pith. sign in

arxiv: 2604.05778 · v1 · submitted 2026-04-07 · 🧮 math.DS · physics.chem-ph· stat.ML

Effective Dynamics and Transition Pathways from Koopman-Inspired Neural Learning of Collective Variables

Alexander Sikorski , Luca Donati , Marcus Weber , Christof Sch\"utte This is my paper

Pith reviewed 2026-05-10 19:03 UTC · model grok-4.3

classification 🧮 math.DS physics.chem-phstat.ML
keywords Koopman operatorcollective variablesmetastable transitionseffective dynamicsneural networkstransition pathwaysinvariant subspaces
0
0 comments X

The pith

Koopman neural learning extracts collective variables that yield effective dynamics and transition rates from molecular simulation data.

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

The paper shows how to learn dominant invariant subspaces of the Koopman operator with neural networks, then derive reduced effective dynamics on those variables that reproduce the original system's metastable transitions. This approach connects the learned variables directly to committor functions, transition pathways, and rate constants without requiring hand-crafted reaction coordinates. A reader would care because it offers a data-driven route to analyze rare events in high-dimensional systems such as molecules, where direct simulation of transitions is impractical. The method is demonstrated on one-, two-, and three-dimensional benchmark potentials that include both enthalpic and entropic barriers, recovering coarse-grained kinetics and transition times.

Core claim

Starting from data, the ISOKANN framework identifies collective variables as dominant invariant subspaces of the Koopman operator via neural networks, constructs the corresponding reduced dynamical model on the latent space, and uses that model to compute transition rates, mean transition times, committor functions, and transition pathways for metastable systems.

What carries the argument

The ISOKANN framework, which learns dominant invariant subspaces of the Koopman operator with artificial neural networks to identify collective variables and then derives the effective dynamics on the resulting latent space.

If this is right

  • Transition rates and pathways become computable directly from trajectory data once the collective variables are learned.
  • The same reduced model reproduces both enthalpic and entropic barrier crossings on benchmark potentials.
  • Committor functions and transition pathways emerge as direct outputs of the effective dynamics on the latent space.
  • Krylov-like subspace methods can be combined with the neural learning step to maintain consistency with the underlying operator theory.

Where Pith is reading between the lines

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

  • The framework could be applied to longer molecular-dynamics trajectories where manual collective-variable design is difficult.
  • It may allow systematic comparison of different neural architectures for subspace learning on the same dataset.
  • Extending the reduced dynamics to include explicit memory terms could improve accuracy for systems with slower hidden variables.

Load-bearing premise

The learned dominant invariant subspaces must accurately represent the collective variables that control the slow, metastable transitions of interest.

What would settle it

On a standard double-well potential with known exact transition time, the method's predicted mean transition time deviates by more than a small percentage from the analytic value or from long direct simulations.

Figures

Figures reproduced from arXiv: 2604.05778 by Alexander Sikorski, Christof Sch\"utte, Luca Donati, Marcus Weber.

Figure 1
Figure 1. Figure 1: Illustration of the transformation from eigen- to [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the three components of χ : R 3 → R 3 obtained from either ISOKANN (blue) or direct discretization through the SqRA (red) for the potential from Section 5.3. Each plot displays one component of χ evaluated at one of 10 000 training samples, where the samples displayed on the horizontal axis were sorted by the respective χ value (different for each subfig￾ure) obtained from SqRA to allow for a… view at source ↗
Figure 3
Figure 3. Figure 3: 1D system. (a) Potential energy function; (b) [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 1D system. (a) MFPT in full space; (b) MFPT in latent space; (c) Committor in full [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 2D system. (a) Potential energy function; (b) [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 2D system. (a) MFPT in full space; (b) MFPT in latent space; (c) Committor in full [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 3D system. (a) Potential energy function; (b,c,d) Three [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: 3D system. (a) Eigenvalues of L and L˜; (b) Committor in full space; (c) Committor q projected on ∆¯ 2 ; (d) Effective committor q˜; (e) TPT rates of full and effective dynamics. 5.4 Discussion Across all examples, the CV χ of ISOKANN allows to consistently reconstruct the dominant eigenmodes and corresponding kinetic observables of the underlying stochastic process. The one-dimensional case validates the … view at source ↗
read the original abstract

