acceptodds
Under review as a conference paper at ICLR 2027

From Potentials to Ensembles: Boltzmann Sampling from Learned Energy Surfaces via Stochastic Optimal Control

Abstract

Equilibrium ensembles governed by the Boltzmann distribution are fundamental to understanding the thermodynamics of ligand binding and conformational landscapes. While generative models have proven effective at sampling molecular configurations, they do not inherently recover the underlying physical energy landscape. Bridging this gap between empirical density estimation and physical thermodynamics remains an open challenge. We introduce *Entropica*, a training framework that guides flow- and diffusion-based molecular generators toward equilibrium distributions determined by physical energies and temperature. *Entropica* identifies the optimal transport path over cumulative informational and physical cost and anchors the predicted energy to the potential energy surface (PES). Supervision therefore comes from energies and forces rather than from sample density alone. We first validate the framework on a two-dimensional prototype system: trained on its PES alone, without access to sample data points, the model recovers both the terminal energy landscape and the Boltzmann distribution it defines, reproducing the reference data. Trained on molecules and density functional theory obtained energies and forces, *Entropica* recovers the PES more accurately than a classical force field, while simultaneously learning to sample the associated equilibrium ensemble. Generated conformations lie in low-energy regions of the learned surface and the sampled density matches the Boltzmann statistics of molecular dynamics ground truth. *Entropica* opens a route toward physics based molecular generation via PES guidance, with potential applications from equilibrium ensemble generation to scoring of molecular interactions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular generative models, equilibrium sampling, Boltzmann generators, energy-based diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_eMaLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ENTROPICA, an energy-based diffusion model for molecules with an energy head and a force head. The authors start from the reverse-KL Boltzmann generator objective, reinterpret it as an optimal control problem and derive a "HJB" PDE whose residual is used as a training loss. They show that this residual is equivalent to the log Fokker–Planck residual of Lai et al. The physics enters through boundary losses at t=1t=1 which regress the energy head and the force head onto DFT energies and forces. The method is tested on a 2D Swiss roll where only the potential is given, on three rMD17 molecules at 500 K, on azobenzene and on ANI-1x.

Critical strengths

Nothing significant enough to support acceptance.

Critical weaknesses
  • The main theoretical claim does not hold up. Eq. 5 with F=−∇uF=-\nabla u and u=−ln⁡pu=-\ln p is simply the continuity equation ∂tp+∇⋅(pv)=0\partial_t p + \nabla\cdot(pv)=0 written for the log-density. There is no minimization over a control anywhere: in eqs. 17–19 the "min over θ\theta" is carried along and then just dropped. So this is not an HJB equation, there is no "optimal transport path", and Proposition 1 is an algebraic identity. What is actually implemented (Algorithms 3–4) is denoising score matching + the Fokker–Planck regularization of Lai et al. / Plainer et al. + DFT boundary terms. The stochastic optimal control story in the title, abstract and first contribution is, to my understanding, misleading.
  • The actual stochastic optimal control samplers are not discussed at all, i.e., Path Integral Sampler, Denoising Diffusion Samplers, Improved Sampling via Learned Diffusions and in particular Adjoint Sampling, which does Boltzmann sampling from a learned MLIP energy for molecular conformers via SOC. These methods use the energy as terminal cost and have a real HJB equation. The claim that the terminal condition of Berner et al. "is unavailable in practice" is only true for the generative modeling setting, not for the sampling setting which this paper also claims to consider (at least on a 2D toy target). Furthermore, the paper misses other related works for training energy based diffusion models such as Dual Score Matching, or Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences.
  • On equilibrium data the full method loses to its own ablations. In Table 6, "DSM only" has lower JS and lower W1W_1 on the ϕ1\phi_1 torsion than ENTROPICA for all three molecules, and "no HJB" has lower energy MAE for all three. Fine-tuning with a stronger HJB weight makes salicylic acid worse (JS 0.0077 → 0.0146). So the HJB term does not help with the torsion statistics, and the abstract's claim that "the sampled density matches the Boltzmann statistics of molecular dynamics ground truth" is not supported by the paper's own tables.
  • For molecules τ\tau is not an input of the network, and sampling at a different temperature just rescales the force by 1/kBτ1/k_B\tau at every noise level. This is only correct at t=1t=1 since the noised marginals of a Boltzmann density at a different temperature are not the rescaled noised marginals.
  • The methods has six loss weights that differ per experiment, HJB warm-up schedules, an optimizer switch from AdamW to SOAP during training, hand-stopped training, single seeds, no compute budget and no code. This raises serious concerns about the practicality of the approach.
Questions for authors
  • Do the authors agree that eq. 5 is the log-form continuity equation of the probability-flow ODE and that no control is optimized? If so, what does the control framing add beyond Lai et al. (2023) / Plainer et al. (2026)?
  • Why does "DSM only" beat the full model on the torsion statistics and "no HJB" on the energy MAE for all three molecules? What is the HJB term actually buying on equilibrium data?
  • How is the τ\tau rescaling at intermediate tt justified for molecules without τ\tau conditioning?
  • What do the authors mean by “The learning problem it entails is that of identifying the optimal transport path ff.”. Why would the method lead to an optimal transport?
2 comments
AuthorsOpenReview ↗

