Pith. sign in

REVIEW 4 major objections 4 minor 48 references

This paper shows that a generative eigensolver plus selected configuration interaction reproduces exact full-configuration references through 30 qubits on an EUV resist molecule and through 32 on a tin-oxide ladder, and the same pipeline ru

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-07-31 23:19 UTC pith:ZMG6Q33K

load-bearing objection Solid scaling demonstration with honest limitations; the top-rung accuracy claims need seed statistics before you quote them, but the verification ladder is real. the 4 major comments →

arxiv 2607.23988 v1 pith:ZMG6Q33K submitted 2026-07-27 quant-ph physics.chem-ph

Scaling a CUDA-Q GQE + QSCI pipeline to 40 qubits for EUV photoresist chemistry

Karim Elgammal , Marc Mau{\ss}ner This is my paper
classification quant-ph physics.chem-ph
keywords generative quantum eigensolverquantum-selected configuration interactionEUV photoresisttin oxo-hydroxidequantum chemistryvariational upper boundactive-space selectionbond dissociation energy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Using a generative quantum eigensolver coupled to quantum-selected configuration interaction, the paper tries to establish that a single, unchanged training protocol can track exact CASCI/FCI references across active spaces of 14 to 40 qubits, staying below 1.6 millihartree error through 30 qubits on methyltin trihydroxide and through 32 on the SnO ladder (best seed at each top rung), and that the same pipeline, depth-truncated, runs on a physical 54-qubit processor. The application is EUV photoresist chemistry: the tin-carbon bond whose breaking on ionisation flips solubility is multireference bond-breaking chemistry where single-determinant DFT fails. A classical language-model policy emits unitary coupled-cluster operator sequences; sampled bitstrings become determinants, diagonalised classically, and a cross-circuit generalised-eigenvalue step makes every reported energy a variational upper bound. The paper reports that ionisation collapses the classical UCCSD(T) Sn-C bond dissociation energy from 72.6 to 21.2 kcal/mol, quantifying the solubility switch. If the central claim holds, the result is a hardware-ready route to correlated energies at scales beyond exact statevector simulation, with a measured boundary at 40 qubits where sampling support rather than device fidelity is the limiting factor.

Core claim

The central claim: a pipeline pairing a generative eigensolver with quantum-selected configuration interaction gives chemical accuracy (below 1.6 mHa) against exact CASCI/FCI through 30 qubits on the methyl resist and 32 on SnO, with best-seed errors of 1.21 and 1.45 mHa. Depth-truncated, it reaches +0.330 mHa on a real 54-qubit processor at 14 qubits and +3.92 mHa at 22 qubits for the industrial ligand. Circuit depth is the scaling lever: raising the operator count from 10 to 80 cuts SnO-32q error from 10.06 to 1.45 mHa while the refined subspace stays 0.017% of the 166M-determinant space. The 40-qubit rung reaches only 22.8 mHa — a support-limited demonstration — and classical recovery rea

What carries the argument

The machinery is a trained generative policy feeding quantum-selected configuration interaction. A transformer emits sequences of L unitary coupled-cluster operators; each sequence defines a circuit whose sampled bitstrings, post-selected for particle number and spin, provide determinants for classical diagonalisation. A generalised-eigenvalue step merges the determinant spaces of all circuits, making every reported energy a variational upper bound. This refined subspace grows near-linearly with L yet stays a vanishing fraction of the full determinant space (0.017% at 32 qubits), so depth buys coverage without classical enumeration. The active-space descriptor — atomic-valence projection for

Load-bearing premise

The pipeline's chemical conclusions rest on the assumption that the automatically selected active spaces (atomic-valence projection for the resist, energy-window selection for SnO) contain the orbitals that actually drive the ionisation-induced Sn-C bond weakening; the paper itself flags the 3.6 Å stretched-geometry point where 'the active-space descriptor is about to dissolve,' and every quantum energy is compared only against CASCI within the same active space.

