acceptodds
Under review as a conference paper at ICLR 2027

Accelerating Fokker–Planck Regularization For Force Field Learning

Abstract

Diffusion models can learn molecular equilibrium distributions without force labels, but they often predict inaccurate forces that cause subsequent Langevin simulations to diverge. While standard Fokker–Planck (FP) regularization mitigates this inconsistency by constraining the density evolution, FP regularization heavily relies on expensive Hessian computations, which makes training computationally prohibitive. To address this, we introduce a novel weak FP regularization that achieves both consistency and computational efficiency. Rather than relying on expensive higher-order derivatives, our method approximates standard FP using a small set of efficiently-computable test functions. To optimize this objective correctly using stochastic gradients, we then derive an unbiased U-statistic estimator for minibatched moments. Through extensive experiments on real-world and analytic potentials using augmented as well as SE(3)-equivariant graph transformers, we demonstrate that our objective achieves the accuracy of the state-of-the-art FP regularization at a fraction of the memory and runtime.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Fokker-Planck, Molecular Dynamics, Force Field, Diffusion Model, DSM, Graph Transformer, SE3

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

This paper proposes an efficient approach to Fokker–Planck (FP) regularization for learning molecular force fields from equilibrium configurations without force labels. Existing FP regularization methods improve the consistency of diffusion-based force fields but incur substantial computational overhead due to expensive higher-order derivatives. To address this issue, the authors introduce two weak-form FP objectives: Galerkin FP, which enforces consistency across diffusion times, and Stein FP, which constrains endpoint forces. By leveraging integration by parts, finite-dimensional test functions, and an unbiased U-statistic estimator, the proposed approach avoids expensive Hessian computations. Experiments demonstrate that Galerkin FP achieves simulation quality comparable to existing FP regularization methods while significantly reducing training time and memory consumption.

Critical strengths
  1. Clear motivation and practical relevance. The paper addresses the substantial computational overhead of higher-order derivatives in FP regularization. Improving training efficiency is practically important, particularly for scaling force-field learning to larger models and molecular systems.
  2. Principled and computationally efficient formulation. The proposed weak-form FP objectives leverage Galerkin projections and integration by parts to avoid expensive Hessian computations. The unbiased U-statistic estimator further provides a theoretically justified formulation rather than a purely heuristic approximation.
  3. Significant efficiency improvements. Galerkin FP achieves simulation quality comparable to existing FP regularization methods while providing approximately 2.5–3.5× speedups and substantially reducing memory consumption. The matched-compute comparisons further demonstrate its practical benefits.
  4. Useful insights from comparing Galerkin FP and Stein FP. The experiments suggest that enforcing consistency across diffusion times is more effective for training from scratch, whereas endpoint constraints can be beneficial for refining pretrained DSM models. This distinction provides useful insights into regularization strategies at different training stages.
Critical weaknesses
  1. Insufficient evaluation of force-field accuracy. Despite the paper's focus on force-field learning, the evaluation primarily emphasizes equilibrium distributions, PMF errors, and simulation stability, with limited direct analysis of force prediction errors.
  2. Limited guarantees from finite-dimensional weak constraints. Galerkin FP only constrains the projection of the FP residual onto a finite-dimensional test-function space, potentially missing important residual components that could affect local force accuracy and long-term simulation stability.
  3. Limited evaluation on larger molecular systems. Most experiments use small coarse-grained systems, including Chignolin represented by only 10 Cα beads. The largest all-atom system contains just 42 atoms, leaving scalability to more complex molecular systems insufficiently validated.
  4. Unclear scalability of the test-function representation. The proposed RBF test functions rely on pair-distance features whose dimensionality may grow substantially with molecular size. The sensitivity of this representation to increasing molecular complexity remains insufficiently explored.
  5. Direct force accuracy and comparison with conventional force fields. Could the authors report direct force MAE/RMSE against reference forces, particularly on MD22? This would help clarify whether the proposed method improves actual force prediction accuracy or primarily improves the equilibrium distributions produced by the learned dynamics. Additionally, could the authors include conventional label-based force fields as reference baselines to contextualize the trade-off between force-free learning and supervised force-field methods, in terms of both force accuracy and equilibrium simulation quality?
  6. Scalability to larger molecular systems. Can the authors provide experiments on larger all-atom systems to demonstrate that the proposed efficiency improvements and simulation quality are maintained as molecular size increases?
  7. Test-function dimensionality and expressiveness. The test functions are constructed from pair-distance features. How does the dimensionality of these features scale with the number of atoms, and how are the relevant atom pairs selected? Although the authors study sensitivity to the number of test functions, it remains unclear how increasing feature dimensionality affects RBF bandwidth selection, test-function expressiveness, and the conditioning of the Gram matrix. Can a fixed number of test functions adequately constrain important FP residuals in larger molecular systems?
  8. Practical importance of the U-statistic estimator. Although the U-statistic estimator is theoretically unbiased, Table 7 shows nearly identical simulation performance with and without the correction. Could the authors clarify under what conditions this correction provides a meaningful practical advantage, and why it is emphasized as a main contribution?
Questions for authors

