pith. sign in

arxiv: 2606.04100 · v1 · pith:HQ2W6LEOnew · submitted 2026-06-02 · 💻 cs.LG · physics.comp-ph

Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials

Joanna Zou , Fraser Birks , Dallas Foster , Youssef Marzouk This is my paper

Pith reviewed 2026-06-28 10:37 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-ph
keywords Stein kernelized molecular dynamicsactive learninginteratomic potentialsmolecular dynamicsenhanced samplingmachine learning potentialsStein variational gradient descent
0
0 comments X

The pith

Stein kernelized molecular dynamics acquires training data for machine learning interatomic potentials by using interacting particles that converge to the Boltzmann distribution.

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

The paper introduces Stein kernelized molecular dynamics (SKMD) to generate informative configurations for active learning of machine learning interatomic potentials. It adapts Stein variational gradient descent into a molecular dynamics framework with asynchronous particle updates and a kernel on global atomic descriptors that respects symmetries. This construction keeps the long-run distribution of sampled configurations equal to the Boltzmann distribution, balancing exploration of new states with focus on likely regions. An adaptive stopping rule then picks non-redundant samples during the run. Experiments on the Müller-Brown potential and alanine dipeptide show the resulting models reach higher accuracy than standard active-learning baselines when trained on the same number of points.

Core claim

SKMD corresponds to a stochastic variant of Stein variational gradient descent adapted for molecular dynamics by incorporating asynchronous particle updates and a kernel of global atomic descriptors, which provides a symmetry-aware measure of configurational similarity. Unlike other enhanced samplers used in molecular dynamics, SKMD preserves the Boltzmann distribution as the asymptotic distribution of the dynamics. This property enforces a balance between the exploration of diverse configurations and attraction toward high-probability regions of the energy landscape. We further propose an approach to efficient online data acquisition using an adaptive stopping criterion that selects non-red

What carries the argument

Interacting particle dynamics driven by a kernel of global atomic descriptors, adapted from Stein variational gradient descent with asynchronous updates, that generates samples converging to the Boltzmann distribution while promoting useful diversity for active learning.

If this is right

  • SKMD balances exploration of diverse configurations with attraction to high-probability regions of the energy landscape.
  • The method preserves the Boltzmann distribution as its long-time limit, unlike many other enhanced samplers.
  • An adaptive stopping criterion allows online selection of non-redundant training configurations.
  • The approach yields higher model accuracy than standard active-learning baselines on both the Müller-Brown potential and alanine dipeptide with the same number of acquired samples.

Where Pith is reading between the lines

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

  • The same particle-interaction structure could be tested on larger molecular systems to measure how the kernel scales with system size.
  • Replacing the kernel with other symmetry-aware descriptors might further improve sampling efficiency in periodic or crystalline materials.
  • The online adaptive acquisition rule could be combined with uncertainty estimates from the current MLIP to prioritize even more informative points.

Load-bearing premise

The chosen kernel of global atomic descriptors must supply a symmetry-aware similarity measure that produces useful particle interactions while still letting the dynamics converge to the Boltzmann distribution.

What would settle it

Run the SKMD dynamics on the Müller-Brown potential for many steps and check whether the histogram of visited configurations converges to the independently computed Boltzmann distribution; alternatively, compare final model error after a fixed number of acquired samples against a non-interacting baseline sampler.

Figures

Figures reproduced from arXiv: 2606.04100 by Dallas Foster, Fraser Birks, Joanna Zou, Youssef Marzouk.