What would settle it

Enlarge the active space for CH3Sn(OH)3 at the stretched Sn-C geometries (e.g., include Sn 4d and the C-H σ/σ* manifold), recompute the neutral and cation dissociation curves with the same QSCI pipeline, and check whether the six chemically accurate points below 3.6 Å stay below 1.6 mHa and whether the 3.6 Å point drops into the chemical-accuracy band. If the ionisation-induced bond-energy collapse (72.6 to 21.2 kcal/mol at UCCSD(T)) shifts by more than the paper's own 0.13 eV channel-ordering margin, the active-space descriptor is the limiting approximation rather than the quantum solver.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the accuracy claims hold, correlated quantum-chemistry benchmarks on near-term hardware can target active spaces of roughly 30 spin-orbitals with exact verification, not just estimates.
  • Because every GQE+QSCI energy in the ladder is a variational upper bound, any future tightening of the sampled subspace strictly lowers the error toward the exact CASCI limit without re-running the training.
  • Circuit depth, not training length, is the effective scaling lever: the measured near-linear growth of the refined subspace with operator count means deeper circuits are the direct path to chemical accuracy at 32+ qubits.
  • For resist design, the quantified 72.6 to 21.2 kcal/mol collapse of the Sn-C bond energy upon ionisation gives a concrete target: tuning that ratio controls the solubility switch.
  • Classical configuration recovery reaches the 32-40 qubit spaces from the pipeline's own seeds and even from scratch (1.48 mHa at 32q, 1.25 mHa at 40q), so the hybrid pipeline's near-term value may be as a seed generator for classical selected-CI rather than a standalone quantum sampler.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's own caveat about the 3.6 Å point suggests a concrete testable extension: re-running the same protocol in active spaces enlarged by Sn 4d or C-H σ/σ* orbitals would show whether the active-space descriptor dissolves universally or only for this selection window.
  • The near-linear subspace growth with L implies a rough cost law for the next rung: at 32 qubits the subspace is 0.017% of the determinant space, so reaching chemically accurate coverage at 40+ qubits likely needs L in the thousands or a different conditioning mechanism than raw depth.
  • The hardware result that noisy circuits broaden determinant support — improving on ideal sampling by 15.5 mHa at 22 qubits — suggests that shallow noisy circuits plus per-circuit readout calibration could replace explicit error mitigation for QSCI if the effect reproduces across devices.
  • The observation that classical S-CORE reaches chemical accuracy from scratch at 32 and 40 qubits hints that the real frontier for quantum advantage in this chemistry is not active-space size but generating determinants that classical selected-CI cannot cheaply reach; a budget-matched comparison on a harder instance would settle where that frontier lies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper describes a hybrid quantum-classical workflow (GQE + QSCI) implemented in CUDA-Q and applies it to active spaces of 14 to 44 qubits for tin-containing EUV photoresist model systems. A GPT-2 policy autoregressively emits UCCSD operator sequences; the resulting circuits are sampled, the sampled bitstrings are mapped to determinants, and the Hamiltonian is diagonalized classically, with a cross-circuit GEVP refinement that yields variational upper bounds. Exact CASCI/FCI references are computed at every rung through 40 qubits. The authors report chemical accuracy (<1.6 mHa) through 30 qubits on CH3Sn(OH)3 and through 32 on SnO, with the 40-qubit resist result explicitly labeled as support-limited at 22.8 mHa. Real-hardware QSCI on IQM Emerald is reported at 14 and 22 qubits, and a classical S-CORE baseline reaches chemical accuracy at 32/40q without a quantum sampler. The paper also quantifies the ionisation-induced weakening of the Sn-C bond using classical coupled-cluster references. The manuscript is unusually explicit about its limitations: the 40q result is a demonstration, the full hardware ansatz is depth-limited, and classical subspace recovery maps the boundary rather than being beaten.

