acceptodds
Under review as a conference paper at ICLR 2027

Langevin Flow Maps: Efficient Molecular Dynamics and Transition Path Sampling

12 views

Abstract

Molecular dynamics simulations proceed by integrating the Langevin equations over many small femtosecond timesteps. This poses a challenge for estimating ensemble properties and transition dynamics that occur on much longer timescales. We introduce Langevin Flow Maps, which extend machine-learned force-fields to additionally learn the stochastic Langevin integrator. We show that Langevin Flow Maps enable large-timestep molecular dynamics and recover accurate dynamical properties of the system, while running an order of magnitude faster than current machine-learned force fields. Further, by training on a diverse molecular dataset, we demonstrate a path towards transferable Langevin Flow Maps.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow maps, molecular dynamics, langevin dynamics, transition path sampling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xtuyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper learns a large-timestep integrator for Langevin dynamics. Given the current state and the random noise path over the step, a network predicts the next position and velocity for steps of up to about 10 fs. The noise enters exactly, and the network learns only the time-averaged velocity and force. Training uses shuffled equilibrium configurations and forces rather than time-ordered trajectories. It combines force and velocity regression at zero step size with a self-consistency loss that ties one step of size h to two steps of size h/2. This extends recent stochastic flow-map work (McCallum et al.) to Langevin dynamics. Tests cover small molecules (MD17), the kinetics of alanine dipeptide, and a transferable model for short peptides.

Critical strengths

The design is clean. Learning the Langevin step directly is meant to remove the splitting bias that Hamiltonian flow maps have when combined with a separate noise step. The paper argues this but does not test it head-to-head on dynamics. On small molecules the accuracy is convincing. On ethanol, a 10 fs step is about as accurate as a standard integrator at about 1 fs (Fig. 2). Structural errors are below both the earlier Hamiltonian flow-map method and the reference ML force field (Fig. 3). In the peptide setting, large steps beat Amber14 on state-occupancy and transition-probability error in 10 of 12 cases at 4 and 8 fs (Table 3).

Critical weaknesses
  1. The practical speed advantage isn't shown. The abstract's "order of magnitude faster than current machine-learned force fields" is real, but Fig. 5 shows it is simply the 10 fs vs 1 fs step ratio (App. C.2). Against classical MD on the same molecule it is only about 13% faster. For larger peptides, LFM at 8 fs becomes slower than classical MD at 1 fs beyond about 70 atoms (Fig. 14). No baseline uses the 2–4 fs steps common in practice, and training wall-clock time isn't reported. The convergence comparisons (Figs. 4 and 12, Table 3) are made at equal network evaluations, which is not equal wall-clock, since one LFM step costs several classical steps (Fig. 14).
  2. The large-step kinetics (transition times and fluxes) are checked on one small system, and the peptide model is biased. Transition times and fluxes are compared with MD only for alanine dipeptide, with no error bars, and all three LFM transition times come out about 6–8% shorter than MD (Fig. 5a). For peptides, the only dynamics metrics are averaged over three peptides per length with no spread (Table 3). At 8 fs the energy distribution is 6–12× further from the reference than standard MD (torsions 1.2–2.3×). There is no Metropolis or reweighting step to remove this bias. An inference-time temperature corrector is used (App. C.3.1), but the main text never mentions it. The authors propose exploring with large steps and then relaxing with small ones (Sec. 4.2), which is reasonable, but it is shown only as a single-peptide example (Fig. 13), with no comparison of discovery rates against plain MD at matched cost. Sec. 4.2's "consistently outperform" also fails in 2 of the 12 large-step cells.

Overall: Convincing single-molecule accuracy, but the practical speed advantage and large-step accuracy on peptides aren't yet established. Error bars, plus claims narrowed to what is shown, would move me to 3; so would a wall-clock win over tuned classical MD.