see weakness

2 comments

Thank you for the careful review. We added force measurements, scaling results, and test-space diagnostics to address your questions.

Summary. We report direct force errors alongside a supervised reference (Tbl. 13), measure the strong residual (Fig. 4), report cost and memory across systems (Fig. 13), and analyse the test space and metric (App. F). Galerkin FP reduces the full residual by 4-7 times on three systems. Although it trains on a finite projection, its cost stays at 1.8–2.8 times DSM for varying number of atoms, and it matches or outperforms recent energy regularizers (Tbl. 2).

Force accuracy (W1, Q5). We computed RMSE, MAE, and cosine similarity against DFT forces on MD22 and mapped all-atom forces on alanine dipeptide. We also trained the same backbone with force labels as a supervised reference (Tbl. 13).

  • Under continuation with uniform t, Galerkin FP reduces MD22's RMSE from 18.4 to 17.5 kcal/mol/Å, comparable to Pointwise FP.
  • On coarse-grained ALA2, Galerkin FP achieves RMSE 19.7 (without force labels) which is close to the 18.1 for the supervised reference. Each label is the all-atom force of one frame mapped onto the beads. It scatters around the mean force, because the atoms that were coarse-grained away sit differently in every frame. Even a perfect model therefore keeps a large error against these labels.
  • With DFT labels, the supervised reference reaches RMSE 3.1 on MD22, which is more accurate than the label-free methods we know of.
  • Started from reference frames, the supervised reference simulates ALA2 at JS 0.170 and MD22 at 0.334 with 2-55% broken-bond states. From the same frames, Galerkin FP reaches 0.142 and 0.346 with below 0.1% broken-bond states and without labels.

FP regularization improves the distributions (Q5). While force error and RSDE sample quality are nearly the same for models, we see how their simulations differ by up to 3.5 times in JS. That means, surprisingly neither tells us which model simulates well. Fu et al. (2023) also found the same for supervised force fields. However, the FP residual at small t does tell them apart. Galerkin FP reduces it by 4-7 times on three of four systems (Fig. 4), and offers a good indicator.

Finite test space (W2). Because Prop. 1 controls only the part of the residual inside the test span, we measured the full residual directly (Fig. 4). At t=10−3t = 10^{-3}, we reduce it by 4-7 times relative to DSM on three datasets and by 2.7 times on Chignolin. On KS dipeptide and MD22, we match Pointwise FP, which has no projection (4.4 vs. 3.1 and 5.3 vs. 6.1 times). Because the two methods also agree in force error (17.5 vs. 17.6 kcal/mol/Å on MD22, Tbl. 13) and in broken bonds (at most 2.4% vs. 2.7% of states, Tbl. 15), we see no sign that the projection misses a component that matters for forces or stability.

Test functions and scaling (W4, Q7). We use all N(N−1)/2 pair distances, as Sec. 4 now states. On MD22, test functions, moments, the Gram matrix, and its solve take 12 ms of a 491 ms step. The main additional cost actually comes from a backward pass through the network. Across various number of atoms, Galerkin FP costs 1.8–2.8 times DSM per step and only about 2.7-3 times its memory, compared with 5.6–6.6 times the step time for Pointwise FP (Fig. 13) using only few Hutchinson probes.

Bandwidth and conditioning. We note that pairs further away may 'dominate' the RBF bandwidth. Thus, we use the median bandwidth. We re-estimate G each step with a relative ridge and a condition cap (10^5). Because the Gram metric makes the loss independent of the scale and correlation of the tests, we see one default weight stay stable on all four dataset (Tbl. 10). When t is almost 0, the covariance has an effective rank of about 1.5 which is rising to ~10 at larger t (Tbl. 8).

Larger systems (W3, Q6). MD22 is our largest dataset / system with 42 atoms. Across our four systems, the cost relative to DSM stays at 1.8–2.8 times (Fig. 13). We tested additive radial functions that keep the number of test functions independent of molecule size and give similar quality to the default (Tbl. 9). Evaluating all pair distances still scales quadratically with the number of particles.

U-statistic (Q8). Because the bias of the naive estimator scales with 1/n1/n, the two estimators agree at batch size 1024, which is the setting of the ablation (Table 7 in the submission, now Tbl. 18). At an exact model, the naive objective reads 31.7 instead of zero at n=16n = 16 and 1.8 at n=256n = 256, while the U-statistic stays at zero within its sampling error (Tbl. 17). The correction therefore matters when memory limits the batch, as on MD22 with n=128n = 128 and on larger systems. It needs no extra model evaluations and guarantees that the loss vanishes in expectation at an FP-consistent model for every batch size.

We ask you to consider the new force, residual, and scaling evidence when updating your assessment.

Thank you for your reply; I keep my score.

Reviewer_8iSBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a method for accelerating Fokker–Planck regularization via a weak Galerkin formulation. Compared to existing denoising score matching models with Fokker–Planck regularization, the weak Galerkin formulation significantly reduces training time while achieving comparable accuracy.

Critical strengths

The idea of applying a weak formulation to the regularization term is concrete and technically solid. The accuracy and performance results support the claims of the paper.

Critical weaknesses

Limited novelty.

While an unbiased estimator and Stein FP are also introduced in this paper, the main contribution appears to be Galerkin FP. However, using a weak formulation to satisfy physical constraints is not new. For example, see VPINN (https://arxiv.org/pdf/1912.00873).

The main idea is to apply a weak formulation to the existing FP regularization method (https://arxiv.org/pdf/2506.17139).

While applying methods from other fields to a new domain can be helpful and lead to stronger research, the novelty and significance of this approach should be justified more convincingly.

Questions for authors
  1. Could the authors provide strong-form FP residuals for models trained with weak-form FP regularization, strong-form FP regularization, and no FP regularization?

  2. Could the authors further discuss the importance of learning coarse-grained MD force fields without force labels and compare their method with other recent approaches addressing this problem?

4 comments

Thank you for the review and the questions about novelty, residuals, and label-free learning.

Summary. We measured the strong-form FP residual. Over three seeds, Galerkin FP reduces it by 4-7 times relative to DSM on three systems, comparable to the strong-form estimator (Fig. 4). We compared with three recent energy regularizers, which Galerkin FP matches or outperforms at similar cost (Tbl. 2), and we expanded the discussion of VPINNs and label-free coarse graining in Sec. 5.

Novelty. While weak formulations are classical, they remain underexplored for FP consistency in MLFF. We now cite and discuss VPINßNs. A VPINN solves a specified PDE by quadrature over a domain. In contrast, we adapt weak residual constraints to a sample-based diffusion objective, which is different in many aspects:

  1. We constrain density evolution in 3N dimensions using samples from the noised data distribution.
  2. We combine integration by parts with the denoising identity to remove the network's second spatial derivatives. This yields the measured 2.5–3.5× speedup over the FP regularization of Plainer et al. (Smoothed FP in our tables).
  3. Because squaring introduces minibatch bias, we provide the U-statistic that removes that sampling bias (Prop. 2), which deterministic quadrature does not introduce.

