Pith. sign in

REVIEW 4 major objections 5 minor 72 references

By learning the laser-driven population dynamics of hydronium with a Fourier neural operator, this paper designs pulse sequences that purify a 20 K thermal mixture to 0.98 target population with up to 86.2% success, demonstrating operator-l

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 →

A Fourier Neural Operator surrogate predicts H3O+ population dynamics, and its stochastic planner designs pulses that reach 0.98 target purity with up to 86.2% success.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Well-built FNO surrogate-control pipeline, but the headline success rate rests on an unvalidated assumption that motional measurement erases molecular coherences; the paper needs a density-matrix check before the numbers can be trusted. the 4 major comments →

arxiv 2608.03702 v1 pith:YBYDEAO4 submitted 2026-08-04 quant-ph

Inverse Design of Quantum Control Sequences with Fourier Neural Operators

classification quant-ph MSC 81Q9368T07 PACS 03.65.-w07.05.Mh
keywords quantum optimal controlFourier neural operatoroperator learninginverse designstate purificationmolecular ionhydroniumquantum logic spectroscopy
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.

The reading

The paper claims that a Fourier neural operator can learn the laser-driven population dynamics of hydronium accurately enough to replace direct quantum simulation during control design, and that this lets a planner find state-preparation pulse sequences in minutes rather than hours. Its stochastic planner, FNO-SPMP, builds pulse sequences that purify a 20 K Boltzmann mixture of H3O+ into a single molecular state, reaching a target population of 0.98 in up to 86.2% of Monte Carlo runs. The same surrogate is differentiable, so the continuous pulse parameters can be further polished by gradient descent. The paper's broader claim is that operator-learning surrogates make inverse design tractable in Hilbert spaces too large for repeated direct optimization.

Core claim

The central discovery is that a control-conditioned population propagator can be learned for a large molecular subspace, then used for inverse design. In an 888-dimensional hydronium subspace, a Fourier neural operator predicts basis-state population trajectories over a single 4 ms pulse in one forward pass, trained with a physics-informed detuning embedding and an activity-weighted loss. The trained surrogate is applied to both Raman polarization channels through a frequency-reflection symmetry, and the block-diagonal Hamiltonian structure means only six unique FNOs need be trained. The stochastic pulse-measurement planner then scores candidate pulses by selective transfer, branch purity, a

What carries the argument

The FNO-based population propagator: for each Hamiltonian block, a Fourier neural operator—a neural network that learns maps between functions by transforming in frequency space—maps the current population vector plus a physics-informed embedding of detuning and drive frequency to full population trajectories in both motional manifolds over the 0-4 ms pulse window. It carries the argument because it turns expensive repeated Schrödinger propagation into a fast, differentiable batched evaluation that the planner can score, rank, sample, and gradient-refine.

Load-bearing premise

The pipeline rests on the H3O+ hyperfine energies and Raman coupling rates being correct and complete—the paper defers their source to an article in preparation—and on motional measurement plus recooling removing coherences exactly, so the population-only state update is valid.

What would settle it

Take the best FNO-SPMP sequence and rerun the Monte Carlo validation with an independently computed H3O+ Hamiltonian parameter set; if the 0.98 target population and roughly 86% success rate do not survive, the central claim collapses. A cheaper in-simulation test is to query the surrogate at a drive frequency near a transition excluded by the detuning or coupling cutoff and check predictions against direct propagation.

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

If this is right

  • For any molecule whose exact dynamics can be simulated offline, pulse-sequence design becomes a fast surrogate-guided search; the 888-dimensional hydronium example is the demonstration.
  • Continuous pulse parameters can be refined by gradient descent through the surrogate, giving a route to high-precision calibration that discrete reinforcement-learning policies do not naturally offer.
  • On a shared discrete action space, FNO-SPMP reaches 79.4% convergence with a mean of 27.9 pulses versus the RL baseline's 42.8% and 49.0 pulses, indicating that large control spaces favor surrogates over learned policies.
  • Surrogate errors stay low across the pulse window, with median population infidelity near 10^-4, so multi-pulse decisions built from single-pulse predictions do not compound appreciably.
  • Because the surrogate is trained on one polarization and reused for the other through the reflection symmetry, control libraries for both sigma+ and sigma- channels cost no additional training.