Response to Reviewer eMaL (1/2)

We thank the reviewer for the careful reading and the detailed, constructive comments, in particular on the theoretical framing and the related literature. We address each point below.

Weakness 1, Question 1,4: HJB claim and optimal control.

We thank the reviewer for this point and agree with it. For a general velocity field, Eq. 5 is the log-density form of the continuity equation, and App. A.1 did not perform a minimization. We have rewritten App. A.1 accordingly. Eq. 5 is now derived as the cost-to-go of a given flow, without any minimization, and Proposition 1 is stated as an algebraic identity.

The control view applies only when the diffusion velocity of Eq. 6 is inserted (VE, Dt=−gt2>0D_t = -g_t^2 > 0). Eq. 5 then becomes

∂tuθ+12Dt,Δuθ−12Dt,∣∇xuθ∣2=0,uθ(⋅,1)=u.\partial_t u_\theta + \tfrac{1}{2} D_t \\,\Delta u_\theta - \tfrac{1}{2} D_t \\,\\|\nabla_x u_\theta\\|^2 = 0, \qquad u_\theta(\cdot, 1) = u.

As Berner et al. (2022) show, this is the HJB equation of the control problem

V(x,t)=min⁡a;E[∫t112∣as∣2,ds+u(X1);∣;Xt=x],dXs=Ds,as,ds+Ds,dWs,V(x, t) = \min_{a}\\; \mathbb{E}\Big[\int_t^1 \tfrac{1}{2}\\|a_s\\|^2 \\, ds + u(X_1) \\;\Big|\\; X_t = x\Big], \qquad dX_s = \sqrt{D_s}\\, a_s \\, ds + \sqrt{D_s}\\, dW_s,

here with the reduced potential as terminal cost. The minimum over the quadratic control cost is

\min_a \Big\\{\tfrac{1}{2}\\|a\\|^2 + \sqrt{D_t}\\, a^\top \nabla_x u_\theta\Big\\} = -\tfrac{1}{2} D_t \\|\nabla_x u_\theta\\|^2, \qquad a^* = \sqrt{D_t}\\, F_\theta .

The optimally controlled process is our reverse SDE and its value function is the noise-smoothed reduced potential.

Entropica does not optimize a control or a policy, as stochastic-control samplers such as PIS or Adjoint Sampling do. The control problem is an interpretation of the PDE we solve which identifies uu as a value function and the reverse SDE as its optimal control. This binds the solution of the PDE to physical energies. The biased-data experiments (Sec. 4.3, 4.4) show that this is what corrects biased training data. We have revised the abstract, contributions, Sec. 3.2 and the conclusion to describe the stochastic optimal control connection as an interpretation of the HJB residual.

To answer question 4, we removed the phrasing "optimal transport path" and state that the learning problem is to find the time-intermediate energy surfaces which satisfy the HJB loss to generate equilibrium distributed samples. Wasserstein optimal transport was not intended.

AuthorsOpenReview ↗

Response to Reviewer eMaL (2/2)

Weakness 2: missing references.

Thank you for pointing this out. We have now added baselines (energy based samplers iDEM (Akhound-Sadegh et al. (2024)), LIB (Máté and Fleuret, 2023) for energy-only experiments; ScoreMD for alanine dipeptide experiments). PIS, DDS, DIS, Adjoint Sampling and ASBS are now discussed in Related Work. However, they address a different task since they learn a drift towards a given energy and evaluate that energy at their own samples during training. Entropica learns the energy itself from labels at fixed configurations, without an energy oracle or simulation during training. We also correct our sentence on Berner et al.: the terminal condition is unavailable only in generative modelling.