Figure 1
Figure 1. Figure 1: Contours of the neural network potential at iterations {1, 2, 4, 8} of active learning by overdamped Langevin dynamics (top row), UDD (middle row), and a-SKMD (bottom row). The accumulated training data are shown in red, the queried data at the current iteration in cyan, and the path from the previous stopping time to the current stopping time in dark blue. The reference Müller–Brown potential and the init… view at source ↗
Figure 2
Figure 2. Figure 2: Results from active learning with the Müller–Brown potential as the reference. Root mean square error (RMSE) in potential energy (left) and forces (right) of the neural network potential across training iterations for the Langevin (blue), UDD (orange), SKMD (green), and a-SKMD (purple) schemes. The solid lines show the median error and the shaded regions show the 25th to 75th percentile range of the error … view at source ↗
Figure 3
Figure 3. Figure 3: Results from alanine dipeptide enhanced sampling and fine-tuning. (a) A contour map of the E(ψ, ϕ) Ramachandran surface of alanine dipeptide using the MACE-OFF-23-small foundation model. (b) Three minimum energy configurations of alanine dipeptide numbered 1 to 3. The Ramachandran angles (ψ, ϕ) are labeled on 1. (c)–(e) Heat maps of all 1000 samples taken during the 10 iterations of active learning with bo… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the quality of samples from overdamped Langevin dynamics and SKMD with varying stopping time ℓ. Quality is measured in terms of a sample-based estimator of the Wasserstein-2 distance with respect to the Boltzmann distribution. high-dimensional state space. When implementing SKMD with a variable kernel bandwidth, offline data acquisition tends to perform better than online data acquisition. Th… view at source ↗
Figure 5
Figure 5. Figure 5: The neural network potential (contours), previous training data (red), and selected data (cyan) corresponding to 1 iteration of active learning by a-SKMD (a) and UDD (b). The corresponding contours of the acquisition function, selected data (cyan), and threshold for the acquisition criterion (red line) for a-SKMD (c) and UDD (d). the simulation after 32 points have been collected. SKMD is implemented accor… view at source ↗
read the original abstract

Machine learning interatomic potentials (MLIPs) enable efficient and accurate atomistic simulations but depend critically on the quality and diversity of the training data. We introduce Stein kernelized molecular dynamics (SKMD), an enhanced sampling method that uses interacting particle dynamics to acquire informative training configurations for the active learning and fine-tuning of MLIPs. SKMD corresponds to a stochastic variant of Stein variational gradient descent that is adapted for molecular dynamics by incorporating asynchronous particle updates and a kernel of global atomic descriptors, which provides a symmetry-aware measure of configurational similarity. Unlike other enhanced samplers used in molecular dynamics, SKMD preserves the Boltzmann distribution as the asymptotic distribution of the dynamics. This property enforces a balance between the exploration of diverse configurations and attraction toward high-probability regions of the energy landscape. We further propose an approach to efficient online data acquisition using an adaptive stopping criterion that selects non-redundant training data over the course of simulation. We demonstrate SKMD for the active learning of a neural network model of the M\"uller-Brown potential and the fine-tuning of a MACE interatomic potential for alanine dipeptide. Compared to active learning baselines, our method achieves higher model accuracy in fewer training iterations with the same number of acquired training samples.

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

0 major / 3 minor

Summary. The manuscript introduces Stein Kernelized Molecular Dynamics (SKMD), a stochastic variant of Stein variational gradient descent adapted for molecular dynamics via asynchronous particle updates and a kernel defined on global atomic descriptors. The central claims are that the resulting dynamics preserve the Boltzmann distribution as the unique stationary measure (providing a balance between exploration and attraction to high-probability regions) and that an adaptive stopping criterion enables efficient online acquisition of non-redundant training data. On the Müller-Brown potential and alanine dipeptide, SKMD is reported to yield higher MLIP accuracy in fewer training iterations than baselines while using the same number of acquired samples.

