Pith. sign in

REVIEW 3 major objections 4 minor 78 references

The Itô-residual loss (32) has the committor as its unique minimizer; trained with neural networks and iterated with AMS, it approximates the committor from short trajectories without pre-computed collective variables.

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 07:28 UTC pith:GHMRYXR5

load-bearing objection The continuous-time Itô-residual loss is a genuine new idea, but the discrete loss actually implemented in Eq. (37) drops the boundary term, so the paper's uniqueness guarantee does not hold for the objective they train. the 3 major comments →

arxiv 2607.21425 v2 pith:GHMRYXR5 submitted 2026-07-23 physics.chem-ph

Approximating committor functions: Objective functions and training data sampling

Thomas Pigeon , Gabriel Stoltz , Tony Leli\`evre This is my paper
classification physics.chem-ph MSC 60J6060H1065C30
keywords committor functionneural networkItô formularare eventsadaptive multilevel splitting (AMS)reaction coordinatetransition pathoverdamped Langevin dynamics
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.

This paper introduces a new objective function for training neural networks to approximate the committor function — the probability that a molecular system reaches the product state before the reactant state. The objective is built from the Itô formula applied to the committor, and the authors argue that its unique minimizer is exactly the committor, with boundary conditions included implicitly. On the Müller–Brown test potential, the proposed loss matches or slightly beats existing residual and variational losses given the same number of configurations. The paper also couples the loss with adaptive multilevel splitting (AMS), alternately using the current approximate committor as a reaction coordinate and retraining on the newly sampled reactive and non-reactive trajectories. The method requires only the definitions of the reactant and product states, no pre-computed collective variables.

Core claim

The central claim is that for any strictly increasing C² function h, the committor function χ is the unique minimizer of the loss (32), which is the squared expectation of the residual of Itô's formula for h(aχ+b) over short trajectories stopped at the first hitting time of R∪P. Because h is strictly monotone and the boundary conditions f=1 on ∂P and f=0 on ∂R are enforced through the boundary term h(a1_P+b)1_{τ<ϑ}, the unique minimizer over all real-valued functions is exactly χ. The paper argues that minimizing the discretized analogue (37) with a neural network, using the same Euler–Maruyama Gaussian increments that generate the trajectories, approximates the committor; and that iterating

What carries the argument

The central object is the Itô-residual loss (37): for each short trajectory Q launched from a sample point, the loss compares the change of h(af(Q)) along the trajectory (including boundary term if R or P is hit) against the sum of the drift term (a²/β) h''|∇f|² dt and the Itô-integral term a√(2/β) h'∇f·dW, with the noise integral replaced by the same Gaussian increments used to generate the trajectory. The workhorse identity is that for the true committor this residual is zero in expectation, and the strict monotonicity of h makes the zero level set single out the committor. The second mechanism is the AMS coupling: the approximate committor serves as the reaction coordinate ξ for splitting

Load-bearing premise

The discretized Itô-residual loss (37) is assumed to be a faithful surrogate for the continuous objective (32), whose unique minimizer is the committor — but the paper asserts uniqueness for (32) without a full proof, and replaces the Itô integral with the trajectory's own Euler–Maruyama Gaussian increments without a consistency or discretization-error analysis (Section II B; see also SI.3).

What would settle it

For a one-dimensional double well with a known analytic committor, minimize the discretized loss (37) with h=Id and h=ln for shrinking Δt and fixed total duration, and check whether the minimizer converges to the analytic committor in L²; if the minimizer drifts away as Δt→0, the discretization is biased and the surrogate assumption fails. Alternatively, on the Müller–Brown benchmark, compare the neural-network minimizer of (37) with the finite-element committor over a sweep of Δt values.

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

If this is right

  • With the same number of training configurations, the new loss (37) yields slightly lower errors than the variational PDE loss (14) and the transition-operator losses (25)/(28) on the Müller–Brown potential; with equal force-call budgets it substantially outperforms the fixed-point loss (23).
  • Using h=ln with ε just above 0 improves log-scale accuracy near both reactant and product basins relative to h=Id.
  • The AMS-coupled iterative scheme converges on the Z-potential after four iterations, and AMS runs using the learned committor as reaction coordinate give 95% confidence intervals for the transition probability that contain the finite-element estimate, whereas a linear interpolation coordinate yields intervals that miss it.
  • The formulation extends to biased dynamics through the modified loss (35), allowing exploration to be accelerated with a biasing potential without changing the target function.
  • Only knowledge of the reactant and product states is required for the iterative procedure; no pre-computed collective variables or long equilibrium trajectories are needed to start.

