Pith.

Chemical Physics · arXiv:2609.01871 v1 · submitted 1 Sep 2026 · 3 references

Latent unified smooth Hamiltonians for excited state chemistry

David Juergens, Martin Stöhr, Andreas E. Hillers-Bendtsen, O. Jonathan Fajen, Todd J. Martínez

The paper says

“We describe a neural network architecture and training procedure designed to model electronic ground and excited states of arbitrary molecular systems.”

more of the abstract

By indirectly learning a latent, implicit basis representation of the electronic-state Hamiltonian, the model offers a unified treatment of multiple electronic states, conical intersections, and non-adiabatic couplings. The formalism can be further extended to learn consistent latent representations of additional operators such as transition dipole moments, for example. To demonstrate the general capabilities of our architecture, we train and evaluate networks on two realistic photochemical systems, thymine and azobenzene. The resulting models accurately reproduce energies and oscillator strengths for the ground- and low-lying excited states relevant to the photochemistry of these systems. We highlight the performance of the trained networks by studying critical molecular geometries, including conical intersections and excited state minima. By construction, the proposed framework also recovers the emergence of Berry phase accumulation around conical intersections. By pairing key mathematical structure from quantum chemistry with the representation learning power of transformers, the presented architecture offers a qualitatively new path toward fast and accurate ground- and excited-state simulations.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 8 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Genuinely new architecture for ML excited states; the thymine CI result is striking. Main soft spots are missing code and limited NAC validation, but the paper is honest about its limitations. This line is wrong

3 major objections · 7 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 3 references say.

The paper says, at Electronic (transition) properties; Figure 2C

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The quantitative accuracy of the two quantities that differentiate LUSH from energy-only excited-state models, the NAC vectors and the transition-dipole vectors, is not established, and this is load-bearing for the abstract's claim of a unified treatment that includes non-adiabatic couplings.

more of objection 1

The authors correctly state that the Hellmann-Feynman formula d_ij = <psi_i|dH/dR|psi_j>/(E_j-E_i) is exact only for a geometry-independent implicit basis and very accurate when the geometry dependence is weak, but nothing in the loss of Eqs. (2)-(5) directly constrains the geometry dependence of the learned basis, and the only NAC evidence offered, the vortex plot in Figure 2C, is qualitative and is compared with no ab initio derivative-coupling reference. Likewise, the oscillator-strength loss constrains only the rotationally invariant magnitudes |mu_ij|, at fixed gaps, and leaves transition-dipole orientations unsupervised. Please add (i) a quantitative comparison of predicted NAC magnitudes and directions against TDDFT or FOMO-hh-TDA derivative couplings at held-out thymine geometries, for example along the optimization path of Figure S3 and in a loop around the LUSH-found MECI; (ii) a diagnostic of latent-basis stability, such as the absolute values of the overlaps of the adiabatic pseudo-wavefunctions between neighboring geometries; or (iii) a substantially softened statement of the NAC and dipole-direction claims in the abstract, introduction, and conclusion.

The paper says, at Excited state landscape of azobenzene; Figures 3-4

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The azobenzene case, one of only two photochemical demonstrations, supports the architecture in a substantially weaker form than the text claims. The manuscript concedes that the learned surfaces have finite S1/S0 gaps at the two reference conical intersection geometries (Figure 3C,D), and the LUSH-optimized MECIs deviate from the reference structures by up to 0.473 Å RMSD with relative-energy errors reaching -1.8 eV at twisted geometries (Figure 4).

more of objection 2

The stated cause, the near absence of cis-like and twisted geometries in the training set documented in Figure S6, is plausible and the authors are transparent about it, but a reader cannot conclude from these data that LUSH can provide accuracy on par with quantum chemical reference data, including conical intersections, for azobenzene. Please either re-train with a more balanced sample, ideally including gradients and NAC targets, or explicitly scope the azobenzene conclusions as a demonstration of qualitative CI reproduction subject to a known sampling bias.