Significance. If the numerical claims hold, this is a useful scaling demonstration: one training pipeline is run across 14-40 qubits with exact references at every verified rung, the GEVP-refined energies are variational upper bounds, and code and persisted hardware counts are provided. The honest separation of solver gap from active-space gap, the explicit admission that classical selected-CI solves the largest spaces, and the reproducing entry points are strengths that make the paper valuable as a benchmark. The main caveat is that the headline accuracy at the largest chemically accurate rungs rests on a single seed or an unspecified best seed, so the statistical reproducibility of the stochastic generative pipeline is not yet established. The paper's own limitations are stated clearly and qualify the scope appropriately.

major comments (4)
  1. [Table 1 footnote / Abstract / Conclusion] The central claim that the pipeline is chemically accurate through 30 qubits on the resist and 32 on SnO rests on a single seed (resist 30q, L=600, seed 137) and an unspecified best seed (SnO 32q). Because GPT-2 training is stochastic and the reported values are variational upper bounds, a favorable draw cannot be ruled out. The footnote reports two seeds at several rungs, but not at these two top-rung accuracy points. Please report seed counts and seed-to-seed statistics (mean/min/max, number of seeds) at least at the 30q resist and 32q SnO rungs, or revise the claim throughout to 'one seed reached chemical accuracy' rather than 'the pipeline is chemically accurate.' This is load-bearing because the scaling claim is precisely a capability claim at the largest scales.
  2. [Section 3, Table 3 and 'real QPU' paragraph] The paper reports two different real-QPU 14q SnO results: +0.330 mHa (5000 shots, 42 CNOT) in Table 3, and later +0.575 mHa (36,000 shots, 31 of 49 determinants) described as 'an independent second route to the same instance's +0.330 mHa.' These are not reconciled. Additionally, Table 3 lists 'SnO GPU sim 14 +1.01 mHa' with 'GEVP-refined: 0 mHa' while Table 1 gives 0.000 mHa for the same rung. Please clarify whether the later number is a different circuit/protocol or a reproduction, and how the GPU-simulation number relates to the ladder's 0.000 mHa entry. As written, a reader cannot identify the canonical reported hardware accuracy.
  3. [Section 3, Table 3 / Section 4, configuration recovery] The 22q hardware result after per-circuit readout calibration is +3.92 mHa (81% of active-space correlation), but the same raw counts under classical configuration recovery are reported as +0.18 to +0.21 mHa. The text should state which of these is the headline hardware accuracy and whether the 'variational upper bound' property is claimed for the raw hardware QSCI energy or only after the classical recovery step. The distinction matters for the hardware section's conclusions, since classical recovery changes the reported energy substantially.
  4. [Section 4, limitations (i)-(iv)] The listed limitations are explicit and appropriate, but the Abstract and Conclusion still say the pipeline 'is chemically accurate' through 30/32q without immediately reminding the reader that this is best-seed/single-seed and that 40q is not chemically accurate. Consider adding a one-sentence qualifier in the Conclusion to match the Abstract's 'on the best seed at the top rungs', so the final summary does not overstate the seed robustness of the result.
minor comments (4)
  1. [Section 2, loss A/B study] The sentence 'both reach chemical accuracy on SnO at 14 and 18 qubits, GEVP error 0.000 mHa in all eight runs' should specify what the eight runs are (number of seeds x number of losses); otherwise the claim is hard to audit.
  2. [Table 1 footnote] The phrase 'It is dashed at 14 and 18 qubits' is ambiguous; 'It' refers to the Subspace column, but the antecedent is not clear. Rephrase.
  3. [Figure 4(b)] The caption mentions 'the best of four seeds where run' but the seed protocol for the dissociation curve is not given in the table or Methods. Adding seed counts for the 22q curve would help.
  4. [Section 4, S-CORE paragraph] The statement that S-CORE from a Hartree-Fock seed reaches 1.48 mHa at 32q and 1.25 mHa at 40q is very relevant to the quantum-advantage question. It is presented clearly, but it could be moved earlier in the Discussion so the reader sees the classical baseline before the quantum claims.

Circularity Check

0 steps flagged

No circular derivation: GQE+QSCI energies are variational bounds checked against independent exact CASCI/FCI references; only a minor non-load-bearing self-citation ([33]) keeps this above zero.