In addition, we also compare with RNE, Dual SM, and stNCE, which also avoid costly second derivatives. Our Galerkin FP achieves similar performance on the coarse-grained systems and the lowest JS among stable (non-bond-breaking) regularizers on MD22, at an equivalent cost to those (Tbl. 2).

Strong-form residuals (Q1). We evaluated the strong residual on held-out frames for DSM, Pointwise FP, and Galerkin FP (Fig. 4). At t=10−3t = 10^{-3}, we reduce it by 4-7 times relative to DSM on three systems and by 2.7 times on Chignolin (three seeds). We see that these gains approach Pointwise FP, which targets this residual directly at a high computational cost. For larger t≥0.1t \geq 0.1, all models have the same residual, including DSM without regularization. That means, we achieve better force predictions without reducing the prediction quality at higher diffusion times.

Label-free (Q2). Since coarse-grained models learn mean forces (no simulation produces them) force matching has to train on instantaneous all-atom forces that are projected onto remaining beads. Those, however, scatter widely around the average, which we can see on ALA2 where models trained on force labels cannot get below a force RMSE of 18.1 kcal/mol/Å. Predicting zero force gives 26.2, while Galerkin FP reaches 19.7 without a single label (Tbl. 13). That means, labels therefore help little for coarse-grained models.

Other approaches. We compare with the three energy regularizers above and discuss relative-entropy minimization, flow-matching coarse graining, and contrastive learning in Sec. 5. Relative-entropy minimization is quite different in that it simulates the model during training, while flow-matching coarse graining still learns a force field from a flow using a second stage. Our diffusion model learns from configurations without either step.

We ask you to reconsider the rating in light of the residual measurements and the expanded comparisons.

Thank you for the additional experiments. Figure 4 provides useful evidence that Galerkin FP reduces the spatial variance of the strong-form residual. However, my original question concerned the magnitude of the strong-form FP residual, particularly the squared residual objective defined in Eq. (7).

Figure 4 reports $ Var_x[r_\theta] $ , rather than Ex[rθ2]\mathbb E_x[r_\theta^2]. These quantities differ by the squared mean residual. Could the authors clarify the motivation for using variance and, if possible, report the mean and mean-squared strong-form residual for DSM, Pointwise FP, and Galerkin FP?

Thank you for the follow-up!

The energy can be shifted by any function of time, Eθ(x,t)→Eθ(x,t)+C(t)E_\theta(x,t) \to E_\theta(x,t) + C(t), without changing the density, the score, or the samples. Because Varx[rθ]\mathrm Var_x[r_\theta] is invariant to the offset and Ex[rθ2]\mathbb E_x[r_\theta^2] is not, we used the variance to compare models whose offsets were trained differently. DSM never constrains the offset, while Eq. (7) does.

In short, the offset shifts the mean residual Ex[rθ]\mathbb E_x[r_\theta] by −dC/dt-\mathrm{d}C/\mathrm{d}t, Ex[rθ2]=Varx[rθ]+(Ex[rθ])2\mathbb E_x[r_\theta^2] = \mathrm Var_x[r_\theta] + (\mathbb E_x[r_\theta])^2 changes through the squared mean, and Varx[rθ]\mathrm Var_x[r_\theta] remains invariant.

In the table below and Fig. 32 (last page), we now report Ex[rθ]\mathbb E_x[r_\theta] and Ex[rθ2]\mathbb E_x[r_\theta^2] in the same setup as Fig. 4. Both reduce Ex[rθ2]\mathbb E_x[r_\theta^2] relative to DSM by an order of magnitude or more around t=0.1t = 0.1 on all four systems, and Galerkin FP matches or improves over Pointwise FP. For example, on MD22, we improve over DSM at every diffusion time, by 2-5 times at t≤10−2t \leq 10^{-2} and with Galerkin FP by more than 100 times around t=0.1t = 0.1.