Significance. If the claims hold, the work supplies a principled enhanced-sampling tool for active learning of interatomic potentials that maintains the correct equilibrium distribution while promoting configurational diversity through interacting particles. The direct numerical verification of the stationary distribution together with matched-budget active-learning comparisons constitute a concrete advance over heuristic samplers commonly used in this domain.

minor comments (3)
  1. [§3] §3 (Methods): the precise definition of the global atomic descriptor kernel and the proof that the asynchronous update rule leaves the Boltzmann measure invariant should be stated explicitly in the main text rather than deferred entirely to the supplement.
  2. [Figure 4, Table 2] Figure 4 and Table 2: the error bars on the active-learning curves are not described; clarify whether they represent standard deviation over independent runs or another measure.
  3. [§4.2] The adaptive stopping criterion is introduced without a formal statement of its convergence properties; a short remark on why it terminates with high probability would improve clarity.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of our work, recognition of its significance as a principled enhanced-sampling tool for active learning of interatomic potentials, and recommendation for minor revision. No major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained adaptation with external verification

full rationale

The manuscript introduces SKMD as a direct stochastic adaptation of Stein variational gradient descent (a pre-existing method) incorporating asynchronous updates and a global atomic descriptor kernel. The central invariance claim (Boltzmann measure as stationary distribution) follows from the construction of the dynamics and is checked numerically on Müller-Brown and alanine dipeptide examples rather than being presupposed. Active-learning gains are reported against matched-sample baselines. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided derivation chain. The method is externally falsifiable via the reported distribution checks and performance metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are detailed beyond the stated properties of the dynamics.

axioms (1)
  • domain assumption The stochastic dynamics preserve the Boltzmann distribution as the asymptotic distribution.
    Central property asserted in the abstract to distinguish SKMD from other samplers.

pith-pipeline@v0.9.1-grok · 5754 in / 1122 out tokens · 28430 ms · 2026-06-28T10:37:52.151748+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

67 extracted references · 23 canonical work pages

  1. [1]

    Active learning of linearly parametrized interatomic potentials

    Evgeny V . Podryabinkin and Alexander V . Shapeev. “Active learning of linearly parametrized interatomic potentials”. In:Computational Materials Science140 (Dec. 2017), pp. 171–180. ISSN: 09270256.DOI:10.1016/j.commatsci.2017.08.031

    work page doi:10.1016/j.commatsci.2017.08.031 2017

  2. [2]

    npj Comput Mater , volume =

    Noam Bernstein, Gábor Csányi, and V olker L. Deringer. “De novo exploration and self-guided learning of potential-energy surfaces”. In:npj Computational Materials5.1 (2019), p. 99.DOI: 10.1038/s41524-019-0236-6

    work page doi:10.1038/s41524-019-0236-6 2019

  3. [3]

    Performant implementation of the atomic cluster expansion (PACE) and application to copper and silicon

    Yury Lysogorskiy et al. “Performant implementation of the atomic cluster expansion (PACE) and application to copper and silicon”. In:npj Computational Materials7.1 (2021), pp. 1–12

    2021

  4. [4]

    Active learning strategies for atomic cluster expansion models

    Yury Lysogorskiy et al. “Active learning strategies for atomic cluster expansion models”. In: Phys. Rev. Mater.7 (4 Apr. 2023), p. 043801.DOI: 10.1103/PhysRevMaterials.7.043801

    work page doi:10.1103/physrevmaterials.7.043801 2023

  5. [5]

    On-the-fly machine learning force field generation: Application to melting points

    Ryosuke Jinnouchi, Ferenc Karsai, and Georg Kresse. “On-the-fly machine learning force field generation: Application to melting points”. In:Phys. Rev. B100 (1 July 2019), p. 014105. DOI: 10.1103/PhysRevB.100.014105 .URL: https://link.aps.org/doi/10.1103/ PhysRevB.100.014105

    work page doi:10.1103/physrevb.100.014105 2019

  6. [6]

    On-the-fly active learning of interpretable Bayesian force fields for atomistic rare events

    Jonathan Vandermause et al. “On-the-fly active learning of interpretable Bayesian force fields for atomistic rare events”. In:npj Computational Materials6.1 (2020), pp. 1–11

    2020

  7. [7]