The paper says, at Methods: Additional details on LUSH Pairformer

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The production of the three Cartesian components of the transition dipole tensor is not fully specified, and as written the architecture cannot supply lab-frame vector-valued dipoles. The feature backbone is SE(3)-invariant (RBF distance embeddings, invariant message passing adapted from Ref. 56, attention and Pairformer updates), and the dipole head is described as a resnet plus linear layer mapping invariant pair …

more of objection 3

… features to R^3 per state pair; an invariant map cannot output a quantity that transforms as a vector under rotations, so either an equivariant mechanism is missing from the description or the three components refer to an unspecified molecule-fixed frame in which the reference dipoles would have to be expressed. In either reading, the oscillator-strength loss of Eq. (5) constrains only the magnitude of each transition dipole, leaving orientations unlearned. Please describe the equivariant or frame-based construction explicitly and restrict the claims about transition dipole vectors accordingly.

7 minor objections

The paper says, at Results: Thymine conical intersections with LUSH

Pith says: minor objection 1 of 7

The reported CI-search result reads as a TDDFT S1/S2 gap of 3.1e8 eV, which appears to be a corrupted glyph; this number is the pivotal out-of-sample prediction of the paper and must be rendered unambiguously as approximately 3.1e-8 eV, with the uncertainty of the extrapolation stated.

The paper says, at Training procedure and loss definition

Pith says: minor objection 2 of 7

Equation (3) and its surrounding text use E_0_ref for the energy of the highest (m-th) state, whereas E_0 elsewhere denotes the ground state; please reindex this quantity, for example as E_m_ref, to avoid ambiguity.

show 5 more

The paper says, at Results: Chemical accuracy for ground state energetics

Pith says: minor objection 3 of 7

The text says the QM9 model approaches chemical accuracy, but Figure S1 reports a final validation MAE of 0.059 eV (1.36 kcal/mol) on U-U_iso; please state the exact number and avoid wording that implies chemical accuracy was reached.

The paper says, at Results: Chemical accuracy for ground state energetics

Pith says: minor objection 4 of 7

The description of the QM9 experiment, which says to take the lowest eigenvalue E_7 while the remaining eigenvalues E_6..9 are unused, is internally inconsistent for m=8 eigenstates; please clarify the indexing of the diagonalized eigenvalues.

The paper says, at Results: Recovering conical intersections and NAC fields

Pith says: minor objection 5 of 7

The starting geometry of the Adam CI search on the LUSH thymine surfaces is not stated; without knowing whether the search was initialized at the EOM-CCSD S1/S2 MECI geometry, the reader cannot assess how out-of-sample the found intersection actually is.

The paper says, at Methods and Computational Details

Pith says: minor objection 6 of 7

The Methods section reports no training hyperparameters (number of Pairformer blocks and attention heads, batch size, learning-rate schedule, epochs) and no inference timings, so the claim of accuracy at a fraction of the computational cost of quantum chemistry is not supportable from the manuscript alone.

The paper says, at Methods: Details of loss definition

Pith says: minor objection 7 of 7

In Eq. (5) the symbol used for the generic error is applied to both the excitation-energy error and the oscillator-strength error; please disambiguate the notation so that the two losses are separately identifiable.

What would settle it

Compute NAC vectors for 50 random geometries around the thymine S1/S2 conical intersection (within a 0.1 Å RMSD of the MECI) using TDDFT at the same level as the training data (CAM-B3LYP/def2-TZVP). Compare the LUSH-predicted NAC vectors (computable via automatic differentiation of the latent Hamiltonian and eigenvectors) to the ab initio NAC vectors using the cosine similarity and relative norm error. If the median cosine similarity is below 0.9 or median norm error exceeds 30%, then the Hellmann-Feynman route is unreliable for quantitative dynamics.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 8 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 23 s · 24,224 tokens read, 2,648 written
  3. Editor wrote the verdict above 13 s · 25,166 tokens read, 1,295 written
  4. Referee major revision: 3 major objections, 7 minor 2 min 37 s · 24,808 tokens read, 17,659 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The central idea here is genuinely new: learn a smooth latent Hamiltonian in an implicit basis, then diagonalize to get adiabatic states and rotate other operators into that basis. That is a cleaner approach than learning characteristic polynomials or separate network heads for NACs. It gives you conical intersections and Berry phase topology for free, which is the right inductive bias.