SystemMethodE[r], t=0.001E[r^2], t=0.001E[r], t=0.01E[r^2], t=0.01E[r], t=0.1E[r^2], t=0.1
Ala. dipeptideDSM-1.3e32.2e67.1e15.3e32.2e15.1e2
Pointwise FP-3.5e21.6e5-2.8e14.8e3-2.9e-11.0e0
Galerkin FP-1.0e22.7e46.4e14.6e31.4e02.1e0
KS dipeptideDSM-1.2e32.1e61.0e11.4e4-9.1e01.0e2
Pointwise FP-6.3e25.7e5-6.5e17.7e3-9.4e-13.4e0
Galerkin FP-8.3e28.7e5-1.8e17.3e2-1.2e06.9e0
ChignolinDSM-1.8e33.4e6-2.0e02.3e3-1.0e11.3e2
Pointwise FP-9.2e29.3e5-1.1e21.4e4-2.1e06.2e0
Galerkin FP-1.0e31.1e6-5.2e14.2e3-1.3e04.6e0
MD22DSM-2.3e35.5e6-2.2e25.8e4-2.0e14.2e2
Pointwise FP-9.5e29.8e5-1.6e22.7e4-1.8e05.1e0
Galerkin FP-1.1e31.5e6-1.4e22.0e4-8.6e-27.6e-2

Because the variance in Fig. 4 is almost identical around t=0.1t = 0.1 while Ex[rθ2]\mathbb E_x[r_\theta^2] differs by 1-2 orders of magnitude, the gain comes from the mean residual, that is, from the offset. At small tt, In Fig. 4 we show that the variance, and thus the density estimate, improves as well.

This is very interesting, thanks again for asking!

We hope this answers your question. If our responses and the new experiments have addressed your concerns, we would be grateful if you would consider updating your score.

Thank you for the detailed clarification and additional experiments. The new results on the strong-form FP residuals address my follow-up question, and I appreciate the expanded comparisons and discussion of novelty.

I am satisfied with the authors' responses and will raise my score.

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

This paper aim to learn energy-parameterized diffusion model with Fokker planck regularization. To address the FK high cost, it proposed two approach: Galerkin FP and Stein FP. The experiment shows that the proposed approach can effectively regularize the learned energy.

Critical strengths

The paper proposed an interesting approach to reduce the calculation of FP regularization. This approach might also have potential to be inspire other application where divergence can be expensive in this field.

Critical weaknesses

The experments in this paper is not convincing enough.

  • Some experimental result (Stein) looks suprising to me (see Q1)

  • Even the proposed method presents a substantial improvement over previous FPE, there are many other energy regularization methods that is much cheaper. However, the paper does not compare, or even discuss with them (see my question 2)

  • the simply way to make it fast is using Hutchinson trace estimator. This can be unbaised in the same sence as slice score matching. However, the paper seems not discuss / compare with this simple setting. This make the experimental results less convincing.

  • The proposed method does not have clear ablation / analysis. For example, i suggest to analyze the number of probe function vs performance (the appendix only analyze a small range, what if ranging from 4 to larger?) is there a threshold it start to work? is there a threshold it shart to outperform FP? Another example: Eq 10 use G to rescale the m. This is reasonable choice. But what if we just minimize L2 of m? or we set G to I? Will the method completely fail or it is less effective? Also, how much influence the phi choice influences results? These are all important research questions that need to be analyze for better understanding of the method.

  • The combination of GFP and stein is also unclear (see Q3)

Slightly Minor: the presentation need to be improved. For example, the paper never really clear define stein FP. Also, I believe Stein FP is not really FP. It is not derived from the FPE and hence I woud suggest the author change the name.