Where Pith is reading between the lines

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

  • A consistency analysis of the discretization (37) is not given; the paper itself notes that analyzing the error of the Euler–Maruyama approximation of the transition operator is beyond its scope (SI.3), so the practical validity of the minimizer hinges on an unquantified discretization bias.
  • Because the loss consumes only short trajectory segments and their noise increments, the training strategy could be applied to any data source where the dynamics and the injected noise are known — for example, enhanced-sampling runs or experimental trajectories with modelled noise — not just freshly simulated Euler–Maruyama paths.
  • The freedom in h (identity, log, and the two-sided log combination) defines a tunable family of attention-weighted losses; h could be adapted during training to focus on regions where the current approximation is worst, or chosen to approximate the optimal importance function for AMS, though the paper does not do this.
  • The same Itô-residual construction carries over to underdamped Langevin dynamics, where the committor depends on momenta; the paper lists this as a future perspective, and the loss would then learn a phase-space function from trajectory data.

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 reviews neural-network methods for approximating the committor function of overdamped Langevin dynamics and proposes a new family of loss functions based on Itô's formula. The continuous objective (32) is designed so that the committor is its unique minimizer, with boundary conditions encoded through a terminal indicator term. The authors discretize this objective as (37), compare it with existing losses on the Müller–Brown potential, and combine it with AMS in an iterative active-learning scheme on the Z-potential. The central claims are that the new loss can be trained from trajectory data without pre-computed collective variables, and that the AMS-coupled iteration yields accurate committor approximations.

Significance. The continuous-time derivation in Section II A is a clean and potentially useful addition to the existing toolbox: it gives a residual formulation with an arbitrary sampling measure, allows emphasis on different committor ranges through the choice of h, and extends to biased dynamics. The review of existing methods, especially Table I and the taxonomy in Section I, is a helpful synthesis. The AMS coupling idea is attractive, and the authors are careful to state that the numerical study is limited to two 2D examples. The code is referenced, which supports reproducibility. However, the paper's main theoretical guarantee is currently tied to Eq. (32), while the implemented loss (37) omits the boundary term; as written, the objective actually minimized has every constant function as a global minimizer. This is a central issue, not a presentation issue.

major comments (3)
  1. [II B, Eq. (37)] Equation (37) does not faithfully discretize (32). The continuous loss contains h(a1_P(q_τ)+b)1_{ϑ≥τ}, which encodes the boundary values f=1 on ∂P and f=0 on ∂R. In (37), the first term is h(af_θ(Q_{k,L∧ℓ_k})+b) for every trajectory, with no indicator and no replacement of f_θ by 1_P at the hitting time. Consequently, for any constant function f_θ≡c, the two h-terms cancel and the two sums vanish, so L(θ)=0 exactly for every dataset and every h,a,b, including the logarithmic variants described in Section II B. Thus every constant function is a global minimizer of (37). The 'unique minimizer' statement in Section II A is therefore not inherited by the objective actually minimized. This is not a higher-order discretization error; it is the removal of the boundary mechanism that made (32) identifiable. The numerical results in Section III cannot be attributed to the stated loss as written;
  2. [II A, Eq. (32)] The paper asserts that the unique minimizer of (32) among functions on (R∪P)^c is the committor, 'thanks to the strict monotonicity of h' and the implicit boundary conditions. No proof is given in the main text or SI, unlike the other minimization problems (14), (20), (25), (28). This is a central claim and should be proved. The argument is not immediate because the objective contains a stochastic integral and the admissible class of functions, and their traces on ∂R∪∂P, are not specified. A rigorous statement is needed before the discretization can be meaningfully analyzed.
  3. [II B, Eq. (37)] Even after correcting the boundary term, the relationship between minimizers of the continuum objective (32) and the Euler–Maruyama discretized loss (37) is not established. The Itô integral is replaced by a sum with the same Gaussian increments used to generate the path, but no consistency or stability analysis is supplied (the paper explicitly declines such an analysis for a related transition-operator discretization in SI.3). Since the loss is a squared residual evaluated on discretized paths, the time step appears inside the square and can bias the minimizer. The authors should either provide a consistency argument (e.g., convergence of minimizers as Δt→0 and L→∞) or at least a numerical study of sensitivity to Δt and L on the Müller–Brown test case.
minor comments (4)
  1. [III A] Only the best of 100 trainings is reported in each configuration. Selecting the model with the smallest independent validation loss can bias the reported RMSEs. Please report the mean and standard deviation over seeds, or show that the qualitative conclusions are stable across initializations.
  2. [II B] The definition of ℓ_k as the smallest integer such that Q^{k,ℓ_k}∈R∪P should include the case ℓ_k=∞ (or ℓ_k>L) for trajectories that never hit R∪P; otherwise the notation L∧ℓ_k in (37) is undefined for such trajectories.
  3. [II A] The indicator notation is inconsistent: the display preceding (31) uses 1_{ϑ⩾τ}, while (32) uses 1_{t⩾τ}. Please harmonize.
  4. [Introduction] There are several typographical errors, e.g. 'writes writes' after Eq. (6); 'Itô' is typeset inconsistently as 'It¯o' in the abstract. Please proofread.