The paper does two things well. First, the thymine demonstration is real: the LUSH model finds a conical intersection with a TDDFT gap below 10−8 eV, even though the training set has no CI geometries. That is not trivial. The branching plane and NAC vortex plot are qualitative but structurally correct. Second, the azobenzene results show that the model captures the main potential energy surfaces and critical geometries, including MECIs, despite being trained only on energies and oscillator strengths—no gradients. That is a serious test and it mostly passes.

The soft spots are the ones the authors acknowledge. The withheld code is the biggest practical barrier; the paper cannot be reproduced until it is released. The quantitative validation of NAC vectors is missing—only a quiver plot for thymine, no comparison to ab initio NACs. The stress-test concern about the implicit basis changing with geometry is technically valid, but the paper's own evidence partially mitigates it: if the basis were wildly geometry-dependent, the CI gap would not close to 10−8 eV. Still, the absence of quantitative NAC comparison means the claim of "unified treatment" is only partially demonstrated. The azobenzene training set bias toward trans geometries is an acknowledged data issue, not an architectural flaw.

Who is this for? Researchers building ML models for excited-state dynamics. The paper deserves a serious referee. The central idea is sound, the execution is careful, and the limitations are stated. I would accept a peer-review invitation. The authors need to release code and provide quantitative NAC validation before I would cite it in my own work, but the paper itself is on the right track.

The strongest objection

LUSH predicts a fixed-size latent Hamiltonian H(R) in an implicit basis and computes nonadiabatic couplings (NACs) via the Hellmann-Feynman relation d_ij = <psi_i|dH/dR|psi_j>/(E_j-E_i). The paper states this is exact when the implicit basis is geometry-independent and very accurate when geometry dependence is weak. However, the basis is learned from data with no explicit regularization to enforce geometry independence. The network is trained only on energies, excitation energies, and oscillator strengths (scalar quantities); there is no supervision or validation of the latent basis stability. Consequently, the predicted NAC vectors and transition dipole orientations—which enter the core promise of a unified treatment—are unverified. The paper provides only a qualitative vortex plot for thymine (Figure 2C) and no quantitative comparison to ab initio NACs. For azobenzene, only oscillator strengths (scalar magnitudes) are compared, not vector directions. If the implicit basis changes appreciably with geometry, the Hellmann-Feynman NACs will be systematically wrong, undermining the claimed ability to simulate nonadiabatic dynamics. This is the most load-bearing concern because the entire framework’s value over prior methods rests on delivering consistent, accurate properties beyond energies, especially NACs and transition dipoles.

The reader's weakest assumption is precisely that the implicit basis is geometry-independent enough for the Hellmann-Feynman NAC formula to be accurate. This is the most load-bearing concern because the entire contribution—unified treatment of NACs and transition properties—depends on it. The paper acknowledges the assumption and provides no quantitative test of its validity. The concrete test proposed would directly assess whether the assumption holds for the trained model on thymine, a benchmark system. This does not change the verdict (CONDITIONAL) because the concern is already recognized in the reader's rationale (lack of quantitative validation for NACs and transition dipoles). The verdict remains unchanged; the condition that code and data be released and additional validation be performed is reinforced. No other concern is more fundamental: the architecture is sound, the energy predictions are credible, and the data coverage issue for azobenzene is a separate matter that future work can address. The honest non-finding for a verdict change is appropriate here.

The referee's report

If the quantitative claims survive revision, this is a significant methodological advance: LUSH provides, from one smooth latent operator, adiabatic energies, conical intersections, NACs, and transition dipoles, with the topology of the adiabatic representation (branching plane, singular NAC line field, Berry phase) guaranteed by the eigenvalue construction rather than fitted. The out-of-sample CI search on thymine is a genuine, falsifiable prediction, and the held-out tracking of states through an avoided crossing gives a convincing test that the learned H(R) is the right object to fit. I also credit the authors for stating the validity condition of the Hellmann-Feynman NAC formula explicitly, for training only on phase-invariant targets, and for candidly reporting the azobenzene sampling limitation. The principal caveat, reflected in the major comments, is that the quantities that go beyond prior energy-focused schemes, namely NAC vector magnitudes and directions and transition-dipole orientations, are validated only qualitatively or not at all, so the unifying claim currently runs ahead of the evidence.

