Pith. sign in

REVIEW 1 major objections 1 minor 59 references

ATLAS is a neural sampler that generates independent configurations from the Boltzmann distribution of an amorphous material directly from its energy function; on a two-dimensional Kob-Andersen glass former it reproduces parallel-tempering

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

2026-08-01 13:08 UTC pith:4DIOMKMQ

load-bearing objection ATLAS is a serious, well-benchmarked attempt at reference-free Boltzmann sampling for amorphous materials, but the central claim of thermodynamic fidelity rests on empirical convergence of a bootstrapped loop rather than the uniqueness theorem, and the main structural validation uses a churn sampler with a fitted strength. the 1 major comments →

arxiv 2607.19198 v1 pith:4DIOMKMQ submitted 2026-07-21 cond-mat.mtrl-sci cs.LGphysics.comp-ph

ATLAS: A Foundation Neural Sampler for Amorphous Materials

Mouyang Cheng , Denis Blessing , Botao Yu , Gerhard Neumann , Mingda Li , Carles Domingo-Enrich , Yuanqi Du This is my paper
classification cond-mat.mtrl-sci cs.LGphysics.comp-ph
keywords neural sampleramorphous materialsBoltzmann samplingdiffusion generative modelfree energy estimationmetallic glassesfixed-point bootstrappingstochastic interpolant
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.

The paper sets out to show that a generative diffusion sampler can replace slow equilibration in glass physics: ATLAS learns the Boltzmann ensemble of an amorphous material directly from its interatomic potential, without any reference structures. The central validation is on a two-dimensional Kob-Andersen glass former, where the trained sampler reproduces parallel-tempering Monte Carlo structural distributions, free energies, and entropies, achieving below 0.2% free energy error in the low-temperature glass regime while using over 500-fold fewer energy evaluations. If this holds, it means the barriers that trap conventional glass simulations no longer need to be crossed: once trained, the model produces independent equilibrium configurations on demand, estimates thermodynamic quantities, and can steer toward target structural or mechanical properties. The paper further claims the same framework generalizes across temperature, system size, and composition, enabling multi-objective inverse design in metallic-glass chemical spaces.

Core claim

The central claim is that a stochastic interpolant on the periodic torus, connecting a uniform prior to the Boltzmann distribution, can be learned self-consistently from forces alone. Because the prior is uniform on the torus, the score of the interpolant marginal is exactly the conditional expectation of the rescaled interatomic force at the endpoint, βF(x1), with no time-dependent prefactor. Since true Boltzmann endpoints are unavailable, training uses a fixed-point bootstrapping loop: the current sampler generates configurations, the potential labels them with forces, and the network regresses its forward and backward drifts against these force-informed targets. The paper shows that the B

What carries the argument

The load-bearing object is a pair of learned forward and backward drift fields, velocity bθ and score sθ, of a stochastic interpolant on the flat torus, connected to the target by the identity s(x,t)=E[βF(x1)|xt=x]. Together with the path-weight relation that makes the partition function an expectation of forward-backward likelihood ratios, these two fields support generation, free-energy estimation, entropy calculation, and inference-time steering from a single network. Training proceeds by a fixed-point loop—generate, label with forces, interpolate, regress—whose idealized fixed point is uniquely the Boltzmann distribution, and the same fields give Tweedie-style estimates of clean terminal

Load-bearing premise

The load-bearing premise is that the bootstrapped fixed-point iteration converges in practice—under finite network capacity, a finite replay buffer, and stochastic gradients—to the Boltzmann distribution; the proof in the paper guarantees uniqueness of the fixed point but not convergence from a random initialization.

What would settle it

Run the self-training loop from multiple random initializations on a small, exactly solvable two-basin potential whose Boltzmann weights can be enumerated, and compare the sampler's reweighted basin populations and relative free energies to the exact values: if the loop consistently settles at non-Boltzmann basin weights, or the free-energy error fails to shrink as training is extended, the fixed-point convergence premise fails.

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

If this is right

  • For amorphous systems where equilibrated reference ensembles are scarce or biased, ATLAS can generate independent Boltzmann-weighted configurations with no training structures at all.
  • A single temperature-conditioned or composition-conditioned model amortizes sampling across many thermodynamic states, so an expensive machine-learned potential's evaluations are reused across many conditions instead of being spent on repeated simulations.
  • Free energy, entropy, and potentials of mean force follow from the forward-backward path weights, with the absolute scale anchored at a high temperature and propagated to the glass regime by multistate reweighting.
  • Inference-time tilting steers the ensemble toward prescribed short-range order or higher bulk modulus without retraining, and it can accommodate non-differentiable or terminal-only rewards.
  • Composition-amortized pretraining plus an agentic optimizer can search multi-element glass composition spaces, yielding a converged Pareto frontier within 480 evaluations of the interatomic potential.

