acceptodds
Under review as a conference paper at ICLR 2027

Foundation Neural-Network Quantum States for Molecular Potential Energy Surfaces in Second Quantization

Abstract

Second-quantized neural-network quantum states have achieved accurate molecular energies, but extending them across molecular geometries requires a shared representation of the geometry-dependent wavefunction coefficients. We introduce geometry-conditioned foundation neural-network quantum states for molecular electronic structure in second quantization. A single autoregressive model learns a family of ground states from sparse anchor geometries and provides wavefunctions at untrained geometries without further optimization. Orbital alignment matches orbital identities and transports their phases, establishing an aligned orbital basis across geometries. Frozen energies reach chemical accuracy at every untrained query geometry for N, CO, and H. On additional molecular paths, the energy-trained wavefunctions yield dipoles, quadrupoles, and natural occupations without property labels. Across three paired N training seeds, orbital alignment lowers the mean absolute energy error over all untrained query geometries from 34-37 mHa to 0.049-0.085 mHa. At approximately 1 mHa mean absolute error, frozen evaluation reduces the per-geometry cost by relative to independent optimization, yielding an estimated end-to-end GPU-cost reduction on a 161-point N grid.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural-network quantum states, potential energy surfaces, second quantization, variational Monte Carlo

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CXoaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a geometry-conditioned neural-network quantum state that shares parameters across different geometries of a fixed molecule in second quantization. It combines orbital identity and phase alignment with a conditioned autoregressive wavefunction, jointly trained at sparse anchor geometries through variational energy minimization. The trained model predicts wavefunctions at unseen geometries without further parameter optimization. Experiments on small molecules along one-dimensional geometry paths demonstrate chemically accurate energies, electronic properties without property-specific training, and computational savings over independent optimization, with more modest gains over sequential warm starts.

Critical strengths

The orbital-alignment ablation provides strong evidence for the method’s central design choice. Across three paired N₂ seeds, alignment reduces the unseen-geometry energy MAE from 34–37 mHa to 0.049–0.085 mHa under matched architectures and training budgets. This connects the improvement to a specific representation issue rather than additional model capacity.

The frozen wavefunctions also achieve chemical accuracy at all tested query geometries for N₂, CO, and H₄, and reproduce electronic observables without property-specific training. These results support the method’s ability to recover useful electronic structure beyond fitting an energy curve.

Critical weaknesses

The advance over existing shared-wavefunction methods is not sufficiently established. Geometry-conditioned parameter sharing already exists in methods such as PESNet, so the contribution rests on its realization in second quantization through orbital alignment and conditioning. Experiments limited to small molecules along one-dimensional interpolation paths do not establish whether this extension offers a substantial advantage in more demanding settings.

The efficiency evidence is also limited against sequential warm starts. On the 161-point N₂ grid, the method reduces total GPU cost by about 16%; on the 41-point grid, it is more expensive. The comparison uses fixed training budgets rather than costs to reach matched accuracy and excludes CPU orbital and integral preparation. These results leave the claimed practical efficiency advantage insufficiently demonstrated.

Questions for authors

Clarification: Beyond adapting geometry-conditioned parameter sharing to second quantization, what methodological advance distinguishes this work from existing shared-wavefunction approaches such as PESNet? In particular, which aspects of the orbital-alignment procedure are new relative to prior molecular wavefunction-transfer methods?

Evaluation question: How does the computational advantage over sequential warm starts change when both methods are tuned to reach the same target accuracy, including orbital and integral preparation costs? This would help determine whether the reported savings reflect a practical efficiency advantage rather than the selected training budgets.

1 comment

We thank Reviewer CXoa for the valuable comments and for recognizing the strong alignment evidence and the chemically accurate frozen predictions.

The contribution combines occupation-space conditioning with a consistent HF-MO representation for shared variational learning. Below, we distinguish this construction from real-space transfer and SAO-based coefficient transfer, then address the accuracy-aware cost comparison. The new Li₂Se and H₄ results extend the tested scale and geometric coverage.

W1 / Q1. Methodological advance over existing shared wavefunctions

"Beyond adapting geometry-conditioned parameter sharing to second quantization, what methodological advance distinguishes this work from existing shared-wavefunction approaches such as PESNet?"