full rationale

The paper's central quantity is the GQE+QSCI energy, obtained by sampling determinants from GPT-2-generated UCCSD circuits and diagonalizing the Hamiltonian classically, then GEVP-refined. The target of comparison is an exact CASCI/FCI reference computed independently with PySCF; the text states 'Every reported error is against an exact solution' and the GEVP construction makes each reported energy a variational upper bound. The GPT-2 policy is trained on its own QSCI-energy rewards, but the exact reference is not used to set parameters or post-select results, so this is an optimization loop, not a fit of the target. The paper explicitly separates the solver gap from the active-space gap ('Every claim separates the solver gap ... from the active-space gap'), so benchmarking QSCI against CASCI in the same AVAS space is an acknowledged model boundary, not a hidden definition of the claim. Scaling laws in Table 6 are labelled as fits, not predictions. The only self-citation by the current authors is [33], used to support a side observation about device-noise broadening on the same QPU ('The same broadening appeared ... in a separate QSCI study ... [33]'); it is not load-bearing for the reported energies or the chemical-accuracy claim. The base framework [14] is by different authors. Single-seed/best-seed reporting is a statistical-robustness concern, not circularity. No self-definitional reduction, fitted-input-as-prediction, imported uniqueness theorem, or ansatz-by-citation chain was found. Score 2 reflects the minor non-load-bearing self-citation only.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced; S-CORE, SKQD, and the GPT-2 policy are algorithms, not entities. The paper's central claims rest on the representativeness of the AVAS active-space model, the correctness of exact CASCI/FCI references, and the validity of GEVP variational bounds.

free parameters (5)
  • Circuit depth L (per rung) = 10-800 operators
    Selected per active space (L=80 for SnO 24-32q, L=600 for resist 26-30q, L=10 for 40q). Accuracy depends strongly on L (Figure 3a); it is a computational budget chosen by hand, not derived.
  • Determinant cap = 1e5
    Used in the fixed-budget scaling study; limits the sampled subspace and affects the coverage percentages.
  • MPS bond dimension chi = 64 (tested 64/128/256)
    Used for the 40q sampler; the paper checks chi=64/128/256 returns identical unions, so it is not the binding limit.
  • Training iterations = 200-600
    200 for most rungs, 400-600 for larger rungs; the 30q result improves after 400 iterations and reaches 1.21 mHa at 600. This is a compute budget, not a fitted constant.
  • ENPT2 screening epsilon = 1e-5
    Threshold for the Epstein-Nesbet perturbative tail; that column is not a variational bound.
axioms (6)
  • standard math GEVP over sampled determinant subspaces yields a variational upper bound on the exact ground-state energy.
    Used for every GQE+QSCI energy (Table 1); relies on the subspace diagonalization/GEVP theory from Motta et al. [15].
  • domain assumption The Born-Oppenheimer electronic Hamiltonian with the chosen basis and AVAS-selected active space is a faithful model for the EUV photoresist chemistry.
    All 'chemistry' claims (the Sn-C bond weakening and ionisation switch) depend on the active space capturing the correlation that drives the bond breaking; the paper notes the 3.6 Å point where the descriptor 'is about to dissolve'.
  • standard math PySCF exact CASCI/FCI calculations used as references are correct.
    Used at every rung 14-40q; PySCF is a standard, widely benchmarked package.
  • domain assumption Per-circuit readout self-calibration on the QPU converts measured bitstrings to ideal counts without residual bias.
    Used for the 22q hardware result (+3.92 mHa); the paper validates emulation-vs-hardware at 14q but does not provide a residual-bias analysis for the 22q calibration.
  • domain assumption The neutral dissociation curve is approximated as a rigid stretch with both fragments frozen.
    The paper states this makes the neutral bond energy a lower bound (Table 4 caption), so the bond-energy comparison is partially an artifact of geometry constraint.
  • domain assumption Classical UCCSD(T) values for the full molecule are reliable references for the ionisation and bond energies.
    The 72.6->21.2 kcal/mol numbers are UCCSD(T), not quantum; the n-butyl channel ordering rests on a 0.13 eV margin the paper calls indicative.

pith-pipeline@v1.3.0-alltime-deepseek · 13139 in / 16875 out tokens · 150815 ms · 2026-07-31T23:19:45.995416+00:00 · methodology