Weakness 3, Question 2: ablations on equilibrium data.

The reviewer is right for torsion JS/W1W_1 and energy MAE. On equilibrium data, DSM slightly outperforms the energy based, PDE regularized methods here. We argue that PDE regularized training is a harder learning task than only score matching. It requires backpropagation through multiple derivatives of the network and finding a non-trivial solution to the PDE residual. Krishnapriyan, Aditi, et al. discuss this in their article "Characterizing possible failure modes in physics-informed neural networks." Advances in neural information processing systems 34 (2021): 26548-26560.

What the Entropica regime improves is the training on data with a distributional bias through energy supervision (Sec. 4.4). The score matching alone matches the torsions as closely, the Entropica setup instead calibrates the sampled energies, at effective temperatures of 539–592 K against 551–608 K for MD and 616–673 K without the HJB term (Table 12).

Weakness 4, Question 3: temperature rescaling.

We thank the reviewer for pointing out this concern: rescaling the force by 1/kBτ1/k_B\tau is exact only at t=1t = 1. We now state it as an approximation. In the azobenzene sweep, the effective temperature tracks the reference near the training temperature. The bias experiment is unaffected, since training and sampling both use τ=500\tau = 500 K. The principled route is temperature conditioning with the HJB residual at each τ\tau, as in the Swiss-roll experiment and Müller–Brown (Sec. 4.1, App. C.1) For molecules we leave it to future work.

Weakness 5.

On the new alanine-dipeptide experiment, Entropica uses the default two-branch setting with one fixed set of loss weights and schedules for all three training sets, and every model is evaluated after a fixed training budget, without early stopping or checkpoint selection. On one A100 GPU, the full two-branch model trains in 2.5 GPU-hours, against 4.4 for ScoreMD. We now report this Appendix B.5 (Table 6).

Reviewer_AyDwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces ENTROPICA, which combines energy/force supervision with HJB objective to learn a generative neural sampler for sampling Boltzmann distribution. Experiments on a toy two-dimensional potential and small molecules, showing recovery of boltzmann distribution and temperature transfer.

Critical strengths

The paper address an important problem, learning an equilibrium sampler when physical labels such as forces/energies are available, but data distribution might be biased. Jointly learning a potential and a sampler is potentially useful approach, while the experimental support is not enough for the central claims.

Critical weaknesses
  1. My major concern is the reported observables largely appear unimodal. Matching these may demonstrate reproduction of local fluctuations but does not establish the method can cover mutiple metastable basins and recover relative equilibrium probabilities. Consequently, the current evaluation provides limited evidence about mode coverage, mode collapse and correct population weighting. The authors should evaluate established multimodal benchmarks such as muller-brown potential and alanine dipeptide at least. The issue is not simply the current molecules are small but that the experiments do not clarify identify the sampling challenge or demonstrate it has been solved. The evaluation should report basin populations, free-energy differences against converged MD references. Pointwise energy/force accuracy, energy histograms, and fitted effective temperatures do not establish correct conformational probabilities. Effective sample size should also be reported if feasible.

  2. Energy-based sampling is already established. The authors should compare with Adjoint Sampling [1] and ASBS [2] which are directly relevant energy-based neural samplers grounded in stochastic optimal control. Also should include comparison with fokker-planck consistent sampling [3] and Target Score Matching [4]. The combination of physical endpoint supervision and intermediate PDE consistency may constitute a useful contribution. However, the manuscript does not convincingly isolate the benefit of this coupling.

[1] Aaron Havens et al. Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching.

[2] Guan-Horng Liu, Jaemoo Choi, Yongxin Chen, Benjamin Kurt Miller, and Ricky T. Q. Chen. Adjoint Schrödinger Bridge Sampler.

[3] Michael Plainer, Hao Wu, Leon Klein, Stephan Günnemann, and Frank Noé. Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models.

[4] Valentin De Bortoli, Michael Hutchinson, Peter Wirnsberger, and Arnaud Doucet. Target Score Matching.

Questions for authors
  1. Please distinguish the specific new ingredient from the already established use of energies and forces to train a sampler without equilibrium data, as shown in previous work such as iDEM, FAB, PITA, Adjoint sampling, ASBS and etc..
  2. For figure 2, show marginal plot for x_1/x_2 axis.
  3. Can the authors run md simulation with the learned score? or the score can only be used for diffusion sampling?
2 comments
AuthorsOpenReview ↗

Response to Reviewer AyDw

We thank the reviewer for the careful review. We agree that the experimental support needed strengthening, and address each point below.