Circularity Check

0 steps flagged

No significant circularity: the Itô-residual objective is a self-contained construction and the FE comparison is an independent benchmark.

full rationale

The paper's central derivation is not circular. The continuous loss (32) is constructed as the squared residual of an exact Itô identity satisfied by the committor (Eqs. (30)-(31)); the committor minimizing that residual is a consistency property of the construction, not a parameter fitted to the target. The neural network parameters are trained on trajectory data, while the reported errors are computed against a finite-element committor (Eqs. (40)-(41), Table II, Fig. 9), so no fitted parameter is relabeled as a prediction. The choices of h, a, b, epsilon, time-lag, and trajectory length are hyperparameters, and the log loss is borrowed from Ref. [32] only as a practical option; the core identity works for any strictly monotone h. Self-citations to AMS theory (Refs. [13,38]) are used for the sampling bootstrap, but the bootstrap is validated externally against the FE committor, so they are not load-bearing circularity. The skeptical observation about Eq. (37) is a genuine discretization/boundary-consistency concern: (37) drops the h(a1_P(q_tau)+b)1_{tau>=vartheta} term of (32), so constant functions have zero discrete loss and the implemented objective is not equivalent to the continuous one. Under the rubric given here, however, that is a correctness risk (an omitted boundary condition / missing error analysis, partly acknowledged in SI.3), not a circular reduction: it does not make the reported committor equal to the training input by construction. The scientific claim would be strengthened by fixing (37) to include the boundary indicator or a boundary penalty, but the circularity score remains low.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The proposed loss rests on standard stochastic calculus plus several unproved methodological choices: the uniqueness of the minimizer, the discretization of the Itô integral, and the AMS weighting of training trajectories. No physical entities or fitted physical constants are introduced.

free parameters (4)
  • Choice of transform h and scale (a,b) = h=Id (a=1,b=0); h=ln (a=1,b=ε) and (a=-1,b=1+ε)
    h,a,b are hand-selected to emphasize low/high committor regions; ε is set from prior knowledge of probability range, not fitted to target errors.
  • Time-lag ϑ / trajectory length L = varied in time steps; L=400 for Z-potential
    Loss (37) requires choosing trajectory length; no criterion is given and results depend on ϑ.
  • Regularization ε in log loss = 0.1, 0.01, 1e-5
    Prevents log(0) and sets the scale of probability resolution; chosen ad hoc.
  • AMS and stopping hyperparameters = Nrep=20/100; stop when linear regression of log committor between iterations ≈0.998
    Heuristic stopping criterion; not error-controlled.
axioms (6)
  • domain assumption The committor for overdamped Langevin dynamics solves Lχ=0 with Dirichlet boundary conditions (Eq. 6).
    Standard transition path theory; used to derive the zero-residual identity (31).
  • standard math Itô's formula applies to h(aχ+b) and h(af+b) along stopped trajectories.
    Core derivation of (30)-(32); requires smoothness, which the sigmoid NN satisfies.
  • domain assumption Euler-Maruyama discretization (26) is a consistent surrogate for the dynamics and loss (37) is a faithful surrogate for (32).
    No consistency or error analysis is provided in Section II B.
  • ad hoc to paper The objective (32) has the committor as its unique minimizer.
    Asserted in Section II A after Eq. (32); no proof in SI for this objective (SI proves analogous uniqueness only for (20), (25), (28)).
  • domain assumption AMS samples reactive trajectories with computable weights (Ref. 38), so cutting AMS trajectories into weighted sub-trajectories gives an unbiased training measure μ.
    Needed for the iterative scheme in Section II C; supported by prior work but not re-derived.
  • domain assumption Finite element committor is an accurate reference for validation.
    Used to define RMSE metrics; standard benchmark in the field.

pith-pipeline@v1.3.0-alltime-deepseek · 24673 in / 16944 out tokens · 155517 ms · 2026-08-01T07:28:18.794012+00:00 · methodology

0 comments
read the original abstract

Many molecular dynamics simulations aim at studying transitions between two states (from reactants to products). In this context, the committor function (which gives for a given molecular configuration the probability to reach the product state before the reactant state) is a pivotal quantity, in particular because it is the optimal importance function for rare event simulation methods such as importance sampling or splitting techniques. These methods are used to sample the reactive path ensemble, and estimate for example the transition rate. However, learning such a function is generally a challenging task due to the high dimensionality of the configuration space. In this work, after reviewing the existing methodologies to construct approximate committor functions, a new loss function based on the application of It\={o}'s formula is proposed to learn the committor function with a minimization procedure on the parameters of a neural network. After comparing this novel approach to existing procedures on the M\"uller--Brown potential, we introduce a coupling strategy with the Adaptive Multilevel Splitting method to better approximate the committor function using a better sampling of the reactive trajectories. This methodology in which the committor function is iteratively learned only requires initially the knowledge of the reactant and product states.