Where Pith is reading between the lines

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

  • If the fixed-point loop converges as reliably for other differentiable potentials, the same force-labeling scheme should transfer to molecular liquids, polymers, or disordered crystals; a natural test is to apply ATLAS to a small molecule with an exactly computable reference ensemble.
  • The 500-fold advantage counts energy evaluations, not wall-clock time; the trade-off becomes more favorable as the one-time training is amortized over many temperatures, compositions, or design queries, so the honest comparison point is total cost to a converged design campaign.
  • The paper's own importance-weighted training and churn/corrector refinements suggest that free-energy accuracy tracks sampler quality; monitoring the effective sample size of the path weights could serve as a cheap, online diagnostic for whether the self-training loop has converged.
  • Because the method is formulated at fixed NVT, an immediate extension is conditioning the sampler on pressure or volume and testing whether density and local packing respond consistently, which the paper itself flags as a limitation.

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

1 major / 1 minor

Summary. The paper introduces ATLAS, a neural sampler that aims to draw independent samples from the Boltzmann distribution of a given interatomic potential without using reference configurations. The method is built on stochastic interpolants on the flat torus, with a graph-neural-network parameterization of the forward and backward drifts. Training is by fixed-point bootstrapping: the model generates configurations, labels them with forces from the target potential, and regresses its score and velocity heads against force-informed targets. The authors prove a target-score identity on the torus and a uniqueness result for the Boltzmann distribution as the idealized fixed point. They validate the method on a two-dimensional 80:20 Kob-Andersen mixture against non-reversible parallel tempering MCMC, reporting agreement in pair correlation functions, energy distributions, free energies, and entropies, with a claimed below-0.2% free-energy error at T=0.2 at over 500-fold fewer energy evaluations. They further demonstrate temperature and composition amortization, system-size transferability, inference-time steering for potentials of mean force and short-range-order targets, and an LLM-guided inverse-design workflow over multi-component metallic glasses using EAM and MACE-MPA-0 potentials. A substantial supplementary information section contains proofs, algorithms, and hyperparameter details.

Significance. If the central claims are correct, ATLAS represents a significant advance: it offers a way to sample Boltzmann-distributed amorphous configurations directly from an energy function, without training on biased reference ensembles, and it couples this with path-weight-based free-energy estimation and amortization over temperature and composition. The paper is unusual in providing an external PT-MCMC benchmark, a formal uniqueness theorem for the training fixed point, detailed algorithms, and honest statements about limitations. The 2D KA validation, if robust, would be a strong demonstration of thermodynamic fidelity in a glassy regime where direct MCMC is very costly. The metallic-glass and inverse-design results are plausible but are secondary to the central Boltzmann-correctness claim. The main reason the paper cannot be accepted as is is that the load-bearing convergence property of the bootstrap loop is not established, and the strongest reported validation results are produced with an empirically tuned churn correction whose effect is not disentangled from the learned model itself.

major comments (1)
  1. No additional major issues.
minor comments (1)
  1. [SI 1.7] The churn sampler is introduced as using the learned score head, but the connection between the learned score and the target Boltzmann score is not stated explicitly. Since this is a central inference-time correction, a short sentence clarifying that the corrector is approximate to the extent that s_theta differs from the true score would help.

Circularity Check

0 steps flagged

No significant circularity: the central Boltzmann-correctness claim is validated against an external PT-MCMC reference and the key identities are re-derived in the SI.

