [33] and are minimized against analytically defined interpolants; no target metric enters the loss as a fitted constant. The torus extension (Sec. III.B.1, App. B.5) is an asserted construction, not a theorem: unwrapping x1 along the shortest geodesic, computing the interpolant in Euclidean space, and wrapping back is a definition, and the paper does not prove that the wrapped nonlinear/SDE process has the correct torus marginal. That is an omitted proof and a correctness risk, but it does not assume the paper's conclusions. Lattice base distributions and center-of-mass removal follow Miller et al. [32] (FlowMM), an external method. CSP/DNG benchmarks use DiffCSP and FlowMM public code, MatterGen checkpoints, SMACT 3.0, MatterSim, and DFT relaxation via the MatterGen codebase; none of these are fitted by OMatG. The S.U.N. metric is evaluated against the external Alex-MP-20 reference set. Tables II-III show no single OMatG configuration dominates all metrics (e.g., MatterGen-MP has RMSD 0.1038 Å vs. best OMatG 0.4187 Å; DiffCSP structural validity 99.91% vs. best OMatG 99.04%), and each reported number is a best-of-HPO single run without seed variance; this weakens the Section I.B 'state of the art' claim as an empirical overclaim, but it does not make any prediction equivalent to its input by construction. Self-citations [17,18] support background statements about simulation databases and are not load-bearing. Score 1 reflects only these minor non-circular caveats.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The SI framework, CSPNet, and DFM are imported from prior work; OMatG contributes their combination and per-dataset tuning. The main additional load-bearing elements are the ad hoc velocity annealing, the torus interpolation procedure, and the evaluation pipeline choices. No new physical entities are introduced.

free parameters (8)
  • lambda_x,b (fractional-coordinate velocity loss weight) = 0.9729 (perov-5 Linear ODE); 0.9994 (MP-20 Linear ODE)
    Relative weight of the SI velocity loss for positions; tuned per dataset by Bayesian HPO and strongly affects match rate (Tables VII, IX, XI).
  • Velocity annealing coefficients s (positions and cell) = 14.11/2.90 (perov-5 Linear ODE); 8.20/1.46 (perov-5 Linear SDE)
    Inference-time rescaling of the learned velocity; the authors state in Appendix B7 that this lacks formal theoretical justification.
  • gamma(t) latent amplitude a = 0.034 (perov-5 Linear ODE); 0.258 (MP-20 Linear ODE)
    Amplitude in gamma = sqrt(a t(1-t)); controls path stochasticity and is tuned per interpolant and dataset.
  • DFM species-noise eta = 7.08 (MP-20 Linear ODE DNG); 0.19 (MP-20 Linear SDE DNG)
    Stochasticity added to the discrete rate matrix via detailed balance; tuned per model.
  • VP SBD base width sigma0 = 0.28 (perov-5 VP SBD ODE); 0.22 (MP-20 VP SBD ODE)
    Width of the Gaussian base for the score-based diffusion interpolant; tuned.
  • VE SBD schedule endpoints (sigma_min, sigma_max) = 0.0078, 0.5165 (perov-5); 0.0047, 0.9967 (MP-20)
    Noise schedule endpoints for the variance-exploding interpolant; tuned.
  • Integration steps = 820 (perov-5 Linear ODE); 130 (MP-20 VP SBD ODE)
    Number of ODE/SDE integrator steps; tuned per model and affects quality and cost.
  • Encoder-decoder switch time and exponent (T_switch, p) = (0.80, 1) perov-5 Enc-Dec ODE; (0.65, 1) MP-20
    Define the encoder-decoder interpolant; tuned.
assumptions (7)
  • standard math Stochastic interpolant theory (Albergo et al.): for interpolants satisfying Eq. (B1), the losses in Eqs. (2)-(3) learn the velocity and denoiser of a process transporting rho0 to rho1.
    Imported background result; used throughout Section III and Appendix B.
  • standard math CSPNet equivariance: message passing with sinusoidal positional embeddings yields permutation-equivariant, rotation-equivariant, translation-invariant embeddings (Jiao et al.).
    Used in Appendix C1 to predict velocity, denoiser, and discrete composition logits.
  • standard math Discrete flow matching / CTMC theory (Campbell et al.): the conditional rate matrix in Eq. (B6) and the update in Eq. (B9) sample from the target discrete distribution.
    Used for atomic species generation in Section III.B.3.
  • domain assumption Shortest-path geodesic unwrapping gives a well-defined periodic interpolant for all interpolant choices, including the latent variable gamma(t)z (Section III.B.1, Appendix B.5).
    Asserted, not derived; non-linear interpolants on the torus are not covered by the original SI theory.
  • domain assumption The MatterSim potential followed by DFT relaxation gives stability labels accurate enough to compare methods (Appendix D.4).
    S.U.N. rates rely on this pipeline; DFT relaxation is applied only to 1000 of 10,000 structures.
  • domain assumption Benchmark splits used for hyperparameter optimization do not leak into the reported test results (Appendix C.3).
    The paper does not explicitly state which split is used for HPO; standard practice is validation-based tuning.
  • ad hoc to paper Velocity annealing improves generation and does not invalidate the learned flow (Appendix B.7).
    Explicitly acknowledged by the authors to lack a formal justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Open Materials Generation with Stochastic Interpolants." pith.science (2026). https://pith.science/paper/YKNCFFTS

