acceptodds
Under review as a conference paper at ICLR 2027

Kolmogorov-Chapman Maps: Trajectory-Free Learning for Simulating Molecular Dynamics

11 views

Abstract

Stochastic differential equations (SDEs) describe stochastic processes across a wide range of scientific domains, including physics, biology, and finance. Although individual numerical integration steps are inexpensive, observing meaningful events or transitions can require millions of sequential steps, resulting in substantial wall-clock simulation time and poor hardware utilization. Deep learning offers an alternative by amortizing many integration steps into a single neural network forward pass, potentially enabling much faster simulation. However, existing supervised approaches typically require long simulated trajectory data, which restricts their applicability. In this work, we introduce Kolmogorov-Chapman Maps (KC-maps), a method for learning the transition kernel of an SDE that relies only on the trajectory-free marginal samples together with the known drift and diffusion coefficients. Given an initial time and state of the process, KC-maps models its distribution integrated to any given time, corresponding to the weak solution of the SDE. To recover the correct dynamics, KC-maps satisfies two key properties: (i) it satisfies the boundary conditions, which we incorporate directly into the neural network architecture, and (ii) it respects the Chapman–Kolmogorov equation, which provides an efficient training objective. We empirically demonstrate that KC-maps learns to integrate the dynamics of peptide molecular systems and outperforms classical molecular dynamics in terms of GPU-hours. The implementation is available at https://anonymous.4open.science/r/kcm-accel-md-anon-5E1Danonymous.4open.science/r/kcm-accel-md-anon-5E1D.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Trajectory-free learning, normalizing flows, accelerated molecular dynamics, SDEs, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7zHeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method called KC maps, which learns a finite-time transition density for an SDE from equilibrium samples and forces alone, without needing time-correlated samples like trajectories. It does so by training conditional normalising flows that map from the current state to the time-warped distribution. The loss is based on two-hop self-consistency, and the architecture smartly incorporates agreement between the known force and diffusion coefficients. Excluding training, it records higher throughput than MD at the cost of moderate error in the molecular systems.

Critical strengths

I think the theory is very elegant and easy to follow, and the paper is quite well written.

The method closes a literature gap. Furthermore, they modify the normalising flow coupling-layer architecture to satisfy most constraints before training, leaving only the two-hop consistency to the loss. For the toy potential, they also find they can use higher-temperature configurations to train the model, which is very promising for the future of a method like this. This would drastically speed up data acquisition, since data collection and training on these single systems make the method a bit moot right now. That said, this is of course understandable, given that this is early-stage method development. I wouldn't have expected the offer to train this on multiple systems, though that would have been a great plus.

Critical weaknesses

In my view, the biggest weakness is the lack of kinetics validation. Given that the data come from equilibrium simulations and the model outputs the same quantities, it would be interesting to include kinetics, especially since one can derive the rates from the reference data. One might imagine training data coming from umbrella sampling, which forsakes kinetics, thus already giving the method a use case, since it could supply those in principle.

Then the presented setup is a bit of a chicken-and-egg problem. Given that one now only trains on one system, it would have been nice to see the authors train on multiple systems, where a DL sampling method would make more sense, since the method would then amortise.

Questions for authors

Could you compute rates using this method? It should be possible, because you can get first passage times and the like just by sampling with this method.

Similarly, for the AL6 experiments, you reach a mode much faster than MD, and MD does not discover the state in 7 times as much simulation time. Is this a statistical accident, or is it related to any kinetic problems you might have? If so, where do they come from? The paper should expand on this. For instance, because it's mass-covering, the KL objective could make the kernel too wide, and clipping could smooth the landscape.

Can you set up a situation where you remove an entire mode from the training data? If so, can the method discover that mode? expectation is no, and that the method like this only shines once one amortises over different systems, but I'm still very curious to find out

There is an energy mismatch in the molecular systems, and I think a Metropolis step could alleviate it. In that case, how would the results change, and what would the acceptance rate be?

6 comments
AuthorsOpenReview ↗

Response #1 to Reviewer 7zHe

We thank the reviewer for their valuable feedback and greatly appreciate their comment on the theory being elegant and paper being easy to follow! We aim to resolve the reviewer’s concerns in the following:

“Lack of kinetics validation.”