Where Pith is reading between the lines

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

  • If the Hamiltonian parameters are correct, the real bottleneck shifts to offline data generation: each new molecule or control configuration requires exact simulation to build training sets, so the advertised speedup is an amortized one, not a free lunch.
  • The planner's choice to follow the nu=0 measurement branch is a heuristic; a balanced branch-aware objective could plausibly improve worst-case behavior when excited-motional outcomes dominate.
  • The 20 K initial temperature absorbs blackbody-heating uncertainty; if real cryogenic experiments run colder, the thermally occupied subspace is smaller, potentially making experimental state preparation easier than the simulation's worst case.
  • The same surrogate-plus-planner recipe could be benchmarked on a small driven quantum system against classical optimal control to identify the Hilbert-space dimension at which operator-learning surrogates become the cheaper route.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. This paper presents a Fourier Neural Operator (FNO) surrogate for the single-pulse, laser-driven population dynamics of a hyperfine-resolved H3O+ molecular ion coupled to a shared motional mode, and an inverse-design protocol (FNO-SPMP) that uses the surrogate to construct pulse sequences for purifying a 20 K Boltzmann distribution. The surrogate is trained on CUDA-Q propagator data with physics-informed detuning embeddings and an activity-weighted loss; validation on held-out frequencies and random initial states gives median population infidelities near 1e-4 and speedups up to 1.84e7 in batched forward evaluation. FNO-SPMP selects pulses by scoring, stochastic active-pool sampling, and optional gradient refinement, and its sequences are validated by direct Monte Carlo rollouts. The authors report target-state population 0.98 with up to 86.2% sequence success, about 30-60x faster sequence generation than an RL baseline, and a roughly 2x higher success rate in a shared discrete action space.

Significance. If the underlying physical assumptions hold, this is a credible demonstration that an operator-learning surrogate can replace many expensive propagations in a quantum-control search over a large molecular Hilbert space. Strengths include held-out frequency/initial-state testing, direct CUDA-Q Monte Carlo validation (so the reported success rates do not merely inherit surrogate errors), and a block-diagonal Hamiltonian decomposition that is verified against full propagation. The main caveats are the unproven population-only state update after motional measurement and the absence of the molecular Hamiltonian parameters; both affect the specific hydronium numbers rather than the general architecture. The paper is likely to be of interest to the quantum-control and machine-learning-for-quantum communities, but the application-level claims need to be made more robust.

major comments (4)
  1. [Sec. II.B, Eq. (9); Sec. II.D.5] The population-only update after motional measurement is not generally exact. Projection onto a motional manifold retains coherences between molecular states in that manifold whenever a single initial state has amplitude into two molecular states (or two pathways meet). The FNO input Eq. (16) and the MC rollouts in Sec. II.D.5 use only diagonal populations, so neither planning nor validation can reveal errors from such coherences. The paper asserts that measurement and recooling remove coherences but gives no derivation or check. Please justify this (e.g., selection rules that make final states orthogonal per initial state, plus a demonstrated decoherence mechanism) or extend the surrogate/validation to conditional density-matrix updates. A direct comparison of Eq. (9) rollouts with P_nu rho P_nu / Tr(P_nu rho) rollouts for the best sequences would settle whether the 86.2% success rate s
  2. [Appendix A1] All numerical results depend on the hyperfine-resolved energies and Raman Rabi couplings of H3O+, but these are not reported; the text says the computational details appear in a subsequent article [48], which is marked 'In preparation'. The 888-dimensional demonstration and the 86.2% success rate therefore cannot be reproduced or independently checked. Please include the level list and coupling matrix (or a stable data file) or make [48] available with the parameters referenced explicitly.
  3. [Sec. III.C.c and Fig. 5a] The RL comparison is ambiguous. The text reports 'Within this shared grid, FNO-SPMP reaches 79.4%' but earlier says each grid includes the result after local gradient refinement. If 79.4% is the gradient-refined value, the actions are no longer on the shared discrete grid; if it is the unrefined value, say so explicitly and report the refined value separately. The 'nearly twice the success rate' claim must be based on the same action space.
  4. [Sec. II.D.5] The Monte Carlo validation is 'on the truncated decision tree constructed by the planner.' Please specify what a rollout does when a sampled history reaches a leaf of that tree (e.g., more than three consecutive nu=1 outcomes). Are such rollouts counted as failures? If new controls are generated on the fly, the validation is no longer of the designed sequence. Without this, the 86.2% success rate is not fully interpretable.
minor comments (5)
  1. [Sec. III.A] The conclusion states population infidelity is 'consistently below 3e-3', but Fig. 3c shows a few isolated test frequencies with larger errors. The text itself says 'most test frequencies'; please align the conclusion with the data.
  2. [Eq. (31)] Typo: T_CUDA- should be T_CUDA-Q.
  3. [Ref. [9]] 'Astropysical' should be 'Astrophysical'.
  4. [Sec. II.B] The protocol restricts the motional basis to nu=0,1. Please quantify (or justify via Lamb-Dicke suppression) that higher motional manifolds do not acquire non-negligible population over a 4 ms pulse for the strongest Rabi couplings used.
  5. [Sec. II.D.2] The score components S_tr and S_br are described in words but never defined by equations. Since the planner's behavior depends on them, explicit definitions would improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the FNO is an explicitly fitted surrogate, and the reported inverse-design success metrics are anchored by independent direct-propagation Monte Carlo validation.