Figures

Figures reproduced from arXiv: 2607.21425 by Gabriel Stoltz, Thomas Pigeon, Tony Leli\`evre.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p024_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Errors on the approximate committor function as a function of the time-lag [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Errors on the approximate committor function as a function of the time-lag [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Errors on the approximate committor function as a function of the time-lag [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Z–potential [PITH_FULL_IMAGE:figures/full_fig_p029_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Result at iteration 0. (a) Initial conditions of the trajectories in the training [PITH_FULL_IMAGE:figures/full_fig_p030_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Results at iteration 1. (a) Initial conditions of the trajectories in the training [PITH_FULL_IMAGE:figures/full_fig_p031_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Results at iteration 2. (a) Initial conditions of the trajectories in the training [PITH_FULL_IMAGE:figures/full_fig_p032_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Best model obtained after the iterative training procedure using AMS. (a) (and [PITH_FULL_IMAGE:figures/full_fig_p032_9.png] 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

78 extracted references · 20 canonical work pages

  1. [7]

    Hartmann , author C

    author author C. Hartmann , author C. Sch \"u tte , author M. Weber , \ and\ author W. Zhang ,\ @noop journal journal Probab. Theory Relat. Fields \ volume 170 ,\ pages 177 ( year 2018 ) NoStop

    2018

  2. [14]

    Pigeon , author G

    author author T. Pigeon , author G. Stoltz , author M. Corral-Valero , author A. Anciaux-Sedrakian , author M. Moreaud , author T. Lelièvre , \ and\ author P. Raybaud ,\ 10.1021/acs.jctc.3c00280 journal journal Journal of Chemical Theory and Computation \ volume 19 ,\ pages 3538 ( year 2023 ) ,\ http://arxiv.org/abs/https://doi.org/10.1021/acs.jctc.3c0028...

    work page doi:10.1021/acs.jctc.3c00280 2023

  3. [16]

    Leli \`e vre \ and\ author G

    author author T. Leli \`e vre \ and\ author G. Stoltz ,\ @noop journal journal Acta Numer. \ volume 25 ,\ pages 681 ( year 2016 ) NoStop

    2016

  4. [18]

    Lelièvre , author M

    author author T. Lelièvre , author M. Rousset , \ and\ author G. Stoltz ,\ 10.1142/p579 title Free Energy Computations \ ( publisher Imperial College Press ,\ year 2010 )\ http://arxiv.org/abs/https://www.worldscientific.com/doi/pdf/10.1142/p579 https://www.worldscientific.com/doi/pdf/10.1142/p579 NoStop

    doi:10.1142/p579 2010

  5. [20]

    author author E. Vanden-Eijnden ,\ title Transition path theory , \ in\ 10.1007/3-540-35273-2_13 booktitle Computer Simulations in Condensed Matter Systems: From Materials to Chemical Biology Volume 1 ,\ editor edited by\ editor M. Ferrario , editor G. Ciccotti , \ and\ editor K. Binder \ ( publisher Springer Berlin Heidelberg ,\ address Berlin, Heidelber...

    doi:10.1007/3-540-35273-2_13 2006

  6. [22]

    Jung , author R

    author author H. Jung , author R. Covino , author A. Arjun , author C. Leitold , author C. Dellago , author P. G. \ Bolhuis , \ and\ author G. Hummer ,\ @noop journal journal Nat. Comput. Sci. \ volume 3 ,\ pages 334 ( year 2023 ) NoStop

    2023

  7. [26]

    author author G. M. \ Rotskoff , author A. R. \ Mitchell , \ and\ author E. Vanden-Eijnden ,\ in\ https://proceedings.mlr.press/v145/rotskoff22a.html booktitle Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference ,\ series Proceedings of Machine Learning Research , Vol.\ volume 145 ,\ editor edited by\ editor J. Bruna , editor J....

    2022

  8. [27]

    Kang , author E

    author author P. Kang , author E. Trizio , \ and\ author M. Parrinello ,\ @noop journal journal Nat. Comput. Sci. \ volume 4 ,\ pages 451 ( year 2024 ) NoStop

    2024

  9. [31]

    author author A. R. \ Mitchell \ and\ author G. M. \ Rotskoff ,\ 10.1021/acs.jctc.4c00997 journal journal Journal of Chemical Theory and Computation \ volume 20 ,\ pages 9378 ( year 2024 ) ,\ http://arxiv.org/abs/https://doi.org/10.1021/acs.jctc.4c00997 https://doi.org/10.1021/acs.jctc.4c00997 NoStop

    doi:10.1021/acs.jctc.4c00997 2024

  10. [32]

    Li , author Y

    author author H. Li , author Y. Khoo , author Y. Ren , \ and\ author L. Ying ,\ in\ https://proceedings.mlr.press/v145/li22a.html booktitle Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference ,\ series Proceedings of Machine Learning Research , Vol.\ volume 145 ,\ editor edited by\ editor J. Bruna , editor J. Hesthaven , \ and\ ...