Pith's reading of the paper (a machine's summary, not the authors')

The paper introduces LUSH, a neural network architecture that learns an implicit, smooth electronic Hamiltonian matrix as a function of molecular geometry. Diagonalizing this matrix yields the energies of multiple electronic states, including correct conical intersections. The eigenvectors allow direct computation of nonadiabatic coupling vectors and transition dipole moments within the same representation, avoiding explicit diabatization and derivative discontinuities. The approach is demonstrated on thymine and azobenzene, showing accurate reproduction of critical photochemical features such as conical intersections and Berry phase accumulation.

What carries the argument. The key machinery is the latent Hamiltonian matrix H(R) in $ℝ^{{m×m}}$, learned in an implicit basis via a combination of message passing, cross-attention to a fixed number of trainable latents, a pair-lift and Pairformer block for pairwise interactions, and a differentiable eigendecomposition. This matrix is smooth in nuclear coordinates, but its eigenvalues produce the required derivative discontinuities at conical intersections. The implicit basis is neither adiabatic nor fully diabatic, allowing flexibility while enabling the Hellmann–Feynman formula for nonadiabatic couplings.

Core claim. The central discovery is that by training a transformer-based neural network to predict a fixed-size latent Hamiltonian matrix H(R) in an implicit (neither strictly adiabatic nor diabatic) basis, and then diagonalizing it, one obtains adiabatic state energies, conical intersections, and nonadiabatic couplings (via the Hellmann–Feynman relation) and transition properties (via eigenvector rotation) all from a single smooth representation. The network uses cross-attention with trainable latents to handle molecules of variable size, a pair-lift and Pairformer block to encode matrix structure, and a differentiable eigensolver that handles degeneracies. A weighted loss function with buffer states mitigates errors from truncation at the spectrum boundary.

The premise it rests on. The implicit electronic-state basis must be nearly geometry-independent for the Hellmann–Feynman NAC formula to be accurate; if the basis rotates strongly with nuclear coordinates, predicted nonadiabatic couplings may be unreliable.

If the paper is right (Pith's inferences, not the authors')

The LUSH framework enables on-the-fly nonadiabatic molecular dynamics simulations with accurate NACs and transition dipoles at machine-learning speed, potentially extending accessible time and length scales.

The model can be extended to learn other electronic operators (e.g., spin-orbit couplings, permanent dipoles) by adding additional latent operator heads, providing a unified property prediction framework.

Because the latent basis is implicit, the method avoids the need for global diabatization, making it scalable to many coupled states without constructing complex locally-diabatic representations.

The correct reproduction of Berry phase accumulation around conical intersections is guaranteed by construction, ensuring topological consistency in nonadiabatic dynamics.

The approach uses buffer states to shield lower-lying states from artifacts at the boundary of the truncated Hamiltonian, allowing a fixed-size matrix to describe a limited number of states reliably.

If the implicit basis rotates rapidly with geometry, the Hellmann–Feynman NAC formula may become inaccurate; the paper's observation that twisted azobenzene geometries (under-sampled in training) show larger energy errors suggests that careful training set coverage is critical for reliable couplings.

The requirement for a fixed-size Hamiltonian means the number of reliably described states is limited by the buffer dimension; for systems with many near-degenerate states, this may require a larger latent space or a different truncation strategy.

The Pairformer component from AlphaFold3 may introduce unnecessary complexity; an ablation study comparing simpler pairwise attention schemes could clarify whether the added computational cost is justified.