"In particular, which aspects of the orbital-alignment procedure are new relative to prior molecular wavefunction-transfer methods?"

A: Our contribution addresses a distinct representation problem in transferable wavefunction learning. We distinguish between real-space transfer and second-quantized wavefunction transfer because they require different learning constructions.

Real-space transfer. PESNet couples a MetaGNN that generates selected parameters to a FermiNet-style real-space wavefunction. Electron–nucleus and electron–electron relative vectors and distances, determinants, nuclear-centered decay envelopes, and an equivariant frame provide its physical inductive biases. In a second-quantized representation, these electron-coordinate features cannot be constructed directly from an occupation string, which contains no electron positions.

Second-quantized transfer. Rath and Booth express anchor coefficients in consistently indexed, symmetrically orthogonalized atomic orbitals for EC. Our autoregressive NNQS is trained and sampled in HF molecular orbitals. These can concentrate probability on relatively few occupation strings, which batched autoregressive sampling exploits by combining repeated configurations. Rath and Booth, Barrett et al., Malyshev et al.

How we use orbital alignment. Matching orbital character and aligning signs are established operations. Using cross-geometry AO overlaps, we match HF orbital identities within compatible occupation and symmetry classes and transport their signs along the training anchors. Hamiltonian integrals and orbital descriptors follow the same transformation; each new query is aligned directly to a training anchor. In these aligned coordinates, geometry supplies the global condition and aligned Fock diagonals supply orbital-specific descriptors. Both modulate occupation-prefix features through post-block FiLM; the phase network receives geometry and the full occupation configuration. Designing the coordinate convention and conditioning interface together lets one shared coefficient function learn physical changes in the electronic state without also fitting arbitrary orbital permutations and sign flips.

The specific advance is to make HF-MO correspondence a representation constraint on shared autoregressive learning, linking coefficient coordinates, orbital-indexed conditioning, and frozen-query attachment. The matched intervention establishes its learning consequence: more than two orders of magnitude lower N₂ query error across three seeds, with model capacity and training budget fixed.

W2 / Q2. Accuracy-aware cost comparison with sequential warm-start

"How does the computational advantage over sequential warm starts change when both methods are tuned to reach the same target accuracy, including orbital and integral preparation costs?"

A: The original submission already reports both full-grid MAEs in Appendix E.4. On the 161-point N₂ grid, frozen evaluation attained 0.822 mHa at 1.644 GPU-hours, and sequential warm starts attained 0.993 mHa at 1.948 GPU-hours. This comparison uses the warm-start cost at its higher error, giving the baseline the more lenient endpoint. Frozen evaluation still used about 16% fewer GPU-hours while achieving a lower MAE. Both results come from fixed schedules targeting about 1 mHa. A comparison at an identical target error remains to be measured.

RHF, orbital alignment, and integral preparation take 6.91 s for the full 161-point grid, less than 0.3% of the elapsed time for either workflow. At the measured rates, the lower cost per frozen query repays joint training at about 130 geometries. Sequential warm starts require the geometries to be optimized in path order. Once the jointly trained parameters are frozen, the queries can be evaluated in parallel (Section 4.4 and Appendix E.4).

We thank Reviewer CXoa again for the valuable comments.

Reviewer_pdZ5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a geometry-conditioned neural-network quantum state for molecular electronic structure in second quantization. The authors match RHF orbitals and align their phases across geometries to establish consistent meanings for occupation-basis wavefunction coefficients. An autoregressive wavefunction is jointly optimized at sparse anchor geometries and subsequently evaluated at other geometries without updating its parameters. Experiments cover six molecular paths, electronic observables, component ablations, and a computational-cost comparison on N₂. The predictions achieve chemical accuracy relative to same-basis FCI references on the tested paths.

Critical strengths

The orbital-alignment experiment provides convincing evidence for the proposed mechanism. Across three paired N₂ training seeds, alignment reduces query MAE from approximately 34–37 mHa to 0.049–0.085 mHa with matched architecture, optimization, and sampling budgets. This clearly demonstrates the importance of consistent orbital identities and phases when sharing an occupation-basis wavefunction across geometries.