    2022

  11. [36]

    Chen , author B

    author author H. Chen , author B. Roux , \ and\ author C. Chipot ,\ 10.1021/acs.jctc.3c00028 journal journal Journal of Chemical Theory and Computation \ volume 19 ,\ pages 4414 ( year 2023 ) ,\ http://arxiv.org/abs/https://doi.org/10.1021/acs.jctc.3c00028 https://doi.org/10.1021/acs.jctc.3c00028 NoStop

    doi:10.1021/acs.jctc.3c00028 2023

  12. [37]

    \ Br \'e hier , author M

    author author C.-E. \ Br \'e hier , author M. Gazeau , author L. Gouden \`e ge , author T. Leli \`e vre , \ and\ author M. Rousset ,\ @noop journal journal Ann. Appl. Probab. \ volume 26 ,\ pages 3559 ( year 2016 ) NoStop

    2016

  13. [39]

    author author S. Barschkis ,\ https://arxiv.org/abs/2310.02548 title Exact and soft boundary conditions in physics-informed neural networks for the variable coefficient poisson equation , \ ( year 2023 ),\ http://arxiv.org/abs/2310.02548 arXiv:2310.02548 [cs.LG] NoStop

    Pith/arXiv arXiv 2023

  14. [40]

    Giuseppe Chen , author C

    author author C. Giuseppe Chen , author C. Tang , author A. Megías , author R. A. \ Talmazan , author S. Contreras Arredondo , author B. Roux , \ and\ author C. Chipot ,\ 10.1021/acs.jctc.6c00007 journal journal Journal of Chemical Theory and Computation \ volume 22 ,\ pages 1258 ( year 2026 ) ,\ note pMID: 41544196 ,\ http://arxiv.org/abs/https://doi.org...

    work page doi:10.1021/acs.jctc.6c00007 2026

  15. [41]

    Meg \' as , author S

    author author A. Meg \' as , author S. Contreras Arredondo , author C. G. \ Chen , author C. Tang , author B. Roux , \ and\ author C. Chipot ,\ @noop journal journal Nat. Comput. Sci. \ volume 5 ,\ pages 592 ( year 2025 ) NoStop

    2025

  16. [42]

    Contreras Arredondo , author C

    author author S. Contreras Arredondo , author C. Tang , author R. A. \ Talmazan , author A. Meg \' as , author C. G. \ Chen , \ and\ author C. Chipot ,\ @noop journal journal Nat. Comput. Sci." \ volume 6 ,\ pages 350 ( year 2026 ) NoStop

    2026

  17. [43]

    M \"u ller \ and\ author L

    author author K. M \"u ller \ and\ author L. D. \ Brown ,\ @noop journal journal Theoretica Chimica Acta \ volume 53 ,\ pages 75 ( year 1979 ) NoStop

    1979

  18. [44]

    Pigeon ,\ @noop title 2dpotentials , \ howpublished https://github.com/thomaspigeon/2dpotentials ( year 2025 ) NoStop

    author author T. Pigeon ,\ @noop title 2dpotentials , \ howpublished https://github.com/thomaspigeon/2dpotentials ( year 2025 ) NoStop

    2025

  19. [46]

    author author M. K. \ Cameron ,\ @noop title transition\_path\_theory\_fem\_distmesh , \ howpublished https://github.com/mar1akc/transition_path_theory_FEM_distmesh ( year 2023 ) NoStop

    2023

  20. [47]

    Glasserman , author P

    author author P. Glasserman , author P. Heidelberger , author P. Shahabuddin , \ and\ author T. Zajic ,\ @noop journal journal IEEE Trans. Automat. Contr. \ volume 43 ,\ pages 1666 ( year 1998 ) NoStop

    1998

  21. [49]

    Laura J. S. Lopes and Tony Leli. Journal of Computational Chemistry , title =. 2019 , month =. doi:10.1002/jcc.25778 , publisher =

    doi:10.1002/jcc.25778 2019

  22. [50]

    e3nn: Euclidean Neural Networks , publisher =

    Geiger, Mario and Smidt, Tess , keywords =. e3nn: Euclidean Neural Networks , publisher =. 2022 , copyright =. doi:10.48550/ARXIV.2207.09453 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.09453 2022

  23. [51]

    Mario Geiger and Tess Smidt and Alby M. and Benjamin Kurt Miller and Wouter Boomsma and Bradley Dice and Kostiantyn Lapchevskyi and Maurice Weiler and Michał Tyszkiewicz and Simon Batzner and Dylan Madisetti and Martin Uhrin and Jes Frellsen and Nuri Jung and Sophia Sanborn and Mingjian Wen and Josh Rackers and Marcel Rød and Michael Bailey , howpublished...

    doi:10.5281/zenodo.6459381 2022

  24. [52]

    2023 , eprint=

    Exact and soft boundary conditions in Physics-Informed Neural Networks for the Variable Coefficient Poisson equation , author=. 2023 , eprint=

    2023

  25. [53]

    , editor =

    Vanden-Eijnden, E. , editor =. Transition Path Theory , year =. Computer Simulations in Condensed Matter Systems: From Materials to Chemical Biology Volume 1 , doi =