The ISOKANN (Invariant Subspaces of Koopman Operators Learned by Artificial Neural Networks) framework provides a data-driven route to extract collective variables (CVs) and effective dynamics from complex molecular systems. In this work, we integrate the theoretical foundation of Koopman operators with Krylov-like subspace algorithms, and reduced dynamical modeling to build a coherent picture of how to describe metastable transitions in high-dimensional systems based on CVs. Starting from the identification of CVs based on dominant invariant subspaces, we derive the corresponding effective dynamics on the latent space and connect these to transition rates and times, committor functions, and transition pathways. The combination of Koopman-based learning and reduced-dimensional effective dynamics yields a principled framework for computing transition rates and pathways from simulation data. Numerical experiments on one-, two-, and three-dimensional benchmark potentials illustrate the ability of ISOKANN to reconstruct the coarse-grained kinetics and reproduce transition times across enthalpic and entropic barriers.

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 paper claims to introduce the ISOKANN framework that combines Koopman operator theory with neural networks to identify dominant invariant subspaces serving as collective variables for metastable transitions. It then derives effective dynamics in this reduced space and links them to committor functions, transition rates, and pathways. The framework is illustrated with numerical experiments on benchmark potentials in one to three dimensions, where it reconstructs coarse-grained kinetics and reproduces transition times for both enthalpic and entropic barriers.

Significance. This result, if the subspaces accurately represent the governing CVs, offers a systematic data-driven method for analyzing rare events and transition mechanisms in complex systems without requiring predefined reaction coordinates. The manuscript's strengths include the coherent theoretical integration of invariant subspace learning with reduced dynamical modeling and the concrete numerical validation on standard test cases that confirms consistency with known transition times. The potential circularity concern regarding subspace selection being tuned to transition data does not appear to land, as the rates are derived quantities from the learned dynamics rather than direct fits. These elements support the potential utility of the approach in molecular dynamics applications.

minor comments (3)
  1. The notation for the Koopman operator and its invariant subspaces could be introduced more clearly in the early sections for readers unfamiliar with the field.
  2. Ensure all acronyms like ISOKANN are defined at first use in the main text.
  3. The abstract mentions integration with Krylov-like subspace algorithms; the main text should provide a brief explicit link to how this is implemented within the neural learning procedure.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and accurate summary of our manuscript on the ISOKANN framework. The assessment correctly highlights the integration of Koopman operator theory with neural networks for learning invariant subspaces as collective variables, the derivation of effective dynamics, and the connections to transition rates, times, and pathways. We also appreciate the recognition of the numerical experiments on benchmark potentials demonstrating consistency with known kinetics for both enthalpic and entropic barriers. No specific major comments requiring clarification or changes were raised in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper presents ISOKANN as a data-driven method that first identifies dominant invariant subspaces via neural networks to obtain collective variables, then derives reduced effective dynamics on that latent space, and finally connects those to committor functions, rates, and pathways using standard Markovian and Koopman theory. No step reduces a claimed prediction or rate to a fitted parameter by construction, nor does any load-bearing premise rest on a self-citation whose content is itself unverified or tautological. The numerical experiments on benchmark potentials serve as external validation rather than internal fitting loops. The central claims therefore remain independent of the inputs they are derived from.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard Koopman operator theory and neural network approximation capabilities; no new physical entities are introduced, but the choice of latent dimension and network architecture function as free parameters.

free parameters (1)
  • latent space dimension
    Selected to isolate dominant invariant subspaces; its value directly affects which transitions are captured.
axioms (1)
  • domain assumption The underlying molecular dynamics admits a Koopman operator representation whose dominant invariant subspaces correspond to the slow collective variables.
    Invoked at the start of the CV identification step.

pith-pipeline@v0.9.0 · 5474 in / 1310 out tokens · 93490 ms · 2026-05-10T19:03:53.217539+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1):015102, 2013

    Gabriel P´ erez-Hern´ andez, Frank Paul, Thomas Giorgino, Gianni De Fabritiis, and Frank No´ e. Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1):015102, 2013

    work page 2013

  2. [2]

