The reviewed record of science sign in
Pith

arxiv: 2607.05266 · v1 · pith:2WU3N3VX · submitted 2026-07-06 · cond-mat.mtrl-sci

A Distributional Framework for Generative Modeling of Molecular Crystals

Michael Kilgour , Alex Dong , Mark E. Tuckerman , Jutta Rogal This is my paper

Reviewed by Pith2026-07-07 20:33 UTCglm-5.2pith:2WU3N3VXopen to challenge →

classification cond-mat.mtrl-sci
keywords crystalsdistributionstructurescrystaldistributionsmoleculargenerativemxtalgflow
0
0 comments X

The pith

Generative model samples Boltzmann distribution of molecular crystals

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

The paper introduces MXtalGFlow, a framework that pairs a canonical 12-dimensional parameterization of molecular crystal structures with energy-based GFlowNet training to sample an approximation of the Boltzmann distribution over crystal packings for a given molecule and space group. The central claim is that, unlike standard crystal structure prediction which produces ranked lists of isolated low-energy candidates, this approach yields a physically interpretable probability distribution over the entire crystal landscape, where the probability mass assigned to each structural basin reflects its thermodynamic weight. The paper demonstrates the framework on two molecules (MIPCAS and NEHZOR) under two energy functions (Lennard-Jones and the Universal Model for Atoms), recovering known experimental polymorphs within high-probability basins and identifying previously unreported packing modes that, under the UMA potential, carry competitive or higher basin probabilities than the experimentally observed structures. The load-bearing mechanism is the combination of a physically grounded parameterization (eliminating unit-cell degeneracy and mapping to a bounded latent space with Jacobian corrections) with a three-stage GFlowNet training protocol that separates mode discovery from distributional equilibration, enabling the model to approximate Boltzmann weights across a rugged, high-dimensional energy surface.

Core claim

By training a diffusion-sampler GFlowNet on a canonical crystal parameterization with Boltzmann rewards, the authors produce a sampled distribution over molecular crystal structures where known polymorphs are recovered as basin maxima and previously unreported packing modes are identified with basin probabilities competitive with or exceeding the experimental structures, as observed for both MIPCAS (P-1) and NEHZOR (P21/c) under the UMA potential. The distributional view reveals that basins with higher minimum energies can still carry higher thermodynamic probability than lower-energy basins due to a larger configurational density of states, a distinction invisible to standard ranked-list晶体C

What carries the argument

The framework rests on three components: (1) a canonical 12-parameter crystal representation (box vectors, centroid position, rotation vector) reduced to the asymmetric unit and mapped to a bounded latent space with Jacobian corrections for volume and orientation; (2) a diffusion-sampler GFlowNet trained via a three-stage protocol (MLE support expansion, backward TB thermalization on a noised prior, forward-backward TB global equilibration) using a Huber-type trajectory balance loss; (3) a distributional analysis pipeline using radial distribution function earth mover's distance (RDF EMD) to define a kernel density estimate P(x), hill-climbing to find probability maxima, and Gaussian basin-7

If this is right

  • The distributional framework could replace ranked-list CSP outputs with thermodynamic probability maps, enabling direct comparison of basin weights under different energy functions and identification of packing modes that are thermodynamically competitive but not energy-minimal.
  • Swapping energy functions (LJ vs UMA vs future machine-learned potentials) becomes a tool for understanding how specific physical interactions (hydrogen bonding, dispersion, sterics) reshape the crystal landscape, rather than merely re-ranking structures.
  • The three-stage training protocol (support expansion, prior thermalization, global equilibration) may transfer to other rugged, high-dimensional Boltzmann sampling problems beyond molecular crystals.
  • Conditional generalization to multiple molecules and space groups in a single model, combined with flexible intramolecular degrees of freedom, could yield an efficient lead generator and probability density estimator across broad chemical spaces.
  • The separation of mode discovery (classical optimization) from distributional equilibration (GFlowNet) acknowledges the complementary strengths of each approach and suggests hybrid pipelines as a practical path forward.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The claim that basin 1 of MIPCAS is thermodynamically preferred over the experimental structure in basin 2 depends critically on the GFlowNet having converged sufficiently to approximate Boltzmann weights; the paper itself acknowledges systematic over-weighting of high-energy states and reports Pearson correlations only on the top 99% of samples, suggesting the quantitative basin rankings should b
  • The kernel density bandwidth (sigma = d_cut/3, d_cut at 15% quantile) and basin assignment threshold (normalized weight > 0.8) are heuristic choices that could affect which packing modes are identified and how probability mass is partitioned; sensitivity analysis on these parameters would strengthen the robustness of the basin rankings.
  • If the framework is extended to flexible molecules (Z' > 1 or multiple conformers), the dimensionality increase may substantially worsen GFlowNet convergence, and the separation of mode discovery from equilibration may need to be revisited.
  • The comparison between LJ and UMA distributions being nearly disjoint for MIPCAS raises the question of whether the framework could be used to benchmark energy functions themselves: if an experimental polymorph falls in a low-probability region under a given potential, that potential may be missing key physics for that system.