The evaluation tests wavefunction information beyond energy. The energy-trained models predict NH₃ dipoles, CO₂ quadrupoles, and ethylene natural occupations without property-specific supervision. These complementary observables strengthen the evidence that frozen predictions retain meaningful electronic-structure information.

Critical weaknesses

The advantage over existing wavefunction interpolation methods is not established. Rath and Booth’s eigenvector-continuation approach [1] already evaluates correlated wavefunctions and observables at new geometries without repeating full nonlinear wavefunction optimization. Shared neural wavefunctions also have precedent in PESNet [2]. The paper discusses potential noise and conditioning difficulties when applying EC to sampled NQSs, but provides no direct comparison quantifying their consequences. Warm-start and fine-tuning comparisons leave the relative accuracy and cost of these alternative approaches unresolved.

The generalization boundary of frozen predictions is not characterized. The experiments use separately trained models for six small systems and test only interpolation along one-dimensional paths. There are no tests beyond the anchor ranges or under off-path geometric perturbations, whereas Rath and Booth [1] examine configurations displaced away from their one-dimensional training geometries. This does not establish that the proposed method cannot extrapolate; it leaves its robustness outside the sampled paths unknown. Such tests would help determine how broadly the central claim of wavefunction reuse applies.

The practical efficiency advantage remains conditional and narrowly tested. Efficiency is evaluated only for N₂/STO-3G. Against sequential warm-start, the total GPU-cost advantage is approximately 1.18× on the 161-point grid, while the proposed method is more expensive on the 41-point grid. CPU orbital and integral preparation is excluded. These results support a specific use case with sufficiently many queries, but provide limited evidence for broader practical advantages.

References

[1] Yannic Rath and George H. Booth. Interpolating numerically exact many-body wave functions for accelerated molecular dynamics. Nature Communications 16, 2005 (2025). Paper

[2] Nicholas Gao and Stephan Günnemann. Ab-Initio Potential Energy Surfaces by Pairing GNNs with Neural Wave Functions. ICLR 2022. Paper

Questions for authors

Can you provide a controlled accuracy–cost comparison with Rath and Booth’s wavefunction interpolation approach [1], accounting for anchor-state preparation and cross-state preprocessing? If this comparison is infeasible, what concrete obstacle arises on the small systems studied here?

With the training anchors and network parameters fixed, how does accuracy change beyond the anchor interval and under geometric perturbations away from the training path? Can you distinguish failures of orbital correspondence from failures of the learned wavefunction to generalize? These tests would clarify the applicability of frozen inference beyond the reported interpolation setting.

How does including RHF, orbital alignment, and integral preparation affect the total-cost comparison with warm-start? How many queries are required before the proposed approach becomes advantageous?

1 comment

We thank Reviewer pdZ5 for the detailed review and for recognizing the controlled alignment evidence and the electronic-property predictions.

W1 / Q1. Direct accuracy–cost comparison with EC

"The advantage over existing wavefunction interpolation methods is not established."

"Can you provide a controlled accuracy–cost comparison with Rath and Booth’s wavefunction interpolation approach [1], accounting for anchor-state preparation and cross-state preprocessing?"

A: The four workflows in Figure 6, which varies the number of N₂ geometries, retain their sampling procedure in larger determinant spaces. We report EC separately because full enumeration is a natural choice for small systems and becomes impractical for larger systems. We compared EC and direct frozen evaluation using the same model and query geometries for each molecule. In N₂ and CO, each anchor has 14,400 coefficients. We express these vectors in consistently indexed Löwdin-orthogonalized atomic orbitals, compute overlaps by inner product, and store their transition densities. Contraction and the eigensolve then take 1.29 and 1.61 ms per query, or 0.116 and 0.138 s including RHF and integral conversion. Table 9 gives the shared training and matrix preparation costs alongside the attained errors.

For Li₂Se, full enumeration would require 9.66 × 10¹² coefficients per anchor. We sample the anchors in their own aligned HF orbitals and build the cross-state tensors using nonorthogonal determinant matrix elements. Beyond the shared training, the tensor kernels take 2.669 eight-thread CPU task-hours in total. The tensors can then be reused at each new geometry for 0.294 s per query including RHF and integral conversion. Direct evaluation of the 16 queries uses 0.799 A100 GPU-hours, plus 0.242 s per geometry for CPU preparation.