@misc{pith2026250202582,
  author       = {Pith},
  title        = {Pith review of: Open Materials Generation with Stochastic Interpolants},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKNCFFTS}},
  note         = {Machine review of arXiv:2502.02582}
}
read the original abstract

The discovery of new materials is essential for enabling technological advancements. Computational approaches for predicting novel materials must effectively learn the manifold of stable crystal structures within an infinite design space. We introduce Open Materials Generation (OMatG), a unifying framework for the generative design and discovery of inorganic crystalline materials. OMatG employs stochastic interpolants (SI) to bridge an arbitrary base distribution to the target distribution of inorganic crystals via a broad class of tunable stochastic processes, encompassing both diffusion models and flow matching as special cases. In this work, we adapt the SI framework by integrating an equivariant graph representation of crystal structures and extending it to account for periodic boundary conditions in unit cell representations. Additionally, we couple the SI flow over spatial coordinates and lattice vectors with discrete flow matching for atomic species. We benchmark OMatG's performance on two tasks: Crystal Structure Prediction (CSP) for specified compositions, and 'de novo' generation (DNG) aimed at discovering stable, novel, and unique structures. In our ground-up implementation of OMatG, we refine and extend both CSP and DNG metrics compared to previous works. OMatG establishes a new state of the art in generative modeling for materials discovery, outperforming purely flow-based and diffusion-based implementations. These results underscore the importance of designing flexible deep learning frameworks to accelerate progress in materials science. The OMatG code is available at https://github.com/FERMat-ML/OMatG.

Figures

Figures reproduced from arXiv: 2502.02582 by the authors.

Figure 1
Figure 1. FIG. 1. Visualization of the tunable components of the SI [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Illustration of crystal structure prediction (CSP) and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Histogram of the computed energies above the convex [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Training pipeline of the OMatG framework: A batch of structures is drawn from a dataset with target distribution [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Numerical integration pipeline of the OMatG framework: An initial structure [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Extending interpolants to incorporate periodic boundary conditions. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. We show here the effect of making matching more difficult by decreasing the length tolerance used by Pymatgen’s [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Histogram of the computed energies above the convex hull for structures generated by the five OMatG DNG models [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Qualitative performance of the distribution of [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. Qualitative performance of the distribution of average coordination number by structure for [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. Qualitative performance of the distribution of average coordination number by species (listed left to right in order of [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12. Qualitative performance of the distribution of crystal system by structure for [PITH_FULL_IMAGE:figures/full_fig_p030_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Multimodal Crystal Flow: Any-to-Any Modality Generation for Unified Crystal Modeling

    cs.LG 2026-02 unverdicted novelty 6.0 of 10

    MCFlow uses decoupled flow time axes for atom types and crystal structures so a single model handles crystal structure prediction, de novo generation, and atom-type generation.

