Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Equivariant Riemannian stochastic interpolants can generate equilibrium configurations for amorphous particle systems, and importance-sampling estimates of energy, specific heat, and radial distribution converge with fewer samples when peri

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 15:28 UTC pith:FAVTOS64

load-bearing objection A genuinely new equivariant stochastic-interpolant framework for amorphous systems, with correct-looking theory and honest ablations, but the abstract's own warning about likelihood-integration error is never quantified, leaving the central importance-sampling results less secure than they appear. the 3 major comments →

arxiv 2512.16607 v2 pith:FAVTOS64 submitted 2025-12-18 stat.ML cond-mat.stat-mechcs.LGphysics.comp-ph

Boltzmann generators for amorphous particle systems

Louis Grenioux , Leonardo Galliano , Ludovic Berthier , Giulio Biroli , Marylou Gabri\'e This is my paper
classification stat.ML cond-mat.stat-mechcs.LGphysics.comp-ph
keywords Boltzmann generatorsamorphous materialsstochastic interpolantsequivariant graph neural networksimportance samplingperiodic boundary conditionsglass transitionflow matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Amorphous materials like glasses are notoriously hard to sample at equilibrium because local Monte Carlo and molecular dynamics relax slowly, and existing Boltzmann generators do not respect the periodic boundaries and symmetries (permutations, translations, signed axis flips) that define these disordered systems. This paper closes that gap by building those constraints into Riemannian stochastic interpolants: it proves that if the base and target distributions are invariant under the symmetry group and the interpolation path is equivariant, then the time-marginals and optimal velocity field inherit the symmetry, and it implements the velocity with a torus-adapted equivariant graph neural network. On a two-dimensional inverse-power-law glass former, importance-sampling estimates of potential energy, specific heat, and radial distribution function converge to reference values with fewer samples than an equivariant Euclidean flow or a non-equivariant torus flow. The paper also reports a central limitation: numerical errors accumulated while integrating the likelihood break time-reversibility, which compromises exact thermodynamic reweighting.

Core claim

The central claim is that a Boltzmann generator for amorphous particle systems can be built by taking a stochastic interpolant on the flat torus, where particle coordinates are wrapped by periodic boundary conditions, and enforcing the full symmetry group of the Boltzmann measure: particle permutations, translations, and signed permutations of coordinate axes. The paper proves that a G_C-equivariant interpolation between G_C-invariant base and target distributions yields G_C-invariant marginals and a G_C-equivariant optimal velocity field, and that the graph-neural-network velocity parameterization is Lipschitz-bounded and G_C-equivariant. With this construction, generated configurations car

What carries the argument

The central object is the equivariant Riemannian stochastic interpolant: an interpolation path on the flat torus chosen to be G_C-equivariant, with the optimal velocity field v*(t, x) = E[∂t I(t, X0, X1) | Xt = x]. The paper proves that when the interpolant is equivariant and the endpoint distributions are invariant, the marginals and this conditional-expectation velocity are equivariant. The velocity is parameterized by an equivariant graph neural network operating on the torus, and the log-likelihood is obtained by integrating the divergence along the ODE flow. This identity carries the argument: it converts group symmetry of the data into group symmetry of the learned generative process,

Load-bearing premise

The method only works if the generated distribution covers every region where the Boltzmann target has significant probability, and if the numerically integrated log-likelihoods are accurate enough that time-reversibility holds; the paper itself states that accumulated numerical errors currently break this reversibility.