0 comments
read the original abstract

We scale a CUDA-Q-native pipeline coupling a generative quantum eigensolver (GQE) to quantum-selected configuration interaction (QSCI) across active spaces of 14 to 44 qubits, applied to the extreme-ultraviolet (EUV) photoresist chemistry of monoalkyltin oxo-hydroxides. A GPT-2 policy emits UCCSD operator sequences; sampled bitstrings become determinants, diagonalised classically, and a cross-circuit generalised-eigenvalue refinement makes every reported GQE+QSCI energy a variational upper bound. Every rung from 14 to 40 qubits carries an exact CASCI or FCI reference, up to 166 million determinants for SnO at 32 qubits. The pipeline is chemically accurate, below 1.6 mHa, through 30 qubits on methyltin trihydroxide and through 32 on SnO, on the best seed at the top rungs. Circuit depth rather than training length is the scaling lever; the refined subspace grows near-linearly with the operator count while staying a vanishing fraction of the determinant space, 0.017% at the 32-qubit SnO rung. It also runs on the 54-qubit IQM Emerald processor, at the shallow depths its routed two-qubit gates allow, reaching +0.330 mHa for SnO at 14 qubits from a CCSD-amplitude-ordered pool prefix and +3.92 mHa for the industrial n-butyltin ligand at 22 qubits from depth-truncated trained circuits under per-circuit readout self-calibration, 81% of the active-space correlation; classical configuration recovery on those counts tightens the 22-qubit result to +0.18 to 0.21 mHa. For the methyl resist, ionisation collapses the classical UCCSD(T) Sn-C bond dissociation energy from 72.6 to 21.2 kcal/mol, the switch that flips solubility on exposure. Against that, the 40-qubit result is support-limited at 22.8 mHa, the full trained ansatz on hardware awaits better fidelities, and classical subspace expansion reaches the 32 to 40-qubit spaces with no quantum sampler, so that boundary is mapped, not beaten.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 9 canonical work pages

  1. [1]

    Kazazis, J

    D. Kazazis, J. Garcia Santaclara, J. van Schoot, I. Mochi, and Y . Ekinci, Extreme ultraviolet lithography, Nat. Rev. Methods Primers, DOI:10.1038/s43586-024-00361-Z 4, 84 (2024)

    work page doi:10.1038/s43586-024-00361-z 2024

  2. [2]

    L. Li, X. Liu, S. Pal, S. Wang, C. K. Ober, and E. P. Giannelis, Extreme ultraviolet resist materials for sub-7 nm patterning, Chem. Soc. Rev., DOI:10.1039/c7cs00080d 46, 4855 (2017)

    work page doi:10.1039/c7cs00080d 2017

  3. [3]

    Manouras and P

    T. Manouras and P. Argitis, High Sensitivity Resists for EUV Lithography: A Review of Material Design Strategies and Performance Results, Nanomaterials, DOI:10.3390/nano10081593 10, 1593 (2020)

    work page doi:10.3390/nano10081593 2020

  4. [4]

    J. H. Ma and others, Mechanistic Advantages of Organotin Molecular EUV Photoresists, ACS Appl. Mater. Interfaces, DOI:10.1021/acsami.1c12411 14, 5514 (2022)

    work page doi:10.1021/acsami.1c12411 2022

  5. [5]

    Du and others, Tin-oxo nanoclusters for extreme ultraviolet photoresists: Effects of ligands, counterions, and doping, J

    T. Du and others, Tin-oxo nanoclusters for extreme ultraviolet photoresists: Effects of ligands, counterions, and doping, J. Chem. Phys., DOI:10.1063/5.0200630 160, 154307 (2024)

    work page doi:10.1063/5.0200630 2024

  6. [6]

    J. Li, Z. Wang, and H. Wang, Computational Study of Organotin Oxide Systems for Extreme Ultraviolet Photore- sist, J. Phys. Chem. A, DOI:10.1021/acs.jpca.4c07585 129, 1420 (2025)

    work page doi:10.1021/acs.jpca.4c07585 2025

  7. [7]

    A. J. Cohen, P. Mori-Sánchez, and W. Yang, Challenges for Density Functional Theory, Chem. Rev., DOI:10.1021/ cr200107z 112, 289 (2012)

    2012

  8. [8]

    McArdle and others, Quantum computational chemistry, Rev

    S. McArdle and others, Quantum computational chemistry, Rev. Mod. Phys. 92, 015003, DOI:10.1103/revmod- phys.92.015003 (2020)

    doi:10.1103/revmod- 2020

  9. [9]