I think this paper can be an interesting paper to read and might also be inspiring for the field. however, in its current form, it is not clear of its connection and comparison to the baselines, or even to the simple design.And without comprehensive ablation, we cannot build a good understanding which part of the method is working. Therefore, I suggest clear rejection. That said, I would be happy to change my rating if these concern is addressed. I will raise to 2 if baselines are clearly discussed. I will support acceptance if the author provide more comprehensive on the method with ablation and analysis and also proposed method show clear advantage to baselines like Hut trace estimator and other methods like RNE, time score matching, particularly considering its cost.

Questions for authors
  1. The Stein FP only involves score (gradient of the potential). However, it is known that this is blind toward the mode weight, etc. In fact. this is the reason of using FP regularise (or any regularise like dual score matching, RNE etc). Therefore, I am curious how Stein FP recovers the mode weight in the experiments. This sounds like a coincident to me. Could you clarify this?

  2. FPE regularize is expensive. However, for an energy-based diffusion, there are many other approaches for regularization: RNE[1], dual score matching [2] , and Spatiotemporal Differences [3], etc. These approach are typically cheaper. Therefore, I am curious on how the proposed method perform compared to them (esp. with same GPU time).

  3. The paper proposed two approach: GFP applies to high diffusion t, and Stein applies to t=0. However, it seems that the paper never apply them together? Why?

[1] RNE: plug-and-play diffusion inference-time control and energy-based training. [2] Learning normalized image densities via dual score matching. [3] Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences.

4 comments

Thank you for the detailed review and the suggestions for comparisons and ablations.

Summary. We ran RNE, dual score matching, and stNCE on all four systems, varied the Hutchinson probes from 1 to 8, ablated Galerkin FP's components, and tested the mode weights. Our Galerkin FP matches the energy regularizers, has the lowest error among those without broken bonds on MD22, and has the best performance in from scratch training. We match the Hutchinson FP estimator at a third of the cost.

Cheaper energy regularizers (Q2). We now discuss RNE, stNCE, and Dual SM in Sec. 5, train with all three, and compare simulation performance (Tbl. 2), reporting JS distances and runtime overhead. Moreover, we report the percentage of simulated states that have at least one broken bond (below, in parentheses). We use original implementations and selected weights by the same rule as for our method (App. E).

MethodAla. dip.KS dip.ChignolinMD22cost
DSM0.4570.4100.3130.4071.0
RNE0.1790.2290.237 (44% broken)0.316 (100% broken)1.9–2.0
stNCE0.1700.2230.2170.276 (16% broken)2.5–2.8
Dual SM0.1780.2040.2120.3461.7
Pointwise FP (Hutchinson, K=2)0.3070.3680.2990.3805.6–6.6
Galerkin FP0.1790.2050.2130.2901.8–2.4

We see that on the three coarse-grained systems, Galerkin FP, Dual SM, and stNCE achieve similar JS. On MD22, Galerkin FP reaches .29 with .6% broken-bond states. Then, stNCE reaches 0.276 with 16%, RNE got 0.316 with 100%, and Dual SM reaches 0.346 with 0%. From scratch, Galerkin FP reaches 0.144 against 0.169-0.203 for the three energy regularizers on alanine dipeptide, and 0.305 against 0.335 for Dual SM on MD22 (Tbl. 5). Under continuation, and log-uniformly diffusion (Fig. 9), Galerkin FP matches state-of-the-art performance (Tbl. 7).

Cost (same GPU time). We use the same number of steps for all regularizers. We see that per-step costs are a factor of 1.7–2.8 of DSM's cost across the energy regularizers (Fig. 7), so equal steps mean nearly equal GPU time for most methods. Pointwise FP, on the other hand, would likely fail to converge.

Hutchinson estimator. We like to clarify that Pointwise FP already uses a Hutchinson trace estimator, as stated in Sec. 3.1. From scratch, Galerkin FP reaches similar JS (0.144 vs. 0.139 on ALA2, 0.305 vs. 0.295 on MD22) at about a third of the step cost. We also tested probe counts K (1, 4, and 8) under continuation with uniform t. We found that more probes do not result in consistent quality gains, while the cost reaches 16-19 times DSM's cost per step and used 137 GB of VRAM on MD22 (Tbl. 11), clearly demonstrating the tractability issues we addressed successfully.