What would settle it

Take a trained eRSI model, integrate the probability-flow ODE forward and backward on the same samples, and check that log p_forward equals log p_backward within a tight tolerance; then recompute the importance-sampling estimates of U and c_V with a substantially stricter ODE tolerance or an exactly reversible integrator and compare with the reported values. If the estimates shift by more than the statistical error, numerical likelihood error is demonstrably causing the non-exact reweighting described in the paper.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Boltzmann generators can now target amorphous materials: for the 2D IPL model, importance-sampling estimates of U, c_V, and g(r) converge with about 10^3 samples where baselines need 10^4 samples or never converge.
  • Encoding periodic boundaries and G_C symmetries is not cosmetic; the comparisons show it directly prevents unphysical collapsed states and boundary overlaps that bias observables.
  • The equivariance theorems generalize within flow matching and stochastic interpolants: any G_C-invariant base and target with a G_C-equivariant interpolation inherits invariant marginals and an equivariant optimal velocity field.
  • The documented likelihood-integration error implies that continuous-flow Boltzmann generators need an additional mechanism—exactly reversible integrators, discrete flows, or acceptance corrections—before exact thermodynamic reweighting can be guaranteed.
  • The approach scales to larger systems and to other glass formers, as demonstrated by the additional Kob-Andersen experiments, though 3D validation remains open.

Where Pith is reading between the lines

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

  • A decisive test of the paper's stated limitation is to run the same model with forward and backward ODE likelihood integration and compare the importance-sampling estimates of U and c_V; if they disagree beyond Monte Carlo error, the numerical time-reversibility violation is quantitatively responsible for residual bias.
  • The support assumption is the deeper load-bearing condition: at temperatures lower than those tested, target modes may fall outside the generated support, and finite importance-sampling estimates would silently omit part of the Boltzmann distribution.
  • The observed ESS pattern—decaying as 1/R before a plateau—can be used as a likelihood-free model-selection heuristic: the sample count at which ESS departs from 1/R measures how well the proposal overlaps the target.
  • The equivariance mechanism likely transfers to other manifolds with the same affine-plus-modulo group action, which would cover crystal and other periodic materials beyond the two-component glass model studied here.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces equivariant Riemannian stochastic interpolants (eRSI) for sampling equilibrium configurations of amorphous particle systems. It combines Riemannian stochastic interpolants on the flat torus with a G_C-equivariant graph neural network, enforcing periodic boundary conditions and the symmetries of multi-component particle systems. Theoretical contributions include proofs that G_C-invariant base/target distributions and G_C-equivariant interpolants yield invariant marginals (Prop. 4) and equivariant optimal velocity fields (Prop. 9), as well as an equivariance proof for the proposed GNN architecture (Prop. 34). Empirically, the authors compare eRSI against an equivariant-Euclidean flow-matching baseline (eFM) and a non-equivariant RSI baseline on 2D IPL glass models with N=10 and N=44. They report that eRSI produces samples from which importance-sampling estimates of potential energy, specific heat, and radial distribution function converge to ground-truth values with fewer samples than the baselines (Figs. 3–4). The abstract further claims an 'intrinsic limitation' of continuous flows: accumulated numerical errors during likelihood integration break time-reversibility and compromise exact thermodynamic reweighting.

Significance. If the empirical claims hold, this is a useful step toward generative sampling of glass-forming systems, a domain where existing Boltzmann generators do not directly apply. The theoretical equivariance results (Props. 4, 9, 33, 34) are clearly original and provide a principled foundation for architecture design on the torus. The paper also ships reproducible code, uses independently generated Monte Carlo training data, and evaluates against ground-truth observables rather than only visual fidelity. The central claim that explicit symmetry and geometry constraints improve sample efficiency is plausible and supported by the N=10 results. However, the abstract-level claim about an 'intrinsic limitation' from likelihood-integration error is not substantiated in the body, and the numerical accuracy of the likelihoods used in importance sampling is not quantified, which is load-bearing for the unbiasedness of the reported estimates.

