- 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 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 and is simply the continuity equation written for the log-density. There is no minimization over a control anywhere: in eqs. 17–19 the "min over " 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 on the 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 is not an input of the network, and sampling at a different temperature just rescales the force by at every noise level. This is only correct at 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 rescaling at intermediate justified for molecules without 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
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/ 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 is exact only at . 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 K. The principled route is temperature conditioning with the HJB residual at each , 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).
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>0). Eq. 5 then becomes
∂tuθ+21Dt,Δuθ−21Dt,∣∇xuθ∣2=0,uθ(⋅,1)=u.
As Berner et al. (2022) show, this is the HJB equation of the control problem
V(x,t)=mina;E[∫t121∣as∣2,ds+u(X1);;Xt=x],dXs=Ds,as,ds+Ds,dWs,
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 u 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.