The ability to 'correct' reference data (e.g., recovering conical intersections absent in TDDFT) hinted at in the conclusion suggests a potential use for denoising or improving electronic structure data, but this remains speculative and requires validation on systems with known deficiencies in the reference method.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

We describe a neural network architecture and training procedure designed to model electronic ground and excited states of arbitrary molecular systems. By indirectly learning a latent, implicit basis representation of the electronic-state Hamiltonian, the model offers a unified treatment of multiple electronic states, conical intersections, and non-adiabatic couplings. The formalism can be further extended to learn consistent latent representations of additional operators such as transition dipole moments, for example. To demonstrate the general capabilities of our architecture, we train and evaluate networks on two realistic photochemical systems, thymine and azobenzene. The resulting models accurately reproduce energies and oscillator strengths for the ground- and low-lying excited states relevant to the photochemistry of these systems. We highlight the performance of the trained networks by studying critical molecular geometries, including conical intersections and excited state minima. By construction, the proposed framework also recovers the emergence of Berry phase accumulation around conical intersections. By pairing key mathematical structure from quantum chemistry with the representation learning power of transformers, the presented architecture offers a qualitatively new path toward fast and accurate ground- and excited-state simulations.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 0.95

“By pairing key mathematical structure from quantum chemistry with the representation learning power of transformers, the presented architecture offers a qualitatively new path toward fast and accurate ground- and excited-state simulations.”

How the main claim is stated: asserted 0.78

What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.09 · human participants 0.01 · controlled comparison 0.04 · randomized assignment 0.01 · states quantitative result 0.08 · compares with prior results 0.07 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
We describe a neural network architecture and training procedure designed to model electronic ground and excited states of arbitrary molecular systems. contribution 0.90 0.92 0.08 asserted 0.98
By indirectly learning a latent, implicit basis representation of the electronic-state Hamiltonian, the model offers a unified treatment of multiple electronic states, conical intersections, and non-adiabatic couplings. method 0.79 0.69 0.20 asserted 0.97
The formalism can be further extended to learn consistent latent representations of additional operators such as transition dipole moments, for example. future work 0.87 0.21 0.04 hedged possibility 0.94
To demonstrate the general capabilities of our architecture, we train and evaluate networks on two realistic photochemical systems, thymine and azobenzene. method 0.67 0.97 0.04 asserted 1.00
The resulting models accurately reproduce energies and oscillator strengths for the ground- and low-lying excited states relevant to the photochemistry of these systems. result 1.00 0.95 0.29 asserted 1.00
We highlight the performance of the trained networks by studying critical molecular geometries, including conical intersections and excited state minima. result 0.47 0.93 0.08 asserted 1.00
By construction, the proposed framework also recovers the emergence of Berry phase accumulation around conical intersections. result 0.63 0.78 0.09 asserted 1.00
By pairing key mathematical structure from quantum chemistry with the representation learning power of transformers, the presented architecture offers a qualitatively new path toward fast and accurate ground- and excited-state simulations. interpretation 0.68 0.61 0.77 asserted 0.78

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,231 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

References (3)
  1. [1] 1 Cerullo, G., Polli, D., Lanzani, G., De Silvestri, S., Hashimoto, H. & Cogdell, R. J. Photosynthetic light harvesting by carotenoids: detection of an intermediate excited state. Science 298, 2395-2398 (2002). https://dx.doi.org/10.1126/science.1074685 2 Herbst, J., Heyne, K.... on PitharXiv
  2. [2] J. Amer . Chem. Soc. 143, 5425-5437 (2021). https://dx.doi.org/10.1021/jacs.1c00058 18 Toldo, J. M., do Casal, M. T., V entura, E., do Monte, S. A. & Barbatti, M. Surface hopping modeling of charge and energy transfer in active environments. Phys. Chem. Chem. Phys. 25, 8293-83... on Pithdoi:10.48550/arxiv.2506.23971
  3. [3] Nature 630, 493-500 (2024). https://dx.doi.org/10.1038/s41586-024-07487-w 64 Kasim, M. F. Derivatives of partial eigendecomposition of a real symmetric matrix for degenerate cases. arXiv, arXiv:2011.04366v04361 (2020). https://dx.doi.org/10.48550/arXiv.2011.04366 65 Stöhr, M. ... on PithPith's reviewarXivdoi:10.48550/arxiv.2011.04366
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment