acceptodds
Under review as a conference paper at ICLR 2027

PhysDFTBench: Which Exact Conditions Survive from Architecture to Released Neural Exchange–Correlation Functionals?

Abstract

Whether released neural functionals inherit the exact conditions hand-built functionals secure by construction is an empirical question. We find no third-party audit of these releases: commonly cited audits evaluate no neural network. PhysDFTBench audits seven exact-condition families under declared protocols and a sealed coverage rule. It reads PBE, DM21's four released weights, Skala-1.1, three controls, our own correction and pcNN in part. Deployed identities fail in released weights. DM21 breaks spin permutation and the uniform-gas limit, and the fractional-charge tail and midpoint readings order the columns differently. For the conditions no deployed construction guarantees, which ones survive is a property of the trained weights, not of the architecture. DM21's supplementary material describes strict and approximate spin-symmetrisation routes. We test the released implementation: a reference-free test puts all four released DM21 weights outside their own round-off in the deployed energy, by 0.02–2.2 mHa. A 5,123-parameter neural correction is our controlled testbed. A loss term written directly on a criterion quantity need not reach that criterion even on its own probe. We also give a three-class rule for omitting the SCF response. In the unprotected class the omitted term need not vanish, and in our own loss it did not.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
exact conditions, density functional theory, neural exchange-correlation functionals, benchmark

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_6ZTmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PhysDFTBench, a tool designed to evaluate whether neural exchange-correlation functional models satisfy specific physical exact conditions inherent to Density Functional Theory (DFT). First, the authors directly test whether published model weights satisfy these properties; for instance, they observe that the energy output of the deployed DM21 model changes when spin channels are swapped for a fixed density—a finding that leads them to conclude the released DM21 models fail to satisfy exact spin-permutation symmetry. Second, the authors train their own neural correction models, leveraging various strategies: some properties (such as M4) are guaranteed through specific construction, while others are approximated by incorporating penalties into the loss function. Experiments demonstrate that conditions guaranteed by construction remain satisfied, whereas conditions enforced via the loss function do not necessarily meet the target criteria. Finally, the authors discuss the circumstances under which the SCF response can be ignored during training, categorizing scenarios into "protected," "needs no protection," and "unprotected." This analysis rectifies an implementation error found in an earlier version of their own implementation.

Critical strengths
  1. The physics-informed condition test proposed by the authors is simple and straightforward. For instance, it requires calculating the energy twice for the same fixed density, simply by swapping the two spin channels. This test requires no reference energy, no retraining, and no additional SCF run. Furthermore, the authors found that the results for all four sets of released DM21 weights exceeded the round-off level; thus, at least one point is clear: the released DM21 models do not satisfy exact spin-permutation symmetry.
  2. The paper is careful to clarify exactly what different tests demonstrate. Some tests correspond to exact identities, while others are conducted within the constraints of a finite basis, finite rr, or a one-sided condition. The paper avoids simply labeling these cases as satisfying an "exact condition." For example, in the case of a one-sided test, merely falling within a certain numerical range does not constitute proof that the exact condition holds.
Critical weaknesses
  1. The paper's core conclusion is quite straightforward, and the experiments largely verify things that could have been anticipated from the outset. A condition guaranteed by construction is fundamentally different from a condition incorporated into the training loss as a penalty. If a specific identity is guaranteed by construction for any set of parameters, the training process will not violate it. Conversely, adding a term like λC(θ)2\lambda C(\theta)^2 to the loss function offers no guarantee that the final solution will satisfy C(θ⋆)=0C(\theta^\star)=0. Thus, it is hardly surprising that a condition included in the loss fails to meet the corresponding criterion in the final model. Similarly, the statement ``an architecture is not a warranty'' is quite direct and can already be anticipated without these experiments. I also have reservations regarding the M4 results. The DM21 supplementary material describes both strict two-pass symmetrization and approximate symmetrization based on random spin flipping, yet the paper under review does not clarify which method was actually used for the four sets of released weights. Consequently, while the M4 results demonstrate that the released DM21 models lack exact spin-permutation symmetry, a non-zero M4 residual is to be expected if the approximate scheme was employed during training. It is therefore unclear how the released weights tested here correspond to the strict or approximate symmetrisation described in the original DM21 paper.