    Active learning of reactive Bayesian force fields: Application to heterogeneous hydrogen-platinum catalysis dynamics

    Jonathan Vandermause et al. “Active learning of reactive Bayesian force fields: Application to heterogeneous hydrogen-platinum catalysis dynamics”. In:arXiv preprint arXiv:2106.01949 (2021)

    arXiv 2021

  8. [8]

    Bayesian force fields from active learning for simulation of inter-dimensional transformation of stanene

    Yu Xie et al. “Bayesian force fields from active learning for simulation of inter-dimensional transformation of stanene”. In:npj Computational Materials7.1 (2021), pp. 1–10

    2021

  9. [9]

    Uncertainty-aware molecular dynamics from Bayesian active learning: Phase Transformations and Thermal Transport in SiC

    Yu Xie et al. “Uncertainty-aware molecular dynamics from Bayesian active learning: Phase Transformations and Thermal Transport in SiC”. In:arXiv preprint arXiv:2203.03824(2022)

    arXiv 2022

  10. [10]

    An entropy-maximization approach to automated training set generation for interatomic potentials

    Mariia Karabin and Danny Perez. “An entropy-maximization approach to automated training set generation for interatomic potentials”. In:The Journal of Chemical Physics153.9 (2020), p. 094110

    2020

  11. [11]

    Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials

    Aparna P. A. Subramanyam and Danny Perez. “Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials”. In:npj Computational Materials11.1 (2025), p. 218.DOI:10.1038/s41524-025-01602-9

    work page doi:10.1038/s41524-025-01602-9 2025

  12. [12]

    High-dimensional neural network potentials for metal surfaces: A prototype study for copper

    Nongnuch Artrith and Jörg Behler. “High-dimensional neural network potentials for metal surfaces: A prototype study for copper”. In:Phys. Rev. B85 (4 Jan. 2012), p. 045439.DOI: 10 . 1103 / PhysRevB . 85 . 045439.URL: https : / / link . aps . org / doi / 10 . 1103 / PhysRevB.85.045439

    2012

  13. [13]

    Data Curation for Machine Learning Interatomic Potentials by Determinantal Point Processes

    Joanna Zou and Youssef Marzouk. “Data Curation for Machine Learning Interatomic Potentials by Determinantal Point Processes”. In:ICLR AI4MAT Workshop(2025)

    2025

  14. [14]

    Metadynamics for training neural network model chemistries: A competi- tive assessment

    John E. Herr et al. “Metadynamics for training neural network model chemistries: A competi- tive assessment”. In:The Journal of Chemical Physics148.24 (Mar. 2018), p. 241710.ISSN: 0021-9606.DOI:10.1063/1.5020067.URL:https://doi.org/10.1063/1.5020067. 10

    work page doi:10.1063/1.5020067.url:https://doi.org/10.1063/1.5020067 2018

  15. [15]

    Uncertainty-driven dynamics for active learning of interatomic potentials

    Maksim Kulichenko et al. “Uncertainty-driven dynamics for active learning of interatomic potentials”. In:Nature Computational Science(Mar. 2023).ISSN: 26628457.DOI: 10.1038/ s43588-023-00406-5

    2023

  16. [16]

    Hyperactive Learning (HAL) for Data-Driven Interatomic Potentials

    Cas van der Oord et al. “Hyperactive Learning (HAL) for Data-Driven Interatomic Potentials”. In: (Oct. 2022).URL:http://arxiv.org/abs/2210.04225

    arXiv 2022

  17. [17]

    Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm

    Qiang Liu and Dilin Wang. “Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm”. In:NeurIPS29 (2016)

    2016

  18. [18]

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

    Frank Noé et al. “Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning”. In:Science365.6457 (2019), eaaw1147.DOI: 10 . 1126 / science . aaw1147

    2019

  19. [19]

    Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models

    Michael Plainer et al. “Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models”. In:Advances in Neural Information Processing Systems. arXiv:2506.17139. 2025

    arXiv 2025

  20. [20]

    Flow matching for accelerated simulation of atomic transport in crystalline materials

    Juno Nam et al. “Flow matching for accelerated simulation of atomic transport in crystalline materials”. In:Nature Machine Intelligence(2025).DOI:10.1038/s42256-025-01125-4

    work page doi:10.1038/s42256-025-01125-4 2025

  21. [21]

    Empirical interatomic potential for silicon with improved elastic properties

    J. Tersoff. “Empirical interatomic potential for silicon with improved elastic properties”. In: Physical Review B38.14 (1988), pp. 9902–9905.DOI:10.1103/PhysRevB.38.9902

    work page doi:10.1103/physrevb.38.9902 1988

  22. [22]

    Computer simulation of local order in condensed phases of silicon

    F. H. Stillinger and T. A. Weber. “Computer simulation of local order in condensed phases of silicon”. In:Physical Review B31.8 (1985), pp. 5262–5271.DOI: 10.1103/PhysRevB.31. 5262

    work page doi:10.1103/physrevb.31 1985

  23. [23]

    Embedded-atom method: Derivation and application to impuri- ties, surfaces, and other defects in metals

    M. S. Daw and M. I. Baskes. “Embedded-atom method: Derivation and application to impuri- ties, surfaces, and other defects in metals”. In:Physical Review B29.12 (1984), pp. 6443–6453. DOI:10.1103/PhysRevB.29.6443

    work page doi:10.1103/physrevb.29.6443 1984

  24. [24]

    Modified embedded atom potentials for HCP metals

    M. I. Baskes and R. A. Johnson. “Modified embedded atom potentials for HCP metals”. In: Modelling and Simulation in Materials Science and Engineering2 (1994), pp. 147–163.DOI: 10.1088/0965-0393/2/1/011

    work page doi:10.1088/0965-0393/2/1/011 1994

  25. [25]

    E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

    Simon Batzner et al. “E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials”. In:Nature Communications13.1 (May 2022), p. 2453.DOI: 10.1038/ s41467-022-29939-5

    2022

  26. [26]

    Learning local equivariant representations for large-scale atomistic dynamics

    A. Musaelian et al. “Learning local equivariant representations for large-scale atomistic dynamics”. In:Nature Communications14.579 (2023).DOI: 10.1038/s41467-023-36329- y

    work page doi:10.1038/s41467-023-36329- 2023

  27. [27]

    MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields

    Ilyes Batatia et al. “MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields”. In:Advances in Neural Information Processing Sys- tems. Ed. by S. Koyejo et al. V ol. 35. Curran Associates, Inc., 2022, pp. 11423–11436. URL: https : / / proceedings . neurips . cc / paper _ files / paper / 2022 / file / 4a36c3c51af11...

    2022

  28. [28]

    Optical elements for anisotropic spin-wave propagation,

    Ilyes Batatia et al. “A foundation model for atomistic materials chemistry”. In:The Journal of Chemical Physics163.18 (Nov. 2025), p. 184110.ISSN: 0021-9606.DOI: 10.1063/5. 0297006

    work page doi:10.1063/5 2025

  29. [29]

    Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons

    Albert P Bartók et al. “Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons”. In:Physical review letters104.13 (2010), p. 136403

    2010

  30. [30]

    Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials

    Aidan P Thompson et al. “Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials”. In:Journal of Computational Physics285 (2015), pp. 316–330

    2015

  31. [31]

    Atomic cluster expansion for accurate and transferable interatomic potentials

    Ralf Drautz. “Atomic cluster expansion for accurate and transferable interatomic potentials”. In:Physical Review B99.1 (2019), p. 014104

    2019

  32. [32]

    Lelievre, M

    T. Lelievre, M. Rousset, and G. Stoltz.Free Energy Computations: A Mathematical Perspective. World Scientific Publishing Company, 2010.ISBN: 9781908978752