full rationale

The paper's derivation chain is not circular. The FNO is explicitly a fitted surrogate: its reference trajectories are computed with CUDA-Q (Sec. II.C.c, Eq. 19), and its accuracy is tested on held-out frequencies and random mixed initial states (Sec. III.A). The inverse-design step uses the FNO to score and refine candidate pulses, but the reported success rates are obtained by 'direct numerical propagation of the selected pulse sequences over 1000 Monte Carlo simulation runs, ensuring that the validation is independent of accumulated FNO prediction errors' (Sec. III.C.b). Thus the central 86.2% success-rate claim does not reduce to the surrogate's own training output. The population-only update in Eq. 9 is stated as an assumption ('We assume that the measurement and subsequent motional cooling remove coherences'), not derived from the conclusion; this is a physical-model limitation that both the surrogate and the validation share, but it is not a definitional equivalence between inputs and outputs. Self-citations to Refs. [29], [38], and [41] are contextual or benchmark references and are not load-bearing for the main results. The deferred H3O+ Hamiltonian parameters (Appendix A1, Ref. [48]) are an external-input correctness risk, not a circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The central claim rests on the fidelity of the CUDA-Q simulator and the correctness of the molecular Hamiltonian, on the assumed validity of population-only state updates after measurement and recooling, on the two-motional-manifold truncation, and on the exactness of the sigma+/sigma- symmetry. The planner also introduces several hand-chosen parameters (cutoffs, embedding scales, score weights), none of which are optimized or justified by sensitivity analysis. No new physical entities are invented.

free parameters (9)
  • FNO network weights = not enumerated
    Learned from CUDA-Q simulated training data; the surrogate accuracy claim depends on this fit.
  • Detuning cutoff Delta_max/2pi = 10^4/(2pi) kHz per text
    Hand-chosen cutoff in Eq. 11 discards far-detuned transitions.
  • Coupling cutoff Omega_min/2pi = 1/(2pi) kHz
    Hand-chosen cutoff in Eq. 11 removes weak couplings.
  • Embedding scale s_emb = 0.05
    Hand-chosen in Eq. 13 to match the population-input scale.
  • Detuning suppression beta = 0.01
    Hand-chosen parameter in Eq. 13.
  • Activity weighting lambda = not stated
    Eq. 20 controls activity-weighted loss; the value is not given in the main text.
  • Score weights w_tr, w_br = 0.5, 1.5
    Fixed heuristically in Eq. 25 and not optimized.
  • Initial molecular temperature = 20 K
    Choice motivated by blackbody heating estimates; affects the initial Boltzmann distribution.
  • Target purity threshold P_target = 0.98
    Sets the success condition in Eq. 10.
axioms (6)
  • domain assumption The H3O+ hyperfine energies and Raman Rabi couplings used to generate training data are correct
    Appendix A1 says computational details are reported in a subsequent article [48]; without these parameters the central demonstration cannot be reproduced.
  • domain assumption Population-only state update after motional measurement and recooling is valid
    Eq. 9 assumes measurement and cooling remove coherences; no error model is given for this assumption.
  • domain assumption Truncation to motional manifolds nu=0,1 is sufficient
    Section II.B says higher excited motional states are omitted; sideband drive may excite nu=1 only but higher states are assumed negligible.
  • domain assumption sigma- dynamics follow from sigma+ by frequency reflection and state permutation
    Sec. II and Eq. 3; symmetry is assumed exact and used to avoid training a second surrogate.
  • domain assumption CUDA-Q numerical propagation is ground truth for training and validation
    All data and final MC validation use CUDA-Q; the surrogate cannot correct simulator errors.
  • standard math FNO universal approximation and generalization to unseen frequencies and initial states
    Relies on neural operator approximation results [44] and held-out frequency/state testing in Sec. III.A.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Inverse Design of Quantum Control Sequences with Fourier Neural Operators." pith.science (2026). https://pith.science/paper/YBYDEAO4