    Mardt, L

    A. Mardt, L. Pasquali, H. Wu, and F. No´ e. Vampnets for deep learning of molecular kinetics.Nature Communications, 9(5), 2018

    work page 2018

  3. [3]

    Wu and F

    H. Wu and F. No´ e. Reaction coordinate flows for model reduction of molecular kinetics. The Journal of Chemical Physics, 160:044109, 2024

    work page 2024

  4. [4]

    R. J. Rabben, S. Ray, and M. Weber. Isokann: Invariant subspaces of Koopman operators learned by a neural network.The Journal of Chemical Physics, 153(11):114109, 2020

    work page 2020

  5. [5]

    On finding optimal collective variables for complex sys- tems by minimizing the deviation between effective and full dynamics.Multiscale Modeling & Simulation, 23(2):924–958, 2025

    Wei Zhang and Christof Sch¨ utte. On finding optimal collective variables for complex sys- tems by minimizing the deviation between effective and full dynamics.Multiscale Modeling & Simulation, 23(2):924–958, 2025

    work page 2025

  6. [6]

    Spectral properties of effective dynamics from conditional expectations.Entropy, 23(2):134, 2021

    Feliks N¨ uske, P´ eter Koltai, Lorenzo Boninsegna, and Cecilia Clementi. Spectral properties of effective dynamics from conditional expectations.Entropy, 23(2):134, 2021

    work page 2021

  7. [7]

    Exact dynamical coarse-graining without time-scale separation.The Journal of Chemical Physics, 141(4):044109, 2014

    Jianfeng Lu and Eric Vanden-Eijnden. Exact dynamical coarse-graining without time-scale separation.The Journal of Chemical Physics, 141(4):044109, 2014

    work page 2014

  8. [8]

    Effective dynamics along given reaction coordinates and reaction rate theory.Faraday Discussions, 195:365–394, 2016

    Wei Zhang, Carsten Hartmann, and Christof Sch¨ utte. Effective dynamics along given reaction coordinates and reaction rate theory.Faraday Discussions, 195:365–394, 2016

    work page 2016

  9. [9]

    Legoll, T

    F. Legoll, T. Leli` evre, and S. Olla. Pathwise estimates for an effective dynamics.Stochastic Processes and their Applications, 127:2841–2863, 2017. 27

    work page 2017

  10. [10]

    Leli` evre and W

    T. Leli` evre and W. Zhang. Pathwise estimates for effective dynamics: The case of nonlinear vectorial reaction coordinates.Multiscale Modeling & Simulation, 17, 2019

    work page 2019

  11. [11]

    Overcoming the timescale barrier in molecular dynamics: Transfer operators, variational principles, and machine learning

    Christof Sch¨ utte, Stefan Klus, and Carsten Hartmann. Overcoming the timescale barrier in molecular dynamics: Transfer operators, variational principles, and machine learning. Acta Numerica, 32:517–673, 2023

    work page 2023

  12. [12]

    P. R. Vlachas, J. Zavadlav, M. Praprotnik, and P. Koumoutsakos. Accelerated simulations of molecular systems through learning of effective dynamics.Journal of Chemical Theory and Computation, 18(1):538–549, 2022

    work page 2022

  13. [13]

    Bovier, M

    A. Bovier, M. Eckhoff, V. Gayrard, and M. Klein. Metastability and low lying spectra in reversible Markov chains.Comm. Math. Phys., 228:219–255, 2002

    work page 2002

  14. [14]

    Sch¨ utte, A

    C. Sch¨ utte, A. Fischer, W. Huisinga, and P. Deuflhard. A direct approach to conformational dynamics based on hybrid Monte Carlo.J. Comp. Physics Special Issue on Computational Biophysics, 151:146–168, 1999

    work page 1999

  15. [15]

    Sch¨ utte and M

    C. Sch¨ utte and M. Sarich.Metastability and Markov State Models in Molecular Dynamics: Modeling, Analysis, Algorithmic Approaches. Courant Lecture Notes No. 32. American Mathematical Society, 2014

    work page 2014

  16. [16]

    S. Klus, F. N¨ uske, P. Koltai, H. Wu, I. Kevrekidis, C. Sch¨ utte, and F. No´ e. Data- driven model reduction and transfer operator approximation.Journal of Nonlinear Science, 28:985–1010, 2018

    work page 2018

  17. [17]

