acceptodds
Under review as a conference paper at ICLR 2027

Physical Knowledge as Side Information for Neural Fermionic Wavefunctions

Abstract

Neural wavefunctions provide a flexible approach to fermionic ground-state problems, but learning physically relevant correlations can remain difficult under finite optimization budgets. Existing methods often encode physical knowledge through architecture-specific modifications, making it hard to separate the value of the physical information from the way it is incorporated. We introduce a modular side-information architecture that makes fixed, Hamiltonian-derived physical computations available to a general fermionic neural wavefunction through trainable readers, while preserving its backbone architecture and determinant readout. The readers can act on particle features, pair communication, or orbitals, making the physical function, reader, and injection site explicit design choices. We evaluate this design on a three-dimensional attractive Fermi gas at infinite scattering length and a two-dimensional moiré electron system. In the Fermi gas, combining a two-body scattering-informed orbital correction with opposite-spin pair communication lowers the backbone energy by 1.82 ℏ²k_F²/m at 10,000 updates. With the reader and spectral modes fixed, the scattering relation lowers the energy relative to a matched constant relation. In the moiré system, four cross-attention readers using Hamiltonian-derived relations lower the energy by 0.12 to 0.19 E_h* at the same budget; density modes also help when read in every block. Across the two systems, the gain depends on the physical functions, their readers, and the injection sites.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum physics, Neural Quantum States, Variational Monte Carlo (VMC)

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Kq3fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper enhances a QERNEL-based fermionic neural wavefunction with fixed physical computations accessed through trainable readers. It separates the physical resource, the reader, and the point where the result enters the network. The resources include low-energy orbitals, two-body scattering relations, density modes, and particle-pair relations. The authors study a 14-particle attractive Fermi gas and a 24-electron moiré system, training each model from scratch. At 10,000 updates, the combined orbital and pair reader lowers the gas energy by about 1.82ℏ2kF2/m1.82\hbar^2 k_F^2/m relative to the coordinate backbone. The four moiré relation-attention readers lower the reported energy by 0.120.12–0.19Eh∗0.19E_h^\ast. All quoted uncertainties are conditional Monte Carlo standard errors for one trained model per setting.

Critical strengths

The experimental interface is well organized. Separating the physical computation, its reader, and its insertion point makes several comparisons easier to interpret. The residual readers recover the parent model at initialization, and the construction preserves the required exchange symmetry. Appendix D also correctly includes derivatives through configuration-dependent resources in the kinetic energy.

Critical weaknesses
  1. The improvement at fixed computational cost is not established. The paper compares energies at equal update counts but reports no wall-clock curves, per-update timings, or systematic complexity analysis of the added resources and readers. Several resources are recomputed at every sampling proposal and contribute coordinate derivatives to the kinetic energy. The occupied-state moiré reader also uses a configuration-dependent matrix inverse. Equal update counts therefore need not represent equal compute. Since the reported states are unconverged and the empirical contribution concerns finite-budget optimization, these measurements are central to assessing the method.

  2. The distinctive physical components have limited support over simpler alternatives. In the gas, the full pair reader, Fourier-pair control, and direct-only matching control give energies of 8.7389±0.01178.7389\pm0.0117, 8.7356±0.01258.7356\pm0.0125, and 8.7323±0.01278.7323\pm0.0127, respectively (Table 15). The no-history variant also performs at least as well, although it changes more than the retained pair state alone. In the moiré system, occupied-state relations and lattice-geometry relations give −7.4829±0.0296-7.4829\pm0.0296 and −7.4888±0.0342-7.4888\pm0.0342 (Table 2). These results support added relation-based communication, but provide little evidence for the particular exchange-matching or pair-history construction. The methodological advance is mainly the organization and controlled comparison of physical conditioning.

  3. Training-seed variability is not evaluated. The reported error bars quantify Monte Carlo sampling uncertainty for the trained wavefunctions. Repeating the baseline–best-reader comparison with independent training seeds would help assess the robustness of the reported gains, particularly the smaller differences between readers and injection sites.

  4. The gas benchmark does not establish competitive absolute performance. At 20,000 updates, the pair reader gives 4.55813±0.014224.55813\pm0.01422, while the simple closed-shell Slater reference has energy 4.016327814.01632781 under the same Hamiltonian, in units of ℏ2kF2/m\hbar^2 k_F^2/m (Table 12 and Appendix F.1). The reported gains depend on the chosen starting point, which uses neither plane-wave initialization nor reference pretraining. Their value on top of a physically initialized, competitive baseline remains untested.

Questions for authors
  1. Please report one-time preparation costs, per-update training time, and total training time for the coordinate backbone and the strongest reader on each system, with hardware and walker counts. Separate reference-resource preparation from configuration-dependent evaluation and derivative costs. Provide an energy-versus-wall-clock comparison for the principal models and a scaling discussion of the dominant added operations.

  2. The most useful additional experiment would be one additional independent training seed for the baseline–best-reader comparison on one selected system, evaluated with the same frozen-state readout protocol. This should take priority over adding further reader variants.

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

This paper proposes a modular way to add physical knowledge to a neural fermionic wavefunction and to test which part of the addition matters. A fixed computation derived from the Hamiltonian (a “resource”, e.g., low-energy plane waves, a short-range scattering profile, pair-overlap matrices, or Fourier density modes) is evaluated at every Monte Carlo configuration. A trainable “reader” maps it into the network, and an “injection site” (the input, a particle token inside a block, a pair state, the orbital matrix, or a log-amplitude factor) sets where it enters. Readers start at zero, so the augmented network starts as the original QERNEL-style backbone with a determinant readout. Exact antisymmetry is preserved (Proposition 1), the augmented model class contains the baseline (Proposition 2), and the coordinate derivatives of the resources are kept in the local energy.

The framework is tested on two systems, each at one Hamiltonian point. For a 3D attractive Fermi gas (N=14, stochastic reconfiguration), adding a scattering-informed orbital correction and opposite-spin pair communication lowers the energy by 1.82 ℏ^2 k_F^2/m at 10,000 updates, about 30% of the gap to a specialized Pfaffian–Jastrow–backflow reference. Controls show that simpler pair relations keep most of the communication gain. For a periodic moiré system (N=24, Adam), four cross-attention “relation” readers lower the energy by 0.12–0.19 E_h^*, and density modes help only when read in every block. The authors conclude that the gain depends on the physical function, the reader, and the injection site. They state that each model was trained once, that errors exclude training variability, and that all energies are far from converged.