  2. FlowMol3: Flow Matching for 3D De Novo Small-Molecule Generation

    cs.LG 2025-08 conditional novelty 6.0 of 10

    Combining self-conditioning, fake atoms, and late-stage geometry distortion lets a compact flow-matching model generate nearly always valid 3D drug-like molecules and match training-data chemistry better than existing...

Reference graph

Works this paper leans on

94 extracted references · 48 canonical work pages · cited by 2 Pith papers

  1. [1]

    Boeri, R

    L. Boeri, R. Hennig, P. Hirschfeld, G. Profeta, A. Sanna, E. Zurek, W. E. Pickett, M. Amsler, R. Dias, M. I. Eremets, C. Heil, R. J. Hemley, H. Liu, Y. Ma, C. Pier- leoni, A. N. Kolmogorov, N. Rybin, D. Novoselov, V. Anisimov, A. R. Oganov, C. J. Pickard, T. Bi, R. Arita, I. Errea, C. Pellegrini, R. Requist, E. K. U. Gross, E. R. Margine, S. R. Xie, Y. Qu...

    2022

  2. [2]

    Gludovatz, A

    B. Gludovatz, A. Hohenwarter, D. Catoor, E. H. Chang, E. P. George, and R. O. Ritchie, Science 345, 1153 (2014)

    2014

  3. [3]

    Gludovatz, A

    B. Gludovatz, A. Hohenwarter, K. V. S. Thurston, H. Bei, Z. Wu, E. P. George, and R. O. Ritchie, Nat. Commun. 7, 10602 (2016)

    2016

  4. [4]

    E. P. George, D. Raabe, and R. O. Ritchie, Nat. Rev. Mater. 4, 515 (2019)

    2019

  5. [5]

    Strmcnik, P

    D. Strmcnik, P. P. Lopes, B. Genorio, V. R. Stamenkovic, and N. M. Markovic, Nano Energy Electrocatalysis, 29, 29 (2016)

    2016

  6. [6]

    Nakaya and S

    Y. Nakaya and S. Furukawa, Chem. Rev.123, 5859 (2023)

    2023

  7. [7]

    C. Liu, F. Li, L.-P. Ma, and H.-M. Cheng, Advanced Materials 22, E28 (2010)

    2010

  8. [8]

    G. J. Snyder and E. S. Toberer, Nature Mater 7, 105 (2008)

    2008

Show all 94 references
  1. [9]

    Cantor, Progress in Materials Science 120, 100754 (2021)

    B. Cantor, Progress in Materials Science 120, 100754 (2021)

    2021

  2. [10]

    Potyrailo, K

    R. Potyrailo, K. Rajan, K. Stoewe, I. Takeuchi, B. Chisholm, and H. Lam, ACS Comb. Sci. 13, 579 (2011)

    2011

  3. [11]

    W. F. Maier, ACS Comb. Sci. 21, 437 (2019)

    2019

  4. [12]

    A. Jain, G. Hautier, C. Moore, S. Ong, C. Fischer, T. Mueller, K. Persson, and G. Ceder, Comput. Mater. Sci. 50, 2295 (2011)

    2011

  5. [13]

    Curtarolo, G

    S. Curtarolo, G. L. W. Hart, M. B. Nardelli, N. Mingo, S. Sanvito, and O. Levy, Nat. Mater. 12, 191 (2013)

    2013

  6. [14]

    Bergerhoff, R

    G. Bergerhoff, R. Hundt, R. Sievers, and I. D. Brown, J. Chem. Inf. Comput. Sci. 23, 66 (1983)

    1983

  7. [15]

    M. J. Mehl, D. Hicks, C. Toher, O. Levy, R. M. Hanson, G. Hart, and S. Curtarolo, Comput. Mater. Sci. 136, S1 (2017)

    2017

  8. [16]