Load-bearing premise

The quantitative basin probability rankings that predict new thermodynamically preferred packing modes depend on the GFlowNet having converged closely enough to the true Boltzmann distribution, but the paper acknowledges systematic over-weighting of high-energy states and reports convergence metrics only on the top 99% of samples, without error bars or sensitivity analysis on the basin probability comparisons.

What would settle it

Train the GFlowNet to convergence on a system where the free energy ranking of polymorphs is known from independent methods (e.g., quasi-harmonic vibrational free energy calculations). If the basin probabilities from MXtalGFlow disagree with the known thermodynamic ranking, or if the predicted new preferred packing modes are destabilized upon DFT-level refinement, the distributional claims about thermodynamic preference are not reliable.

Figures

Figures reproduced from arXiv: 2607.05266 by Alex Dong, Jutta Rogal, Mark E. Tuckerman, Michael Kilgour.

Figure 1
Figure 1. Figure 1: Illustration of our crystal parameterization scheme and crystal building workflow. In energy [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diagrams of MIPCAS and NEHZOR conformers in (a) and (b). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Summary energy and density statistics for distributions of crystals sampled from trained [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP embeddings of RDF EMD matrices for 5k samples each from LJ and UMA GFlowNet [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Landscape analysis for the learned P¯1 MIPCAS distribution under UMA. (a) A UMAP embedding of the RDF EMD matrix, with colors identifying assignments to the top basins. (b-f) Energy vs. density plots for each basin, with colors normalized against the probability maximum. (g) Summary statistics for each basin, with the normalized probability densities evaluated at x ∗ , the basin maxima under UMA, and for t… view at source ↗
Figure 6
Figure 6. Figure 6: Landscape analysis for the learned P21/c NEHZOR distribution under UMA. (a) A UMAP embedding of the RDF EMD matrix, with colors identifying assignments to the top basins. (b-e) Energy vs. density plots for each basin, with colors normalized against the probability maximum. (f) Summary statistics for each basin, with the normalized probability densities evaluated at x ∗ , the basin maxima under UMA, and for… view at source ↗
Figure 7
Figure 7. Figure 7: Distributions of asymmetric unit lengths, normed asymmetric unit lengths, and log normed [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Schematic potential energy surface with sample probability distributions for each training [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

Molecular crystals are a highly polymorphic class of materials, with a single molecule commonly crystallizing via multiple packing patterns, making structure and property prediction very challenging. Crystal structure prediction typically comprises the production of sets of promising candidate structures, each considered in isolation rather than as samples in a thermodynamic distribution. Likewise, modern generative approaches to this problem, despite naturally sampling distributions of crystals, lack a concrete formulation of the distributions being sampled. Two components are required to impart meaning to the distributions of crystals generated under such models: a canonical parameterization, and a loss function which equilibrates the generated samples to some target distribution. We develop such a parameterization, and train energy-based generative flow networks (GFlowNets) to approximate the Boltzmann distribution over crystal structures for target molecules and space groups. Combined, these components comprise our MXtalGFlow framework for molecular crystal modeling. Going beyond sampling disconnected sets of low-energy structures, MXtalGFlow yields a thermodynamic distribution over crystal structures. We sample and analyze distributions of crystals for two molecules, each under two energy functions, a Lennard-Jones potential and the Universal Model for Atoms. We characterize the local structural basins about the known polymorphs, and identify additional as-yet un-reported packing modes with competitive probabilities to the known experimental structures. With MXtalGFlow, we illustrate how to define and train a model to sample a thermodynamically meaningful distribution of molecular crystals, and analyze such a distribution to glean useful information.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 5 minor

Summary. The manuscript presents MXtalGFlow, a framework for generative modeling of molecular crystal structures that combines a canonical crystal parameterization with energy-based GFlowNet training. The central claim is that the sampled distribution approximates the Boltzmann distribution over crystal structures, enabling thermodynamically meaningful comparison of structural basins. The authors demonstrate the approach on two molecular systems (MIPCAS and NEHZOR) under two energy functions (Lennard-Jones and UMA), recovering known experimental polymorphs and identifying new competitive packing modes. The framework rests on three pillars: a physically grounded parameterization, a three-phase GFlowNet training protocol, and distributional analysis via RDF-EMD-based probability density estimation.