Nearly dependent anchors make the sampled overlap matrix sensitive to noise. At the primary cutoff, 158 of 256 single-fit/query pairs passed the independent positive-norm check. Pooling the matrices gave positive norm intervals at all 16 queries. Appendix E.5 also reports the cost of this independent validation.

W2 / Q2. Generalization beyond the original paths

"The generalization boundary of frozen predictions is not characterized."

"With the training anchors and network parameters fixed, how does accuracy change beyond the anchor interval and under geometric perturbations away from the training path?"

A: We perturbed the original angle-only H₄ model off its training path. At 87°, changing the radius by ±1%, ±3%, and ±5% gives 0.373 mHa MAE and 0.829 mHa maximum error with the original weights fixed. The minimum matched orbital overlap across these six queries is 0.966 (Appendix F.2).

Two further experiments test how training coverage affects frozen predictions. In Fig. 9, separate N₂ and CO models trained out to 2.20 Å reach chemical accuracy at all 69 and 60 test points, respectively (Appendix F.3). In Fig. 3, an H₄ model trained on nine separate angular or radial distortions gives 0.062 mHa MAE and 0.138 mHa maximum error at 36 combined-distortion queries, including 12 outside the anchor convex hull. Appendix F.2 gives the split and the matched nine-anchor tensor-grid control.

W3 / Q3. Preparation cost, query count, and warm-start comparison

"How does including RHF, orbital alignment, and integral preparation affect the total-cost comparison with warm-start? How many queries are required before the proposed approach becomes advantageous?"

A: On eight CPU threads, RHF, orbital alignment, and integral preparation take 1.87 s for the 41-point N₂ grid and 6.91 s for the 161-point grid. CPU preparation accounts for 0.23% and 0.10% of the 161-point elapsed times for joint training plus frozen evaluation and for sequential warm starts, respectively.

On the 161-point grid the totals are 1.644 GPU-hours for joint training plus frozen evaluation and 1.948 for warm starts. At the measured rates, savings per frozen query repay joint training at about 130 evaluated geometries, including anchors. Each warm-start optimization needs the preceding geometry's optimized parameters. After joint training, frozen queries at different geometries can run in parallel (Section 4.4 and Appendix E.4, Table 7).

Reviewer_6XTFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This works studies the neural network quantum state transferable across geometry in the second quantization. They provide the methodology to make the definition of the molecular orbital basis consistent across geometries. They demonstrate that the amortized training successfully produces the model that predicts the ground state energies of the molecules with unseen geometry at the level of chemical accuracy. They also show the successful property prediction without the dedicated training for the task.

Critical strengths
  • The consistent / end-to-end optimizable learning of neutral quantum state across geometries in the second quantization and its demonstration
  • The demonstration of the power of the orbital alignment (Figure. 3)
Critical weaknesses
  • The contribution in terms of the model / machine learning perspective is weak; the alignment between orbitals are well-known in quantum chemistry. In the CO ablation (Fig. 7, Table 6), the proposed FiLM conditioning with per-orbital descriptors does not improve over simply adding the geometry to the input embeddings beyond seed-to-seed variation, as the authors acknowledge.
  • The molecules chosen in this paper are relatively small (the total number of determinants is up to 10^4~10^6); in my understanding, all of them can be solved with the full configuration interactions. The N2 and CO results are shown only for the regime scans stop at 1.40 Å and 1.60 Å, respectively, and do not extend toward dissociation. The region would probably be the most practical use case of NQS.

Overall: I recognize the contribution of the paper as mentioned in the strength. However, given the existing works in the first quantized NQS across geometry / molecule, ML-contribution and/or strong benchmark is necessary for the ICLR acceptance. I therefore consider the paper to be marginal.

Questions for authors

Could you extend the geometry region of N2 and CO toward dissociation?

1 comment

We thank Reviewer 6XTF for the valuable comments and for recognizing the cross-geometry learning formulation and the effect of orbital alignment.