    2010

  33. [33]

    Enhanced Sampling Methods for Molecular Dynamics Simulations

    Jérôme Hénin et al. “Enhanced Sampling Methods for Molecular Dynamics Simulations”. In:Living Journal of Computational Molecular Science4.1 (Dec. 2022), p. 1583.DOI: 10. 33011/livecoms.4.1.1583

    2022

  34. [34]

    A bound for the error in the normal approximation to the distribution of a sum of dependent random variables

    Charles Stein. “A bound for the error in the normal approximation to the distribution of a sum of dependent random variables”. In:Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability2 (1972), pp. 583–602

    1972

  35. [35]

    Stein variational gradient descent as gradient flow

    Qiang Liu. “Stein variational gradient descent as gradient flow”. In:NeurIPS30 (2017). 11

    2017

  36. [36]

    On the Mean-Field Limit of Stein Variational Gradient Descent

    Jianfeng Lu, Yulong Lu, and James Nolen. “On the Mean-Field Limit of Stein Variational Gradient Descent”. In:SIAM Journal on Mathematical Analysis51.5 (2019), pp. 3611–3640

    2019

  37. [37]

    LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales,

    Aidan P. Thompson et al. “LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales”. In:Computer Physics Communications (2022).DOI:10.1016/j.cpc.2021.108171

    work page doi:10.1016/j.cpc.2021.108171 2022

  38. [38]

    Stochastic Gradient MCMC with Repulsive Forces

    Victor Gallego and David Rios Insua. “Stochastic Gradient MCMC with Repulsive Forces”. In: (Nov. 2018).URL:http://arxiv.org/abs/1812.00071

    arXiv 2018

  39. [39]

    On the geometry of Stein variational gradient descent

    Andrew Duncan, Lukasz Szpruch, and Nikolas Nusken. “On the geometry of Stein variational gradient descent”. In:Journal of Machine Learning Research24.56 (2023), pp. 1–39.ISSN: 1532-4435

    2023

  40. [40]

    Stein Self-Repulsive Dynamics: Benefits From Past Samples

    Mao Ye, Tongzheng Ren, and Qiang Liu. “Stein Self-Repulsive Dynamics: Benefits From Past Samples”. In: (Feb. 2020).URL:http://arxiv.org/abs/2002.09070

    arXiv 2020

  41. [41]

    Bayesian experimental design using regularized determinantal point processes

    Michał Derezi´nski, Feynman Liang, and Michael W. Mahoney. “Bayesian experimental design using regularized determinantal point processes”. In:AISTATS(2020)

    2020

  42. [42]

    Stein Points

    Wilson Ye Chen et al. “Stein Points”. In:Proceedings of the 35th International Conference on Machine Learning. V ol. 80. Proceedings of Machine Learning Research. PMLR, 2018, pp. 844–853

    2018

  43. [43]

    A Stein variational Newton method

    Gianluca Detommaso et al. “A Stein variational Newton method”. In: (June 2018).URL: http://arxiv.org/abs/1806.03085

    Pith/arXiv arXiv 2018

  44. [44]

    Stein variational gradient descent with matrix-valued kernels

    Dilin Wang et al. “Stein variational gradient descent with matrix-valued kernels”. In:NeurIPS 32 (2019)

    2019

  45. [45]

    A stochastic version of Stein variational gradient descent for efficient sampling

    Lei Li et al. “A stochastic version of Stein variational gradient descent for efficient sampling”. In:Communications in Applied Mathematics and Computational Science15.1 (2020), pp. 37– 63

    2020

  46. [46]

    Multilevel Stein variational gra- dient descent with applications to Bayesian inverse problems