As the reviewer requested, we provide results evaluating the kinetics for alanine dipeptide for KC-Maps against the reference MD where we measure the mean first passage times (see figure 4.b), autocorrelation of mode occupancy (see figure 4.c), and the transition probabilities between modes (see Appendix E.2 and Figure A2). We added a paragraph discussing the results (see Kinetics in Section 4.3) and a description of the metrics in Appendix D.4. From the results, we see that our method, with a larger step size of τ=30 \tau=30\,fs, is closely aligned with MD, especially in the larger modes where training data is abundant. We thank the reviewer for bringing up this point and for their suggestion as we believe this substantially strengthens our work.

It would have been nice to see the authors train on multiple systems

Per the reviewer's request, we ran a preliminary study on amortization. We trained KC-Maps on 200 dipeptides from the ManyPeptidesMD dataset and evaluated it on 30 unseen dipeptides. Table A6 (Appendix E.4) reports the energy, torsion, and TICA W2, averaged over the 30 test dipeptides. The model generalizes well to unseen peptides. At τ=30 fs\tau = 30\,\mathrm{fs}, its average energy error is 0.48 kBT0.48\,k_BT. Its TICA error (0.1120.112) is of the same order as MD’s (0.0580.058), showing that the model recovers the slow conformational modes of molecules it has never seen.

We also compare against two transferable baselines, Timewarp [5] and Prose [4], with numbers taken from Table 2 of [4] (the paper does not report TICA for dipeptide). We see KC-Maps outperforms Timewarp on both energy (0.480.48 vs. 4.534.53) and torsion (0.550.55 vs. 0.840.84) W2\mathcal{W}_2. On energy it is competitive with Prose (0.480.48 vs. 0.370.37), even though Prose relies on self-normalized importance sampling (SNIS) and KC-Maps uses no reweighting. Finally, 16 MD chains converge quite well to the equilibrium within one GPU-hour, as we also observe for AL2 in the single-peptide experiments (Table 3). The advantage of KC-Maps appears on larger systems where MD mixes slowly. Namely, on AL4 and AL6, KC-Maps reaches lower TICA error than MD at the same compute budget (Table 3) and explores the reference modes faster (Figure 3). Given the time constraints, we only ran this study for dipeptides, and we leave a more focused study on amortization and scaling to future work.

Similarly, for the AL6 experiments, you reach a mode much faster than MD, and MD does not discover the state in 7 times as much simulation time. Is this a statistical accident, or is it related to any kinetic problems you might have? If so, where do they come from?

We appreciate the question and hope to add clarification – the methods differ in simulated time and KC-Maps simulate for longer given the same compute budget (single GPU-hour on L40S). As we can see in figure 3, for 16 chains, we have ~505 total ns / GPU-Hour for KC-Maps (w/ step-size \tau=40fs) vs ~183 total ns / GPU-hour for MD, which results in a per-chain simulation time of ~31 ns and ~11 ns respectively. The increased throughput is primarily a result of two reasons: (i) the larger step size of KC-Maps, and (ii) the better utilization of the GPU of neural networks versus the force field. As for MD, the low per-chain simulation time can be explained by force fields poorly utilizing the GPU, especially with increasing number of chains as seen in figure 3.b where MD saturates at 4 chains for AL6.

In contrast, KC-Maps has a higher throughput ceiling as the number of chains continues to increase. This results in our method reaching a longer time horizon per chain whereas MD may not have just had enough time to reach that mode. Indeed, we see in the reference MD with longer trajectory in figure 3.a, MD eventually does recover that mode. Moreover, as we can see from the kinetic results in an earlier answer, we see the issue is not because of kinetics.

AuthorsOpenReview ↗

Response #2 to Reviewer 7zHe

We continue addressing the reviewer concerns in the following:

Can you set up a situation where you remove an entire mode from the training data?

As the reviewer requested, we ran a longer simulation on AL6 for KC-Maps beyond the single GPU-hour and observed a mode recovered that was not present in the training data (sourced from [6]). To verify, we ran a larger MD run where we ran 16 chains of 2 microseconds per chain and indeed this mode is not an artifact of our method. To be clear, we kept the training data the same (sourced from Tan et al. [6]) which does not include that additional mode and illustrate this in Figure A1. As a result, we see our method is able to discover modes not present in the training data. We also show updated equilibrium metrics for AL6 using the new reference MD in Table A5.

There is an energy mismatch in the molecular systems, and I think a Metropolis step could alleviate it. In that case, how would the results change, and what would the acceptance rate be?