major comments (3)
  1. [Section F.3, Eq. (8), Figs. 3–4] The IS estimates use weights w = p*/q, which require the exact model density q. The log-density is obtained by numerical integration of Eq. (8) with dopri5 at atol=rtol=1e-5, yet the abstract itself states that 'accumulated numerical errors during likelihood integration break time-reversibility, compromising exact thermodynamic reweighting.' The body never quantifies this error, gives no error bars on the IS curves, and does not show that the error is small relative to Monte Carlo sampling error. Without such evidence, the observed convergence to ground truth could be affected by biased weights. Please add (i) a tolerance-sensitivity study, for example comparing atol/rtol 1e-4, 1e-5, 1e-6, (ii) a forward/backward log-likelihood consistency check, and (iii) error bars or confidence intervals on the U and c_V curves, at least for the N=10 system.
  2. [Abstract vs. Section 6] The abstract's strong claim that the experiments 'reveal an intrinsic limitation of the continuous-flow formulation' is not supported by any experiment or analysis in Sections 5–6. Section 6 only mentions the computational cost of likelihood evaluation. The phrase 'intrinsic limitation' is a much stronger statement than 'numerical integration at the chosen tolerance introduces error.' Either provide a systematic demonstration that the error is inherent to continuous-flow likelihood integration (e.g., cannot be reduced by tighter tolerances or better integrators) or temper the claim to describe a practical limitation of the current implementation.
  3. [Fig. 4 and Section 5 (N=44)] For N=44, the eFM baseline is evaluated only up to R=8e3 samples, while eRSI is evaluated up to R=1e6 (Section F.4). The text states that eFM 'stabilizes away from the ground truth,' but given that a large fraction of eFM samples are discarded due to boundary overlaps and that no error bars are reported, the apparent failure may be a small-sample effect. Please report the effective sample size for eFM at R=8e3 and include confidence intervals on the U and c_V estimates, or extend the eFM run to a larger R if computationally feasible. Otherwise the N=44 comparison between eFM and eRSI is not fully conclusive.
minor comments (4)
  1. [Abstract (arXiv metadata vs. full text)] The abstract quoted in the manuscript metadata ('Boltzmann generators for amorphous particle systems... intrinsic limitation...') differs substantially from the abstract at the top of the full text. The two should be harmonized, especially regarding the 'intrinsic limitation' claim.
  2. [Eq. (13)] In Eq. (13), the expression 'dM(X(i), X(j)) + 1 \hat{\phi}_d(M^k_{ij})' appears to be missing a division symbol or parentheses. It should likely read '(dM(...) + 1) / \hat{\phi}_d(...)' or similar. Please clarify.
  3. [Section 5 / Table 1] The text says all architectures have the same number of trainable parameters, but Table 1 lists different hyperparameter choices for different systems. It would help to state the exact parameter counts for eRSI, eFM, and RSI in one place.
  4. [Section F.4] The notation R is used both for the number of generated samples and for the radial coordinate in g(r). This is slightly confusing in Figures 3–4 and the ESS discussion; consider renaming the sample size to N_s or R_s.

Circularity Check

0 steps flagged

No significant circularity: theoretical equivariance proofs are self-contained and the empirical benchmarks compare against independently generated ground-truth MC data.

full rationale

The paper's central derivation chains are self-contained. Propositions 4, 9, 33, and 34 are proved in the appendices from the definitions of G_C invariance, equivariant interpolants, the conditional-expectation velocity field, and the GNN update equations; they do not assume the conclusions. The training objective (9) and the IS estimator (Section E) are standard and do not define the target observables in terms of the model output. The reported U, c_V, and g(r) values are compared against ground truth computed from independently generated Monte Carlo samples (Zenodo dataset, Grenioux et al. 2025), and the reweighted estimates use the fitted model density inside the importance weights, so the comparison is not forced by construction. The abstract's statement that accumulated numerical errors in likelihood integration 'break time-reversibility, compromising exact thermodynamic reweighting' is an acknowledged correctness/robustness limitation, not a circularity: it concerns whether the implemented weights equal the intended model density, not whether the result is built into the input. The only self-citations (Jung et al. 2024, Gabrié et al. 2022, Grenioux et al. 2025) appear as related work, future-work suggestions, or dataset provenance; none is load-bearing for the central equivariance results or the empirical comparisons. No specific reduction of a claimed prediction to a fitted parameter or to a self-citation chain can be exhibited, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced; the framework adds a learned equivariant velocity field on configuration space. The main free choices are numerical tolerances and network hyperparameters, plus the minibatch-OT training coupling. The axioms are mostly standard manifold/IS assumptions, but the support condition and the numerical likelihood accuracy are load-bearing for the reweighting claim.