Questions for authors
  1. Can you compare wall-clock time against classical MD at 2–4 fs, and report training time?
  2. Can you add error bars to Fig. 5 and Table 3? Are the shorter transition times outside MD sampling error, and what friction does the reference MD use?
  3. Table 3 compares at equal network evaluations (10⁷), so the 8 fs model simulates 8× more time. How do the results compare at equal simulated time and at equal wall-clock?
  4. Which results use the temperature corrector? App. C.3.1 says the main correction just scales the noise, with a fallback to velocity rescaling, but gives no α values, no fallback rate, and no effect on transition statistics.
  5. Fig. 13: the new state survives 7–16 ns in three Amber14 runs but is absent from the reference. Can you estimate its equilibrium population, to rule out an artefact of the biased large-step ensemble?
3 comments
AuthorsOpenReview ↗

Response to xtuy - part 1

Thank you for your review and suggestions for improvement. We reply to each question in turn below.

  1. Can you compare wall-clock time against classical MD at 2–4 fs, and report training time?

We have measured the integration throughput of classical MD at 2-4fs, using hydrogen bond constraints at 4fs. Please see Figure 18, or Table R6 below for a summary. The throughput of classical MD is higher than the LFM in the 2-4fs timestep range. The LFM obtains a slightly higher throughput at 8fs compared to OpenMM at 1fs for small and medium systems.

The transferable LFM does not obtain a speedup over classical MD. This is now highlighted within the text in the integration throughput discussion, conclusions and future work sections. We note that the LFMs do obtain a significant speedup over MLFFs and this is the application where we hope LFMs can provide a contribution.

To further validate this improvement over MLFFs and previous Hamiltonian Flow Map (HFM) architectures, we have trained the LFM and HFM baseline up to 50fs steps on alanine dipeptide. We report the transition dynamics predicted by the HFM and LFM models in Figure 13. At 50fs steps, the LFM obtains accurate transition probabilities between meta-stable states within +/- 2pp, while the HFM has diverged due to integration bias. This suggests the possibility for scaling the LFM model to much larger timesteps to obtain further integration speedups over MLFFs.

Table R6: Integration throughput in ns/hour.

MethodTimestep (fs)Small: 23 atomsMedium: 68 atomsLarge: 121 atoms
OpenMM141.6836.5831.03
OpenMM259.0243.3450.18
OpenMM471.3460.0161.41
LFM0.54.883.641.76
LFM19.187.033.46
LFM216.5713.086.46
LFM429.0421.8812.01
LFM842.7737.1021.43

The training time for the transferable LFM is reported in Table R7 below.

Table R7: Transferable training time.

ModelTotal time (h / GPU-h)
9M, 100 atoms/param10.6 / 21
9M, 200 atoms/param23.2 / 46
9M, 480 atoms/param54.1 / 108
  1. Can you add error bars to Fig. 5 and Table 3? Are the shorter transition times outside MD sampling error, and what friction does the reference MD use?

Error bars have been added to Figure 5 and Table 3. The reference MD uses the same friction of γ=10\gamma = 10ps$^{-1}$ as the LFM. See Table R8 for the uncertainty estimates for the alanine dipeptide MFPTs. The shorter transition times of the LFM are slightly outside MD sampling error.

Table R8: Alanine dipeptide MFPTs plus 95% uncertainty ranges.

TransitionMDMLFFLFM
α → β27.14 [26.31, 28.21]26.75 [25.94, 27.54]24.42 [23.65, 25.25]
α → ppII28.37 [27.53, 29.48]28.01 [27.22, 28.81]25.93 [25.14, 26.74]
β → ppII5.77 [5.66, 5.89]5.77 [5.63, 5.93]5.25 [5.15, 5.36]
  1. Table 3 compares at equal network evaluations (10⁷), so the 8 fs model simulates 8× more time. How do the results compare at equal simulated time and at equal wall-clock?

We have computed the transferable LFM Table 3 results for iso-simulation and iso-wallclock time. Please see Tables 8 and 9, reproduced below in Tables R9 and R10 respectively (comment 2).

At iso-simulation time, the large-timestep LFM maintains the improved meta-stable state convergence across all system sizes and comparable transition dynamics accuracy for 4AA and 8AA systems.

At iso-wallclock time, the large-timestep LFM achieves improved meta-stable state convergence across the 4AA and 8AA systems and is comparable to the Amber14 force-field on the 2AA systems.