    Blaiszik, K

    B. Blaiszik, K. Chard, J. Pruyne, R. Ananthakrishnan, S. Tuecke, and I. Foster, JOM 68, 2045 (2016)

    2016

  9. [17]

    J. A. Vita, E. G. Fuemmeler, A. Gupta, G. P. Wolfe, A. Q. Tao, R. S. Elliott, S. Martiniani, and E. B. Tadmor, J. Chem. Phys. 159, 154802 (2023)

    2023

  10. [18]

    Fuemmeler, G

    E. Fuemmeler, G. Wolfe, A. Gupta, J. A. Vita, E. B. Tadmor, and S. Martiniani, inAI for Accelerated Materials Design-NeurIPS 2024 (2024)

    2024

  11. [19]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications 13, 2453 (2022)

    2022

  12. [20]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Csanyi, Advances in Neural Information Processing Systems 35, 11423 (2022)

    2022

  13. [21]

    Chen and S

    C. Chen and S. P. Ong, Nature Computational Science 2, 718 (2022)

    2022

  14. [22]

    G. H. Booth, A. Gr¨ uneis, G. Kresse, and A. Alavi, Nature 493, 365 (2013)

    2013

  15. [23]

    N. Zaki, H. Park, R. M. Osgood, A. J. Millis, and C. A. Marianetti, Phys. Rev. B 89, 205427 (2014)

    2014

  16. [24]

    E. B. Isaacs and C. A. Marianetti, Phys. Rev. B 102, 045146 (2020)

    2020

  17. [25]

    C. J. Pickard and R. J. Needs, J. Phys.: Condens. Matter 23, 053201 (2011)

    2011

  18. [26]

    W. W. Tipton and R. G. Hennig, Journal of Physics: Condensed Matter 25, 495401 (2013)

    2013

  19. [27]

    A. R. Oganov, C. J. Pickard, Q. Zhu, and R. J. Needs, Nat Rev Mater 4, 331 (2019)

    2019

  20. [28]

    Merchant, S

    A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Nature 624, 80 (2023)

    2023

  21. [29]

    T. Xie, X. Fu, O.-E. Ganea, R. Barzilay, and T. Jaakkola, Crystal Diffusion Variational Autoencoder for Periodic Material Generation (2022), arXiv:2110.06197 [cs]

    2022 arXiv

  22. [30]

    R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu, and Y. Liu, Crystal Structure Prediction by Joint Equivariant Diffusion (2023)

    2023

  23. [31]

    C. Zeni, R. Pinsler, D. Z¨ ugner, A. Fowler, M. Hor- ton, X. Fu, Z. Wang, A. Shysheya, J. Crabb´ e, S. Ueda, R. Sordillo, L. Sun, J. Smith, B. Nguyen, H. Schulz, S. Lewis, C.-W. Huang, Z. Lu, Y. Zhou, H. Yang, H. Hao, J. Li, C. Yang, W. Li, R. Tomioka, and T. Xie, Nature 10.10...

    2025 doi

  24. [32]

    B. K. Miller, R. T. Q. Chen, A. Sriram, and B. M. Wood, FlowMM: Generating Materials with Rieman- nian Flow Matching (2024), arXiv:2406.04713 [cond-mat, physics:physics, stat]

    2024 arXiv

  25. [33]

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden, Stochastic Interpolants: A Unifying Framework for Flows and Diffusions (2023), arXiv:2303.08797 [cond-mat]

    2023 arXiv

  26. [34]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, Deep Unsupervised Learning using Nonequi- librium Thermodynamics (2015), arXiv:1503.03585 [cond- mat, q-bio, stat]

    2015 arXiv

  27. [35]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-Based Generative Mod- eling through Stochastic Differential Equations (2021), arXiv:2011.13456 [cs]

    2021 arXiv

  28. [36]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Den- ton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mah- davi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi, Photorealistic Text-to-Image Diffu- sion Models with Deep Language Understanding (2022), arXiv:2205.11487 [cs]

    2022 arXiv

  29. [37]