Significance. The conceptual contribution is timely and well-motivated: framing molecular crystal structure prediction as distributional sampling rather than enumeration of disconnected candidates is a meaningful shift. The canonical parameterization (Section 4.1) is a solid building block, and the three-phase training protocol (Section 4.5) addresses real practical difficulties in GFlowNet convergence on rugged landscapes. The comparison of crystal distributions under different energy functions (Section 2.5) is illustrative. However, the quantitative basin thermodynamic predictions that constitute the most novel claims rest on un-reweighted sample densities from imperfectly converged models, which limits the reliability of the central results as currently presented.

major comments (4)
  1. Section 2.6, panel (g) discussion and Figure 6 panel (f): The claim that 'basin 1 is slightly thermodynamically preferred over the known experimental structure in basin 2' (for MIPCAS) and that 'basins 1 and 2 are thermodynamically preferred compared to the known polymorphs' (for NEHZOR) rests on comparing peak probability densities P(x*) across basins. However, P(x) as defined in Eq. (4) is a KDE over raw model samples without importance reweighting, and the paper explicitly acknowledges that 'GFN models of this type tend to over-weight higher energy states systematically' (Section 2.5). The paper states: 'We omit manual sample reweighting and, instead, present the learned distributions as they are.' This means the basin probability comparisons directly inherit the model's convergence errors. Since the differences are described as 'slightly' preferred, even modest systematic bias could翻
  2. Section 2.3, Eq. (4): P(x*) is the peak density, not the integrated basin probability. Basins with different shapes or widths could have high peak density but low total thermodynamic weight. The paper does not report integrated basin probabilities, which would be the more appropriate thermodynamic quantity. This is a load-bearing distinction for the claim that certain basins are 'thermodynamically preferred.'
  3. Section 2.3, Eq. (4) and surrounding text: The KDE parameters (sigma = d_cut/3, d_cut at 15% quantile) and the basin assignment threshold (0.8) are heuristic choices. No sensitivity analysis is provided to show that the basin probability rankings are robust to these choices. Given that the thermodynamic preference claims are described as marginal ('slightly preferred'), the absence of error bars or sensitivity analysis on these parameters undermines the reliability of the quantitative predictions.
  4. Figure S7 and Section S7: The TB parity plots show visible deviation from the diagonal, and Pearson correlations are reported only on the top 99% of samples by reward. The paper should more transparently report the full convergence metrics (including the full-sample correlations) and discuss quantitatively how the residual convergence error propagates into the basin probability estimates.
minor comments (5)
  1. Section 2.1: The phrase 'specifically specifically' appears to be a typo in the first paragraph of Section 2.2.
  2. Section 2.3: The definition of d_step and the scanning procedure over increasing d_step could be described more precisely — it is unclear what 'stable plateau' means quantitatively.
  3. Figure 5, panel (g): The table would benefit from explicit numerical values for the integrated basin probabilities (or at least basin sample counts) in addition to peak densities.
  4. Section 4.4: The LJ stiffness factor k=2.5 is stated without justification; a brief comment on sensitivity to this choice would help.
  5. The paper would benefit from a clearer statement upfront (e.g., in the abstract or introduction) that the thermodynamic preference claims are conditional on the GFlowNet having converged, and that this convergence is imperfect.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive reading of our manuscript. The referee correctly identifies that the most novel quantitative claims—basin thermodynamic preferences—rest on analysis choices and convergence levels that warrant greater scrutiny than we provided. We address each comment below and outline concrete revisions.

read point-by-point responses

  1. Referee: Basin probability comparisons use un-reweighted sample densities from imperfectly converged models; marginal differences ('slightly preferred') could be flipped by systematic bias.