Critical strengths
  1. A clear way to separate three design choices. Splitting each addition into a prescribed computation, a trainable reader, and an injection site, with zero-initialized readers and proofs of exact antisymmetry and neutral recovery, makes comparisons inside one backbone clean (Section 2.2, Appendix A). Keeping the coordinate derivatives of the resources in the local energy (Appendix D) is a correct and important detail.
  2. Informative controls, including negative ones. The controls in the gas (Fourier-pair content, bare kernel, direct matching, no-history, body-only) and in the moiré system (front versus in-block reading, generic Fourier features, free-particle versus moiré spectrum) show which parts of a reader carry the effect. Several show that the physical content is not what drives the gain (Sections 3.1–3.2, Appendix F).
  3. Open reporting of uncertainty and limitations. The paper labels many effects as unresolved, reports readers that hurt (bilinear reading, the body-and-amplitude reader), reports the wider width-96 model where the density benefit disappears, and states that errors exclude training variability (Section 4, Appendices D–F).
  4. Careful physics and two different systems. The healed scattering profile, the periodic continuation, the Ewald conventions, and the moiré one-body spectra are described in enough detail to be checked (Appendices B, C, F).
Critical weaknesses
  1. The statistical support for the main comparisons is thin. Every model is trained once, and the error bars (conditional MCSEs) capture only Monte Carlo sampling noise, not differences between training runs (Section 4). The main moiré gains are only 2.6-4.6 error bars (Table 2), and the in-block density gain, 0.097±0.046E_h^*, is about 2.1. Table 9 suggests that training noise may be larger than these error bars: the density-side minus front difference moves from -0.335±0.065 at 4,000 updates to -0.047±0.062 at 6,000, and the only repeated gas run (the two scattering-relation rows of Table 1) has the same seed, so it does not measure seed variability. Many comparisons are made (Tables 2, 6, 11, 14-19), and the “main gains” highlighted in Figure 2 appear to be chosen from them without a correction for multiple comparisons.
  2. The experiments stop before convergence. At 20,000 updates the best gas model (4.558 ℏ^2 k_F^2/m, Table 12) is still above the free Slater state (4.016, Appendix F.1), and the curves are still falling (Figures 3, 5). The gains change with training: the pair increment over the orbital reader drops from 0.549±0.020 at 10,000 updates to 0.175±0.019 at 20,000, and the total gain over the backbone from 1.82 to about 1.39 (Table 12), although the share of the gap to the pairing reference closed by the pair model rises from 31.8% to 52.7% at 16,000 updates (Appendix F.5). In the moiré system, the density benefit is absent in the width-96 model at 20,000 updates (+0.047±0.041, Appendix E) and the spectral benefit is unresolved (Appendix E.1). It is therefore unclear whether the readers lower the final energy or only reach it sooner, and the cost per update is not reported (Section 4).
  3. The gains are only partly shown to come from physical knowledge. In the gas, replacing the scattering-specific pair content with Fourier functions, removing the scattering factor, or removing the exchange assignment changes the energy by less than 0.01 (-0.003±0.017, +0.003±0.017, -0.007±0.017; Section 3.1), while the pair reader alone on the bare backbone already gives -0.816±0.024. The largest gain therefore seems to come from the added pair structure and parameters (+30,912, Table 13), as the paper notes (Section 1), although the title frames the result in terms of physical knowledge. In the moiré system, occupied-state and lattice-geometry pair relations differ by 0.006±0.045, which is too imprecise to tell whether the content matters, because the gain itself is only 0.134. No control uses learned, random, or shuffled inputs with the same added parameters (27,747 and 13,827 for the two relation readers, Appendix B.2), so extra capacity is not ruled out. The best evidence for physical content is the gas scattering relation (0.254±0.043, about six times the difference between the two scattering runs in Table 1), but it comes from one seed; the moiré spectrum effect (0.167±0.049) reverses when the same modes are read inside the blocks.
  4. The reference and baseline make the size of the gains hard to judge. The specialized gas reference has 8,530 parameters and ran with a different adaptive damping (it reached 10, against 10^(-3) for the general models); the authors say the protocol does not reproduce the original converged results (Appendix F.1). Even so, the best augmented model closes only about 30% of the gap to it, which suggests that the determinant readout may be the main limit for a paired state. A Pfaffian or geminal readout combined with the readers is not tested. For the moiré system I found no reference energy (e.g., Hartree–Fock or a converged result), so the best gain of 0.19 E_h^* (about 2.6% of the total energy) cannot be compared with the remaining error.
  5. The scope is narrow and gives little guidance. In the main experiments each system is studied at one Hamiltonian point, weights are not transferred across Hamiltonians, and the reader must be rebuilt by hand for each system (Section 4). In the only other couplings, a historical width-96 comparison, the effect of the density-power reader changes sign between λ=0 (+0.076±0.009) and λ=2 (-0.210±0.018, Appendix G), and the paper gives no rule to predict this. The conclusion that the gain “depends on the physical functions, their readers, and the injection sites” therefore does not tell a user how to choose a resource, reader, or site for a new system. The main contribution is a comparison protocol whose value depends on the reliability of the comparisons (Weakness 1).
Questions for authors
  1. Training variability. Can you report several independent training seeds (e.g., three to five) for the key comparisons: the orbital readers, pair communication and its controls in the gas, and density-in-every-block and the four relation readers in the moiré system? Which differences remain larger than the spread between seeds?
  2. Convergence and cost. Do the gains remain when training runs to a plateau? Please show energy against wall-clock time or FLOPs, including the cost per update of the readers, and say whether the readers lower the final energy or only reach the same energy faster.
  3. Controls without physics. Could you add controls with the same reader and number of added parameters, but with learned, random, or shuffled inputs (matched in mean and variance), for the gas pair reader and the moiré relation readers? This would separate physical content from extra capacity.
  4. Reference energies. What are the converged or best-known energies for the N=14 gas and the N=24 moiré system? Relating the gains to the remaining gap would make them easier to interpret.
  5. Determinant readout. Is the remaining gap to the pairing reference caused by the determinant readout? Would a Pfaffian or geminal readout with the proposed readers, or the readers added to the specialized ansatz, change the conclusions?
  6. Fair optimization. Were the learning rate and damping tuned separately for the baseline and the augmented models, and why did the adaptive damping of the reference and the general model differ so much? How sensitive are the main differences to these settings?
  7. Selection and uncertainty. How many reader and site variants were tried in total, and how were the highlighted configurations in Figure 2 chosen? For the moiré readers, can the pairing of runs be verified, or can longer readouts be used to check the errors, given the heavy-tailed local energies (Appendix F.5)?
Reviewer_Lv3eOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies injecting side information for neural wavefunctions and evaluate which side information paths help the most for Fermi gas and moire systems.

Critical strengths

This paper studies how to make existing physical knowledge injection more effective in the problem of neural wavefunctions for fermionic ground-state problems.

Critical weaknesses
  • writing and presentation can be improved substantially. ICLR is not a physics/chemistry venue, so the problem domain should be explained very clearly for the general AI audience. the paper they cite (Glehn etal. 2023) does an excellent job in this.
  • a lot of the terms are unclear: i have never heard of the term like "trainable reader" or "physical reader," which authors use interchangeably. if this is a term they came up with, they should define it clearly or better, USE THE TERMS THAT ALREADY EXIST in the field, like physics encoder or physics adapter.
  • their contributions are summarized in terms that are not familiar to the AI audience. Thier contributions should be clear without having temrs lie h^2 k_F^2/m or "MCSEs" which was never defined before (they define it later in the experiments section).
  • Theta was never defined.
  • I also don't know what "physical resources" mean
  • Figure 1 is more of an appendix figure with all the details, instead of conveying the main idea/contribution of the paper.
  • There is no clear take-away from the results. The question itself made sense: how should we introduce physical knowledge into the network? There is no clear take-away messages of interesting discoveries from the paper.
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.