Weakness 1: multimodality and population weights.

We agree that the rMD17 torsions are close to unimodal and do not test mode coverage or population weighting. We add two multimodal benchmarks:

  • Müller-Brown potential with three basins and report the basin populations, free-energy differences and effective sampling sizes.
  • Alanine dipeptide: basin populations against converged MD, Wasserstein distances and ESS, for training on equilibrium data and on data with reweighted basin populations and temperature reweighted as well.

Weakness 2: comparisons and isolating the coupling.

  • Comparable baselines. We add direct comparisons with Fokker–Planck-consistent sampling (ScoreMD, Plainer et al.). By Proposition 1, our "DSM + HJB" ablation is a re-implementation of ScoreMD in our backbone, we additionally report results with the original ScoreMD code.
  • Isolating the coupling. The biased-data experiment (Sec. 4.3, 4.4) is a 2 × 2 design that isolates the coupling of endpoint supervision and intermediate PDE consistency. Trained on azobenzene data resampled to a higher temperature, models with score matching alone, with only the residual added, or with only the boundary terms added all reproduce the bias of their training data. Only the combination of energy supervision at the endpoint and the HJB residual largely removes it and recovers the equilibrium ensemble. The benefit therefore comes from the coupling itself, not from either ingredient alone. On alanine dipeptide (Sec. 4.4), score matching, score matching with the HJB residual and ScoreMD keep 101–102 % of the basin bias; the labels with the HJB residual in one network keep 68 %, and the two-branch correction 9 %. On Müller–Brown with biased samples, putting the force labels into the score regression instead, by target score matching, keeps the bias too (91 and 108 % of the temperature and basin bias retained); the two-branch correction of a TSM base removes most of it (18 and 22 %, one seed; App. C.1, Table 11).

Share of the azobenzene temperature bias retained (Sec. 4.3):

(%)without HJB residualwith HJB residual
without boundary terms95 ± 3 (score matching)115 ± 4
with boundary terms104 ± 311 ± 2 (Entropica)

Weakness 2 / Question 1: relation to Adjoint Sampling, ASBS, iDEM, FAB and PITA.

Adjoint Sampling and ASBS. These methods solve a stochastic control problem of the same structure as ours but on-policy, in a different setting. They learn a drift towards a given energy function, simulate the controlled process in every training step, and evaluate the energy and its gradient at the generated samples. ASBS extends this to a Schrödinger bridge from an arbitrary source distribution. Neither uses configuration data, but both need an energy oracle throughout training. Entropica solves the problem off-policy and without simulation. It learns the energy itself, i.e. the value function, from DFT labels on a fixed set of configurations, and evaluates no energy during training. A direct comparison therefore requires a target with an energy oracle. Because the two settings differ in what is given (an energy function to query versus a fixed set of labelled configurations), a benchmark against Adjoint Sampling and ASBS would compare methods for different problems rather than different solutions to the same one.

Entropica addresses a different setting: it learns the PES itself from DFT energy and force labels on a fixed set of configurations, with no energy evaluation during training. This gives two ingredients in the setting of a denoising diffusion model that we believe are new:

  • configuration data with biased population weights can be used for training, and the physical terms largely correct these weights (Sec. 4.3, alanine dipeptide above),
  • the learned energy is a by-product usable as a potential, and the model samples unseen molecules without any energy evaluation (ANI-1x, held-out molecules).

We discuss these methods in Related Work and discuss the difference of our setting in discussion.

Question 2: Fig. 2 marginals.

Added: marginals along x1x_1 and x2x_2 against the exact reference, for Entropica, LIB and iDEM at every temperature, with their total variations (App. C.1, Fig. 6).

Question 3: MD with the learned model.

Yes. At t=1t = 1 the energy head is a conservative potential trained on DFT energies and forces, so it could drive MD like any MLIP. Our aim, however, is to sample the equilibrium ensemble without MD.

I appreciate the authors’ reply. The manuscript has improved with the added results, and the claims are now more reasonable. However, some important issues remain.

For example, the alanine dipeptide model is trained only on dihedral angles, but the energy labels come from the potential energy of the full molecular system. After coarse-graining, the relevant potential becomes a potential of mean force, which is a free energy. These are different quantities, and I do not think this mismatch has been properly addressed.