free parameters (3)
  • ODE solver tolerances (atol/rtol) = 1e-5
    Chosen by hand for likelihood integration; the abstract claims accumulated numerical errors at this tolerance break time-reversibility, so reported IS accuracy depends on this choice.
  • GNN hyperparameters (K=3 layers, width 32) = K=3, HF=32
    Selected via validation loss; larger variants (K=5, HF=128) improve fidelity (Fig. 10), so headline results use the smallest architecture.
  • Minibatch-OT coupling (Hungarian per species)
    Training detail affecting learned paths; only particle-permutation invariance is used in the OT coupling, and the authors note other invariances are ignored for computational budget.
axioms (5)
  • domain assumption The flat torus M with geodesic interpolant exp_x(t log_x(y)) is the correct configuration-space geometry for amorphous particle systems under PBC.
    Section 2.3; inherited from Wu et al. (2025) and Chen & Lipman (2024); the cut-locus non-smoothness of d_M/log is not addressed.
  • domain assumption The Boltzmann distribution p* (and its fixed-composition conditionals) is G_C-invariant.
    Proved in Appendix A (Prop 21/22) from the pair-potential form of W; if W had multi-body orientation-dependent terms this would fail.
  • domain assumption The neural velocity field is Lipschitz-bounded so ODE solutions are unique and transport maps are diffeomorphisms.
    Proposition 34 asserts this for the GNN, but it is not verified for the trained networks.
  • ad hoc to paper The learned flow's output distribution has support containing the target support (supp(p*) ⊆ supp(model)).
    Required for unbiased IS (Section E); the ESS/energy-overlap analysis in Figures 8–9 suggests the model places little mass on low-energy target modes at finite R, so this is not guaranteed.
  • ad hoc to paper The log-likelihood from the instantaneous change-of-variables formula is accurately approximated by numerical integration with dopri5 at tolerance 1e-5.
    The abstract states this is compromised ('accumulated numerical errors during likelihood integration break time-reversibility'), so the IS weights may be biased.

pith-pipeline@v1.3.0-alltime-deepseek · 28429 in / 14018 out tokens · 138139 ms · 2026-08-03T15:28:15.518486+00:00 · methodology

0 comments
read the original abstract

Sampling configurations in thermodynamic equilibrium is a long-standing challenge in statistical physics. Boltzmann generators address this problem by employing generative models to propose independent configurations, which are then reweighted via importance sampling using exact likelihood evaluations. Recent Boltzmann Generators based on continuous normalizing flows and flow matching have achieved significant success for particle systems and biomolecules. However, these approaches have not been extended to amorphous materials (glasses), for which equilibrium sampling is notoriously slow. Because of their disordered structure, the invariances and geometrical constraints of amorphous materials differ from those of crystals and biomolecules, preventing the direct use of existing generative models. Here, we develop Boltzmann Generators tailored to amorphous materials by building the required equivariances directly into Riemannian stochastic interpolants. Our framework incorporates periodic boundary conditions and particle symmetries using equivariant graph neural networks. Numerical experiments demonstrate that enforcing physical symmetries significantly improves the accuracy of Boltzmann Generators, but also reveal an intrinsic limitation of the continuous-flow formulation: accumulated numerical errors during likelihood integration break time-reversibility, compromising exact thermodynamic reweighting. These results reveal a fundamental challenge for continuous-flow generative models in statistical mechanics and call for alternative approaches that preserve exact thermodynamic consistency.

Figures

Figures reproduced from arXiv: 2512.16607 by Giulio Biroli, Leonardo Galliano, Louis Grenioux, Ludovic Berthier, Marylou Gabri\'e.