  26. [54]

    Reactive trajectories and the transition path process , year =

    Jianfeng Lu and James Nolen , journal =. Reactive trajectories and the transition path process , year =. doi:10.1007/s00440-014-0547-y , publisher =

    work page doi:10.1007/s00440-014-0547-y

  27. [55]

    Solving for high-dimensional committor functions using artificial neural networks , year =

    Yuehaw Khoo and Jianfeng Lu and Lexing Ying , journal =. Solving for high-dimensional committor functions using artificial neural networks , year =. doi:10.1007/s40687-018-0160-2 , publisher =

    doi:10.1007/s40687-018-0160-2

  28. [56]

    Computing committor functions for the study of rare events using deep learning , year =

    Qianxiao Li and Bo Lin and Weiqing Ren , journal =. Computing committor functions for the study of rare events using deep learning , year =. doi:10.1063/1.5110439 , publisher =

    doi:10.1063/1.5110439

  29. [57]

    Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =

    A semigroup method for high dimensional committor functions based on neural network , author =. Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =. 2022 , editor =

    2022

  30. [58]

    and Rotskoff, Grant M

    Mitchell, Andrew R. and Rotskoff, Grant M. , title =. Journal of Chemical Theory and Computation , volume =. 2024 , doi =

    2024

  31. [59]

    and Lu, Deyu and Schulten, Klaus , title = "

    Park, Sanghyun and Sener, Melih K. and Lu, Deyu and Schulten, Klaus , title = ". The Journal of Chemical Physics , volume =. 2003 , month =. doi:10.1063/1.1570396 , url =

    work page doi:10.1063/1.1570396 2003

  32. [60]

    , title = "

    Lechner, Wolfgang and Rogal, Jutta and Juraszek, Jarek and Ensing, Bernd and Bolhuis, Peter G. , title = ". The Journal of Chemical Physics , volume =. 2010 , month =. doi:10.1063/1.3491818 , url =

    work page doi:10.1063/1.3491818 2010

  33. [61]

    String method with swarms-of-trajectories, mean drifts, lag time, and committor , year =

    Benoît Roux , journal =. String method with swarms-of-trajectories, mean drifts, lag time, and committor , year =. doi:10.1021/acs.jpca.1c04110 , publisher =

    work page doi:10.1021/acs.jpca.1c04110

  34. [62]

    Theoretica Chimica Acta , year=

    Location of saddle points and minimum energy paths by a constrained simplex optimization procedure , author=. Theoretica Chimica Acta , year=

  35. [63]

    Journal of Chemical Theory and Computation , volume =

    Chen, Haochuan and Roux, Benoît and Chipot, Christophe , title =. Journal of Chemical Theory and Computation , volume =. 2023 , doi =

    2023

  36. [64]

    Iterative variational learning of committor-consistent transition pathways using artificial neural networks , author =. Nat. Comput. Sci. , publisher =. 2025 , copyright =

    2025

  37. [65]

    Learning the committor without collective variables , author =. Nat. Comput. Sci." , publisher =. 2026 , copyright =

    2026

  38. [66]

    and Contreras Arredondo, Sergio and Roux, Benoît and Chipot, Christophe , title =

    Giuseppe Chen, Cheng and Tang, Chenyu and Megías, Alberto and Talmazan, Radu A. and Contreras Arredondo, Sergio and Roux, Benoît and Chipot, Christophe , title =. Journal of Chemical Theory and Computation , volume =. 2026 , doi =

    2026

  39. [67]

    and Moroni, Daniele and Bolhuis, Peter G

    van Erp, Titus S. and Moroni, Daniele and Bolhuis, Peter G. , title =. The Journal of Chemical Physics , volume =. 2003 , month =. doi:10.1063/1.1562614 , url =

    doi:10.1063/1.1562614 2003

  40. [68]

    Allen and Patrick B

    Rosalind J. Allen and Patrick B. Warren and Pieter Rein ten Wolde , journal =. Sampling rare switching events in biochemical networks , year =. doi:10.1103/physrevlett.94.018104 , publisher =

    doi:10.1103/physrevlett.94.018104

  41. [69]

    Optimal control for sampling the transition path process and estimating rates , journal =

    Jiaxin Yuan and Amar Shah and Channing Bentz and Maria Cameron , keywords =. Optimal control for sampling the transition path process and estimating rates , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.cnsns.2023.107701 , url =

    arXiv 2024

  42. [70]

    Risk-neutral limit of adaptive importance sampling of random stopping times , journal =

    Carsten Hartmann and Annika Jöster , keywords =. Risk-neutral limit of adaptive importance sampling of random stopping times , journal =. 2024 , note =. doi:https://doi.org/10.1016/j.ifacol.2024.10.166 , url =

    work page doi:10.1016/j.ifacol.2024.10.166 2024

  43. [71]

    A large deviations perspective on the efficiency of multilevel splitting

    Glasserman, P and Heidelberger, P and Shahabuddin, P and Zajic, T. A large deviations perspective on the efficiency of multilevel splitting. IEEE Trans. Automat. Contr