full rationale

The load-bearing claim — that ATLAS generates Boltzmann-distributed configurations directly from a target energy — is not equivalent to any input. Training uses only pointwise forces on self-generated configurations; the fixed-point iteration's unique fixed point is proved in SI 1.3 (Proposition 2), not merely imported from the co-authored bridge-matching sampler [33]. The target-score identity (Eq. 11) is re-derived in SI 1.2. The path-weight estimator (Eq. S22) is exact for any learned drift because the forward/backward kernels telescope to the partition function; its use as an absolute high-temperature anchor plus MBAR is an internal self-consistency scheme, and the low-temperature free-energy curves are checked against an external non-reversible PT-MCMC reference (SI 2.1). The structural and energetic benchmarks are also against PT-MCMC, and the empirically chosen churn strength lambda(T)=4T^{-5/4} (SI 2.2) is a sampler hyperparameter that uses the learned score rather than reference energies; it does not reduce the Boltzmann claim by construction. The main caveat is a correctness/rigor gap: SI 1.3 proves uniqueness of the fixed point only under exact regression, unbounded capacity, and exact expectations, not convergence of the stochastic-gradient loop actually run. That is an unproven assertion that could threaten the empirical claims, but it is not a circular step and is not hidden by an equivalent input — it is openly testable against the external PT-MCMC benchmark.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

ATLAS introduces no new physical entities, particles, forces, dimensions, or conserved quantities; all novel content is algorithmic. The central claims rest on the force-only fixed-point training convergence, the accuracy of input potentials, and several hand-chosen sampler and potential-modification parameters. These are the items a reader 'did not pay for upstream' beyond the method itself.

free parameters (4)
  • Churn strength λ(T) = 4T^{-5/4} = 4T^{-5/4}
    Empirically determined temperature-dependent correction strength used for the reported KA structural/energetic distributions; introduced in SI §2.2, not derived from theory.
  • Churn hyperparameters r and h_max = r = 0.16, h_max = 0.05
    Chosen by hand for the SNR-matched churn corrector; directly affects the quality of the structural/energy validation in Fig. 2a (SI §2.2).
  • Noise schedule endpoints = KA: σ_min = 0.01, σ_max = 1.0; metallic: σ_min = 0.02, σ_max = 0.8, ρ_EDM = 7
    Hand-selected diffusion schedules that define the transport; different schedules change the difficulty of learning and the integration error (SI §2.2).
  • Soft-core modification cutoffs of the target potential = 1.5 Å for metallic systems; 0.8σ_AA for KA
    Ad hoc regularization of the potential below short distances. The paper asserts it does not affect the equilibrium region, but provides no direct quantitative check of the perturbed ensemble (Methods, Target systems).
axioms (6)
  • domain assumption Uniform prior on the torus plus zero-center-of-mass restriction makes the target score identity exact.
    Proposition 1 in SI 1.2 relies on p0 = Unif(T^D); if the prior were not uniform, the force-only identity s(x,t) = E[βF(x1)|xt=x] would acquire a different form or fail.
  • domain assumption The fixed-point iteration converges in practice despite finite capacity, finite replay buffer, and stochastic optimization.
    SI 1.3 proves uniqueness of the fixed point under idealized regression, not convergence. Training uses finite batches, a finite buffer, and an approximate network, so convergence is an unproved premise supported only by the benchmarks.
  • domain assumption The target interatomic potentials (EAM, MACE-MPA-0) are sufficiently accurate for the thermodynamic and mechanical claims.
    ATLAS samples the ensemble of a given U(x); any inaccuracy of U is inherited. The authors explicitly caveat that MACE-MPA-0 mechanical predictions are model-guided hypotheses (Discussion).
  • ad hoc to paper Soft-core continuation of U below short distances does not perturb the equilibrium ensemble.
    The potential is modified below 1.5 Å or 0.8σ_AA to prevent collapsed configurations; if low-temperature equilibrium has non-negligible weight in that region, the sampled Boltzmann distribution is for a different energy function.
  • ad hoc to paper The straight-through surrogate gradient for bulk modulus provides a valid reward direction.
    The gradient of the EOS-based surrogate is approximated by ignoring the relaxation trajectory; the paper states the approximation is 'sufficient' for guidance but does not quantify the bias (Methods, Evaluation of target properties).
  • domain assumption Vegard-style elemental-volume mixing rule sets the fixed NVT cell volume.
    All metallic-glass results use volumes from Eq. (16). If the mixing rule deviates from equilibrium amorphous density, the sampled NVT ensemble differs from experimental conditions; the paper reports ~1% agreement on tested configurations but this is not a rigorous guarantee.

pith-pipeline@v1.3.0-alltime-deepseek · 35428 in / 15003 out tokens · 163610 ms · 2026-08-01T13:08:27.914837+00:00 · methodology