@misc{pith2026260803702,
  author       = {Pith},
  title        = {Pith review of: Inverse Design of Quantum Control Sequences with Fourier Neural Operators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YBYDEAO4}},
  note         = {Machine review of arXiv:2608.03702}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Quantum optimal control is a key tool for steering quantum dynamics, but its computational cost grows rapidly with the Hilbert space dimension. Here, we introduce a Fourier Neural Operator (FNO)-based framework for learning high dimensional molecular quantum dynamics and accelerating the inverse design of control protocols. Given an initial molecular population distribution, laser frequency, and polarization, the FNO predicts molecular-motional population dynamics up to $10^7$ times faster than GPU-accelerated numerical propagation with CUDA-Q Dynamics. Using this fast and differentiable surrogate, we develop the FNO stochastic pulse-measurement planner (FNO-SPMP), which constructs pulse sequences to purify an initially mixed Boltzmann distribution. We demonstrate the protocol in an 888-dimensional subspace of the hydronium molecule at 20 K, achieving a target-state population of 0.98 with a sequence success rate of up to 86.2%. In a shared discrete control space, FNO-SPMP achieves nearly twice the success rate of a reinforcement-learning baseline while using roughly half as many quantum control pulses and reducing pulse-sequence generation time from approximately 10 hours to 10-20 minutes. These results show that operator-learning surrogates can enable inverse design in quantum systems whose Hilbert spaces are too large for conventional direct optimization.

Figures

Figures reproduced from arXiv: 2608.03702 by Anastasia Pipi, Anima Anandkumar, Emily Been, Prineha Narang, Taylor L. Patti, Valentin Duruisseaux, Xuecheng Tao.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Architecture of the Fourier Neural Operator (FNO), [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

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

72 extracted references · 39 canonical work pages

  1. [1]

    Evaluation on a discrete action grid.We first de- fine a fixed discrete action grid over frequency and pulse duration, use the FNO to propagate the dynamics for each candidate control on this grid, so that they can later be scored. Frequencies are sampled within the FNO training window using a predefined spacing∆ω, ωa =ω min +a∆ω, ω a ≤ω max,(23) whereω m...

  2. [2]

    We write the score as S(α, sn) =w trStr(α, sn) +w brSbr(α, sn) +w 1(sn)Π1(α, sn)

    Scoring and ranking of candidate controls.Each candidate control,α, is scored using three physically mo- tivated criteria: selective transfer, branch purity, and excited-branch population. We write the score as S(α, sn) =w trStr(α, sn) +w brSbr(α, sn) +w 1(sn)Π1(α, sn). (25) HereS tr rewards selective transfer of the most populated molecular states into t...

  3. [3]

    The ac- tive pool contains the topNpool = 16controls, together with any additional controls whose scores lie withinδS= 0.003of the best score

    Stochastic branch-aware pulse planner.The can- didates are ranked by decreasing scoreS(α, sn). The ac- tive pool contains the topNpool = 16controls, together with any additional controls whose scores lie withinδS= 0.003of the best score. One pulse is then sampled uni- formly from this active pool. Thus, high-scoring controls are favored through the constr...

  4. [4]

    (Optional) Gradient-based optimization refinement. After these sequences are generated, the continuous pulse parameters selected can be optionally refined further us- ing gradient-based optimization, leveraging the differen- tiability of the FNO, as shown in the rightmost box of Fig. 2. For a selected pulseαn = (ωn, τn, σn), the polar- izationσ n is kept ...