    Authors: The referee is correct that our basin probability comparisons inherit the model's convergence errors, and that the differences we report for MIPCAS are described as marginal. We acknowledge that presenting these as thermodynamic predictions without error quantification or reweighting overstates the reliability of the result. We will revise the manuscript in two ways. First, we will implement importance reweighting of samples using the trajectory balance residual as an importance weight estimator (following the approach in the Boltzmann Generator literature, Ref. [9]), and re-evaluate the basin probability rankings under reweighting. If the marginal preference for basin 1 over basin 2 in MIPCAS does not survive reweighting, we will report that explicitly and soften the claim accordingly. Second, we will reframe the language throughout: rather than stating that basins are 'thermodynamically preferred,' we will state that the sampled distribution suggests competitive probability density, subject to the convergence limitations we document. The NEHZOR case, where basins 1 and 2 have substantially higher probability densities than the polymorph basins, is more robust to modest bias, but we will still apply reweighting and report both raw and reweighted results. revision: yes

  2. Referee: P(x*) is peak density, not integrated basin probability; basins with different shapes could have high peak but low total thermodynamic weight.

    Authors: This is a valid and important distinction. We used peak density P(x*) as a computationally convenient proxy, but the referee is right that integrated basin probability is the more appropriate thermodynamic quantity. We will add integrated basin probabilities to the summary tables in Figures 5(g) and 6(f), computed by summing the normalized P(x) over all samples assigned to each basin (using the Gaussian basin membership weights already defined in Section 2.3). We will report both peak density and integrated probability, and we will base our thermodynamic preference claims on the integrated quantities. If the rankings change when using integrated probabilities rather than peak densities, we will report that transparently. revision: yes

  3. Referee: KDE parameters (sigma, d_cut) and basin assignment threshold (0.8) are heuristic; no sensitivity analysis provided.

    Authors: We agree that a sensitivity analysis is needed, particularly given that the MIPCAS preference is described as marginal. We will add a sensitivity analysis varying d_cut (at the 10%, 15%, and 20% quantiles), sigma (at d_cut/2, d_cut/3, and d_cut/4), and the basin assignment threshold (at 0.7, 0.8, and 0.9). We will report how the basin probability rankings and integrated basin probabilities change across these parameter choices. This will be included as a new supplementary figure and discussed in the main text. If the marginal MIPCAS result is not robust to these variations, we will state so explicitly. revision: yes

  4. Referee: TB parity plots show deviation from diagonal; correlations reported only on top 99% of samples; should report full-sample metrics and discuss error propagation.

    Authors: The referee is right that reporting convergence metrics only on the top 99% of samples by reward gives an incomplete picture. We will add the full-sample Pearson correlations to Figure S7 and the accompanying text, alongside the top-99% values, so that readers can assess the degree of residual convergence error. We will also add a quantitative discussion of how the TB residual (the deviation from the diagonal) propagates into basin probability estimates. Specifically, we will estimate the effective error on basin probability ratios by propagating the per-sample TB residuals through the KDE-based probability density, giving approximate error bars on the integrated basin probabilities. These error bars will be included in the revised summary tables. revision: yes

Circularity Check

0 steps flagged

No significant circularity; one minor self-citation for parameterization tooling that is not load-bearing for the central thermodynamic claims.

full rationale

The paper's derivation chain is largely non-circular. The GFlowNet is trained on an external energy function (UMA from [15,16] or a standard Lennard-Jones potential), and the prior distribution is constructed from classical local optimization on that same external energy — neither is defined in terms of the paper's outputs. The parameterization scheme is adopted from the authors' own prior work [17, 18], but these are tooling/parameterization papers providing a coordinate mapping, not theoretical results that determine the basin rankings. The central quantitative claim — that certain basins are 'thermodynamically preferred' (Section 2.6, panel g) — is an output of the trained model's sampled distribution (Eq. 4 KDE over GFlowNet samples), not a quantity that is defined in terms of itself. The acknowledged imperfect convergence ('GFN models of this type tend to over-weight higher energy states systematically,' Section 2.5) and the absence of error bars or sensitivity analysis on the KDE parameters are correctness and robustness concerns, not circularity: the model's output is not equivalent to its input by construction. No step in the chain exhibits self-definitional equivalence, fitted-input-renamed-as-prediction, or a self-citation chain that forces the conclusion. The self-citation to [17, 18] for parameterization is minor and does not bear on the thermodynamic basin comparison claims.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 3 invented entities

The framework introduces several heuristic choices (bandwidth, cutoffs, energy filter, basin threshold) that function as effectively free parameters in the analysis pipeline. The core axioms are mostly standard (GFlowNet theory) or domain assumptions (rigid molecule, fixed space group), but the assumption that the prior covers all packing modes and that P(x) is a meaningful thermodynamic proxy are specific to this paper and not independently validated.