Figure 1
Figure 1. Figure 1: Illustration of invariance group actions on a configuration of the 2D 10-particle IPL model. The system contains two particle species with different effective diameters (see Section 5). The symmetrized transformation shown corresponds to a 90◦ counterclockwise rotation – equivalently, an axial symmetry with respect to the diagonal from the bottom-left to the top-right corner. Equilibrium distribution. Deno… view at source ↗
Figure 2
Figure 2. Figure 2: Samples from p⋆ and compared generative models on the 44-particle IPL system. eFM denotes a model trained with standard FM that uses an equivariant velocity respecting system symmetries but ignores the torus geometry, thus generating unphysical particle overlaps near bound￾aries. RSI uses a non-equivariant velocity field. eRSI is the proposed approach, combining RSI with an equivariant velocity field. at l… view at source ↗
Figure 3
Figure 3. Figure 3: Results for N = 10 particles. (a) Mean potential energy U and (b) specific heat cV as a function of the number of generated samples R for RSI, eFM, and eRSI. (c–e) Radial distribution function g(r) for the three models, showing target, direct model, and reweighted estimates. RSI fails completely, eFM partially recovers observables, and eRSI performs best. R 100 101 102 103 104 105 106 U 102 104 106 108 101… view at source ↗
Figure 4
Figure 4. Figure 4: Results for N = 44 particles. (a) Mean potential energy U, and (b) specific heat cV as a function of the number of generated samples R for RSI, eFM, and eRSI. (c) Radial distribution function g(r) for eRSI averaged over 1.8 × 106 samples. RSI and eFM deviate, due to collapsed states (RSI) or boundary overlaps (eFM). RSI samples were of too poor quality to produce cV estimates. Only eRSI remains consistent … view at source ↗
Figure 5
Figure 5. Figure 5: Two trajectories of particle configurations generated by ODE integration with an equiv [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional samples for the IPL system with N = 10. Configurations generated by the eRSI model (left) compared with reference samples from the dataset (right). F.2 DESIGNS OF THE VELOCITY FIELDS In the GNN implementation (14), we set K = 3. The networks ϕˆ e and ϕˆ h are 3-layer MLPs with width 32, while ϕˆm is a 2-layer MLP with width 32. Species are embedded using one-hot encoding. The network has a total… view at source ↗
Figure 7
Figure 7. Figure 7: Additional samples for the IPL system with N = 44. Configurations generated by the eRSI model (left) compared with reference samples from the dataset (right). Additional samples from eRSI Additional samples from both the reference models and the dataset are shown in Figures 6 and 7. Effective sample size and energy overlap. A central difficulty in reweighting is that the ESS (see Section E) can be extremel… view at source ↗
Figure 8
Figure 8. Figure 8: Deep dive into the effective sample size (ESS) for the N = 10 system. (Top): Nor￾malized ESS (see Section E) as a function of the number of samples R. This metric reflects the proportion of samples effectively contributing to expectation estimates and provides an upper bound on the estimator variance (Agapiou et al., 2017). (Bottom): Histograms of the target energy observ￾able U⋆ for samples generated by e… view at source ↗
Figure 9
Figure 9. Figure 9: Deep dive into the effective sample size (ESS) for the N = 44 system. (Top): Nor￾malized ESS (see Section E) as a function of the number of samples R. This metric reflects the proportion of samples effectively contributing to expectation estimates and provides an upper bound on the estimator variance (Agapiou et al., 2017). (Bottom): Histograms of the target energy observ￾able U⋆ for samples generated by e… view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of model sizes on the IPL system with N = 44. Variants of the EGNN architecture (14) are evaluated with K ∈ 3, 4, 5 and hidden feature size (i.e., the width of each neural network denoted HF) ∈ 32, 64, 128. All models were trained and selected following the procedure in Section F.3. (Left): distribution of energies from 8192 generated samples per model. (Right): corresponding radial distributio… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of dataset sizes on the IPL system with N = 44. Models were trained on datasets of size 104 and 105 , both generated following the procedure described in Section F.1. The GNN used has the same architecture as in Section F.2. (Left): Energy distribution of 8192 gen￾erated samples. (Right): corresponding radial distribution functions. Larger datasets yield notably improved fidelity in both metric… view at source ↗
Figure 12
Figure 12. Figure 12: Energy and structural statistics for the KA system with N = 44 at two temperatures. The left panels show histograms of the potential energy U⋆ evaluated on 8192 samples from the model, while the right panels display the corresponding radial distribution functions. Results are shown at T = 1.0 (top row) and T = 0.32 (bottom row). We trained a simple eRSI model using the smallest GNN architecture (see Secti… view at source ↗
Figure 13
Figure 13. Figure 13: Additional equilibrium samples for the KA system with N = 44 at two temper￾atures. Configurations generated by the eRSI model (left) and reference dataset samples (right), shown at T = 1.0 (top row) and T = 0.32 (bottom row). 32 [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. ATLAS: A Foundation Neural Sampler for Amorphous Materials