    T. D. Kharazi and others, Quantum Simulations for Extreme Ultraviolet Photolithography, arXiv:2602.20234 (2026)

    arXiv 2026

  10. [10]

    R. M. Cigala, C. De Stefano, A. Giacalone, A. Gianguzza, and S. Sammartano, Hydrolysis of Monomethyl-, Dimethyl-, and Trimethyltin(IV) Cations in Fairly Concentrated Aqueous Solutions at I = 1 mol L-1 (NaNO3) and T = 298.15 K. Evidence for the Predominance of Polynuclear Species, J. Chem. Eng. Data, DOI:10.1021/ je101069y 56, 1108 (2011)

    2011

  11. [11]

    R. T. Frederick, S. Saha, J. T. Diulus, F. Luo, J. M. Amador, M. Li, D.-H. Park, E. L. Garfunkel, D. A. Keszler, and G. S. Herman, Thermal and radiation chemistry of butyltin oxo hydroxo: A model inorganic photoresist, Microelectron. Eng., DOI:10.1016/j.mee.2018.11.011 205, 26 (2019)

    work page doi:10.1016/j.mee.2018.11.011 2018

  12. [12]

    Cardineau, R

    B. Cardineau, R. Del Re, M. Marnell, H. Al-Mashat, M. V ockenhuber, Y . Ekinci, C. Sarma, D. A. Freedman, and R. L. Brainard, Photolithographic properties of tin-oxo clusters using extreme ultraviolet light (13.5 nm), Microelectron. Eng., DOI:10.1016/j.mee.2014.04.024 127, 44 (2014)

    work page doi:10.1016/j.mee.2014.04.024 2014

  13. [13]

    NIST, Tin Monoxide, https://webbook.nist.gov/cgi/cbook.cgi?ID=C21651194&Mask=1000

  14. [14]

    Kemmoku, Q

    R. Kemmoku, Q. Gao, S. Kanno, K. Keithley, I. Hamamura, N. Yamamoto, and K. Nakaji, Generative Circuit Design for Quantum-Selected Configuration Interaction, arXiv:2604.09756 (2026)

    Pith/arXiv arXiv 2026

  15. [15]

    Motta, W

    M. Motta, W. Kirby, I. Liepuoniute, K. J. Sung, J. Cohn, A. Mezzacapo, K. Klymko, N. Nguyen, N. Yoshioka, and J. E. Rice, Subspace methods for electronic structure simulations on quantum computers, Electron. Struct. 6, 013001, DOI:10.1088/2516-1075/ad3592 (2024)

    doi:10.1088/2516-1075/ad3592 2024

  16. [16]

    Robledo-Moreno, M

    J. Robledo-Moreno, M. Motta, and others, Chemistry Beyond the Scale of Exact Diagonalization on a Quantum- Centric Supercomputer, arXiv:2405.05068 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Vaquero-Sabater, A

    N. Vaquero-Sabater, A. Carreras, L. Broers, T. Shirakawa, S. Yunoki, and D. Casanova, Noise and Configuration Recovery Impact on Quantum Selected Configuration Interaction, arXiv:2605.23697 (2026)

    Pith/arXiv arXiv 2026

  18. [18]

    A. A. Holmes, N. M. Tubman, and C. J. Umrigar, Heat-Bath Configuration Interaction: An Efficient Selected Configuration Interaction Algorithm Inspired by Heat-Bath Sampling, J. Chem. Theory Comput., DOI:10.1021/ acs.jctc.6b00407 12, 3674 (2016)

    2016

  19. [19]

    Nakaji and others, The Generative Quantum Eigensolver (GQE) and its Application for Ground State Search, arXiv:2401.09253 (2024)

    K. Nakaji and others, The Generative Quantum Eigensolver (GQE) and its Application for Ground State Search, arXiv:2401.09253 (2024)

    arXiv 2024

  20. [20]

    Alexeev and others, Artificial intelligence for quantum computing, Nat

    Y . Alexeev and others, Artificial intelligence for quantum computing, Nat. Commun., DOI:10.1038/ s41467-025-65836-3 16, 10829 (2025)

    2025

  21. [21]