free parameters (10)
  • kBT = 2.5 kJ/mol
    Sampling temperature set to approximate room temperature; not fitted but chosen.
  • dcut = 15% quantile of all-to-all RDF EMD
    Cutoff for probability density estimation (Eq. 4); empirically chosen.
  • sigma (bandwidth) = dcut/3
    Kernel density bandwidth; derived from dcut but ultimately heuristic.
  • dstep = scanned until plateau
    Graph step size for hill-climbing basin identification; empirically tuned.
  • energy filter threshold = 15 kBT above minimum
    Post-hoc filter removing ~5% of high-energy samples due to imperfect convergence.
  • LJ stiffness k = 2.5
    Stiffness factor for softened LJ potential (Eq. S12); chosen by hand.
  • reward clip Rrange = 100
    Logarithmic reward clipping range; chosen for training stability.
  • Huber loss beta = 10
    Smooth L1 loss cutoff for TB loss (Eq. S11); chosen for stability.
  • packing coefficient bounds = 0.55-0.95
    Density penalty thresholds (Eq. S13); based on CSD statistics.
  • basin assignment threshold = 0.8
    Normalized basin weight threshold for assignment (Section 2.3); chosen by convention.
axioms (5)
  • standard math GFlowNet trajectory balance loss converges to the Boltzmann distribution when the model has full support and TB loss reaches zero (Eqs. 2-3).
    Established result from [10, 12]; invoked in Section 2.2. The paper relies on this theoretical guarantee but acknowledges imperfect convergence in practice.
  • domain assumption Rigid molecule approximation: intramolecular degrees of freedom are frozen and Z'=1.
    Stated in Section 2.1 and 4.1. This restricts the crystal parameterization to 12 dimensions and excludes conformational polymorphism, which is a significant physical limitation for many real systems.
  • domain assumption Fixed space group: the model generates crystals within a single pre-specified space group.
    Stated in Section 4.1. The parameterization includes only the asymmetric unit, so space group transitions are not sampled. The paper acknowledges this is inappropriate for high-barrier solid-solid transitions [37, 38].
  • ad hoc to paper The prior dataset (constructed from classical local optimization) covers all low-energy packing modes.
    Stated in Section 4.3: 'a reference distribution that defines all areas of nonzero probability density, the support of the target distribution.' If the prior misses a packing mode, the GFlowNet will never sample it, undermining the distributional claims.
  • ad hoc to paper P(x) as defined in Eq. 4 is a meaningful proxy for relative basin thermodynamic probability.
    The paper states P(x) is 'not intended as a proxy or replacement for quantitative free energy estimates' (Section 2.3) but then uses it to rank basins thermodynamically (Section 2.6, panel g). The kernel density estimate with heuristic bandwidth choices is the basis for the key predictions.
invented entities (3)
  • MXtalGFlow framework no independent evidence
    purpose: Combined parameterization + GFlowNet + distributional analysis pipeline for molecular crystals.
    The framework is the paper's contribution; its components are validated on two test systems but not independently reproduced.
  • Three-phase training protocol (expand support, thermalize on prior, global equilibration) no independent evidence
    purpose: Stable GFlowNet training on rugged molecular crystal energy landscapes.
    The protocol is novel and engineered for this problem; no independent benchmark or comparison to alternative protocols is provided.
  • RDF EMD-based probability density P(x) and hill-climbing basin identification no independent evidence
    purpose: Carving sampled crystal distributions into structural basins with probability assignments.
    The analysis pipeline is introduced here; the basin probability rankings that drive predictions depend on its specific choices (dcut, sigma, dstep).

pith-pipeline@v1.1.0-glm · 24682 in / 3658 out tokens · 346468 ms · 2026-07-07T20:33:46.963549+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages · 5 internal anchors

  1. [1]

    Oxtal: An all-atom diffusion model for organic crystal structure prediction,

    E. Jin, A. C. Nica, M. Galkin, J. Rector-Brooks, K. L. K. Lee, S. Miret, F. H. Arnold, M. Bronstein, A. J. Bose, A. Tong, and C.-H. Liu, “Oxtal: An all-atom diffusion model for organic crystal structure prediction,” 12 2025

    work page 2025

  2. [2]

    PackFlow: Generative Molecular Crystal Structure Prediction via Reinforcement Learning Alignment,

    A. Subramanian, E. Pan, J. Nam, M. Weiler, S. Qu, C. W. Park, T. S. Jaakkola, E. Olivetti, and R. Gomez-Bombarelli, “PackFlow: Generative Molecular Crystal Structure Prediction via Reinforcement Learning Alignment,”arXiv, 2026

    work page 2026

  3. [3]