  2. Section 5 essentially involves re-applying a standard result to the authors' own early implementation.

The paper itself acknowledges that using the envelope theorem to neglect the SCF response—under appropriate conditions—is a standard shortcut. The issues identified in Section 5 stem from the authors' own earlier training implementation: they applied this shortcut without satisfying the necessary stationarity conditions, thereby omitting the response term. Furthermore, their proposed A/B/C classification appears to derive directly from the expression Ri=⟨∂Q∂κ,∂κ⋆∂θi⟩R_i=\left\langle \frac{\partial Q}{\partial\kappa},\frac{\partial\kappa^\star}{\partial\theta_i}\right\rangle. Class A corresponds to ∂Q/∂κ=0\partial Q/\partial\kappa=0; Class B applies when the density is independent of the parameter with respect to which the derivative is taken; and Class C covers cases where neither condition holds. Properly verifying the assumptions of the chain rule and the envelope theorem naturally leads to these scenarios.

Moreover, it is unclear what the A/B/C classification actually adds compared to the correct application of the existing envelope theorem. The failure case explicitly demonstrated in the paper stems from the authors' own previous implementation. Furthermore, the training run in which the authors corrected the gradient also involved changes to the label set; consequently, the specific impact of the gradient correction on the final training result was not isolated or measured.

Questions for authors

The questions stem from the previously mentioned weakness:

  1. How should the results for M4 be interpreted? The DM21 supplementary material describes both "strict two-pass symmetrization" and "approximate symmetrization based on random spin flipping." Which of these methods was actually used for the four sets of released weights tested in this paper? If this is currently unknown, then aside from demonstrating that the released DM21 implementation is not exactly spin-permutation symmetric, what else does M4 reveal?

  2. Why is the A/B/C classification necessary? The paper notes that neglecting the SCF response via the envelope theorem is a standard shortcut. If the assumptions of the envelope theorem are properly verified, the A, B, and C scenarios appear to follow directly. Does this work introduce any new sufficient or necessary conditions? Apart from the authors' own previous implementation, has any existing neural DFT work been shown to suffer from the same gradient problem? Can these findings be used to improve those existing studies?

Reviewer_z2YuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces PhysDFTBench to audit seven exact-condition families in released neural exchange–correlation functionals, supplemented by an SCF-gradient analysis and a small neural correction to PBE.

Critical strengths
  1. Auditing released neural functionals for physical consistency provides useful information beyond conventional energy-accuracy benchmarks.

  2. The fixed-density spin-swap test separates deployed-energy asymmetry from SCF convergence effects and includes controls for numerical artifacts.

Critical weaknesses
  1. The writing is difficult to follow. Dense, heavily qualified sentences and frequent shifts between findings, implementation details, and caveats obscure the argument, while essential interpretation rules require extensive consultation of the appendix.

  2. The figures and tables lack clarity. Figures 1–2 contain crowded annotations and small text, while Table 1 combines heterogeneous diagnostics and row-specific verdict rules, making the principal findings difficult to extract.

  3. The central scientific contribution remains unclear. The connection between the audit, gradient classification, and PBE correction is insufficiently developed, and the training experiment does not isolate the effect of the fixed-density repair.

Questions for authors
  1. What is the central new scientific insight, and how do the benchmark, SCF-gradient analysis, and PBE correction jointly establish it?

  2. Which conclusions in Table 1 remain robust after accounting for fit-window sensitivity, basis-set effects, and solver uncertainty, rather than merely describing the chosen evaluation protocol?

  3. Could the authors isolate the fixed-density repair under matched labels and training settings, and quantify its effect on constraint satisfaction and held-out accuracy?

1 comment

Thanks for the authors' responses. I will keep my initial score.

Reviewer_bExtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper builds benchmarks for checking the condition matching for neural network based functionals.

Critical strengths

This paper poses potentially an interesting problem of providing physical constraint guarantees not from constructions but from benchmark measurements, but from how it is written, it is very hard to assess its core strengths.

Critical weaknesses

Writing and presentation is very poor. From paragraph 1 and 2 in the introduction, the authors already assume that the readers would know what exact conditions are or what density-functional approximations are. ICLR is for a broad AI audience, and how it is written would not be able to convey its message. Even in the intro, the necessary background information and the identifying the gap in literature are all tangled up, and hard to identify. All of the figures are very poor quality, including Figure 1, which has overlapping texts with boxes/arrows. Overall, this is hastily put together, poorly presented paper, which might require a substantive amount of work before submission.

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.