  5. [5]

    ( ( !( ' ) '! )!%$ '! %$+ ' + ' () () ' !$ # $) () % $' ! * ' #

    Monte Carlo validation of designed sequences.We evaluate each generated pulse sequence using Monte Carlo rollouts on the truncated decision tree constructed by the planner. Each rollout emulates a single experi- mental run. At every pulse, the motional measurement outcome is sampled according to the population in each branch,Π 0,n andΠ 1,n (Eq. 8). The st...

  6. [6]

    Koch, Ugo Boscain, Tommaso Calarco, Gunther Dirr, Stefan Filipp, Steffen J

    Christiane P. Koch, Ugo Boscain, Tommaso Calarco, Gunther Dirr, Stefan Filipp, Steffen J. Glaser, Ron- nie Kosloff, Simone Montangero, Thomas Schulte- Herbrüggen, Dominique Sugny, and Frank K. Wil- helm. Quantum optimal control in quantum technolo- gies. strategic report on current status, visions and goals for research in europe.EPJ Quantum Technology, 9...

  7. [7]

    Shai Machnes, Elie Assémat, David Tannor, and Frank K. Wilhelm. Tunable, flexible, and ef- ficient optimization of control pulses for practical qubits.Phys. Rev. Lett., 120(15):150401, 2018. doi: 10.1103/PhysRevLett.120.150401

  8. [8]

    Motzoi, J

    F. Motzoi, J. M. Gambetta, P. Rebentrost, and F. K. Wilhelm. Simple pulses for elimination of leakage in weakly nonlinear qubits.Phys. Rev. Lett., 103(11): 110501, 2009. doi:10.1103/PhysRevLett.103.110501

  9. [9]

    T. Choi, S. Debnath, T. A. Manning, C. Figgatt, Z.- X. Gong, L.-M. Duan, and C. Monroe. Optimal quan- tum control of multimode couplings between trapped ion qubits for scalable entanglement.Phys. Rev. Lett., 112 (19):190502, 2014. doi:10.1103/PhysRevLett.112.190502

  10. [10]

    Oshnik, Nimba Müller, Si- mone Montangero, Tommaso Calarco, and Elke Neu

    Phila Rembold, Matthias M. Oshnik, Nimba Müller, Si- mone Montangero, Tommaso Calarco, and Elke Neu. Introduction to quantum optimal control for quan- tum sensing with nitrogen-vacancy centers in dia- mond.A VS Quantum Sci., 2(2):2639–0213, 2020. doi: https://doi.org/10.1116/5.0006785

  11. [11]

    Safronova, D

    M.S. Safronova, D. Budker, D. DeMille, Derek F. Jack- son Kimball, A. Derevianko, and Charles W. Clark. Search for new physics with atoms and molecules.Re- views of Modern Physics, 90(2):025008, 2018. doi: 10.1103/RevModPhys.90.025008

  12. [12]

    Quantum sensing and metrology for fundamental physics with molecules.Nature Physics, 20: 741–749, 2024

    David DeMille, Nicholas R Hutzler, Ana Maria Rey, and Tanya Zelevinsky. Quantum sensing and metrology for fundamental physics with molecules.Nature Physics, 20: 741–749, 2024. doi:10.1038/s41567-024-02499-9

  13. [13]

    Perspectives on parity violation in chiral molecules: the- ory, spectroscopicexperimentandbiomolecularhomochi- rality.Chemical Science, 13:10598–10643, 2022

    Martin Quack, Georg Seyfanga, and Gunther Wichmann. Perspectives on parity violation in chiral molecules: the- ory, spectroscopicexperimentandbiomolecularhomochi- rality.Chemical Science, 13:10598–10643, 2022. doi: https://doi.org/10.1039/D2SC01323A

  14. [14]

    Kozlov M. G. and Levshakov S. A. Sensitivity of the h3o+ inversion–rotational spectrum to changes in the electron-to-proton mass ratio.The Astropysical Journal, 726(2):65, 2011. doi:10.1088/0004-637X/726/2/65

  15. [15]

    Weiss and et al

    Leah R. Weiss and et al. A high-resolution molecular spin-photon interface at telecommunication wavelengths. Science, 390:76–81, 2025. doi:10.1126/science.ady8677

  16. [16]