We agree that there is an energy mismatch for AL6 though we argue that energy distribution for AL2 and AL6 is well-matched; see figure 2). Though, we agree that a Metropolis step could help alleviate it. We provide the acceptance rates for all three systems (see Figure A3). We observe that, for AL6, it has low acceptance rates (< 1 %) for \tau > 20 fs so it makes MH-adjusted run not practical. However, we conducted an additional ablation where we incorporate the detailed-balance as a loss objective in the last 15% steps of training, and we observe that this addresses the energy issue seen in AL6 as well as improving the results in general. We report the results in Table 2 and add a dedicated section in Section 3.3.

Then the presented setup is a bit of a chicken-and-egg problem.

We agree that the trajectory-free setting indeed poses a "chicken and egg" problem by not specifying the source of the samples from the equilibrium distribution. Our main contribution is the method (KC-Maps) that can learn the kinetics only from independent samples at the equilibrium distribution. Therefore, we believe that our method would yield a practical solution when paired with any of the following approaches:

  • Getting uncorrelated equilibrium samples from other sampling algorithms like Sequential Monte Carlo [1], Parallel Tempering MCMC [2], Annealed Importance Sampling [3] and then using the proposed method for the kinetics.
  • Learning our method at a higher temperature and reducing the temperature for the inference (see Section 4.2 and Table 2 for Lennard-Jones experiment where training samples come from higher temperature data, but we learn the dynamics for the target temperature/distribution).
  • Learning our method on partially available data, e.g. partial mode coverage. instead of the equilibrium distribution (see experiment in Appendix A.1 Figure A1)
  • Amortizing our method across systems, i.e. it has been demonstrated previously that models can generalize to unseen systems [4, 5]. As suggested by the reviewer, we additionally demonstrate a small study on amortizing across dipeptides (see Appendix E.4).

Our paper does not provide a definitive answer, but we envision multiple possible solutions as described in what follows. Each of the proposed solutions deserves an independent study in future work.

[1] Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B, 68(3), 411–436.

[2] Hukushima, K., & Nemoto, K. (1996). Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6), 1604–1608.

[3] Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing, 11(2), 125–139.

[4] Tan, C. B., Hassan, M., Klein, L., Syed, S., Beaini, D., Bronstein, M. M., Tong, A., & Neklyudov, K. (2025). Amortized sampling with transferable normalizing flows. arXiv:2508.18175.

[5] Klein, L., Foong, A. Y. K., Fjelde, T. E., Mlodozeniec, B., Brockschmidt, M., Nowozin, S., Noé, F., & Tomioka, R. (2023). Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics. NeurIPS 2023.

[6] Tan, C. B., Bose, A. J., Lin, C., Klein, L., Bronstein, M. M., & Tong, A. (2025). Scalable equilibrium sampling with sequential Boltzmann generators. ICML 2025.

Thank you for the extensive revision and the many new experiments. The multi-system training on dipeptides, the AL2 kinetics, the unseen AL6 mode and the detailed-balance loss go well beyond what I asked for and address most of my concerns. Two points remain.

  1. Mode discovery on AL6. My concern is about simulated time rather than compute. By the time KC-maps finds the third mode, it has simulated much less time than MD did over the full hour without finding it (roughly 25 ns against 183 ns). If both followed the same dynamics, this would be unexpected, so I suspect the original model crosses barriers too quickly. The AL2 kinetics use the new L_DB model, while Fig. 3 appears to still show the original one, so they don’t resolve this. Could you rerun the mode-discovery experiment with the L_DB model, ideally with a few replicas? If it no longer finds the mode much faster than MD, the corresponding claims in the abstract and introduction should be adjusted.

  2. Acceptance with and without L_DB. Fig. A3 reports acceptance only for the model trained with Eq. (20). Since L_DB is the expected negative log Metropolis–Hastings ratio, the same curve without it would show directly what the new term contributes. Please also clarify which model the “<1 % for τ > 20 fs” in your response refers to.

Great work in general; I am considering raising my score.

Reviewer_2EDGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KC-maps, which learns SDE transition kernels from unpaired states. The method incorporates the short-time dynamics into a normalizing flow and trains it using Kolmogorov–Chapman consistency. Experiments on Lennard-Jones systems and alanine peptides show promising sampling performance, with faster simulation and improved conformational coverage on peptides.

Critical strengths
  • The paper is generally well written.