We believe that these are already promising results as a demonstration of a transferable MLFF that is capable of large-timestep integration. However, we agree that more work is required to obtain an integrator with comparable performance to classical MD. This sentiment is reflected in the integration throughput discussion, conclusions and limitations and future work section.

AuthorsOpenReview ↗

Response to xtuy - part 2

Table R9: Iso-simulation time transferable LFM sampling.

ForcesIntegratorTimestep2AA E-W₂2AA T-W₂2AA π-TV2AA P-TV4AA E-W₂4AA T-W₂4AA π-TV4AA P-TV8AA E-W₂8AA T-W₂8AA π-TV8AA P-TV
LFMLFM8 fs3.580.840.130.1511.932.330.250.3625.924.190.060.54
LFMLFM4 fs2.760.640.130.158.582.090.190.3121.314.290.100.52
LFMLFM2 fs1.860.610.240.147.131.950.210.3017.784.160.170.48
LFMLFM1 fs1.020.630.300.195.741.980.210.3315.334.370.160.45
LFMBAOAB1 fs0.310.370.190.110.971.620.410.313.042.900.070.48
Amber14BAOAB1 fs0.260.330.150.121.201.260.240.332.383.130.130.47

Table R10: Iso-wallclock time transferable LFM sampling.

ForcesIntegratorTimestep2AA E-W₂2AA T-W₂2AA π-TV2AA P-TV4AA E-W₂4AA T-W₂4AA π-TV4AA P-TV8AA E-W₂8AA T-W₂8AA π-TV8AA P-TV
LFMLFM8 fs3.580.810.100.1111.932.310.240.2625.924.100.040.43
LFMLFM4 fs2.770.630.080.128.572.050.110.2321.314.190.070.38
LFMLFM2 fs1.850.600.230.077.231.940.120.2218.104.050.160.42
LFMLFM1 fs1.020.640.290.115.901.950.170.3115.124.330.150.46
LFMBAOAB1 fs0.350.340.130.140.921.450.350.282.772.840.070.48
Amber14BAOAB1 fs0.140.200.060.040.600.910.150.191.212.340.120.46
  1. Which results use the temperature corrector? App. C.3.1 says the main correction just scales the noise, with a fallback to velocity rescaling, but gives no α values, no fallback rate, and no effect on transition statistics.

Only the transferable LFM model uses the temperature corrector during inference. We have measured the α\alpha values and fallback rate which are shown in Table 7. Reproduced below in Table R11 for convenience. The median α\alpha value reduces the standard deviation of the noise and falls back to velocity rescaling around one third of the time.

Table R11: Temperature corrector statistics.

Timestepα median [IQR]Fallback
1 fs0.45 [−0.32, 1.57]33%
2 fs0.48 [−0.18, 1.33]34%
4 fs0.43 [−0.10, 1.06]34%
8 fs0.34 [−0.10, 0.80]37%
  1. Fig. 13: the new state survives 7–16 ns in three Amber14 runs but is absent from the reference. Can you estimate its equilibrium population, to rule out an artefact of the biased large-step ensemble?

We have estimated the meta-stable state's equilibrium population using the forward and reverse estimation method from Klimovich et al. (2015) using replica-exchange MD. The result is shown in Figure 17 and predicts an occupancy of 0.035±0.0070.035 \pm 0.007%. This shows that the state is rare but stable and present in the classical MD Boltzmann distribution.

Concluding remarks Overall, we believe that the contribution of LFMs is to improve the runtime of MLFFs while maintaining comparable accuracy. We agree that the runtime improvement over tuned classical-force-field MD is not yet established and have updated the manuscript to further reflect this in the integration throughput discussion, conclusions and limitations and future work sections.

We thank the reviewer again for their feedback which has strengthened the contributions of the paper. If there are any further questions, we are happy to continue the discussion.

Dear reviewer xtuy,

We are thankful for your time and thoughtful review, especially the suggestion for further comparison to classical MD force-fields which we have now added.

As the end of the rebuttal period approaches, we are happy to answer any further questions. If our responses and additional experiments have addressed your questions, we would be grateful if you might consider updating your assessment.