    G. Lao, T. Khvorost, A. Macias, and et al. Bottom-up approach to making larger hydrocarbon molecules capa- ble of optical cycling.Nat. Chem., 18:84–91, 2026. doi: https://doi.org/10.1038/s41557-025-01965-y

  17. [17]

    P. O. Schmidt, T. Rosenband, C. Langer, W. M. Itano, J. C. Bergquist, and D. J. Wineland Wineland. Spec- troscopy using quantum logic.Science, 309:749–752,

  18. [18]

    Navin Khaneja, Timo Reiss, Cindie Kehlet, Thomas Schulte-Herbrüggen, and Steffen J. Glaser. Op- timal control of coupled spin dynamics: design of nmr pulse sequences by gradient ascent algo- rithms.J. Magn. Reson., 172(2):296–305, 2005. doi: https://doi.org/10.1016/j.jmr.2004.11.004

  19. [19]

    Leibfried

    D. Leibfried. Quantum state preparation and control of single molecular ions.New J. Phys., 14(023029), 2012. doi:10.1088/1367-2630/14/2/023029

  20. [20]

    Ding and D

    S. Ding and D. N. Matsukevich. Quantum logic for the control and manipulation of molecular ions using a fre- quency comb.New J. Phys., 14(023028), 2012. doi: 10.1088/1367-2630/14/2/023028

  21. [21]

    Cw. Chou, C. Kurz, and D. et al. Hume. Preparation and coherent manipulation of pure quantum states of a single molecular ion.Nature, 545:203–207, 2017. doi: https://doi.org/10.1038/nature22338

  22. [22]

    C. W. Chou, A. L. Collopy, C. Kurz, P. N. Plessow, T. Fortier, S. Diddams Diddams, D. Leibfried, and D. R. Leibrandt. Frequency-comb spectroscopy on pure quan- tum states of a single molecular ion.Science, 367:1458– 1461, 2020. doi:10.1126/science.aba3628

  23. [23]

    F. Wolf, Y. Wan, and J. et al. Heip. Non-destructive state detection for quantum logic spectroscopy of molecular ions.Nature, 530:457–460, 2016. doi: https://doi.org/10.1038/nature16513

  24. [24]

    Ohtsuki, G

    Y. Ohtsuki, G. , Turinici, and H. Rabitz. Generalized monotonically convergent algorithms for solving quan- tum optimal control problems.J. Chem. Phys., 120: 5509–5517, 2004. doi:https://doi.org/10.1063/1.1650297

  25. [25]

    de Fouquières, S

    P. de Fouquières, S. G. Schirmer, S. J. Glaser, and I. Kuprov. Second order gradient ascent pulse en- gineering.J. Magn. Reson., 212(412), 2011. doi: https://doi.org/10.1016/j.jmr.2011.07.023

  26. [26]

    M. H. Goerz, S. C. Carrasco, and V. S. Mali- novsky. Quantum optimal control via semi-automatic differentiation.Quantum, 6:871, 2022. doi: https://doi.org/10.22331/q-2022-12-07-871

  27. [27]

    Machnes, U

    S. Machnes, U. Sander, S. J. Glaser, P. de Fouquières, A. Gruslys, S. Schirmer, and T. Schulte-Herbrüggen. Comparing, optimizing, and benchmarking quantum- control algorithms in a unifying programming frame- work.Phys. Rev. A, 84(2):022305, 2011. doi: 10.1103/PhysRevA.84.022305

  28. [28]

    Chopped random-basis quantum opti- mization.Phys

    Tommaso Caneva, Tommaso Calarco, and Simone Montangero. Chopped random-basis quantum opti- mization.Phys. Rev. A, 84(2):022326, 2011. doi: 10.1103/PhysRevA.84.022326

  29. [29]

    Abdelhafez, D

    M. Abdelhafez, D. I. Schuster, and J. Koch. Gradient- based optimal control of open quantum systems us- ing quantum trajectories and automatic differenti- ation.Phys. Rev. A, 99(052327), 2019. doi: https://doi.org/10.1103/PhysRevA.99.052327

  30. [30]

    Zhang, Z

    X.M. Zhang, Z. Wei, R. Asad, X.C. Yang, and X. Wang. When does reinforcement learning stand out in quantum control? a comparative study on state preparation.npj Quantum Inf, 5:85, 2019. doi: https://doi.org/10.1038/s41534-019-0201-8