    R Bowman, V

    G. R Bowman, V. S Pande, and F. No´ e.An introduction to Markov state models and their application to long timescale molecular simulation, volume 797. Springer Science & Business Media, 2013

    work page 2013

  18. [18]

    Transition pathways in complex sys- tems: Reaction coordinates, isocommittor surfaces, and transition tubes.Chemical Physics Letters, 413(1-3):242–247, 2005

    Weinan E, Weiqing Ren, and Eric Vanden-Eijnden. Transition pathways in complex sys- tems: Reaction coordinates, isocommittor surfaces, and transition tubes.Chemical Physics Letters, 413(1-3):242–247, 2005

    work page 2005

  19. [19]

    Bittracher, P

    A. Bittracher, P. Koltai, S. Klus, R. Banisch, M. Dellnitz, and C. Sch¨ utte. Transition manifolds of complex metastable systems.Journal of Nonlinear Science, 28, 2018

    work page 2018

  20. [20]

    Kappler, J

    J. Kappler, J. O. Daldrop, F. N. Bruenig, M. D. Boehle, and RR Netz. Memory-induced acceleration and slowdown of barrier crossing.J Chem Phys., 148(1), 2018

    work page 2018

  21. [21]

    C. Ayaz, L. Tepper, F. N. Br¨ unig, J. Kappler, J. O. Daldrop, and R. R. Netz. Non- markovian modeling of protein folding.Proceedings of the National Academy of Sciences of the United States of America, 118(4):e2023856118, 2021

    work page 2021

  22. [22]

    Legoll and T

    F. Legoll and T. Lelievre. Effective dynamics using conditional expectations.Nonlinearity, 23(9):2131, 2010

    work page 2010

  23. [23]

    Reliable approximation of long relaxation timescales in molecular dynamics.Entropy, 19(7):367, 2017

    Wei Zhang and Christof Sch¨ utte. Reliable approximation of long relaxation timescales in molecular dynamics.Entropy, 19(7):367, 2017. 28

    work page 2017

  24. [24]

    Legoll, T

    F. Legoll, T. Leli` evre, and U. Sharma. Effective dynamics for non-reversible stochastic differential equations: A quantitative study.Nonlinearity, 32:4779–4816, 2019

    work page 2019

  25. [25]

    Zhang, T

    W. Zhang, T. Li, and C. Sch¨ utte. Solving eigenvalue PDEs of metastable diffusion processes using artificial neural networks.Journal of Computational Physics, 465:111377, 2022

    work page 2022

  26. [26]

    Learning koopman eigenfunc- tions of stochastic diffusions with optimal importance sampling and ISOKANN.Journal of Mathematical Physics, 65(1):013502, 2024

    Alexander Sikorski, Enric Ribera Borrell, and Marcus Weber. Learning koopman eigenfunc- tions of stochastic diffusions with optimal importance sampling and ISOKANN.Journal of Mathematical Physics, 65(1):013502, 2024

    work page 2024

  27. [27]

    Characterization of transition states in conformational dynamics using fuzzy sets

    Marcus Weber and Tobias Galliat. Characterization of transition states in conformational dynamics using fuzzy sets. Technical Report 02-12, ZIB, Takustr. 7, 14195 Berlin, 2002

    work page 2002

  28. [28]

    The kramers turnover in terms of a macro-state projection on phase space.Mol

    Luca Donati, Christof Sch¨ utte, and Marcus Weber. The kramers turnover in terms of a macro-state projection on phase space.Mol. Phys., 0(0):e2356748, 2024

    work page 2024

  29. [29]

    De Gruyter, Berlin, Boston, 2025

    Alexander Sikorski, Robert Julian Rabben, Surahit Chewle, and Marcus Weber.Capturing the macroscopic behavior of molecular dynamics with membership functions, pages 41–58. De Gruyter, Berlin, Boston, 2025

    work page 2025

  30. [30]

    Topological analysis reveals multiple pathways in molecular dynamics.J

    Luca Donati, Surahit Chewle, Dominik St Pierre, Vijay Natarajan, and Marcus Weber. Topological analysis reveals multiple pathways in molecular dynamics.J. Chem. Theory Comput., 21(20):10385–10397, 2025

    work page 2025

  31. [31]