  44. [72]

    The Journal of Chemical Physics , volume =

    Hummer, Gerhard , title =. The Journal of Chemical Physics , volume =. 2004 , month =. doi:10.1063/1.1630572 , url =

    work page doi:10.1063/1.1630572 2004

  45. [73]

    Torrie and J.P

    G.M. Torrie and J.P. Valleau , abstract =. Nonphysical sampling distributions in Monte Carlo free-energy estimation: Umbrella sampling , journal =. 1977 , issn =. doi:https://doi.org/10.1016/0021-9991(77)90121-8 , url =

    doi:10.1016/0021-9991(77)90121-8 1977

  46. [74]

    Journal of Physics: Condensed Matter , abstract =

    Bucko, T , title =. Journal of Physics: Condensed Matter , abstract =. 2008 , month =. doi:10.1088/0953-8984/20/6/064211 , url =

    work page doi:10.1088/0953-8984/20/6/064211 2008

  47. [75]

    Carter and Giovanni Ciccotti and James T

    E.A. Carter and Giovanni Ciccotti and James T. Hynes and Raymond Kapral , abstract =. Constrained reaction coordinate dynamics for the simulation of rare events , journal =. 1989 , issn =. doi:https://doi.org/10.1016/S0009-2614(89)87314-2 , url =

    work page doi:10.1016/s0009-2614(89)87314-2 1989

  48. [76]

    The activated complex in chemical reactions , year =

    Henry Eyring , journal =. The activated complex in chemical reactions , year =. doi:10.1063/1.1749604 , publisher =

    doi:10.1063/1.1749604

  49. [77]

    Yuk and Manh-Thuong Nguyen and Mal-Soon Lee and Vassiliki-Alexandra Glezakou and Roger Rousseau , journal =

    Greg Collinge and Simuck F. Yuk and Manh-Thuong Nguyen and Mal-Soon Lee and Vassiliki-Alexandra Glezakou and Roger Rousseau , journal =. Effect of collective dynamics and anharmonicity on entropy in heterogenous catalysis: Building the case for advanced molecular simulations , year =. doi:10.1021/acscatal.0c01501 , publisher =

    work page doi:10.1021/acscatal.0c01501

  50. [78]

    The Journal of Chemical Physics , volume =

    Roux, Benoît , title = ". The Journal of Chemical Physics , volume =. 2022 , month =. doi:10.1063/5.0084209 , url =

    doi:10.1063/5.0084209 2022

  51. [79]

    Committor-Consistent Variational String Method , year =

    Ziwei He and Christophe Chipot and Benoît Roux , journal =. Committor-Consistent Variational String Method , year =. doi:10.1021/acs.jpclett.2c02529 , publisher =

    work page doi:10.1021/acs.jpclett.2c02529

  52. [80]

    Frassek and A

    M. Frassek and A. Arjun and P. G. Bolhuis , journal =. An extended autoencoder model for reaction coordinate discovery in rare event molecular dynamics datasets , year =. doi:10.1063/5.0058639 , publisher =

    work page doi:10.1063/5.0058639

  53. [81]

    Machine-guided path sampling to discover mechanisms of molecular self-organization , author =. Nat. Comput. Sci. , volume =

  54. [82]

    Unbiasedness of some generalized adaptive multilevel splitting algorithms , author =. Ann. Appl. Probab. , publisher =

  55. [83]

    Analysis of adaptive multilevel splitting algorithms in an idealized case , year =

    Charles-Edouard Bréhier and Tony Lelièvre and Mathias Rousset , journal =. Analysis of adaptive multilevel splitting algorithms in an idealized case , year =. doi:10.1051/ps/2014029 , publisher =

    arXiv

  56. [84]

    Computing the committor with the committor to study the transition state ensemble , author =. Nat. Comput. Sci. , publisher =

  57. [85]

    Estimating committor functions via deep adaptive sampling on rare transition paths , journal =

    Yueyang Wang and Kejun Tang and Xili Wang and Xiaoliang Wan and Weiqing Ren and Chao Yang , keywords =. Estimating committor functions via deep adaptive sampling on rare transition paths , journal =. 2026 , issn =. doi:https://doi.org/10.1016/j.jcp.2025.114439 , url =

    arXiv 2026

  58. [86]

    and Lorpaiboon, Chatipat and Dinner, Aaron R

    Strahan, John and Guo, Spencer C. and Lorpaiboon, Chatipat and Dinner, Aaron R. and Weare, Jonathan , year=. Inexact iterative numerical linear algebra for neural network-based spectral estimation and rare-event prediction , volume=. The Journal of Chemical Physics , publisher=. doi:10.1063/5.0151309 , number=

    work page doi:10.1063/5.0151309

  59. [87]