    Terrence Alsup, Luca Venturi, and Benjamin Peherstorfer. “Multilevel Stein variational gra- dient descent with applications to Bayesian inverse problems”. In: (Apr. 2021).URL: http: //arxiv.org/abs/2104.01945

    arXiv 2021

  47. [47]

    p-Kernel Stein Variational Gradient Descent for Data Assimilation and History Matching

    Andreas S. Stordal et al. “p-Kernel Stein Variational Gradient Descent for Data Assimilation and History Matching”. In:Mathematical Geosciences53 (2021), pp. 375–393.DOI: 10 . 1007/s11004-021-09937-x

    2021

  48. [48]

    ESCORT: Efficient Stein-variational and Sliced Consistency-Optimized Temporal Belief Representation for POMDPs

    Yunuo Zhang et al. “ESCORT: Efficient Stein-variational and Sliced Consistency-Optimized Temporal Belief Representation for POMDPs”. In:Advances in Neural Information Processing Systems. arXiv:2510.21107. 2025

    arXiv 2025

  49. [49]

    Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks

    Daniel Schwalbe-Koda, Aik Rui Tan, and Rafael Gómez-Bombarelli. “Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks”. en. In:Nature Communi- cations12.1 (Aug. 2021), p. 5104.ISSN: 2041-1723.DOI: 10.1038/s41467-021-25342-8

    work page doi:10.1038/s41467-021-25342-8 2021

  50. [50]

    Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials

    Viktor Zaverkin et al. “Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials”. en. In:npj Computational Materials10.1 (Apr. 2024), p. 83.ISSN: 2057-3960.DOI:10.1038/s41524-024-01254-1

    work page doi:10.1038/s41524-024-01254-1 2024

  51. [51]

    MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules

    Dávid Péter Kovács et al. “MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules”. In:Journal of the American Chemical Society147.21 (May 2025), pp. 17598–17611.ISSN: 0002-7863.DOI:10.1021/jacs.4c07099

    work page doi:10.1021/jacs.4c07099 2025

  52. [52]

    Spice, a dataset of drug-like molecules and peptides for training machine learning potentials.Scientific Data, 10 (1):11, 2023

    Peter Eastman et al. “SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials”. en. In:Scientific Data10.1 (Jan. 2023), p. 11.ISSN: 2052-4463. DOI:10.1038/s41597-022-01882-6

    work page doi:10.1038/s41597-022-01882-6 2023

  53. [53]

    Isambard-AI: a leadership- class supercomputer optimised specifically for Artificial Intelligence

    Simon McIntosh-Smith, Sadaf Alam, and Christopher Woods. “Isambard-AI: a leadership- class supercomputer optimised specifically for Artificial Intelligence”. In:Proceedings of the Cray User Group. CUG ’24. Association for Computing Machinery, 2025, pp. 44–54.ISBN: 9798400713286.DOI: 10.1145/3725789.3725794.URL: https://doi.org/10.1145/ 3725789.3725794

    work page doi:10.1145/3725789.3725794.url: 2025

  54. [54]

    Calculating free energies using average force

    Eric Darve and Andrew Pohorille. “Calculating free energies using average force”. In:The Journal of Chemical Physics115.20 (2001), pp. 9169–9183

    2001

  55. [55]

    Riemann Manifold Langevin and Hamiltonian Monte Carlo Methods

    Mark Girolami and Ben Calderhead. “Riemann Manifold Langevin and Hamiltonian Monte Carlo Methods”. In:Journal of the Royal Statistical Society: Series B73.2 (2011), pp. 123–214. 12

    2011

  56. [56]

    Using Perturbed Underdamped Langevin Dynamics to Efficiently Sample from Probability Distributions

    A. B. Duncan, N. Nüsken, and G. A. Pavliotis. “Using Perturbed Underdamped Langevin Dynamics to Efficiently Sample from Probability Distributions”. In:Journal of Statistical Physics169 (6 Dec. 2017), pp. 1098–1131.ISSN: 00224715.DOI: 10.1007/s10955-017- 1906-8

    work page doi:10.1007/s10955-017- 2017

  57. [57]