For ESS, computing it only in dihedra space tells us about that CV distribution, but does not establish accurate sampling of the full configuration space. This limitation should be made clearer. In general, the writing also needs substantial improvement to make the paper easier to follow.

I will maintain my score. I appreciate the authors’ efforts and hope these comments help them further improve the work.

Reviewer_JYkfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ENTROPICA, a framework for jointly learning molecular potential energy surfaces and generating configurations consistent with their Boltzmann distributions. It combines an energy-level HJB/Fokker–Planck residual with energy and force boundary supervision, energy–score consistency, and denoising score matching. Validation includes an energy-only two-dimensional example and molecular experiments on rMD17 and ANI-1x, examining energy accuracy, generated ensembles, and sensitivity to biased training data.

Critical strengths
  1. Jointly learning a potential energy surface and a corresponding generative model is a useful objective. Physical energy and force supervision provides information beyond empirical sample frequencies and can help correct biases in the training distribution.

  2. The biased-data experiment and HJB/PES ablations are informative. They investigate whether physical supervision changes the generated ensemble and help distinguish the roles of the proposed constraints.

Critical weaknesses
  1. The methodological novelty is insufficiently established relative to closely related prior work: "Learning Interpolations between Boltzmann Densities" (LIB; Máté and Fleuret, 2023, https://openreview.net/forum?id=TH6YrEcbth). LIB learns energy interpolations and transport fields by minimizing a continuity-equation residual, but is not cited. The central equation (12) in LIB is closely related to equation (5) here, after identifying the force with the negative energy gradient and accounting for the time-dependent normalization term. Consequently, the HJB interpretation alone does not establish a sufficiently distinct methodological contribution. The practical scope of energy-only training is also unclear: although the residual can be evaluated without trajectory simulation, obtaining informative training points in high dimensions may still require ODE/SDE sampling or another exploration mechanism. The molecular experiments rely on configuration data and therefore do not demonstrate a general solution to this difficulty.

  2. The experimental comparisons are insufficient to establish an advantage over existing methods. There is substantial literature on energy- and data-driven generative modeling, Boltzmann sampling, and free-energy estimation. Here, the generative comparisons are largely restricted to DSM-based baselines and internal ablations. MACE and MMFF94 comparisons concern potential prediction rather than the competitiveness of the sampler. Without direct comparisons to representative energy-based generative methods, it remains difficult to determine whether the proposed combination offers a meaningful advance.

  3. V(x,t) is introduced without a sufficiently explicit mathematical definition.

Questions for authors
  1. What is the substantive mathematical distinction between equation (5) and equation (12) of LIB, beyond the diffusion-specific parameterization, normalization convention, and additional physical supervision?

  2. In a high-dimensional setting with access only to the energy function, how would informative training points be obtained? Is there evidence that the method can work reliably without ODE/SDE-based sampling or externally supplied configuration data?

  3. Can the authors provide a direct comparison with a representative existing energy-based generative method under comparable data and computational budgets?

  4. What is the precise definition of V(x,t), including the controls over which it is optimized, and in what sense does u_\theta approximate this value function?

3 comments
AuthorsOpenReview ↗

Response to Reviewer JYkf (1/2)

We thank the reviewer for the constructive review and for highlighting the biased-data experiment and the HJB/PES ablations. We address each point below.

Weakness 1, Question 1: relation to LIB.

Thank you for pointing us to LIB, which we now cite and discuss. Up to sign conventions, LIB eq. (12) is our Eq. 5 with their interpolation ftf_t as our energy uθu_\theta, their velocity VtV_t as our vθv_\theta, and their normalization term ∂tlog⁡Zt\partial_t \log Z_t. The substantive differences are:

  • In our set up the velocity is tied to the energy. It is the probability-flow velocity of the energy's gradient, vθ=12DtFθv_\theta = \tfrac{1}{2} D_t F_\theta with Fθ=−∇uθF_\theta = -\nabla u_\theta. The residual is a closed, second-order PDE in uu alone, with a unique solution given the terminal energy. In LIB the velocity is a separate network with many (f,V)(f, V) pairs satisfying eq. (12).
  • That unique solution is the diffusion interpolation of the target which LIB notes is not computable for general targets. In our derivation, their normalization term vanishes and does not need to be estimated.
  • Because the interpolation is the noised target, the same network can also be trained by denoising score matching and sampled with the standard reverse SDE.
  • For molecules we learn the terminal energy from a finite set of DFT energy and force labels, whereas LIB evaluates a closed-form target energy.