Reviewer_xN6iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Sampling equilibrium distribution of atomic configurations with Molecular Dynamics can take prohibitively long. The main bottlenekcs have commonly been i) the interatomic potential that defines the energy and forces is expensive ii) Sampling the equilibrium distirbution defined by the interatomic potentials is prohibitively expensive for many systems, common methods for these are Molecular Dynamics. This work aim to accelerate the latter.

This works presents an algorithm that improves efficiency at sampling with Langevin Dynamics by x10 w.r.t to Molecular Dynamics. The novelty of the paper lies in the training strategy. While similar speed ups have been achieved at generation by related works like FlashMD (Bigi et al., 2025) or TimeWrap (Klein et al., 2023), these methods require sequential data at training. In contrast, this work does not require running Langevin dynamics to generate training data. Instead it builds on Hamiltonian Flow Maps (HFM; Ripken et al., 2026), which learn the deterministic Hamiltonian flow over large timesteps and combine it with a separate Langevin thermostat step at inference. LFM instead learns the solution map of Langevin dynamics.

Critical strengths
  • The main strength lies in no trajectories needed for training, while stile capturing kinetics.
  • About x10 speed up over standard sampling with similar accuracy.
  • It transfers to unseen peptides.
  • Self-consistency losses is becoming a relevant direction in accelerating Langevin Dynamics sampling.
Critical weaknesses

Equilibrium training data (experimental weakness)

The method can train on trajectory free data, but the training sets (alanine dipeptide, Many Peptides) are equilibrium samples, which already require long MD simulations to obtain. The paper doesn't seem to show whether the LFM can recover the equilibrium distribution when trained on unconverged data, e.g., configurations from only a subset of the metastable states. Prior work like FlashMD (Bigi et al., 2025) shows this kind of generalization is possible. Since the core motivation is to overcome the slow convergence of MD to equilibrium, an experiment showing generalization to equilibrium distributions when trained on non-equlibrium data as done in FlashMD would really strengthen the paper.

HFM as a Baseline

  • LFM builds heavily on HFM, but HFM is compared only in Fig. 3 (h(r)) and Table 2 (forces).
  • HFM is absent from Boltzmann convergence (Fig. 4), alanine dipeptide kinetics (Fig. 5), transition paths (Fig. 6) and the transferable experiments (Table 3) while a main contribution of the paper is how LFM is better than HFM. The unbiased transition dynamics claimed in Table 1, should also be proven empirically in these experiments.
  • Additionally HFM is originally analyzed in bigger molecules such as Chignolin or Bba, while these analysis are skipped here.
  • These points are relevant to understand as a reader what are the contributions of LFM over HFM.

Missing baseline

No comparison with FlashMD and similar methods, even though it is very related.

Questions for authors

Equilibrium training data

Can the authors show whether LFM can recover the equilibrium distribution when trained on unconverged data, e.g., configurations from only a subset of the metastable states, as done in FlashMD?

HFM as a baseline

Can the authors compare LFM to HFM on both equilibrium sampling and dynamics? This could be done by adding HFM as a baseline to the current experiments (Figs. 4–6, Table 3), or by adding LFM to experiments from the HFM paper (e.g., alanine dipeptide, Chignolin).

If these points are addressed, I am happy to increase my score.

2 comments
AuthorsOpenReview ↗

Response to xN6i

Thank you for your time and constructive feedback. We reply to each suggestion / question in turn below.

Equilibrium training data

Can the authors show whether LFM can recover the equilibrium distribution when trained on unconverged data, e.g., configurations from only a subset of the metastable states, as done in FlashMD?

We have now trained the LFM on the Ac-Ala3-NHMe system from MD22. This training dataset has not fully converged and is missing a conformational mode. We show in Figure 12 that the LFM is able to recover this mode when trained on the unconverged data (or see Table R3 below for convenience).

We additionally train the HFM baseline with matched epochs and architecture hyperparameters on this system. To measure the structural accuracy of the mode recovery we compare the LFM and HFM predictions to their respective h=0h=0 MLFF networks. We see that the LFM obtains a more accurate distribution at both stepsizes than the HFM, especially at the Δt=10\Delta t = 10fs step where the HFM predictions diverge.

Table R3: Boltzmann convergence 𝕋-W2W_2 metric for Ac-Ala3-NHMe.