    Corso, H

    G. Corso, H. St¨ ark, B. Jing, R. Barzilay, and T. Jaakkola, DiffDock: Diffusion Steps, Twists, and Turns for Molecu- lar Docking (2023), arXiv:2210.01776 [q-bio]

    2023 arXiv

  30. [38]

    Liu, Rectified Flow: A Marginal Preserving Approach to Optimal Transport (2022), arXiv:2209.14577 [stat]

    Q. Liu, Rectified Flow: A Marginal Preserving Approach to Optimal Transport (2022), arXiv:2209.14577 [stat]. 11

    2022 arXiv

  31. [39]

    Lipman, R

    Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, Flow Matching for Generative Modeling (2023), arXiv:2210.02747 [cs, stat]

    2023 arXiv

  32. [40]

    M. S. Albergo and E. Vanden-Eijnden, Building Nor- malizing Flows with Stochastic Interpolants (2023), arXiv:2209.15571 [cs, stat]

    2023 arXiv

  33. [41]

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, G. Wolf, and Y. Bengio, Improving and generalizing flow-based generative models with minibatch optimal transport (2024), arXiv:2302.00482

    2024 arXiv

  34. [42]

    R. T. Q. Chen and Y. Lipman, Flow Matching on General Geometries (2024), arXiv:2302.03660 [cs, stat]

    2024 arXiv

  35. [43]

    I. Gat, T. Remez, N. Shaul, F. Kreuk, R. T. Q. Chen, G. Synnaeve, Y. Adi, and Y. Lipman, Discrete Flow Matching (2024), arXiv:2407.15595 [cs]

    2024 arXiv

  36. [44]

    Campbell, J

    A. Campbell, J. Yim, R. Barzilay, T. Rainforth, and T. Jaakkola, Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design (2024), arXiv:2402.04997 [cs, q-bio, stat]

    2024 arXiv

  37. [45]

    D. W. Davies, K. T. Butler, A. J. Jackson, J. M. Skelton, K. Morita, and A. Walsh, Journal of Open Source Software 4, 1361 (2019)

    2019

  38. [46]

    V. G. Satorras, E. Hoogeboom, and M. Welling, in Pro- ceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 139, edited by M. Meila and T. Zhang (PMLR, 2021) pp. 9323–9332

    2021

  39. [47]

    AI4Science, A

    M. AI4Science, A. Hernandez-Garcia, A. Duval, A. Volokhova, Y. Bengio, D. Sharma, P. L. Carrier, Y. Benabed, M. Koziarski, and V. Schmidt, Crystal- GFN: Sampling crystals with desirable properties and constraints (2023), arXiv:2310.04925 [cs]

    2023

  40. [48]

    Z. Cao, X. Luo, J. Lv, and L. Wang, Space Group In- formed Transformer for Crystalline Materials Generation (2024)

    2024

  41. [49]

    R. Zhu, W. Nong, S. Yamazaki, and K. Hippalgaonkar, WyCryst: Wyckoff Inorganic Crystal Generator Frame- work (2024), arXiv:2311.17916

    2024 arXiv

  42. [50]

    Kazeev, R

    N. Kazeev, R. Zhu, I. Romanov, A. E. Ustyuzhanin, S. Yamazaki, W. Nong, and K. Hippalgaonkar, in AI for Accelerated Materials Design - NeurIPS 2024(2024)

    2024

  43. [51]

    R. Jiao, W. Huang, Y. Liu, D. Zhao, and Y. Liu, Space Group Constrained Crystal Generation (2024), arXiv:2402.03992 [cond-mat]

    2024 arXiv

  44. [52]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Computational Materials Science 68, 314 (2013)

    2013

  45. [53]

    MatterGen, Mattergen, Microsoft (2025)

    2025

  46. [54]

    Full workflow details are given in Appendix D 4

    is utilized for initial structural relaxation, which is subsequently followed by a more computationally expen- sive DFT relaxation. Full workflow details are given in Appendix D 4. In addition to the stability-based met- rics, we also report the root mean squared displacement ...