We add LIB as a baseline on the 2D targets. Müller–Brown (Table 8), largest error of the sampled free energies in kBTk_BT with the effective sample size (%) in brackets:

methodτ=0.5\tau=0.5τ=1\tau=1τ=1.5\tau=1.5τ=2\tau=2PES evaluations
LIB0.09 (100)0.04 (100)0.02 (99)0.03 (97)4 × 1.3·10⁸
iDEM1.79 (1)0.30 (64)0.14 (18)0.04 (31)4 × 2.7·10¹⁰
Entropica (one network)0.21 (95)0.04 (99)0.01 (98)0.01 (96)8.8·10⁷

Question 2: energy-only training in high dimensions.

We agree with the reviewer and do not claim energy-only training in high dimensions. The HJB loss is a pointwise residual without simulation in the objective, and it is valid for any collocation distribution that covers the target's support. While the exact solution does not depend on where the residual is evaluated, in practice the collocation points determine where it is resolved accurately.

In two dimensions, uniform collocation already works at high temperatures: for the flow-matching variant on the Swiss roll (App. C.2), for the diffusion model on the Swiss roll with an effective sample size of 97–99 % at τ≥0.5\tau \ge 0.5, and on Müller–Brown with 96–97 % at τ≥1.5\tau \ge 1.5. At low temperatures ESS is lower because the density concentrates in narrow basins that uniform points rarely reach (36 % on the Swiss roll at τ=0.1\tau = 0.1, 45 % on Müller–Brown at τ=0.5\tau = 0.5). Drawing half of the points from detached samples of the model, redrawn every 500 steps in only the last fifth of training raises ESS to 85–99 % at every temperature (Tables 8 and 10; new ablation in Sec. 4.1 and App. C.1). The PES is then also evaluated at these points. LIB and iDEM obtain their training points in the same way across the full training.

ESS (%)Müller–Brown, τ\tau = 0.5 / 1 / 1.5 / 2Swiss roll, τ\tau = 0.1 / 0.5 / 1 / 2
uniform collocation only45 / 83 / 97 / 9636 / 98 / 99 / 97
half from the model's samples95 / 99 / 98 / 9685 / 95 / 95 / 96

In high dimensions uniform collocation is not possible, and exploration from the energy alone remains open. Our molecular experiments therefore use noised labelled configurations as collocation points, together with denoising score matching, and train on their precomputed DFT labels without simulation or energy evaluation. Sampling without SDE/ODE is not in the scope of our manuscript. We are focusing on applying our derived training regime to denoising diffusion models where SDE/ODE sampling is the standard.

AuthorsOpenReview ↗

Response to Reviewer JYkf (2/2)

Weakness 2, Question 3: experimental comparisons are insufficient to establish an advantage over existing methods and direct comparison with a representative existing energy-based generative method under comparable data and computational budgets.

We thank the reviewer for pointing this out. We now added the relevant standard benchmark example of the Alanine Dipeptide equilibrium distribution recovery. We compare our approach against the DSM and the directly related ScoreMD (Plainer et al.) approaches, also for the case of a biased population of the training data, showing that our approach leads to better recovery of the equilibrium distribution compared to the relevant baselines. Alanine dipeptide with basin a up-weighted, 20,000 samples per model (Table 1), with the training GPU-hours (Table 6):

modelbasins a/b/c (%)bias retained (%) ↓W1W_1 / W2W_2 (°) ↓ESS (%) ↑GPU-h
equilibrium (floor)29/29/4101.0 / 4.299
biased training data77/14/1010058.0 / 79.936
DSM only77/13/1010158.0 / 79.7350.35
DSM + HJB77/13/1010157.8 / 79.435
ScoreMD77/13/910259.0 / 80.9344.43
Entropica, one network61/25/146846.7 / 70.6541.20
Entropica, two branches33/29/3797.4 / 19.3972.48

We additionally report the same comparison for the Müller-Brown potential test case and report the budget.

Weakness 3, Question 4: definition of V(x, t).

We now define the cost-to-go explicitly in Sec. 3.2, following Bellman's dynamic programming (Bellman, 1957; Bellman & Lee, 1978): uθ(x,t)u_\theta(x, t) is the cost that remains from state xx at time tt to the terminal time along the given flow. Its minimum over controls, V(x,t)V(x, t), is the value function of the control problem in App. A.1, which satisfies the same HJB equation once the velocity is tied to −∇uθ-\nabla u_\theta.