0 comments
read the original abstract

Amorphous materials exhibit exceptional mechanical and functional properties, yet their rugged energy landscapes are notoriously difficult to sample. Below the glass-transition temperature, conventional molecular dynamics and Monte Carlo become inefficient because equilibration relies on rare barrier-crossing events, while data-driven generative models are constrained by scarce and biased reference ensembles. Here, we introduce ATLAS, an efficient sampler that learns a diffusion process to generate Boltzmann-distributed amorphous structures directly from a target energy function. Parameterized by an equivariant graph neural network, ATLAS generalizes across system size, temperature, and composition. By exploiting the time reversal of the diffusion process, it enables efficient estimation of thermodynamic quantities and steering toward target observables. In two-dimensional Kob-Andersen systems, ATLAS reproduces parallel tempering Markov chain Monte Carlo structural distributions, free energies and entropies, achieving below 0.2% free energy error in the low-temperature glass regime with over 500-fold fewer energy evaluations. In Cu-Zr and Cr-Co-Ni metallic glasses, ATLAS recovers experimentally observed short-range-order trends and steers structures toward prescribed order parameters and optimized bulk moduli. Moreover, composition-amortized pretraining outperforms composition-specific training from scratch, reduces inverse-design costs by several hundred-fold, and enables sampling with expensive universal machine learning interatomic potentials. Coupled to a large language model agent, ATLAS searches an eight-element space for high-entropy metallic glasses balancing stiffness and ductility, identifying a converged Pareto frontier within 480 oracle evaluations. Together, these results establish ATLAS as a foundation model for sampling, steering and designing amorphous materials.

Figures

Figures reproduced from arXiv: 2607.19198 by Botao Yu, Carles Domingo-Enrich, Denis Blessing, Gerhard Neumann, Mingda Li, Mouyang Cheng, Yuanqi Du.

Figure 1
Figure 1. Figure 1: Foundation neural sampling of amorphous materials. a. Molecular dynamics (MD) and Markov chain Monte Carlo (MCMC) can suffer from slow barrier-crossing when traversing rugged energy landscapes. ATLAS instead learns stochastic transport, through forward and backward stochastic differential equations (SDEs), between a uniform prior under periodic boundary conditions and the Boltzmann ensemble, enabling direc… view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Conditional generation and property tilting with ATLAS on metallic glasses. a. Schematic of conditional generation from a pretrained ATLAS model. The sampler can be adapted by direct fine-tuning or steered at inference time without retraining. In the training-free route, particles evolve along the generative trajectory and are progressively biased toward a target reward, with unsuccessful trajectories down… view at source ↗
Figure 4
Figure 4. Figure 4: Agentic inverse design of amorphous materials with ATLAS as a foundation model. a. Schematic of ATLAS-driven composition-space optimization. A pretrained ATLAS model is amortized over chemical compositions and coupled to either classical force fields or MLIPs. For each proposed composition, ATLAS samples amorphous structures and returns ensemble-evaluated target properties to an LLM-based agentic optimizer… view at source ↗

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

59 extracted references · 8 linked inside Pith

  1. [1]

    Mauro, J. C. & Zanotto, E. D. Two centuries of glass research: historical trends, current status, and grand challenges for the future.Int. J. Appl. Glass Sci.5, 313–327 (2014)

    2014

  2. [2]

    S., Simine, L

    Liu, Y ., Madanchi, A., Anker, A. S., Simine, L. & Deringer, V . L. The amorphous state as a frontier in computational materials design.Nat. Rev. Mater.10, 228–241 (2025). 3.Zachariasen, W. H. The atomic arrangement in glass.J. Am. Chem. Soc.54, 3841–3851 (1932)

    2025

  3. [4]

    Cheng, Y . & Ma, E. Atomic-level structure and structure–property relationship in metallic glasses.Prog. materials science56, 379–473 (2011)

    2011

  4. [5]

    Shi, Z., Ge, H. & He, Y . The emergence of amorphous materials approaching the single-layer limit.Chem. Sci. 16, 16392–16413 (2025)

    2025

  5. [6]

    D.et al.Photochemical route for accessing amorphous metal oxide materials for water oxidation catalysis.Science340, 60–63 (2013)

    Smith, R. D.et al.Photochemical route for accessing amorphous metal oxide materials for water oxidation catalysis.Science340, 60–63 (2013). 7.Hong, S.et al.Ultralow-dielectric-constant amorphous boron nitride.Nature582, 511–514 (2020). 8.Tian, H.et al.Disorder-tuned conductivity in amorphous monolayer carbon.Nature615, 56–61 (2023). 9.Cai, Z.et al.Ceramic...