    1917

  47. [55]

    H. Yang, C. Hu, Y. Zhou, X. Liu, Y. Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni, M. Horton, R. Pinsler, A. Fowler, D. Z¨ ugner, T. Xie, J. Smith, L. Sun, Q. Wang, L. Kong, C. Liu, H. Hao, and Z. Lu, MatterSim: A Deep Learning Atomistic Model Across Elements, Tempera- tures an...

    2024 arXiv

  48. [56]

    I. E. Castelli, D. D. Landis, K. S. Thygesen, S. Dahl, I. Chorkendorff, T. F. Jaramillo, and K. W. Jacobsen, Energy & Environmental Science 5, 9034 (2012)

    2012

  49. [57]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Materials 1, 011002 (2013)

    2013

  50. [58]

    S. G. Baird, H. M. Sayeed, J. Montoya, and T. D. Sparks, Journal of Open Source Software 9, 5618 (2024)

    2024

  51. [59]

    Schmidt, N

    J. Schmidt, N. Hoffmann, H.-C. Wang, P. Borlido, P. J. M. A. Carri¸ co, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, Large-scale machine-learning- assisted exploration of the whole materials space (2022), arXiv:2210.00579 [cond-mat]

    2022 arXiv

  52. [60]

    Schmidt, H.-C

    J. Schmidt, H.-C. Wang, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, Scientific Data 9, 64 (2022)

    2022

  53. [61]

    C. J. Pickard, Airss data for carbon at 10gpa and the c+n+h+o system at 1gpa (2020)

    2020

  54. [62]

    Sriram, B

    A. Sriram, B. Miller, R. T. Chen, and B. Wood, Ad- vances in Neural Information Processing Systems 37, 46025 (2024)

    2024

  55. [63]

    J. Yim, A. Campbell, A. Y. K. Foong, M. Gastegger, J. Jim´ enez-Luna, S. Lewis, V. G. Satorras, B. S. Veel- ing, R. Barzilay, T. Jaakkola, and F. No´ e, Fast protein backbone generation with SE(3) flow matching (2023), arXiv:2310.05297

    2023 arXiv

  56. [64]

    A. J. Bose, T. Akhound-Sadegh, G. Huguet, K. Fa- tras, J. Rector-Brooks, C.-H. Liu, A. C. Nica, M. Ko- rablyov, M. Bronstein, and A. Tong, SE(3)-Stochastic Flow Matching for Protein Backbone Generation (2024), arXiv:2310.02391

    2024 arXiv

  57. [65]

    M. S. Albergo, M. Goldstein, N. M. Boffi, R. Ranganath, and E. Vanden-Eijnden, Stochastic interpolants with data- dependent couplings (2024), arXiv:2310.03725 [cs, stat]

    2024 arXiv

  58. [66]

    J. Ho, A. Jain, and P. Abbeel, Denoising Diffusion Prob- abilistic Models (2020), arXiv:2006.11239 [cs]

    2020 arXiv

  59. [67]

    Aranguri, G

    S. Aranguri, G. Biroli, M. Mezard, and E. Vanden-Eijnden, Optimizing Noise Schedules of Generative Models in High Dimensions (2025), arXiv:2501.00988 [cs]

    2025 arXiv

  60. [68]

    A. Q. Nichol and P. Dhariwal, in Proceedings of the 38th International Conference on Machine Learning, Proceed- ings of Machine Learning Research, Vol. 139, edited by M. Meila and T. Zhang (PMLR, 2021) pp. 8162–8171

    2021

  61. [69]

    Song and S

    Y. Song and S. Ermon, in Advances in Neural Informa- tion Processing Systems, Vol. 32, edited by H. Wallach, H. Larochelle, A. Beygelzimer, F. d 'Alch´ e-Buc, E. Fox, and R. Garnett (Curran Associates, Inc., 2019)

    2019

  62. [70]

    R. Liaw, E. Liang, R. Nishihara, P. Moritz, J. E. Gonzalez, and I. Stoica, arXiv preprint arXiv:1807.05118 (2018)

    2018 arXiv

  63. [71]

    Bergstra, D

    J. Bergstra, D. Yamins, and D. D. Cox, TProc. of the 30th International Conference on Machine Learning (ICML

  64. [72]

    A. M. Ganose, H. Sahasrabuddhe, M. Asta, K. Beck, T. Biswas, A. Bonkowski, J. Bustamante, X. Chen, Y. Chi- ang, D. Chrzan, J. Clary, O. Cohen, C. Ertural, M. Gal- lant, J. George, S. Gerits, R. Goodall, R. Guha, G. Hau- tier, M. Horton, A. Kaplan, R. Kingsbury, M. Kuner, B. Li...