  31. [31]

    Gollub, M

    C. Gollub, M. Kowalewski, and R. de Vivie-Riedle. Monotonic convergent optimal control theory with strict limitations on the spectrum of optimized laser fields.Phys. Rev. Lett., 101(073002), 2008. doi: https://doi.org/10.1103/PhysRevLett.101.073002

  32. [32]

    Alexeev, M.H

    Y. Alexeev, M.H. Farag, and T.L. et al. Patti. Artificial intelligence for quantum computing.Nat Commun, 16 (10829), 2025. doi:https://doi.org/10.1038/s41467-025- 65836-3

  33. [33]

    Marin Bukov, Alexandre G. R. Day, Dries Sels, Phillip Weinberg, Anatoli Polkovnikov, and Pankaj Mehta. 15 Reinforcement learning in different phases of quan- tum control.Phys. Rev. X, 8(3):031086, 2018. doi: 10.1103/PhysRevX.8.031086

  34. [34]

    Mackeprang, D.B.R

    J. Mackeprang, D.B.R. Dasari, and J. A Wrachtrup. A reinforcement learning approach for quantum state engineering.Quantum Mach. Intell., 2(5), 2020. doi: https://doi.org/10.1007/s42484-020-00016-8

  35. [35]

    A. Pipi, X. Tao, A. Wu, P. Narang, and R. D. Leibrandt. Molecular quantum control algorithm design by rein- forcement learning.Phys. Rev. Research, 8:033103, 2026. doi:https://doi.org/10.1103/t159-pzlx

  36. [36]

    Principled approaches for extending neural architectures to function spaces for operator learning.Nature Machine Intelligence, 07 2026

    Julius Berner, Miguel Liu-Schiaffini, Jean Kossaifi, Valentin Duruisseaux, Boris Bonev, Kamyar Azizzade- nesheli, and Anima Anandkumar. Principled approaches for extending neural architectures to function spaces for operator learning.Nature Machine Intelligence, 07 2026. doi:10.1038/s42256-026-01267-z

  37. [37]

    From architectures to applications: a review of neural quantum states.Quantum Sci

    HannahLange, AnkaVandeWalle, AtiyeAbedinnia, and Annabelle Bohrdt. From architectures to applications: a review of neural quantum states.Quantum Sci. Technol., 9(4):040501, 2024. doi:10.1088/2058-9565/ad7168

  38. [38]

    Benavides-Riveros, and Lipeng Chen

    Jiaji Zhang, Carlos L. Benavides-Riveros, and Lipeng Chen. Artificial-intelligence-based surrogate solution of dissipative quantum dynamics: Physics-informed reconstruction of the universal propagator.J. Phys. Chem. Lett., 15(13):3603–3610, 2024. doi: 10.1021/acs.jpclett.4c00598

  39. [39]

    Benavides-Riveros, and Lipeng Chen

    Jiaji Zhang, Carlos L. Benavides-Riveros, and Lipeng Chen. Neural quantum propagators for driven-dissipative quantum dynamics.Phys. Rev. Res., 7(1):3603–3610,

  40. [40]

    Z. Qi, Y. Peng, and C. Earls. Fourier neural op- erators for time-periodic quantum systems: Learning floquet hamiltonians, observable dynamics, and oper- ator growth.PRX Quantum, 7(020340), 2026. doi: https://doi.org/10.1103/2tt3-yndr

  41. [41]

    Neural operator: Learning maps between function spaces with applications to pdes.Jour- nal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Aziz- zadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Jour- nal of Machine Learning Research, 24(89):1–97, 2023. URLhttp://jmlr.org/papers/v24/21-1524.html

  42. [42]

    Neural operators for accelerating scientific simu- lations and design.Nature Reviews Physics, pages 1–9,

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anand- kumar. Neural operators for accelerating scientific simu- lations and design.Nature Reviews Physics, pages 1–9,

  43. [43]

    Sørensen and K

    A. Sørensen and K. Mølmer. Entanglement and quantum computation with ions in thermal mo- tion.Phys. Rev. A, 62(022311), 2000. doi: https://doi.org/10.1103/PhysRevA.62.022311

  44. [44]