    Raissi and P

    M. Raissi and P. Perdikaris and G.E. Karniadakis , keywords =. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , journal =. 2019 , issn =. doi:https://doi.org/10.1016/j.jcp.2018.10.045 , url =

    doi:10.1016/j.jcp.2018.10.045 2019

  60. [88]

    2010 , doi =

    Lelièvre, Tony and Rousset, Mathias and Stoltz, Gabriel , title =. 2010 , doi =

    2010

  61. [89]

    2017 , isbn =

    Baron Peters , title =. 2017 , isbn =. doi:https://doi.org/10.1016/B978-0-44-456349-1.00001-5 , url =

    work page doi:10.1016/b978-0-44-456349-1.00001-5 2017

  62. [90]

    Calculating free energies using average force , year =

    Eric Darve and Andrew Pohorille , journal =. Calculating free energies using average force , year =. doi:10.1063/1.1410978 , publisher =

    work page doi:10.1063/1.1410978

  63. [91]

    Laio and M

    A. Laio and M. Parrinello , journal =. Escaping free-energy minima , year =. doi:10.1073/pnas.202427399 , publisher =

    doi:10.1073/pnas.202427399

  64. [92]

    Acta Numer

    Partial differential equations and stochastic methods in molecular dynamics , author =. Acta Numer. , publisher =

  65. [93]

    Multiscale Modeling & Simulation , volume =

    Dupuis, Paul and Spiliopoulos, Konstantinos and Wang, Hui , title =. Multiscale Modeling & Simulation , volume =. 2012 , doi =. https://doi.org/10.1137/110842545 , abstract =

    work page doi:10.1137/110842545 2012

  66. [94]

    Importance sampling in path space for diffusion processes with slow-fast variables , author =. Probab. Theory Relat. Fields , publisher =

  67. [95]

    Journal of Statistical Mechanics: Theory and Experiment , abstract =

    Hartmann, Carsten and Schütte, Christof , title =. Journal of Statistical Mechanics: Theory and Experiment , abstract =. 2012 , month =. doi:10.1088/1742-5468/2012/11/P11004 , url =

    work page doi:10.1088/1742-5468/2012/11/p11004 2012

  68. [96]

    Journal of Chemical Theory and Computation , volume =

    Pigeon, Thomas and Stoltz, Gabriel and Corral-Valero, Manuel and Anciaux-Sedrakian, Ani and Moreaud, Maxime and Lelièvre, Tony and Raybaud, Pascal , title =. Journal of Chemical Theory and Computation , volume =. 2023 , doi =

    2023

  69. [97]

    Journal of Physics: Condensed Matter , abstract =

    Allen, Rosalind J and Valeriani, Chantal and Rein ten Wolde, Pieter , title =. Journal of Physics: Condensed Matter , abstract =. 2009 , month =. doi:10.1088/0953-8984/21/46/463102 , url =

    doi:10.1088/0953-8984/21/46/463102 2009

  70. [98]

    On the Asymptotic Normality of Adaptive Multilevel Splitting , year =

    Cérou, Frédéric and Delyon, Bernard and Guyader, Arnaud and Rousset, Mathias , journal =. On the Asymptotic Normality of Adaptive Multilevel Splitting , year =. doi:10.1137/18M1187477 , eprint =

    work page doi:10.1137/18m1187477

  71. [99]

    Adaptive multilevel splitting for rare event analysis , year =

    Frédéric Cérou and Arnaud Guyader , journal =. Adaptive multilevel splitting for rare event analysis , year =. doi:10.1080/07362990601139628 , publisher =

    doi:10.1080/07362990601139628

  72. [100]

    Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =

    Active Importance Sampling for Variational Objectives Dominated by Rare Events: Consequences for Optimization and Generalization , author =. Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =. 2022 , editor =

    2022

  73. [101]

    Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =

    Active importance sampling for variational objectives dominated by rare events: consequences for optimization and generalization , author =. Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference , pages =. 2022 , editor =

    2022

  74. [102]

    Tal , journal =

    Eric Vanden-Eijnden and Fabio A. Tal , journal =. Transition state theory: variational formulation, dynamical corrections, and error estimates , year =. doi:10.1063/1.2102898 , publisher =

    work page doi:10.1063/1.2102898

  75. [103]

    Dinner and Jonathan Weare , keywords =

    John Strahan and Justin Finkel and Aaron R. Dinner and Jonathan Weare , keywords =. Predicting rare events using neural networks and short-trajectory data , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.jcp.2023.112152 , url =

    arXiv 2023

  76. [104]

    2dpotentials , year =

    Thomas Pigeon , howpublished =. 2dpotentials , year =

  77. [105]

    Cameron , howpublished =

    Maria K. Cameron , howpublished =. transition\_path\_theory\_FEM\_distmesh , year =

  78. [106]

    Sharp error estimates for target measure diffusion maps with applications to the committor problem , journal =

    Shashank Sule and Luke Evans and Maria Cameron , keywords =. Sharp error estimates for target measure diffusion maps with applications to the committor problem , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.acha.2025.101803 , url =

    arXiv 2025