Latent unified smooth Hamiltonians for excited state chemistry
David Juergens, Martin Stöhr, Andreas E. Hillers-Bendtsen, O. Jonathan Fajen, Todd J. Martínez
physics.chem-ph · cs.LG
Quick verdict
Latent-Hamiltonian NN unifies excited-state energies, CIs and NACs; convincing thymine/azobenzene demos, yet no ML baselines, no dynamics, code unreleased.
Summary
The paper introduces LUSH, a neural network architecture that learns a latent electronic Hamiltonian matrix for molecular excited states rather than predicting smooth adiabatic energies directly (which fail at conical intersections) or an explicit diabatic basis (non-unique for polyatomics). An SE(3)-invariant message-passing network and Perceiver-style cross-attention compress arbitrary molecules into a fixed token set; a pair-lift plus AlphaFold3-style Pairformer refines a state-pair representation from which parallel heads emit a symmetrized latent Hamiltonian and a transition-dipole tensor. Diagonalization yields adiabatic energies and pseudo-eigenvectors; other operators are rotated into this frame, and nonadiabatic couplings follow from the Hellmann-Feynman relation, guaranteeing conical-intersection topology (branching plane, NAC vortices, Berry phase) by construction. An adaptive weighting treats the highest states as buffers against truncation-boundary degeneracies. Results: QM9 ground-state MAE ~0.059 eV (~1.36 kcal/mol) in ~48 h on one V100; a 7.01M-parameter model trained on 135,000 CAM-B3LYP/def2-TZVP QeMFi points reproduces thymine spectra and locates an S1/S2 conical intersection whose TDDFT gap is 3.1×10⁻³ eV (smallest gap in QeMFi: 6.53×10⁻³ eV); a five-state model on 780,745 FOMO-hh-TDA azobenzene points reproduces spectra and minima, but twisted MECI relative energies err by up to −1.8 eV owing to sparse twisted-geometry sampling. Main limitations: no ML-baseline comparisons, NACs validated only qualitatively, no dynamics demonstration, and code to be released only upon acceptance.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LUSH learns a smooth latent Hamiltonian whose diagonalization yields adiabatic energies with correct conical-intersection topology — branching plane, NAC vortex fields and Berry phase — by construction.
- Nonadiabatic couplings computed from the latent Hamiltonian via the Hellmann-Feynman relation are exact when the state space is geometry-independent, and additional operators (transition dipoles) can be learned in the same implicit basis and rotated into the adiabatic frame.
- An adaptive state-weighting scheme with buffer states prevents contamination of low-lying states from degeneracies at the truncation boundary of the finite m×m Hamiltonian.
- Trained on QM9, LUSH reaches chemical accuracy (~0.059 eV ≈ 1.36 kcal/mol MAE) for ground-state energies in ~48 h on a single V100.
- On thymine and azobenzene, LUSH reproduces excitation energies, oscillator strengths, spectra and critical geometries, locating a thymine S1/S2 CI whose TDDFT gap (3.1×10⁻³ eV) is below the smallest gap in the training database.
Strengths
Physically grounded inductive bias that guarantees CI topology
Learning a smooth latent Hamiltonian and diagonalizing it makes conical shape (Fig. 2B), NAC vortex fields (Fig. 2C) and Berry phase structural consequences of the eigenvalue problem, whereas separate per-property heads (refs. 31-33, 45) cannot guarantee them; the buffer-state adaptive loss (Eq. 3, 0.5 eV window, top two states unsupervised) rigorously handles the (m)/(m+1) truncation-boundary discontinuity.
Validation spans ground-state universality and realistic photochemistry
QM9 chemical accuracy (~0.059 eV ≈ 1.36 kcal/mol MAE, Fig. S1, ~48 h on one V100 32GB) plus thymine (135k CAM-B3LYP/def2-TZVP QeMFi points) and azobenzene (780,745 FOMO-hh-TDA-BHLYP/def2-SVP points) with held-out critical geometries rather than toy model Hamiltonians.
CI discovery on learned surfaces near the true seam
Consistent multi-operator treatment demonstrated via state tracking
Candid error accounting
Weaknesses
No head-to-head comparison with competing ML methods
NACs and Berry phase validated only qualitatively
The photochemically decisive region is the least accurate
Reproducibility currently gated
Scalability and internal consistency gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
LUSH is a method paper from the Martínez group (Stanford/SLAC; arXiv 2609.01871) proposing a transformer-based architecture that predicts ground- and excited-state electronic structure by learning a latent m×m electronic Hamiltonian H(R) instead of adiabatic energies directly. Central claims: (1) diagonalizing the smooth latent Hamiltonian reproduces conical intersections (CIs), NAC vortices and Berry phase "by construction"; (2) additional operators (transition dipoles) share one implicit basis via rotation into the adiabatic frame; (3) the approach reaches chemical accuracy on QM9 and faithfully reproduces the multi-state landscapes of thymine and azobenzene. Claims (1)-(2) are supported structurally and claim (3) for the tested systems; the weakest links are quantitative NAC accuracy and the absence of competing-ML benchmarks.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Why can't standard machine learning potentials describe conical intersections?
Conventional ML potentials predict each adiabatic energy as a smooth scalar function of geometry, but at a conical intersection the adiabatic surfaces cross with a derivative discontinuity, so smooth models produce artifacts like spurious avoided crossings. LUSH instead learns a latent Hamiltonian matrix and diagonalizes it, so intersections and their coupling topology emerge naturally from the eigenvalue problem.
How accurate is the LUSH model?
On the QM9 ground-state benchmark it reaches a validation MAE of about 0.059 eV (~1.36 kcal/mol) after roughly 48 hours on one V100 GPU. For thymine and azobenzene excited states it reproduces reference spectra and critical geometries (minima, conical intersections) well, though twisted azobenzene intersection energies deviate by up to about 1.8 eV due to sparse training data in that region.
Is the LUSH code or training data available?
Not yet: the paper states the source code will be made public upon journal acceptance. The QM9 and QeMFi training datasets are publicly available, but deposition of the newly generated 780,745-point azobenzene dataset is not specified in the text.
Has LUSH been used to run nonadiabatic molecular dynamics simulations?
No. The paper demonstrates static predictions — spectra, optimized geometries, minimum-energy conical intersections and NAC fields. Dynamics simulations on the LUSH surfaces, plus gradient-inclusive training, are explicitly listed as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Mathematical notation and state labels are mangled by PDF extraction (e.g., S1/S2 rendered as '{6/{0}', superscripts garbled as in '3.1 ⋅ 108$ eV' for 3.1×10⁻³ eV); all numbers quoted in this review were decoded from context present in the text.
- Supplementary figures S1-S6 and Table S1 appear only as captions and in-text descriptions in the extraction; figure-level judgments rely on those captions and the authors' prose rather than the plotted values themselves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.