    Kanno and others, Quantum-Selected Configuration Interaction: Classical Diagonalization of Hamiltonians in Subspaces Selected by Quantum Computers, arXiv:2302.11320 (2023)

    K. Kanno and others, Quantum-Selected Configuration Interaction: Classical Diagonalization of Hamiltonians in Subspaces Selected by Quantum Computers, arXiv:2302.11320 (2023)

    Pith/arXiv arXiv 2023

  22. [22]

    NVIDIA CUDA-Q development team, CUDA-Q: The Platform for Hybrid Quantum-Classical Computing, Version 0.14, https://nvidia.github.io/cuda-quantum

  23. [23]

    Radford and others, Language Models Are Unsupervised Multitask Learners , https://cdn.openai.com/better- language-models/language_models_are_unsupervised_multitask_learners.pdf

    A. Radford and others, Language Models Are Unsupervised Multitask Learners , https://cdn.openai.com/better- language-models/language_models_are_unsupervised_multitask_learners.pdf

  24. [24]

    G. D. Purvis and R. J. Bartlett, A full coupled-cluster singles and doubles model: The inclusion of disconnected triples, J. Chem. Phys., DOI:10.1063/1.443164 76, 1910 (1982)

    doi:10.1063/1.443164 1910

  25. [25]

    Richer and others, PyCI: A Python-scriptable library for arbitrary determinant CI, J

    M. Richer and others, PyCI: A Python-scriptable library for arbitrary determinant CI, J. Chem. Phys., DOI:10.1063/5.0219010 161, 132502 (2024)

    doi:10.1063/5.0219010 2024

  26. [26]

    Shao and others, DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models, arXiv:2402.03300 (2024)

    Z. Shao and others, DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models, arXiv:2402.03300 (2024). 11

    Pith/arXiv arXiv 2024

  27. [27]

    C. C. J. Roothaan, Self-Consistent Field Theory for Open Shells of Electronic Systems, Rev. Mod. Phys., DOI:10.1103/revmodphys.32.179 32, 179 (1960)

    doi:10.1103/revmodphys.32.179 1960

  28. [28]

    J. R. McClean and others, OpenFermion: the electronic structure package for quantum computers, Quantum Sci. Technol. 5, 034014, DOI:10.1088/2058-9565/ab8ebc (2020)

    doi:10.1088/2058-9565/ab8ebc 2058

  29. [29]

    NIST, Hydroxyl Radical, https://webbook.nist.gov/cgi/cbook.cgi?ID=C3352576&Mask=1000

  30. [30]

    E. R. Sayfutyarova and others, Automated Construction of Molecular Active Spaces from Atomic Valence Orbitals, J. Chem. Theory Comput., DOI:10.1021/acs.jctc.7b00128 13, 4063 (2017)

    doi:10.1021/acs.jctc.7b00128 2017

  31. [31]

    Sun and others, Recent developments in the PySCF program package, J

    Q. Sun and others, Recent developments in the PySCF program package, J. Chem. Phys. 153, 024109, DOI:10.1063/5.0006074 (2020)

    doi:10.1063/5.0006074 2020

  32. [32]

    NIST, Methyl Radical, Gas Phase Ion Energetics, https://webbook.nist.gov/cgi/cbook.cgi?ID=C2229074&Mask= 20

  33. [33]

    Elgammal and M

    K. Elgammal and M. Maußner, Additive binding energies in asphalt on a quantum processor via quantum-selected configuration interaction (QSCI), arXiv:2605.27640 (2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Motta and others, Bridging physical intuition and hardware efficiency for correlated electronic states: the local unitary cluster Jastrow ansatz for electronic structure, Chem

    M. Motta and others, Bridging physical intuition and hardware efficiency for correlated electronic states: the local unitary cluster Jastrow ansatz for electronic structure, Chem. Sci., DOI:10.1039/d3sc02516k 14, 11213 (2023)

    doi:10.1039/d3sc02516k 2023

  35. [35]

    Yu et al., Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization, arXiv:2501.09702 (2025)

    J. Yu et al., Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization, arXiv:2501.09702 (2025)

    arXiv 2025

  36. [36]

    Peruzzo and others, A variational eigenvalue solver on a photonic quantum processor, Nat