    Robust Perron cluster analysis in conformation dy- namics.Linear Algebra Appl., 398:161–184, 2004

    Peter Deuflhard and Marcus Weber. Robust Perron cluster analysis in conformation dy- namics.Linear Algebra Appl., 398:161–184, 2004

    work page 2004

  32. [32]

    A coarse graining method for the identification of transition rates between molecular conformations.J

    Susanna Kube and Marcus Weber. A coarse graining method for the identification of transition rates between molecular conformations.J. Chem. Phys., 126(2):024103, 2007

    work page 2007

  33. [33]

    Implications of PCCA+ in Molecular Simulation.Computation, 6(1), 2018

    Marcus Weber. Implications of PCCA+ in Molecular Simulation.Computation, 6(1), 2018

    work page 2018

  34. [34]

    P. G. Bolhuis, D. Chandler, C. Dellago, and P. Geissler. Transition path sampling: throwing ropes over mountain passes, in the dark.Annu. Rev. Phys. Chem., 59:291, 2002

    work page 2002

  35. [35]

    A. K. Faradjian and R. Elber. Computing time scales from reaction coordinates by mile- stoning.J. Chem. Phys., 120:10880–10889, 2004

    work page 2004

  36. [36]

    Moroni, T

    D. Moroni, T. van Erp, and P. Bolhuis. Investigating rare events by transition interface sampling.Physica A, 340:395–401, 2004

    work page 2004

  37. [37]

    W. E and E. Vanden-Eijnden. Metastability, conformation dynamics, and transition path- ways in complex systems. InMultiscale modelling and simulation, volume 39 ofLect. Notes Comput. Sci. Eng., pages 35–68. Springer, Berlin, 2004

    work page 2004

  38. [38]

    W. E and E. Vanden-Eijnden. Towards a theory of transition paths.Journal of statistical physics, 123:503–523, 2006

    work page 2006

  39. [39]

    W. E and E. Vanden-Eijnden. Transition-path theory and path-finding algorithms for the study of rare events.Annual Review of Physical Chemistry, 61:391–420, 2010. 29

    work page 2010

  40. [40]

    A. M. Berezhkovskii and A. Szabo. Committors, first-passage times, fluxes, markov states, milestones, and all that.The Journal of Chemical Physics, 150(5):054106, 2019

    work page 2019

  41. [41]

    Metzner, C

    P. Metzner, C. Sch¨ utte, and E. Vanden-Eijnden. Illustration of transition path theory on a collection of simple examples.J. Chem. Phys., 125(8), 2006. 084110

    work page 2006

  42. [42]

    Noise in Nonlinear Dynamical Systems

    A.A. Andronov, L.S. Pontryagin, and A.A. Vitt. On statistical considerations of dynam- ical systems.Zhurnal `Eksperimental’no˘ ı i Teoretichesko˘ ı Fiziki, 3:165–180, 1933. English translation in “Noise in Nonlinear Dynamical Systems”, F. Moss & P. V. E. McClintock (eds), Cambridge University Press, 1989, Vol.1, p. 329

    work page 1933

  43. [43]

    A square root approximation of transition rates for a Markov State Model.SIAM

    Han Cheng Lie, Konstantin Fackeldey, and Marcus Weber. A square root approximation of transition rates for a Markov State Model.SIAM. J. Matrix Anal. Appl., 34:738–756, 2013

    work page 2013

  44. [44]

    Donati, M

    L. Donati, M. Heida, B. G. Keller, and M. Weber. Estimation of the infinitesimal generator by square-root approximation.J. Phys. Condens. Matter, 30:425201, 2018

    work page 2018

  45. [45]

    Donati, M

    L. Donati, M. Weber, and B. G. Keller. Markov models from the square root approximation of the Fokker–Planck equation: Calculating the grid-dependent flux.J. Phys. Condens. Matter, 33:115902, 2021

    work page 2021

  46. [46]

    Zhang and C

    W. Zhang and C. Sch¨ utte. Understanding recent deep-learning techniques for identify- ing collective variables of molecular dynamics.Proceedings in Applied Mathematics and Mechanics, 23:e202300189, 2023. 30

    work page 2023