    MolCrystalFlow: Molecular Crystal Structure Prediction via Flow Matching,

    C. Zeng, H. W. Sullivan, T. Egg, M. M. Martirossyan, P. Höllmer, J. Jin, R. G. Hennig, A. Roitberg, S. Martiniani, E. B. Tadmor, and M. Liu, “MolCrystalFlow: Molecular Crystal Structure Prediction via Flow Matching,”Open MIND, pp. 1–20, 2 2026

    work page 2026

  4. [4]

    Orgflow: Generative modeling of organic crystal structures from molecular graphs,

    M. Vahediahmar, M. A. McDonald, and F. Liu, “Orgflow: Generative modeling of organic crystal structures from molecular graphs,”arXiv preprint arXiv:2602.20195, 2 2026

    work page arXiv 2026

  5. [5]

    Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching

    A. Lo, L. Mucko, A. H. Cheng, A. Cai, A. J. Price, W. Matusik, and A. Aspuru-Guzik, “Fast organic crystal structure prediction with unit cell flow matching,”arXiv preprint arXiv:2606.03199, 6 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    Bernstein,Polymorphism in molecular crystals, vol

    J. Bernstein,Polymorphism in molecular crystals, vol. 30. Oxford University Press, 5 2020

    work page 2020

  7. [7]

    Generative models for crystalline materials,

    H. Metni, L. Ruple, L. N. Walters, L. Torresi, J. Teufel, H. Schopmans, J. Östreicher, Y. Zhang, M. Neubert, Y. Koide,et al., “Generative models for crystalline materials,”Advanced Materials, vol. 38, p. e23620, 2 2026

    work page 2026

  8. [8]

    Vector field oriented diffusion model for crystal material generation,

    A. Klipfel, Y. Fregier, A. Sayede, and Z. Bouraoui, “Vector field oriented diffusion model for crystal material generation,” inProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intellige...

    work page 2024

  9. [9]

    Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning,

    F. Noé, S. Olsson, J. Köhler, and H. Wu, “Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning,”Science, vol. 365, p. eaaw1147, 9 2019

    work page 2019

  10. [10]

    Gflownet foundations,

    Y. Bengio, S. Lahlou, T. Deleu, E. J. Hu, M. Tiwari, and E. Bengio, “Gflownet foundations,”Journal of Machine Learning Research, vol. 24, no. 210, pp. 1–55, 2023

    work page 2023

  11. [11]

    Flow network based generative models for non-iterative diverse candidate generation,

    E. Bengio, M. Jain, M. Korablyov, D. Precup, and Y. Bengio, “Flow network based generative models for non-iterative diverse candidate generation,”Advances in neural information processing systems, vol. 34, pp. 27381–27394, 6 2021

    work page 2021

  12. [12]

    A theory of continuous generative flow networks

    S. Lahlou, T. Deleu, P. Lemos, D. Zhang, A. Volokhova, A. Hernandez-Garcia, L. N. Ezzine, Y. Bengio, and N. Malkin, “A theory of continuous generative flow networks,” inInternational Conference on Machine Learning(A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, eds.), vol. abs/2301.12594, pp. 18269–18300, PMLR, Cornell Universi...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Biological sequence design with gflownets,

    M. Jain, E. Bengio, A. Hernandez-Garcia, J. Rector-Brooks, B. F. Dossou, C. A. Ekbote, J. Fu, T. Zhang, M. Kilgour, D. Zhang,et al., “Biological sequence design with gflownets,” inInternational Conference on Machine Learning(K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvári, G. Niu, and S. Sabato, eds.), vol. 162, pp. 9786–9801, PMLR, Cornell University, 3...

    work page 2022

  14. [14]

    Unifying Generative Models with GFlowNets and Beyond

    D. Zhang, R. T. Chen, N. Malkin, and Y. Bengio, “Unifying generative models with gflownets and beyond,”arXiv preprint arXiv:2209.02606, 9 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  15. [15]

    Uma: A family of universal models for atoms,

    B. M. Wood, M. Dzamba, X. Fu, M. Gao, M. Shuaibi, L. Barroso-Luque, K. Abdelmaqsoud, V. Gharakhanyan, J. R. Kitchin, D. S. Levine,et al., “Uma: A family of universal models for atoms,” arXiv preprint arXiv:2506.23971, vol. abs/2506.23971, 6 2025

    work page arXiv 2025

  16. [16]

    Open Molecular Crystals 2025 (OMC25) Dataset and Models

    V. Gharakhanyan, L. Barroso-Luque, Y. Yang, M. Shuaibi, K. Michel, D. S. Levine, M. Dzamba, X. Fu, M. Gao, X. Liu,et al., “Open molecular crystals 2025 (omc25) dataset and models,”arXiv preprint arXiv:2508.02651, 8 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Multi-type point cloud autoencoder: a complete equivariant embedding for molecule conformation and pose,