    2025

  65. [73]

    A. Togo, K. Shinohara, and I. Tanaka, Spglib: A software library for crystal symmetry search (2024), arXiv:1808.01590 [cond-mat]

    2024 arXiv

  66. [74]

    Gruver, A

    N. Gruver, A. Sriram, A. Madotto, A. G. Wilson, C. L. Zitnick, and Z. Ulissi, Fine-Tuned Language Models Generate Stable Inorganic Materials as Text (2024), arXiv:2402.04379 [cs]. 12 Appendix A: Acronyms Table IV provides a list of acronyms used throughout this paper for refer...

    2024 arXiv

  67. [75]

    Depending on the specific task, there are several stochastic interpolants at once

    OMatG F ramework Figures 4 and 5 summarize the training and the integration pipeline of the OMatG framework, respectively. Depending on the specific task, there are several stochastic interpolants at once. For CSP, one stochastic interpolant considers lattice vectors L, and an...

  68. [76]

    Interpolant Choice In this work, we are concerned with spatially linear interpolants of the form specified in Eq. (1). The following conditions must be met [33]: α(0) = β(1) = 1, α (1) = β(0) = γ(0) = γ(1) = 0, γ (t) > 0 ∀t ∈ (0, 1). (B1) Under these constraints, the form of t...

  69. [77]

    [33], the loss function can become unstable around t = 0 and t = 1 for certain choices of γ(t)

    Antithetic Sampling As shown by Albergo et al. [33], the loss function can become unstable around t = 0 and t = 1 for certain choices of γ(t). To account for this, we implement antithetic sampling. This requires simultaneously computing the loss at both x+ and x− where x+(t, x...

  70. [78]

    Albergo et al

    Diffusion Coefficient An important inference-time parameter for models integrated with an SDE is the choice of ϵ(t) ≥ 0 which plays the role of a diffusion coefficient. Albergo et al. [33] note that the presence of γ−1(t) in the drift term seen in Fig. 5 can pose a numerical i...

  71. [79]

    We do not attempt to generalize stochastic interpolants (SIs) to arbitrary manifolds as in Riemannian flow matching [42]

    Interpolation with Periodic Boundary Conditions We adopt a task-specific formulation for handling periodic boundary conditions with SIs tailored to flat tori, which are the relevant manifolds for fractional coordinates in crystal generation. We do not attempt to generalize sto...

  72. [80]

    (The geodesic is the same as the linear interpolant wrapped back into the box.) 15 We perform this procedure for all choices of interpolants

    = (1 − t)x0 + tx′ 1, as if in Euclidean space and finally wrap the interpolated path back onto the torus. (The geodesic is the same as the linear interpolant wrapped back into the box.) 15 We perform this procedure for all choices of interpolants. The reason for unwrapping acc...

  73. [81]

    As discussed, a parameterized neural network pθ 1|t(x1|xt) is learned, which attempts to predict the final sequence from the sequence at time t

    DFM Details DFM allows for generative modeling of discrete sequences of tokens while respecting the discrete nature of the design space. As discussed, a parameterized neural network pθ 1|t(x1|xt) is learned, which attempts to predict the final sequence from the sequence at tim...

  74. [82]

    For instance, Miller et al

    V elocity Annealing Velocity annealing—rescaling the learned velocity field during generation to increase velocity over time as bθ(t, x) → (1 +st) bθ(t, x) with s as an hyperparameter during integration—has been empirically shown to improve performance in a number of studies t...