    cond-mat.mtrl-sci 2026-07 conditional novelty 6.0

    ATLAS is a force-trained diffusion sampler that generates Boltzmann-distributed amorphous structures, estimates free energies, and drives multi-objective inverse design of metallic glasses.

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [5]

    Gerhard Jung, Giulio Biroli, and Ludovic Berthier

    URLhttps://openreview.net/forum?id=89QPmZjIhv. Gerhard Jung, Giulio Biroli, and Ludovic Berthier. Predicting dynamic heterogeneity in glass- forming liquids by physics-inspired machine learning.Physical Review Letters, 130(23), June

  2. [7]

    Walter Kob and Hans C

    URLhttps://proceedings.neurips.cc/paper_files/paper/2023/ file/bc827452450356f9f558f4e4568d553b-Paper-Conference.pdf. Walter Kob and Hans C. Andersen. Testing mode-coupling theory for a supercooled binary lennard-jones mixture i: The van hove correlation function.Phys. Rev. E, 51:4626–4641, May

    2023

  3. [9]

    URLhttps://www.science.org/doi/full/ 10.1126/science.adv9817

    doi: 10.1126/science.adv9817. URLhttps://www.science.org/doi/full/ 10.1126/science.adv9817. Publisher: American Association for the Advancement of Science. Honglin Li, Chuhao Liu, Yongfeng Guo, Xiaoshan Luo, Yijie Chen, Guangsheng Liu, Yu Li, Ruoyu Wang, Zhenyu Wang, Jianzhuo Wu, Cheng Ma, Zhuohang Xie, Jian Lv, Yufei Ding, Huabin Zhang, Jian Luo, Zhichen...

    Pith/arXiv arXiv 2025

  4. [10]

    Benjamin Kurt Miller, Ricky T

    URLhttps://proceedings.neurips.cc/paper_files/paper/2023/ file/fa55eb802a531c8087e225ecf2dcfbca-Paper-Conference.pdf. Benjamin Kurt Miller, Ricky T. Q. Chen, Anuroop Sriram, and Brandon M Wood. FlowMM: Generating materials with Riemannian flow matching. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett,...

    2023

  5. [11]

    URLhttps://www.science.org/doi/abs/10

    doi: 10.1126/science.aaw1147. URLhttps://www.science.org/doi/abs/10. 1126/science.aaw1147. Anshul D. S. Parmar, Misaki Ozawa, and Ludovic Berthier. Ultrastable metallic glasses in silico. Phys. Rev. Lett., 125:085505, Aug 2020. doi: 10.1103/PhysRevLett.125.085505. URLhttps: //link.aps.org/doi/10.1103/PhysRevLett.125.085505. 14 Preprint Donna N. Perera and...

    doi:10.1126/science.aaw1147 2020

  6. [12]

    Anuroop Sriram, Benjamin Kurt Miller, Ricky T

    URLhttps://proceedings.neurips.cc/paper_files/paper/2023/ file/01d64478381c33e29ed611f1719f5a37-Paper-Conference.pdf. Anuroop Sriram, Benjamin Kurt Miller, Ricky T. Q. Chen, and Brandon M. Wood. Flowllm: Flow matching for material generation with large language models as base distributions. In A. Glober- son, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. ...