The cost has an entropic running part, whose expectation is the relative entropy and an energetic terminal part, the reduced potential uu. At t=1t = 1 no entropic cost remains, so V(⋅,1)=uV(\cdot, 1) = u is the physical PES.

Thank authors for the detailed response and revisions to the theoretical and experimental aspects. In particular, the inclusion of some important and relevant methods as baseline methods is commendable. The ablation experiments further confirm my point: the performance of this method is highly dependent on the data generation method.

However, judging from the revised draft, the method's innovation is still insufficient. Compared with existing methods such as LIB, the main difference lies in the parameterization method. Additionally, I have two problems on the alanine dipeptide experiment. First, the feature dimension used in the original ScoreMD paper is higher than that used in ENTROPICA in this paper; I am unsure whether this paper guarantees the same input features. Second, the two-dimensional features used in this paper are simply too low-dimensional to illustrate the performance of the proposed method.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers

We thank all reviewers for their careful and constructive reviews. We summarize the main changes here. Changes in the manuscript are marked in blue and the individual responses address each point in detail.

Scope of the method. Entropica targets a specific regime: a set of configurations, labelled with physical energies and forces. From these it learns a transferable PES and an equilibrium sampler jointly. It does not target energy-only sampling of a single given energy function. Stochastic-control samplers such as Adjoint Sampling and Schrödinger-bridge samplers such as ASBS learn a drift towards a given energy, simulate the controlled process during training and evaluate the energy at their own samples. Entropica instead learns the energy itself from labels on fixed configurations; for molecules, without simulation or energy evaluation during training (in 2D, half of the collocation points are the model's own samples, Sec. 4.1). This yields a PES supervised diffusion model and, in contrast to non-PES anchored energy based diffusion models, retains the equilibrium distribution in sampling even when trained on a biased non-equilibrium distribution, which we now additionally evaluate on the alanine dipeptide case. We now discuss this and the above mentioned methods in related work and the discussion section.

Theory. We revised our derivation of the HJB loss - the former value function V(x,t)V(x,t) is now the cost-to-go u(xt,t)u(x_t,t). We define it in 3.2. We add to App. A.1 the connection of Eq. 5 to the stochastic control problem of Berner et al. (2022), and restate Proposition 1 as an algebraic identity. We remove the phrase "optimal transport path", and cite and discuss LIB (Máté and Fleuret, 2023).

Corrected models. All models are retrained with a corrected energy head, and all tables and figures are updated. On rMD17 the accuracy is unchanged (8.4–11.4 meV with the PES terms). On ANI-1x, the per-molecule energy MAE is 172 meV, against 222 meV in the submission. To account for the alanine dipeptide standard benchmark, we trained a DSM model, a DSM+HJB model (the Fokker–Planck term of ScoreMD in our backbone, by Proposition 1), Entropica (DSM+HJB+boundary terms) in torsion angle space and compare with the original ScoreMD implementation (heavy atoms, coordinate space). Trained on equilibrium data, ScoreMD reaches the smallest torus Wasserstein-1 and -2 distances, with all tested approaches in the similar range (2.2° / 8.8°, against 3.8–5.4° / 11.1–14.4° for the other models; Table 14). Trained on population biased data, all non-energy supervised models inherit the training data bias while the energy supervised Entropica two-branch run removes most of it.

New multimodal benchmarks.

  • Müller–Brown potential: basin populations, free-energy differences, effective sample sizes and x1x_1/x2x_2 marginals against the exact reference, and populations as a function of temperature.
  • Alanine dipeptide: basin populations, free-energy differences, Wasserstein distances and effective sample sizes against converged MD. We train on equilibrium data and on data with reweighted basin populations or reweighted temperature, to test equilibrium recovery from biased data.

New baselines.

  • Fokker–Planck-consistent sampling (ScoreMD, Plainer et al.), the closest and comparable method to ours
  • Müller–Brown, energy-only methods: LIB and iDEM.
  • Target score matching (De Bortoli et al., 2024), which puts the force labels into the score regression, on Müller–Brown with biased samples, alone and as the base of the two-branch correction (App. C.1, Table 11).

Reproducibility. We now report for the practical alanine dipeptide benchmark the compute budget for the training of the evaluated models (Entropica as well as the compared reference methods), without early stopping or checkpoint selection (App. B.5).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.