Metric. First, by replacing G with the identity or a diagonal changes the loss scale by about 100 times. After retuning the weight, both reach similar JS in a one-seed screen, though the diagonal gives 4–97% broken-bond states. With G, the default weight is far more stable across different datasets (Tbl. 10).

Test function count (App. F). We saw that four test functions already improve on DSM (0.289 vs. 0.457, continuation) and every test function count M from 64 to 512 gives 0.144–0.160, matching Pointwise FP (from scratch, Fig. 10). Near t=0, many test functions could be collinear (the effective rank is low), which is why few suffice (Tbl. 8). The choice of test functions also matters little. We compared five families whose effective rank differs by 16 times, and all give the same quality within noise (Tbl. 9).

Stein and mode weights (Q1). Because Stein essentially only constrains endpoints, we see that it corrects the forces while retaining the mode weights of the model it starts from. Under continuation, those come from the DSM checkpoint. On Müller-Brown, data connects the basins, so the endpoint score along these paths already fixes their populations. For disconnected modes, Stein cannot do that. As FP connects forces at t=0 to scores at higher diffusion times, Galerkin FP can regulate how much probability mass each mode needs. We see that Galerkin FP recovers 0.78 ±\pm 0.03 for the heavy mode of a two-well target with weights 0.2/0.8 from scratch, against 0.72 ±\pm 0.11 for Stein and 0.70 ±\pm 0.12 for DSM (Fig. 11).

Combining the objectives (Q3). Interesting! Adding Stein to Galerkin FP with uniform t improves JS on three of four systems, including 0.384 →\to 0.210 on alanine dipeptide (Tbl. 12). Here, Stein adds more endpoint constraints that are needed for uniform t. However, with log-uniform t, we see no further gain over Galerkin FP.

Presentation. We now define the Stein objective in Sec. 4 and dropped the 'FP' from its name.

We ask you to reconsider the rating in light of these comparisons and ablations.

Interesting! Why you think Dual SM / Galerkin FP is better in avoiding bond breakage, even though it looks worse in JS in some cases?

Thanks for the follow-up!

Bond stability. Bond stability depends on the forces at t≈0t \approx 0. Stretching a bond costs a lot of energy, so each bond sits in a steep, narrow well. If a loss flattens that well, it becomes easier to leave it, and thus to break the bond.

DSM is already a bit too "flat" at t≈0t \approx 0. In simulation, DSM's bond lengths is 1.3 times (alanine dipeptide) and 1.9 times (MD22) longer than in the reference (Fig. 8), and in one seed on MD22, it breaks bonds in 41% of the states.

RNE is a biased estimator in the second order. That term penalises curvature, so it flattens the wells, making escaping them easier, and, thereby, breaking bonds. On MD22, for example, the bond length grows to 18 times the reference with the RNE weight (Fig. 8), and lowering the weight does not help much.

Dual SM and Galerkin FP have in common that they

  1. are unbiased (the minimiser is the true energy for any curvature of the data distribution), and
  2. use the exact time derivative instead of a finite step.
  3. do not smoothen the curvature.

We believe that FD could matter for the unbiased stNCE as it compares states by taking a finite step which seemingly breaks bonds in 16% of the MD22 states.

In sum, on MD22 Galerkin FP and Dual SM yield much stabler bonds (Tbl. 2).

Why it does not show in JS. We compute JS between simulated and reference states on ϕ/ψ\phi/\psi dihedrals (dipeptides) and on the two slowest TICs (Chignolin, MD22; heavy atoms only on MD22). Because of this projection, JS reacts only weakly to individual bond lengths, whereas a state counts as broken as soon as one bond deviates by more than 0.5 Å.

We hope this answers your question. If our responses and the new experiments have addressed your concerns, we would be grateful if you would consider updating your score.

Interesting observation! Although I feel that much of the issue could be addressed by parameter tuning. Overall, I think this is an interesting addition to the toolbox that could be useful and inspire further work. I will increase my score

Best Reviewer

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.