    M. Kilgour, M. E. Tuckerman, and J. Rogal, “Multi-type point cloud autoencoder: a complete equivariant embedding for molecule conformation and pose,”Machine Learning: Science and Technology, vol. 6, p. 035055, 8 2025

    work page 2025

  18. [18]

    Mxtaltools: A toolkit for machine learning on molecular crystals,

    M. Kilgour, M. E. Tuckerman, and J. Rogal, “Mxtaltools: A toolkit for machine learning on molecular crystals,”Journal of Chemical Information and Modeling, vol. 66, no. 7, pp. 3402–3408, 2026. PMID: 41875147

    work page 2026

  19. [19]

    Spglib: a software library for crystal symmetry search,

    A. Togo, K. Shinohara, and I. Tanaka, “Spglib: a software library for crystal symmetry search,” Science and Technology of Advanced Materials: Methods, vol. 4, p. 2384822, 8 2024

    work page 2024

  20. [20]

    Improved off-policy training of diffusion samplers,

    M. Sendera, M. Kim, S. Mittal, P. Lemos, L. Scimeca, J. Rector-Brooks, A. Adam, Y. Bengio, and N. Malkin, “Improved off-policy training of diffusion samplers,”Advances in Neural Information Processing Systems, vol. 37, pp. 81016–81045, 2 2024

    work page 2024

  21. [21]

    Stable-GFlowNet: Toward Diverse and Robust LLM Red-Teaming via Contrastive Trajectory Balance,

    M. Kwon, S. Baek, M. Kim, J. Yu, D. Han, and J. Kim, “Stable-GFlowNet: Toward Diverse and Robust LLM Red-Teaming via Contrastive Trajectory Balance,”arXiv, 5 2026

    work page 2026

  22. [22]

    Learning Shortest Paths with Generative Flow Networks,

    N. Morozov, I. Maksimov, D. Tiapkin, and S. Samsonov, “Learning Shortest Paths with Generative Flow Networks,”arXiv, 3 2026

    work page 2026

  23. [23]

    Evaluating GFlowNet from partial episodes for stable and flexible policy-based training,

    P. Niu, S. Wu, and X. Qian, “Evaluating GFlowNet from partial episodes for stable and flexible policy-based training,”arXiv, vol. abs/2603.01047, 3 2026

    work page arXiv 2026

  24. [24]

    Partial GFlowNet: Accelerating Convergence in Large State Spaces via Strategic Partitioning,

    X. Yu, X. Wang, R. Zhu, Y. Zhang, and Y. Wang, “Partial GFlowNet: Accelerating Convergence in Large State Spaces via Strategic Partitioning,”Open MIND, 2 2026

    work page 2026

  25. [25]

    Con- trolling Exploration-Exploitation in GFlowNets via Markov Chain Perspectives,

    L. Chen, S. Drapeau, F. Shao, X. Zhu, B. Xue, Y. Song, M. Laurière, and Z. Lin, “Con- trolling Exploration-Exploitation in GFlowNets via Markov Chain Perspectives,”Open MIND, vol. abs/2602.01749, 2 2026

    work page arXiv 2026

  26. [26]

    Loss-Guided Auxiliary Agents for Overcoming Mode Collapse in GFlowNets,

    I. Malek, A. Sharma, and S. Lahlou, “Loss-Guided Auxiliary Agents for Overcoming Mode Collapse in GFlowNets,”arXiv, 5 2025

    work page 2025

  27. [27]

    Thompson sampling for improved exploration in GFlowNets,

    J. Rector-Brooks, K. Madan, M. Jain, M. Korablyov, C.-H. Liu, S. Chandar, N. Malkin, and Y. Bengio, “Thompson sampling for improved exploration in GFlowNets,”arXiv, 6 2023

    work page 2023

  28. [28]

    Reducing overprediction of molecular crystal structures via threshold clustering,

    P. W. Butler and G. M. Day, “Reducing overprediction of molecular crystal structures via threshold clustering,”Proceedings of the National Academy of Sciences, vol. 120, p. e2300516120, 5 2023

    work page 2023

  29. [29]

    Exploration and optimization in crystal structure prediction: Combining basin hopping with quasi-random sampling,