    Well-Tempered Metadynamics: A Smoothly Converging and Tunable Free-Energy Method

    Alessandro Barducci, Giovanni Bussi, and Michele Parrinello. “Well-Tempered Metadynamics: A Smoothly Converging and Tunable Free-Energy Method”. In:Phys. Rev. Lett.100 (2 Jan. 2008), p. 020603.DOI:10.1103/PhysRevLett.100.020603

    work page doi:10.1103/physrevlett.100.020603 2008

  58. [58]

    Enhancing Important Fluctuations: Rare Events and Metadynamics from a Conceptual Viewpoint

    Omar Valsson, Pratyush Tiwary, and Michele Parrinello. “Enhancing Important Fluctuations: Rare Events and Metadynamics from a Conceptual Viewpoint”. In:Annual Review of Physical Chemistry67 (2016), pp. 159–184.DOI: https://doi.org/10.1146/annurev-physchem- 040215-112229

    work page doi:10.1146/annurev-physchem- 2016

  59. [59]

    Lectures in Mathematics ETH Zürich

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré.Gradient Flows: In Metric Spaces and in the Space of Probability Measures. Lectures in Mathematics ETH Zürich. Birkhäuser Basel, 2008.ISBN: 978-3-7643-8722-8

    2008

  60. [60]

    Super-Samples from Kernel Herding

    Yutian Chen, Max Welling, and Alexander J. Smola. “Super-Samples from Kernel Herding”. In: Proceedings of the Twenty-Sixth Conference on Uncertainty in Artificial Intelligence. UAI’10. AUAI Press, 2010, pp. 109–116

    2010

  61. [61]

    Roshan Joseph.Projected Support Points: A New Method for High- Dimensional Data Reduction

    Simon Mak and V . Roshan Joseph.Projected Support Points: A New Method for High- Dimensional Data Reduction. arXiv:1708.06897. 2017. arXiv:1708.06897 [stat.ME]

    Pith/arXiv arXiv 2017

  62. [62]

    Support Points

    Simon Mak and V . Roshan Joseph. “Support Points”. In:The Annals of Statistics46.6A (2018), pp. 2562–2592.DOI:10.1214/17-AOS1629

    work page doi:10.1214/17-aos1629 2018

  63. [63]

    Kernel Thinning

    Raaz Dwivedi and Lester Mackey. “Kernel Thinning”. In:Journal of Machine Learning Research25.152 (2024), pp. 1–77

    2024

  64. [64]

    Compress Then Test: Powerful Kernel Testing in Near-linear Time

    Carles Domingo-Enrich, Raaz Dwivedi, and Lester Mackey. “Compress Then Test: Powerful Kernel Testing in Near-linear Time”. In:Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS). V ol. 206. Proceedings of Machine Learning Research. PMLR, 2023

    2023

  65. [65]

    A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advances

    Brian B. Moser et al. “A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advances”. In:arXiv preprint arXiv:2505.17799(2025)

    arXiv 2025

  66. [66]

    LoRA: Low-Rank Adaptation of Large Language Models

    Edward J Hu et al. “LoRA: Low-Rank Adaptation of Large Language Models”. In:Interna- tional Conference on Learning Representations. 2022.URL: https://openreview.net/ forum?id=nZeVKeeFYf9. 13 A Assumptions, Analysis, and Proofs A.1 Asymptotic analysis of SKMD We show that the asymptotic behavior of (5) maintains fidelity to the Boltzmann distribution πθ, s...

    2022

  67. [67]

    Therefore, the acquisition criterion does not relate to the minimization of KSD, as the objective differs in terms of the Hilbert space norm. In practice, the L2(X) norm of ϕ∗ ˆqn,π is simpler to compute in an online fashion compared to the RKHS norm, as it does not require the calculation of additional gradients or second-order terms at each simulation s...