  • The idea of learning transition kernels through Kolmogorov-Chapman consistency is interesting.

  • The experiments show faster simulation and better exploration of metastable states compared to baselines.

Critical weaknesses
  • The long-run experiments mainly evaluate equilibrium distributions and mode coverage. Since the objective is to learn the correct dynamics, and these metrics could in principle also be matched by a kernel with different kinetics, it would be helpful to see a dynamical evaluation as well. Could the authors evaluate transition rates or relaxation times? I think this is important since the goal is to learn the transition kernel of a given SDE.
  • In Table 3, the proposed method improves TICA-W2 but is somewhat worse on energy and torsion W2 than the baselines across all three peptides. While better mode coverage is useful, the overall picture is a bit mixed. Could the authors clarify this tradeoff?
Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2EDG

We sincerely thank the reviewer for their time and effort in reviewing our work. We are pleased that the reviewer finds the idea interesting and finds the paper to be well written. We aim to resolve the reviewer’s concerns in what follows.

The long-run experiments mainly evaluate equilibrium distributions and mode coverage. Since the objective is to learn the correct dynamics, and these metrics could in principle also be matched by a kernel with different kinetics, it would be helpful to see a dynamical evaluation as well. Could the authors evaluate transition rates or relaxation times? I think this is important since the goal is to learn the transition kernel of a given SDE.

We agree on the importance of evaluating the kinetics of our kernel to demonstrate we learn the correct dynamics. Per the reviewer’s request, we provide results evaluating the kinetics for alanine dipeptide for KC-Maps against the reference MD using three metrics: the mean first passage times (see figure 4.b), autocorrelation of mode occupancy (see figure 4.c), and the transition probabilities between modes (see Appendix E.2 and Figure A2). We added a paragraph discussing the results (see Kinetics in Section 4.3) and a description of the metrics in Appendix D.4. From the results, we see that our method, with a larger step size of τ=30\tau=30fs, is closely aligned with MD, especially in the larger modes where training data is abundant. We thank the reviewer for bringing up this point and for their suggestion as we believe this substantially strengthens our work.

In Table 3, the proposed method improves TICA-W2 but is somewhat worse on energy and torsion W2 than the baselines across all three peptides. While better mode coverage is useful, the overall picture is a bit mixed. Could the authors clarify this tradeoff?

Thank you for pointing out this tradeoff between the metrics, we believe that the difference in the mode coverage performance is of much higher importance than the discrepancy in the energy and torus W2 metrics as we demonstrate in Fig. 3 where MD and HFM are missing modes. Moreover, we see in figure 2, that the energy histograms are quite well matched for AL2 and AL4, but we see this is not the case for AL6. However, we performed an additional ablation where we finetune the model with the detailed balance loss (eq. 20) and bridge the gap in the metrics (see Table 2).

We would like to thank the reviewer again for their feedback. We hope we addressed your concerns and if there is anything else, please let us know.

Thank you for the detailed response. I will keep my score.

Reviewer_U365OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce Kolmogorov-Chapman Maps (KC-maps), a method for learning finite-time transition kernels of stochastic differential equations without requiring reference trajectories for training. The approach combines conditional normalizing flows with architectural constraints that preserve the dynamics of the underlying SDE and a training objective based on the Chapman-Kolmogorov equation. Evaluated on Lennard-Jones systems and three alanine peptides of increasing size, the method achieves higher simulation throughput and faster exploration of conformational states than conventional molecular dynamics and Hamiltonian Flow Maps. However, the experimental evaluation primarily focuses on equilibrium sampling and conformational exploration, without assessing the accuracy of dynamical properties such as transition rates or relaxation times.

Critical strengths

S1. Novel approach to a relevant scientific problem: The paper tackles an important scientific problem in a novel way.

S2. Potential for accelerated molecular dynamics: At least from my high-level understanding, the authors propose a potentially powerful method for accelerating molecular simulations. Unlike approaches such as Boltzmann generators, which enable equilibrium conformational sampling, the proposed method could, in principle, also be used to calculate dynamical properties such as transition rates.

Critical weaknesses

W1. Unclear practical advantages of trajectory-free learning: While the method avoids trajectory-based supervision, meaning that it does not require consecutive molecular configurations, it still relies on hundreds of thousands of molecular configurations for training. These configurations must be generated somehow, typically through expensive MD simulations. Therefore, the practical advantage of trajectory-free learning remains unclear.