Model5 fs10 fs
HFM0.6751.143
LFM0.3110.238

HFM as a baseline

We compare to HFMs on equilibrium sampling in Table R3 on the Ac-Ala3-NHMe system above and to the MD17 organic molecules now included in Figure 4, reproduced below in Table R4. Both LFM and HFM use 9fs timesteps.

We see from Figure 4 that the convergence speed of LFM and HFM at matched timesteps is comparable, as expected. The final conformational accuracy is slightly improved by LFMs over HFMs.

Table R4: Boltzmann convergence T\mathbb{T}-W2W_2 metric at 10410^4 NFEs for MD17 datasets.

SystemMLFFLFMHFM
Aspirin1.499 ± 0.0410.684 ± 0.0460.857 ± 0.034
Ethanol0.180 ± 0.0170.137 ± 0.0180.156 ± 0.033
Naphthalene0.047 ± 0.0010.045 ± 0.0020.056 ± 0.001
Salicylic acid0.101 ± 0.0240.078 ± 0.0040.090 ± 0.010

We compare LFMs to HFMs on transition dynamics on alanine dipeptide. We trained matched architecture-hyperparameter models for 200 epochs with a 50fs maximum timestep and measured the transition probabilities between the three main meta-stable alanine dipeptide states. Please see Figure 13 for the full results and Table R5 below for a summary of this figure showing transition probability errors for LFM and HFM predictions compared to MD.

At small stepsizes (55fs) we see that the LFM and HFM both predict accurate transition dynamics. As the stepsize increases towards 5050fs, the LFM predictions remain comparable to MD while the HFM predicts incorrect dynamics and eventually diverges. We attribute this increasing error to HFMs O(Δt3)O(\Delta t^3) splitting scheme bias introduced in the transition probability density.

Table R5: Mean transition probability errors in percentage points (pp).

Step size (fs)LFM (pp)HFM (pp)
50.940.99
101.2510.14
251.5933.43
500.88NaN

Larger system sizes

Additionally HFM is originally analyzed in bigger molecules such as Chignolin or Bba, while these analysis are skipped here.

The HFM Chignolin and BBA experiments are implemented on coarse-grained systems retaining only the CαC^\alpha atoms. These systems have 10 and 28 atoms per molecule, respectively. The system sizes investigated here are much larger with the Many Peptides dataset containing all-atom octapeptides with, for example, ANKSMIEA containing 121 atoms (evaluated in Table 3 - 8AA).

Concluding remarks

We thank the reviewer for their suggestions of an unconverged training data experiment and to compare the transition dynamics to HFMs directly, which we believe have greatly strengthened the paper. We are happy to continue the discussion if there are any further suggestions.

Dear reviewer xN6i,

We are grateful for your time and insightful review. We have added the unconverged training data experiment as well as the equilibrium sampling and transition dynamics comparison to HFMs.

As the end of the rebuttal period approaches, we are happy to answer any further questions. If our responses and additional experiments have addressed your questions, we would be grateful if you might consider updating your assessment.

Reviewer_LvXaOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Langeving Flow Maps (LFM): a novel method to model molecular dynamics based on a trained stochastic Langevin integrator. LFM provides a modeling for the transition and equilibrium dynamics without requiring long molecular trajectories, but rather using a datasets of sampled molecular configurations and the associated force vectors. An experimental section demonstrates the effectiveness of LFM in modeling several statistics of molecular trajectories and demonstrates the generalization capabilities on the method on peptides not observed during training.

Critical strengths
  • The paper is very clear and well written, providing sufficient background and intuition to understand the problem and the proposed solution.
  • The paper introduces a very flexible method that extends existing approaches in literature and demonstrates the effectiveness on a wide range of tasks.
  • Overall the proposed solution is quite intuitive, simple and elegant, which makes it even more appealing for practical use-cases.
Critical weaknesses

Main concerns

  • The paper clearly refers to [1], however the main differences between the two approaches are not discussed in depth.

Other issues

  • The paper provides little to no experimental detail regarding neural network architectures and optimization procedures, which may limit the reproducibility of the reported experiments.
  • The Langevin Flow Map algorithm introduces a number of hyperparameters which are not discussed in detail and no additional intuition is provided. Similarly no intuition is provided regarding the relation between the integration timescale at inference and the sampled timesteps at training.
  • The colors used in Figure 2, 6 and 8 are almost indistinguishable. Figure 2 (right) is missing units on the x axis.