  75. [83]

    Data-Dependent Coupling SIs have been used with data-dependent couplings [ 64], where a coupling function ν(x0, x1) enables biasing of x0 based on the sampled x1. In OMatG, we incorporate an optional data-dependent coupling that enforces an ordering (i.e., a permutation on the...

  76. [84]

    from structures x0 ∈ ρ0 and x1 ∈ ρ1. We find that the inclusion of this data-dependent coupling is optimal during hyperparameter tuning depending on the type of model: CSP models typically performed better without this coupling, but DNG models (see Tab. XI) can benefit in cert...

  77. [85]

    FlowMM is naturally subsumed by OMatG

    SI unifies CFM and SBDM The SI framework implemented in OMatG unifies the frameworks of CFM, as implemented in FlowMM [ 32], and SBDM, as implemented in DiffCSP [ 30] and MatterGen [ 31]. FlowMM is naturally subsumed by OMatG. For the choice of ODE-based sampling, the velocity...

  78. [86]

    The notable differences between OMatG-Linear models and FlowMM are as follows: (1) Discrete flow matching on species for OMatG vs

    Comparison of OMatG-Linear to FlowMM A subset of OMatG models, specifically those which use linear interpolants for both the fractional coordinates and lattice vectors, map closely onto the conditional flow-matching model FlowMM [ 32]. The notable differences between OMatG-Lin...

  79. [87]

    Graph Neural Network We implement a message-passing graph neural network (GNN) with CSPNet as introduced in Jiao et al. [30]: hi (0) = ϕh(0) (ai) (C1) mij (s) = φm hi s−1, hj s−1, l, SinusoidalEmbedding(xj − xi) (C2) mi (s) = NX j=1 mij (s) (C3) hi (s) = hi (s−1) + φh(h(s−1), ...

  80. [88]

    (C7) For the CSP task, the last line is left out

    Loss F unction With Eqs (2), (3), and (4), we can construct a loss function for the modeling of our joint distribution of interest for the DNG task, L(θ) =Et,z,x0,x1 λx,b |bθ x(t, xt)|2 − 2∂tx(t, x0, x1, z) · bθ x(t, xt) + λx,z |zθ x(t, xt)|2 − 2zθ x(t, xt) · z + λl,b |bθ l (t...

  81. [89]

    Hyperparameter Optimization For every choice of the positional interpolant, sampling scheme, and latent variableγ, an independent hyperparameter optimization was performed using the Ray Tune package [69] in conjunction with the HyperOpt Python library [ 70] for Bayesian optimi...

  82. [90]

    We show in Fig

    Match Rate and RMSE The tradeoff between match rate and RMSE most strongly influences the perov-5 dataset. We show in Fig. 7 how different positional interpolants for the atomic coordinates (trigonometric vs. linear with ODE sampling schemes) learn to generate matched structur...

  83. [91]

    The compositional validity is defined according to the SMACT software package [45]

    V alidity Metrics The structural validity of generated structures is defined according to the bond lengths present in the structure—all lengths must be >0.5 ˚A to be considered valid. The compositional validity is defined according to the SMACT software package [45]. We note t...

  84. [92]

    For structural matching, we use CrystalNN, and for compositional matching, we use Magpie fingerprints

    Coverage and Property Statistics As in [ 29], we evaluate coverage recall and precision (reported as rates) by measuring the percentage of crystals in the test set and in the generated samples that match each other within a defined fingerprint distance threshold. For structura...

  85. [93]

    Rates TABLE XII

    Calculation of S.U.N. Rates TABLE XII. Stability (defined as ≤ 0.1 eV/atom above hull), uniqueness, and novelty results from de novo generation on the MP-20 dataset computed for the same models as in Tab. II. All evaluations are performed with the MatterGen code base [ 53] wit...

    1984

  86. [94]

    Stability and Structural Analysis of Generated Structures In Fig. 8, we show the distribution of computed energies above the convex hull across various OMatG models, showing best stability of generated structures for linear, encoder-decoder, trigonometric, and VP SBD positiona...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.