W2. Limited experimental evaluation: Although the proposed method is potentially quite powerful (S2), particularly because it could provide access to dynamical properties that are often more interesting and scientifically relevant than equilibrium properties, the experimental evaluation focuses on relatively simple equilibrium properties and rather small, simplified molecular systems (short alanine peptides with implicit solvent).

Questions for authors

Q1. Exploration beyond the training data: Can the proposed method discover metastable conformational states that are not represented in the training data? Since training relies on configurations obtained from extensive MD simulations, it would be interesting to understand to what extent the method can explore previously unvisited regions of conformational space, rather than merely reproducing the states already covered by the training data.

5 comments
AuthorsOpenReview ↗

Response to Reviewer U365

We thank the reviewer for their valuable feedback and for their time and effort in reviewing our work. We appreciate their evaluation of our approach, kindly remarking its novelty and potential for accelerated molecular dynamics. In what follows, we address their requested clarifications and concerns:

W1. Unclear practical advantages of trajectory-free learning

Our main contribution is the method (KC-Maps) that can learn the kinetics only from independent samples at the equilibrium distribution. Therefore, we believe that our method would yield a practical solution when paired with any of the following approaches:

  • Getting uncorrelated equilibrium samples from other sampling algorithms like Sequential Monte Carlo [1], Parallel Tempering MCMC [2], Annealed Importance Sampling [3] and then using the proposed method for the kinetics.
  • Learning our method at a higher temperature and reducing the temperature for the inference (see Section 4.2 and Table 2 for Lennard-Jones experiment where training samples come from higher temperature data, but we learn the dynamics for the target temperature/distribution).
  • Learning our method on partially available data, e.g. partial mode coverage. instead of the equilibrium distribution (see experiment in Appendix A.1 and Figure A1)
  • Amortizing our method across systems, i.e. it has been demonstrated previously that models can generalize to unseen systems [4, 5]. Additionally, we demonstrate a small study on amortizing across dipeptides (see Appendix E.4).

Furthermore, even when generating data via MD, our method can be used to achieve significant memory savings by aggressively pruning the trajectory data. Indeed, supervised methods methods require time-ordered pairs at the target time lag which means the frames of the trajectory must be saved no larger than minimum of the range of τ\tau e.g. we learn τ∈[1,100] \tau \in [1, 100]\,fs which means we need to save frames every 1 fs. This results in large amounts of memory storage versus in our case, we can save in 1 ps frames or even larger, so it is storage efficient.

We made adjustments to the text (Lines 52-69) to make this point more clear in our work. We thank the reviewer for this valuable feedback which improves the quality of our work.

W2. Limited experimental evaluation

We agree with the reviewer on the importance of dynamical properties. As such, we provide results evaluating the kinetics for alanine dipeptide for KC-Maps against the reference MD using three metrics: the mean first passage times (see figure 4.b), autocorrelation of mode occupancy (see figure 4.c), and the transition probabilities between modes (see Appendix E.2 and Figure A2). We added a paragraph discussing the results (see Kinetics in Section 4.3) and a description of the metrics in Appendix D.4. From the results, we see that our method, with a larger step size of τ=30 \tau=30\,fs, is closely aligned with MD, especially in the larger modes where training data is abundant. We thank the reviewer for bringing up this point and for their suggestion as we believe this substantially strengthens our work.

Q1. Exploration beyond the training data:

As the reviewer requested, we ran a longer simulation on AL6 for KC-Maps beyond the single GPU-hour and observed a mode recovered that was not present in the training data (sourced from [6]). To verify, we ran a larger MD run where we ran 16 chains of 2 microseconds per chain and indeed this mode is not an artifact of our method. To be clear, we kept the training data unchanged (sourced from Tan et al. [6]), which does not include that additional mode and illustrate this in Figure A1. As a result, we see our method is able to discover modes not present in the training data. We also show updated equilibrium metrics for AL6 using the new reference MD in Table A5. We thank the reviewer for their question/suggestion as we believe this result certainly strengthens our paper.

[1] Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B, 68(3), 411–436.

[2] Hukushima, K., & Nemoto, K. (1996). Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6), 1604–1608.

[3] Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing, 11(2), 125–139.

[4] Tan, C. B., Hassan, M., Klein, L., Syed, S., Beaini, D., Bronstein, M. M., Tong, A., & Neklyudov, K. (2025). Amortized sampling with transferable normalizing flows. arXiv:2508.18175.

[5] Klein, L., Foong, A. Y. K., Fjelde, T. E., Mlodozeniec, B., Brockschmidt, M., Nowozin, S., Noé, F., & Tomioka, R. (2023). Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics. NeurIPS 2023.

[6] Tan, C. B., Bose, A. J., Lin, C., Klein, L., Bronstein, M. M., & Tong, A. (2025). Scalable equilibrium sampling with sequential Boltzmann generators. ICML 2025.

Thank you for your detailed rebuttal and the substantial additional experiments. I believe you have satisfactorily addressed my concerns regarding W1, W2, and Q1. In particular, the new kinetic validation considerably strengthens the paper, although it remains limited to AL2, the smallest molecular system.

Given these improvements, I am inclined to increase my rating to a weak acceptance.

However, the other reviewers' comments and the revisions to the manuscript have raised a couple of follow-up concerns regarding the accuracy of the learned dynamics on larger molecular systems.

In particular, the very low Metropolis-Hastings acceptance rates for AL6 (Figure A3) suggest substantial deviations from the target dynamics at the step sizes used for acceleration. Does Figure A3 evaluate the original model or the detailed-balance-regularized version? Furthermore, does the detailed-balance regularization preserve the kinetic accuracy demonstrated on AL2?

AuthorsOpenReview ↗

Response to Comment by Reviewer U365

We thank the reviewer promptness in their response, and greatly appreciate their inclination to raise their score. We hope to add clarification in the following:

Does Figure A3 evaluate the original model or the detailed-balance-regularized version? Furthermore, does the detailed-balance regularization preserve the kinetic accuracy demonstrated on AL2?

Both Figure A3 and the AL2 kinetics evaluation use the model trained with the additional detailed-balance loss, so the kinetic accuracy shown on AL2 is that of the regularized model. The detailed-balance loss preserves this accuracy because training continues to minimize the Chapman–Kolmogorov loss LKC\mathcal{L}_{\mathrm{KC}}. We made edits to the manuscript to make this more clear.

The very low Metropolis-Hastings acceptance rates for AL6 (Figure A3) suggest substantial deviations from the target dynamics at the step sizes used for acceleration.

Indeed, acceptance rates below 100% mean we did not learn a perfect kernel, nor do we claim this. While theoretically, the Kolmogorov-Chapman objective (with and without additional detailed-balance loss) learn the exact transition kernel at the global optimum. In practice, however, finite capacity and finite training result in an imperfect model, as we stated in Lines 310-321 in the detailed-balance section. However, low acceptance does not imply large errors in the dynamics since very small model errors can give low acceptance.

In particular, MH acceptance tests exactness pointwise in the full joint space of position and momentum. For the acceptance ratio rr with samples from x∼πx\sim \pi and y∼q(⋅∣x,τ)y\sim q(\cdot \mid x, \tau), we have E[log⁡r]=−LDB\mathbb{E}[\log r] = -\mathcal{L}_{\mathrm{DB}}, and that KL grows roughly linearly with the number of degrees of freedom, e.g., 378 dimensions for AL6 against 132 for AL2. At the same accuracy per coordinate, the acceptance falls exponentially with system size [1, 2].

Moreover, the friction γ\gamma influences the acceptance rates since a lower γ\gamma (γ=0.3 ps−1\gamma = 0.3\,\mathrm{ps}^{-1} for AL6 vs γ=1 ps−1\gamma = 1\,\mathrm{ps}^{-1} for AL2) results in less noise being injected in momentum (∝2γτ\propto \sqrt{2\gamma \tau}). Therefore, the transition kernel is sharply concentrated at these step sizes, and a narrow density is sensitive to small errors.

Lastly, the accuracy of the learned dynamics is best assessed through the observables they reproduce, e.g., the equilibrium distributions in Table 3 and the AL2 kinetics in figure 4. On both, our model closely matches or outperforms the MD baseline, measured against long reference MD.

[1] Roberts, G. O., & Rosenthal, J. S. (1998). Optimal scaling of discrete approximations to Langevin diffusions. Journal of the Royal Statistical Society: Series B, 60(1), 255–268.

[2] Beskos, A., Pillai, N., Roberts, G., Sanz-Serna, J.-M., & Stuart, A. (2013). Optimal tuning of the hybrid Monte Carlo algorithm. Bernoulli, 19(5A), 1501–1534.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.