    S. Yang and G. M. Day, “Exploration and optimization in crystal structure prediction: Combining basin hopping with quasi-random sampling,”Journal of Chemical Theory and Computation, vol. 17, pp. 1988–1999, 2 2021

    work page 1988

  30. [30]

    Global analysis of the energy landscapes of molecular crystal structures by applying the threshold algorithm,

    S. Yang and G. M. Day, “Global analysis of the energy landscapes of molecular crystal structures by applying the threshold algorithm,”Communications Chemistry, vol. 5, p. 86, 7 2022

    work page 2022

  31. [31]

    Compack: a program for identifying crystal structure similarity using distances,

    J. A. Chisholm and S. Motherwell, “Compack: a program for identifying crystal structure similarity using distances,”Applied Crystallography, vol. 38, no. 1, pp. 228–231, 2005

    work page 2005

  32. [32]

    First-principles modeling of molecular crystals: structures and stabilities, temperature and pressure,

    J. Hoja, A. M. Reilly, and A. Tkatchenko, “First-principles modeling of molecular crystals: structures and stabilities, temperature and pressure,”Wiley Interdisciplinary Reviews: Computational Molecular Science, vol. 7, p. e1294, 1 2017. 16

    work page 2017

  33. [33]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  34. [34]

    Space Group Conditional Flow Matching,

    O. Puny, Y. Lipman, and B. K. Miller, “Space Group Conditional Flow Matching,”arXiv, 2025

    work page 2025

  35. [35]

    Space Group Equivariant Crystal Diffusion,

    R. Chang, A. Pak, A. Guerra, N. Zhan, N. Richardson, E. Ertekin, and R. P. Adams, “Space Group Equivariant Crystal Diffusion,”arXiv, 5 2025

    work page 2025

  36. [36]

    Crystal-GFN: sampling crystals with desirable properties and constraints,

    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,”arXiv, 10 2023

    work page 2023

  37. [37]

    Solid-solid phase transitions between crystalline polymorphs of organic materials,

    I. B. Rietveld, “Solid-solid phase transitions between crystalline polymorphs of organic materials,” Current Pharmaceutical Design, vol. 29, pp. 445–461, 2 2023

    work page 2023

  38. [38]

    Phase transformations in solids,

    M. J. Buerger, “Phase transformations in solids,”John Wiley, p. 183, 1951

    work page 1951

  39. [39]

    T. Hahn, U. Shmueli, and J. W. Arthur,International tables for crystallography, vol. 1. Reidel Dordrecht, 1983

    work page 1983

  40. [40]

    The cambridge structural database,

    C. R. Groom, I. J. Bruno, M. P. Lightfoot, and S. C. Ward, “The cambridge structural database,” Acta Crystallographica Section B: Structural Science, Crystal Engineering and Materials, vol. 72, pp. 171–179, 4 2016

    work page 2016

  41. [41]

    A unified algorithm for determining the reduced (niggli) cell,

    I. Křivy and B. Gruber, “A unified algorithm for determining the reduced (niggli) cell,”Foundations of Crystallography, vol. 32, pp. 297–298, 3 1976

    work page 1976

  42. [42]

    Topological crystal structure prediction,

    M. Tuckerman and N. Galanakis, “Topological crystal structure prediction,”Research Square, 9 2023

    work page 2023

  43. [43]

    The classical equation of state of gaseous helium, neon and argon,

    R. A. Buckingham, “The classical equation of state of gaseous helium, neon and argon,”Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, vol. 168, pp. 264–283, 10 1938

    work page 1938

  44. [44]

    I. G. Kaplanet al.,Intermolecular interactions: physical picture, computational methods and model potentials, vol. 367. Wiley Online Library, 2006

    work page 2006

  45. [45]

    What has scripting ever done for us? the csd python application programming interface (api),

    R. A. Sykes, N. T. Johnson, C. J. Kingsbury, J. Harter, A. G. Maloney, I. J. Sugden, S. C. Ward, I. J. Bruno, S. A. Adcock, P. A. Wood,et al., “What has scripting ever done for us? the csd python application programming interface (api),”Applied Crystallography, vol. 57, pp. 1235–1250, 7 2024

    work page 2024

  46. [46]

    Generating uniform incremental grids on so (3) using the hopf fibration,

    A. Yershova, S. Jain, S. M. Lavalle, and J. C. Mitchell, “Generating uniform incremental grids on so (3) using the hopf fibration,”The International journal of robotics research, vol. 29, no. 7, pp. 801–812, 2010. 17 S1 Model Details Our GFlowNet model follows closely the implementation in [20]. A forward SDE step is integrated from a constant initial sta...

    work page 2010