References

[1]: McCallum, Sam, et al. "Strong stochastic flow maps." arXiv preprint arXiv:2606.01086 (2026).

Questions for authors
  1. In step 5 of Algorithm 1, the time step hih_i is sampled from a log-uniform distribution from hminh_{min} and hmaxh_{max}. Can the authors clarify what values are used for hminh_{min} and hmaxh_{max} and provide some intuition regarding their choice, the use of a log uniform distribution and the relation between the sampled hh and simulated timestep t−st-s?

  2. Can the authors clarify why the force errors for LFM are larger than the standard MLFF training? Are the architectures considered in the neural network architectures used for the two methods in the table similar? What are the potential implications of a less-accurate force prediction model?

  3. When evaluating the Boltzmann Convergence in Figure 4, the BAOAB and LFM integration are unfolded at different timescales (0.5fs vs 9fs). Can the authors clarify the reason behind this specific choice of time scales? How does LFM integration perform for different integration scales for this experiment? What is the relation between the choice of integration scale and the training parameters?

5 comments
AuthorsOpenReview ↗

Response to LvXa - part 1

Thank you for your review and helpful feedback. We respond to each suggestion in turn below.

The paper clearly refers to [1], however the main differences between the two approaches are not discussed in depth.

Langevin Flow Maps build on the machinery developed by [1]. The extension requires deriving the consistency conditions for the coupled Langevin system, which results in a force matching and velocity matching objective (see Proposition 3.1). Additionally, Strong Stochastic Flow Maps learn the solution map over bounded diffusion time [0,1][0, 1], whereas in Langevin Flow Maps we must extend this to unbounded physical time [0,∞][0, \infty]. This requires consideration of the timestep sampling (please see discussion below).

The paper provides little to no experimental detail regarding neural network architectures and optimization procedures, which may limit the reproducibility of the reported experiments.

The experimental details are discussed in Appendices B and C, which describe the neural network architecture and optimization procedures, respectively. Additionally, the training code is able to be released directly on GitHub with scripts for training each model and reproducing the results.

The Langevin Flow Map algorithm introduces a number of hyperparameters which are not discussed in detail and no additional intuition is provided. Similarly no intuition is provided regarding the relation between the integration timescale at inference and the sampled timesteps at training.

  1. In step 5 of Algorithm 1, the time step  is sampled from a log-uniform distribution from  and . Can the authors clarify what values are used for hminh_\text{min} and hmaxh_\text{max} and provide some intuition regarding their choice, the use of a log uniform distribution and the relation between the sampled hh and simulated timestep t−st-s?

The hyperparameters introduced by LFMs are the timestep sampling (hmin,hmax)h_\text{min}, h_\text{max}), the polynomial degree NN and the EMA decay β\beta.

The degree N=4N=4 was selected for each experiment after the hyperparameter was ablated; please see Figure 9 (or Table R1 below for convenience). This provides sufficient path information without overwhelming the model with many additional features.

Table R1: Polynomial degree ablation.

Polynomial degreeIntegration RMSEh(r) MAE
10.02140.313
20.01370.128
30.00550.071
40.00430.053

The EMA decay β=0.99\beta=0.99 was selected for the experiments following common practice for flow maps. This is described in Section 3.3.

The timestep hyperparameters used for the experiments are hmin=0.01h_\text{min}=0.01fs and hmax=10h_\text{max}=10fs, described in Appendix C. Since this range spans 4 orders of magnitude, we use log-uniform sampling to balance the sampling contribution across these different magnitudes. The sampled hh is the timestep t−st-s, defined on line 236 as h=t−sh=t-s.

To provide a further datapoint on the maximum timestep hmaxh_\text{max}, we additionally train the alanine dipeptide model up to hmax=50h_\text{max}=50fs and demonstrate that the model captures accurate transition dynamics in Figure 13.