    A. Peruzzo and others, A variational eigenvalue solver on a photonic quantum processor, Nat. Commun., DOI:10.1038/ncomms5213 5, 4213 (2014)

    doi:10.1038/ncomms5213 2014

  37. [37]

    Tilly and others, The Variational Quantum Eigensolver: A review of methods and best practices, Phys

    J. Tilly and others, The Variational Quantum Eigensolver: A review of methods and best practices, Phys. Rep., DOI:10.1016/j.physrep.2022.08.003 986, 1 (2022)

    doi:10.1016/j.physrep.2022.08.003 2022

  38. [38]

    H. R. Grimsley and others, An adaptive variational algorithm for exact molecular simulations on a quantum computer, Nat. Commun., DOI:10.1038/s41467-019-10988-2 10, 3007 (2019)

    doi:10.1038/s41467-019-10988-2 2019

  39. [39]

    Keithley and others, Auger Spectroscopy via Generative Quantum Eigensolver: A Quantum Approach to Molecular Excitations, arXiv:2603.12859 (2026)

    K. Keithley and others, Auger Spectroscopy via Generative Quantum Eigensolver: A Quantum Approach to Molecular Excitations, arXiv:2603.12859 (2026)

    arXiv 2026

  40. [40]

    J. R. McClean and others, Barren plateaus in quantum neural network training landscapes, Nat. Commun., DOI:10.1038/s41467-018-07090-4 9, 4812 (2018)

    doi:10.1038/s41467-018-07090-4 2018

  41. [41]

    K. S. V . Anurag, A. K. Patra, M. Mukherjee, and others, Bridging the NISQ and Fault-Tolerant Regimes: Gener- ative-ML-Assisted Quantum Selected CI for Molecular Simulations, arXiv:2606.30551 (2026)

    Pith/arXiv arXiv 2026

  42. [42]

    Rao et al., A Performance Model for Hybrid Quantum-Classical Workflows, arXiv:2607.15426 (2026)

    P. Rao et al., A Performance Model for Hybrid Quantum-Classical Workflows, arXiv:2607.15426 (2026)

    Pith/arXiv arXiv 2026

  43. [43]

    Shirakawa, J

    T. Shirakawa, J. Robledo-Moreno, T. Itoko, V . Tripathi, K. Ueda, Y . Kawashima, and others, Closed-loop calcula- tions of electronic structure on a quantum processor and a classical supercomputer at full scale, arXiv:2511.00224 (2025)

    arXiv 2025

  44. [44]

    Cai and others, Quantum error mitigation, Rev

    Z. Cai and others, Quantum error mitigation, Rev. Mod. Phys. 95, 045005, DOI:10.1103/revmodphys.95.045005 (2023)

    doi:10.1103/revmodphys.95.045005 2023

  45. [45]

    LaRose and others, Mitiq: A software package for error mitigation on noisy quantum computers, Quantum, DOI:10.22331/q-2022-08-11-774 6, 774 (2022)

    R. LaRose and others, Mitiq: A software package for error mitigation on noisy quantum computers, Quantum, DOI:10.22331/q-2022-08-11-774 6, 774 (2022)

    doi:10.22331/q-2022-08-11-774 2022

  46. [46]

    Ramakrishnan and others, Big Data Meets Quantum Chemistry Approximations: The Δ-Machine Learning Approach, J

    R. Ramakrishnan and others, Big Data Meets Quantum Chemistry Approximations: The Δ-Machine Learning Approach, J. Chem. Theory Comput., DOI:10.1021/acs.jctc.5b00099 11, 2087 (2015)

    doi:10.1021/acs.jctc.5b00099 2087

  47. [47]

    D. A. Kreplin and others, sQUlearn: A Python Library for Quantum Machine Learning, IEEE Software, DOI:10.1109/ms.2025.3527736 42, 65 (2025)

    arXiv 2025

  48. [48]

    S. Minami and others, Generative Quantum Combinatorial Optimization by means of a novel Conditional Gener- ative Quantum Eigensolver, Digital Discovery, DOI:10.1039/d5dd00138b 4, 2229 (2025). 12

    work page doi:10.1039/d5dd00138b 2025