    2013

  6. [10]

    & Coslovich, D

    Ninarello, A., Berthier, L. & Coslovich, D. Models and algorithms for the next generation of glass transition studies.Phys. Rev. X7, 021039 (2017)

    2017

  7. [11]

    & Reichman, D

    Berthier, L. & Reichman, D. R. Modern computational studies of the glass transition.Nat. Rev. Phys.5, 102–116 (2023)

    2023

  8. [12]

    & Ninarello, A

    Leoni, F., Ozawa, M., Russo, J., Yanagishima, T. & Ninarello, A. Computational methods towards ultrastable glasses.arXiv preprint arXiv:2605.02679(2026)

    Pith/arXiv arXiv 2026

  9. [13]

    & Binder, K

    V ollmayr, K., Kob, W. & Binder, K. Cooling-rate effects in amorphous silica: A computer-simulation study. Phys. Rev. B54, 15808 (1996)

    1996

  10. [14]

    & Kob, W

    Yamamoto, R. & Kob, W. Replica-exchange molecular dynamics simulation for supercooled liquids.Phys. Rev. E61, 5473 (2000)

    2000

  11. [15]

    & Pastore, G

    Gazzillo, D. & Pastore, G. Equation of state for symmetric non-additive hard-sphere fluids: An approximate analytic expression and new monte carlo results.Chem. physics letters159, 388–392 (1989)

    1989

  12. [16]

    Grigera, T. S. & Parisi, G. Fast monte carlo algorithm for supercooled soft spheres.Phys. Rev. E63, 045102 (2001)

    2001

  13. [17]

    Mater.1–17 (2026)

    Cheng, M.et al.Artificial intelligence-driven approaches for materials design and discovery.Nat. Mater.1–17 (2026)

    2026

  14. [18]

    & Jaakkola, T

    Xie, T., Fu, X., Ganea, O.-E., Barzilay, R. & Jaakkola, T. S. Crystal diffusion variational autoencoder for periodic material generation. InInternational Conference on Learning Representations(2022)

    2022

  15. [19]

    Neural Inf

    Jiao, R.et al.Crystal structure prediction by joint equivariant diffusion.Adv. Neural Inf. Process. Syst.36, 17464–17497 (2023). 20.Zeni, C.et al.A generative model for inorganic materials design.Nature639, 624–632 (2025)

    2023

  16. [21]

    & Schwalbe-Koda, D

    Yang, K. & Schwalbe-Koda, D. A generative diffusion model for amorphous materials.npj Comput. Mater. (2025)

    2025

  17. [22]

    & Gabrié, M

    Grenioux, L., Galliano, L., Berthier, L., Biroli, G. & Gabrié, M. Riemannian stochastic interpolants for amorphous particle systems.arXiv preprint arXiv:2512.16607(2025)

    Pith/arXiv arXiv 2025

  18. [23]

    Li, H.et al.Conditional generative modeling for amorphous multi-element materials.arXiv preprint arXiv:2503.07043(2025). 34/37

    Pith/arXiv arXiv 2025

  19. [24]

    Noé, F., Olsson, S., Köhler, J. & Wu, H. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science365, eaaw1147 (2019)

    2019

  20. [25]

    Gabrié, M., Rotskoff, G. M. & Vanden-Eijnden, E. Adaptive monte carlo augmented with normalizing flows. Proc. Natl. Acad. Sci.119, e2109420119 (2022)

    2022

  21. [26]

    & Chen, Y

    Zhang, Q. & Chen, Y . Path integral sampler: A stochastic control approach for sampling. InInternational Conference on Learning Representations(2022)

    2022

  22. [27]

    Vargas, F., Grathwohl, W. S. & Doucet, A. Denoising diffusion samplers. InInternational Conference on Learning Representations(2023)

    2023

  23. [28]

    & Berner, J

    Richter, L. & Berner, J. Improved sampling via learned diffusions. InInternational Conference on Learning Representations(2024)

    2024

  24. [29]

    & Nüsken, N

    Vargas, F., Padhy, S., Blessing, D. & Nüsken, N. Transport meets variational inference: Controlled monte carlo diffusions. InInternational Conference on Learning Representations(2024)

    2024

  25. [30]

    arXiv preprint arXiv:2502.06685(2025)

    He, J.et al.No trick, no treat: Pursuits and challenges towards simulation-free training of neural samplers. arXiv preprint arXiv:2502.06685(2025)

    Pith/arXiv arXiv 2025

  26. [31]

    J.et al.Adjoint sampling: Highly scalable diffusion samplers via adjoint matching

    Havens, A. J.et al.Adjoint sampling: Highly scalable diffusion samplers via adjoint matching. InForty-second International Conference on Machine Learning(2025)

    2025

  27. [32]

    Liu, G.-H., Choi, J., Chen, Y ., Miller, B. K. & Chen, R. T. Adjoint schrödinger bridge sampler.Adv. Neural Inf. Process. Syst.38, 15673–15708 (2026)

    2026

  28. [33]

    & Neumann, G

    Blessing, D., Richter, L., Berner, J., Malitskiy, E. & Neumann, G. Bridge matching sampler: Scalable sampling via generalized fixed-point diffusion matching. InForty-third International Conference on Machine Learning (2026)

    2026

  29. [34]

    & Wittmer, J

    Li, D., Xu, H. & Wittmer, J. Glass transition of two-dimensional 80–20 kob–andersen model at constant pressure.J. Physics: Condens. Matter28, 045101 (2016)

    2016

  30. [35]

    & Cunningham, J

    Wu, L., Trippe, B., Naesseth, C., Blei, D. & Cunningham, J. P. Practical and asymptotically exact conditional sampling in diffusion models.Adv. Neural Inf. Process. Syst.36, 31372–31403 (2023). 36.He, J., Hernández-Lobato, J. M., Du, Y . & Vargas, F. Rne: plug-and-play diffusion inference-time control and energy-based training. InInternational Conference ...

    2023

  31. [37]

    Robbins, H. E. An empirical bayes approach to statistics. InBreakthroughs in Statistics: Foundations and basic theory, 388–394 (Springer, 1992)

    1992

  32. [38]

    Daw, M. S. & Baskes, M. I. Embedded-atom method: Derivation and application to impurities, surfaces, and other defects in metals.Phys. review B29, 6443 (1984)

    1984

  33. [39]

    Ma, D.et al.Efficient local atomic packing in metallic glasses and its correlation with glass-forming ability. Phys. Rev. B80, 014202 (2009)

    2009

  34. [40]

    Zhang, R.et al.Short-range order and its impact on the crconi medium-entropy alloy.Nature581, 283–287 (2020)

    2020

  35. [41]

    & Caro, M

    Zarrouk, T. & Caro, M. A. Molecular augmented dynamics: Generating experimentally consistent atomistic structures by design.arXiv preprint arXiv:2508.17132(2025)

    arXiv 2025

  36. [42]

    P., Simm, G., Ortner, C

    Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Csányi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields.Adv. neural information processing systems35, 11423–11436 (2022). 43.Batatia, I.et al.A foundation model for atomistic materials chemistry.The J. chemical physics163(2025)

    2022

  37. [44]

    Stabilization of metallic supercooled liquid and bulk amorphous alloys.Acta materialia48, 279–306 (2000)

    Inoue, A. Stabilization of metallic supercooled liquid and bulk amorphous alloys.Acta materialia48, 279–306 (2000). 35/37

    2000

  38. [45]

    & Inoue, A

    Takeuchi, A. & Inoue, A. Classification of bulk metallic glasses by atomic size difference, heat of mixing and period of constituent elements and its application to characterization of the main alloying element.Mater. transactions46, 2817–2829 (2005)

    2005

  39. [46]

    Fragapane, N. L. & Deringer, V . L. Amorphous materials as a frontier challenge for universal interatomic potentials.arXiv preprint arXiv:2607.11384(2026)

    Pith/arXiv arXiv 2026

  40. [47]

    & Sjolander, A

    Bengtzelius, U., Gotze, W. & Sjolander, A. Dynamics of supercooled liquids and the glass transition.J. Phys. C: solid state Phys.17, 5915–5934 (1984). 48.Janssen, L. M. Mode-coupling theory of the glass transition: A primer.Front. Phys.6, 97 (2018)

    1984

  41. [49]

    L.et al.De novo design of protein structure and function with rfdiffusion.Nature620, 1089 – 1100 (2023)

    Watson, J. L.et al.De novo design of protein structure and function with rfdiffusion.Nature620, 1089 – 1100 (2023)

    2023

  42. [50]

    Albergo, M., Boffi, N. M. & Vanden-Eijnden, E. Stochastic interpolants: A unifying framework for flows and diffusions.J. Mach. Learn. Res.26, 1–80 (2025)

    2025

  43. [51]

    & Doucet, A

    De Bortoli, V ., Hutchinson, M., Wirnsberger, P. & Doucet, A. Target score matching.arXiv preprint arXiv:2402.08667(2024)

    Pith/arXiv arXiv 2024

  44. [52]

    & Shaul, N

    Havens, A., Karrer, B. & Shaul, N. Flow sampling: Learning to sample from unnormalized densities via denoising conditional processes.arXiv preprint arXiv:2605.03984(2026)

    Pith/arXiv arXiv 2026

  45. [53]

    Neural Inf

    Du, Y .et al.Feat: Free energy estimators with adaptive transport.Adv. Neural Inf. Process. Syst.38, 142130– 142167 (2026). 54.Jarzynski, C. Nonequilibrium equality for free energy differences.Phys. review letters78, 2690 (1997)

    2026

  46. [55]

    & Jarzynski, C

    Vaikuntanathan, S. & Jarzynski, C. Escorted free energy simulations: Improving convergence by reducing dissipation.Phys. Rev. Lett.100, 190601 (2008)

    2008

  47. [56]

    Shirts, M. R. & Chodera, J. D. Statistically optimal analysis of samples from multiple equilibrium states.The J. chemical physics129(2008). 57.Efron, B. Tweedie’s formula and selection bias.J. Am. Stat. Assoc.106, 1602–1614 (2011)

    2008

  48. [58]

    Torrie, G. M. & Valleau, J. P. Nonphysical sampling distributions in monte carlo free-energy estimation: Umbrella sampling.J. computational physics23, 187–199 (1977)

    1977

  49. [59]

    & Gastegger, M

    Schütt, K., Unke, O. & Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. InInternational conference on machine learning, 9377–9388 (PMLR, 2021)

    2021

  50. [60]

    M., Trautt, Z

    Hale, L. M., Trautt, Z. T. & Becker, C. A. Evaluating variability with atomistic simulations: the effect of potential and calculation methodology on the modeling of lattice and elastic constants.Model. Simul. Mater. Sci. Eng.26, 055003 (2018)

    2018

  51. [61]

    InInternational Conference on Learning Representations, vol

    Wang, H.et al.Efficient evolutionary search over chemical space with large language models. InInternational Conference on Learning Representations, vol. 2025, 51694–51727 (2025)

    2025

  52. [62]

    Novikov, A.et al.Alphaevolve: A coding agent for scientific and algorithmic discovery.arXiv preprint arXiv:2506.13131(2025)

    Pith/arXiv arXiv 2025

  53. [63]

    Parego: A hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems.IEEE transactions on evolutionary computation10, 50–66 (2006)

    Knowles, J. Parego: A hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems.IEEE transactions on evolutionary computation10, 50–66 (2006)

    2006

  54. [64]

    & Hernández-Lobato, J

    Grenioux, L., OuYang, R. & Hernández-Lobato, J. M. A diffusive classification loss for learning energy-based generative models. InForty-third International Conference on Machine Learning(2026)

    2026

  55. [65]

    InInternational Conference on Learning Representations(2021)

    Song, Y .et al.Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations(2021)

    2021

  56. [66]

    & Nemoto, K

    Hukushima, K. & Nemoto, K. Exchange monte carlo method and application to spin glass simulations.J. Phys. Soc. Jpn.65, 1604–1608 (1996). 36/37

    1996

  57. [67]

    & Doucet, A

    Syed, S., Bouchard-Côté, A., Deligiannidis, G. & Doucet, A. Non-reversible parallel tempering: A scalable highly parallel mcmc scheme.J. Royal Stat. Soc. Ser. B (Statistical Methodol.84, 321–350 (2022)

    2022

  58. [68]

    D., Pendleton, B

    Duane, S., Kennedy, A. D., Pendleton, B. J. & Roweth, D. Hybrid monte carlo.Phys. letters B195, 216–222 (1987). 69.Neal, R. M. Mcmc using hamiltonian dynamics.Handb. markov chain monte carlo47–95 (2011)

    1987

  59. [70]

    & Laine, S

    Karras, T., Aittala, M., Aila, T. & Laine, S. Elucidating the design space of diffusion-based generative models. Adv. Neural Inf. Process. Syst.35, 26565–26577 (2022). 37/37

    2022