AuthorsOpenReview ↗

Response to LvXa - part 2

  1. Can the authors clarify why the force errors for LFM are larger than the standard MLFF training? Are the architectures considered in the neural network architectures used for the two methods in the table similar? What are the potential implications of a less-accurate force prediction model?

Thank you for the question, we believe this is an important point raised by learning the integrator directly rather than just the MLFF. We hypothesize that the force errors are larger for the LFM than the MLFF since the finite-capacity model must additionally learn the integrator (minimize two losses). However, as we see in Figure 3, despite the less-accurate force model the LFMs predictions are more structurally accurate than the MLFF. This suggests that achieving low force error alone is not sufficient for learning neural network models of molecular dynamics but rather it is the joint objective of force error and integrator accuracy. Please see [2] for further discussion on this topic.

  1. When evaluating the Boltzmann Convergence in Figure 4, the BAOAB and LFM integration are unfolded at different timescales (0.5fs vs 9fs). Can the authors clarify the reason behind this specific choice of time scales? How does LFM integration perform for different integration scales for this experiment? What is the relation between the choice of integration scale and the training parameters?

The 0.5fs timestep was chosen for BAOAB to ensure numerical stability on the neural network force-field obtained by the h=0h=0 force network. The LFM was trained up to 10fs steps and the 9fs evaluation step was chosen as a sweet spot between accuracy and speed.

We have now computed the predictions for h=1,5,9h=1, 5, 9fs steps for the LFM and present the results in Figure 11 (or Table R2 below for convenience). As the stepsize increases, the convergence is obtained with fewer function evaluations. For a 11fs stepsize, the convergence speed is comparable to the MLFF which suggests the relaxation dynamics are accurately predicted.

Table R2: Boltzmann convergence T\mathbb{T}-W2W_2 for the Aspirin dataset.

MethodStep size (fs)NFE = 10NFE = 100NFE = 1,000NFE = 10,000
MLFF0.52.632.472.071.50
LFM12.582.362.001.25
LFM52.472.091.540.80
LFM92.372.041.280.68

Concluding remarks

We thank the reviewer again for their time and helpful questions which have improved the clarity of the paper. We are happy to continue the discussion and answer any further questions.

References

[1] Sam McCallum et al. "Strong stochastic flow maps." ArXiv 2026.

[2] Xiang Fu et al. "Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations." TMLR 2023.

Dear reviewer LvXa,

We are grateful for your time and helpful review of our submission, particularly for the questions regarding MLFF force errors and suggestion for adding multiple timescales to the Boltzmann convergence evaluation.

As the end of the rebuttal period approaches, we are happy to answer any further questions. If our responses and additional experiments have addressed your questions, we would be grateful if you might consider updating your assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Discussion summary

We thank the reviewers for their helpful and constructive reviews. We provide a summary of the discussions and additional experiments here.

LFM hyperparameters (LvXa). The influence of the timestep, polynomial degree and EMA decay were discussed and further clarified with the polynomial degree ablation (Table R1) and Boltzmann convergence experiment (Table R2).

Unconverged training data (xN6i). We trained the LFM on the Ac-Ala3-NHMe system from MD22, which is missing a conformational mode. The LFM is able to recover this mode during inference and thereby debias it's predictions from the training data (Table R3).

HFMs as a baseline (xN6i). We have compared LFMs to Hamiltonian Flow Maps (HFMs) on equilibrium sampling and transition dynamics. The LFM achieves a tighter convergence to the Boltzmann (Table R4) and significantly more accurate transition dynamics (Table R5) than HFMs.

Comparison to classical MD (xtuy). The comparison to classical MD has been expanded with additional wallclock measurements (Table R6), transition rate error bars (Table 3 & R8) and matched simulation-time and wallclock sampling (Table R9 & R10). The discussion clarified that the integration speedup for LFMs over current MLFFs is significant, while an improvement over classical MD is not yet consistent and will require future work to scale the integration timestep.

Predicted state population (xtuy). The LFM was shown to discover a meta-stable state not covered by the reference MD dataset on the ANKSMIEA octapeptide. The equilibrium population of this state was determined using the forward / reverse estimation method from Klimovich et al. (2015). The state was found to be rare but stable with an occupancy of 0.035±0.0070.035 \pm 0.007%.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.