    2023

  7. [1987]

    URLhttps://link.aps.org/doi/10.1103/ PhysRevA.36.4891

    doi: 10.1103/PhysRevA.36.4891. URLhttps://link.aps.org/doi/10.1103/ PhysRevA.36.4891. Ludovic Berthier and Giulio Biroli. Theoretical perspective on the glass transition and amorphous materials.Reviews of Modern Physics, 83(2):587–645, June 2011. ISSN 1539-0756. doi: 10.1103/revmodphys.83.587. URLhttp://dx.doi.org/10.1103/RevModPhys.83. 587. Ludovic Berth...

    work page doi:10.1103/physreva.36.4891 2011

  8. [1995]

    URLhttps://link.aps.org/doi/10.1103/ PhysRevE.51.4626

    doi: 10.1103/PhysRevE.51.4626. URLhttps://link.aps.org/doi/10.1103/ PhysRevE.51.4626. Jonas K ¨ohler, Leon Klein, and Frank Noe. Equivariant flows: Exact likelihood generative learn- ing for symmetric densities. In Hal Daum ´e III and Aarti Singh (eds.),Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machi...

    doi:10.1103/physreve.51.4626 2020

  9. [2019]

    URLhttps://par.nsf.gov/biblio/10322883

    doi: 10.24963/ijcai.2019/103. URLhttps://par.nsf.gov/biblio/10322883. Louis Grenioux, Leonardo Galliano, Ludovic Berthier, Giulio Biroli, and Marylou Gabri´e. Equilib- rium samples from Inverse Power Law and Kob-Andersen glassy systems, December 2025. URL https://zenodo.org/records/17966995. Emiel Hoogeboom, V´ıctor Garcia Satorras, Cl ´ement Vignac, and ...

    arXiv 2019

  10. [2021]

    Tomas Geffner, Kieran Didi, Zhonglin Cao, Danny Reidenbach, Zuobai Zhang, Christian Dallago, Emine Kucukbenli, Karsten Kreis, and Arash Vahdat

    URLhttps://proceedings.neurips.cc/paper_files/paper/2021/ file/21b5680d80f75a616096f2e791affac6-Paper.pdf. Tomas Geffner, Kieran Didi, Zhonglin Cao, Danny Reidenbach, Zuobai Zhang, Christian Dallago, Emine Kucukbenli, Karsten Kreis, and Arash Vahdat. La-Proteina: Atomistic Protein Genera- tion via Partially Latent Flow Matching, July 2025. URLhttp://arxiv...

    Pith/arXiv arXiv 2021

  11. [2023]

    doi: 10.1103/physrevlett.130.238202

    ISSN 1079-7114. doi: 10.1103/physrevlett.130.238202. URLhttp://dx.doi.org/ 10.1103/PhysRevLett.130.238202. Gerhard Jung, Giulio Biroli, and Ludovic Berthier. Normalizing flows as an enhanced sampling method for atomistic supercooled liquids.Machine Learning: Science and Technology, 5(3): 035053, aug 2024. doi: 10.1088/2632-2153/ad6ca0. URLhttps://dx.doi.o...

    work page doi:10.1103/physrevlett.130.238202 2024

  12. [2024]

    Alexander Tong, Kilian FATRAS, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio

    URLhttps://proceedings.neurips.cc/paper_files/paper/2024/ file/51d317df78eded9eb3c9d3fb1091c279-Paper-Conference.pdf. Alexander Tong, Kilian FATRAS, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative mod- els with minibatch optimal transport.Transactions on M...

    Pith/arXiv arXiv 2024

  13. [2025]

    arXiv:2505.22598 [cond-mat]

    URLhttp://arxiv.org/abs/2505.22598. arXiv:2505.22598 [cond-mat]. Ricky T. Q. Chen. torchdiffeq, 2018. URLhttps://github.com/rtqichen/ torchdiffeq. Ricky T. Q. Chen and Yaron Lipman. Flow matching on general geometries. InThe Twelfth Interna- tional Conference on Learning Representations, 2024. URLhttps://openreview.net/ forum?id=g7ohDlTITL. Ricky T. Q. Ch...

    arXiv 2018