The review asks about the methodological contribution and the scale of the benchmarks. The matched alignment intervention and the new Li₂Se result directly address these concerns.

W1. Methodological advance and conditioning ablation

"The contribution in terms of the model / machine learning perspective is weak; the alignment between orbitals are well-known in quantum chemistry."

"... the proposed FiLM conditioning with per-orbital descriptors does not improve over simply adding the geometry to the input embeddings beyond seed-to-seed variation ..."

A: Orbital correspondence in shared NNQS learning. We use established orbital matching and phase alignment. Zhanserkeev et al., JCTC 2021, for example, track orbital character, ordering, and phase along molecular trajectories. Our innovation lies in using orbital tracking to construct a representation for a correlated NNQS that transfers across geometries.

When an NNQS is optimized independently at each geometry, each network can adapt its coefficients to the local HF orbital convention. Joint training asks one conditional network to learn across all those conventions. Orbital permutations and sign changes then introduce artificial coefficient discontinuities even where the physical state varies smoothly. Supplying nuclear geometry alone does not identify or remove that arbitrary convention. Our methodological step is to establish consistent occupation coordinates for shared variational learning, enabling a single model to predict many-electron states at untrained geometries with frozen parameters.

The CO conditioning ablation documents the particular interface we use; the differences between conditioning variants are comparable to seed variation. The contribution is the transferable second-quantized wavefunction construction: geometry-conditioned coefficient learning in a consistent HF-MO representation. The ablation explains the implementation choice within that construction.

W2. Scale beyond the original small systems

"The molecules chosen in this paper are relatively small ... in my understanding, all of them can be solved with the full configuration interactions."

A: Li₂Se directly addresses the scale objection. The new STO-3G benchmark contains 40 electrons in 28 spatial orbitals, corresponding to approximately 9.66 × 10¹² fixed-Ms determinants. With five training anchors, the frozen model achieves 0.831 mHa MAE and 0.888 mHa maximum error across all 16 untrained queries. The shared-wavefunction result therefore extends well beyond the original enumerated spaces.

Q1. Extension toward dissociation

"Could you extend the geometry region of N2 and CO toward dissociation?"

A: We extended both scans to 2.20 Å, using 12 training points for N₂ and 11 for CO. The frozen models give MAEs of 0.274 and 0.173 mHa across the 69 and 60 test points, respectively. The maximum errors are 1.025 and 1.059 mHa (Fig. 9, Appendix F.3).

We thank Reviewer 6XTF again for the valuable comments.

Comments on the paper

1 comment

We sincerely thank our reviewers for their time and constructive feedback.

All three reviews recognize the controlled orbital-alignment evidence, and Reviewers pdZ5 and CXoa highlight the electronic-property predictions without property-specific supervision. These findings establish the central result: consistent occupation coordinates enable one jointly trained conditional NNQS per molecule to provide accurate wavefunction predictions at untrained geometries with frozen parameters.

The reviews ask us to establish the methodological advance, extend the tested scale and geometric coverage, and clarify the accuracy–cost comparison. We address these questions with the following additions.

Additional Results:

  • Li₂Se beyond the original small spaces. With 40 electrons in 28 spatial orbitals, the all-electron STO-3G model achieves chemical accuracy at all 16 untrained queries using five training anchors (Fig. 2(d), Appendix F.1).
  • H₄ compositional generalization. Training on separate angular and radial distortions produces chemically accurate predictions at all 36 combined-distortion queries, including points outside the training-anchor convex hull (Fig. 3, Appendix F.2).
  • Stretched N₂ and CO bonds. We extended each training interval to 2.20 Å. All 69 N₂ and 60 CO test points are within 1.6 mHa of FCI (Fig. 9, Appendix F.3).
  • Eigenvector continuation. We measured the preparation and query costs for EC using enumeration on N₂ and CO and sampling on Li₂Se (Section 4.4, Appendix E.5).

Other Clarifications:

The methodological contribution combines a consistent HF-MO representation with occupation-space conditioning for shared variational learning. The matched alignment intervention isolates the importance of this representation. Section 4.4 and Appendix E.4 now give the CPU preparation times and the estimated crossover with sequential warm starts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.