    On universal approximation and error bounds for Fourier neural operators.J

    Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On universal approximation and error bounds for Fourier neural operators.J. Mach. Learn. Res., 22(1),

  45. [45]

    Fourier neural operator for parametric partial differential equations.arXiv, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv, 2020. doi:https://doi.org/10.48550/arXiv.2010.08895

  46. [46]

    Fourier neural operators explained: A practical perspective.https://arxiv.org/pdf/2512.01421, 2025

    Valentin Duruisseaux, Jean Kossaifi, and Anima Anand- kumar. Fourier neural operators explained: A practical perspective.https://arxiv.org/pdf/2512.01421, 2025

  47. [47]

    Shah, T.L

    F. Shah, T.L. Patti, J. Berner, and et al. Fourier neural operators for learning dynamics in quantum spin systems.Communication Physics, 2026. doi: https://doi.org/10.1038/s42005-026-02644-1

  48. [48]

    Wu et al

    A. Wu et al. Prospects of Local Position Invariance mea- surement with Quantum Logic Spectroscopy of a Hydro- nium ion.In preparation

  49. [49]

    A. Pipi, N. Gopinath, V. Duruisseaux, M. Marmarelis, T. L. Patti, B. Khailany, P. Narang, and A. Anandku- mar. Inverse design with fourier neural operators for quantum system control.NeurIPS Workshop on Ma- chine Learning and the Physical Sciences (ML4PS),2025. URLhttps://ml4physicalsciences.github.io/2025/ files/NeurIPS_ML4PS_2025_77.pdf

  50. [50]

    Z. Qi, , C. Earls, and Y. Peng. Universal neu- ral propagator: Learning time evolution in many- body quantum systems.arXiV, 2026. doi: https://doi.org/10.48550/arXiv.2605.05299

  51. [51]

    Gege Wen, Zongyi Li, Qirui Long, Kamyar Azizzade- nesheli, Anima Anandkumar, and Sally M. Benson. Real- time high-resolution CO2 geological storage prediction using nested Fourier neural operators.Energy Environ. Sci., 16:1732–1741, 2023. doi:10.1039/D2EE04204E

  52. [52]

    Wong, Costas A

    Luca Ghafourpour, Valentin Duruisseaux, Bahareh Tolooshams, Philip H. Wong, Costas A. Anastassiou, and Anima Anandkumar. Noble – neural operator with biologically-informed latent embeddings to capture ex- perimental variability in biological neuron models, 2025. URLhttps://arxiv.org/abs/2506.04536

  53. [53]

    Physics-informed neu- ral operator for learning partial differential equations

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzade- nesheli, and Anima Anandkumar. Physics-informed neu- ral operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1–27, 2024. doi: https://doi.org/10.1145/3648506. 16

  54. [54]

    CUDA-Q.https:// github.com/NVIDIA/cuda-quantum

    The CUDA-Q development team. CUDA-Q.https:// github.com/NVIDIA/cuda-quantum

  55. [55]

    Chaffee, B

    D. Chaffee, B. Margulis, A. Sheffield, J. Schmidt, A. Reisenfeld, D. R. Leibrandt, D. Leibfried, , and C. W. Chou Chou. High-fidelity quantum state con- trol of a polar molecular ion in a cryogenic environ- ment.Phys. Rev. Lett., 135(240801), 2025. doi: https://doi.org/10.1103/7ypf-91jr

  56. [56]

    Swapnil P. and K. R. Brown. Precise determination of excited state rotational constants and black-body ther- mometryincoulombcrystalsofca+andcah+.The Jour- nal of Physical Chemistry A, 129(16):3624–3629, 2025. doi:10.1021/acs.jpca.5c00229

  57. [58]

    Fourier neural operator for plasma modelling.arXiv,

    Vignesh Gopakumar, Stanislas Pamela, Lorenzo Zanisi, Zongyi Li, Anima Anandkumar, and MAST Team. Fourier neural operator for plasma modelling.arXiv,

  58. [60]

    FourCastNet: Acceleratingglobalhigh-resolution weather forecasting using adaptive Fourier neural opera- tors, 2022

    Thorsten Kurth, Shashank Subramanian, Peter Harring- ton, Jaideep Pathak, Morteza Mardani, David Hall, An- drea Miele, Karthik Kashinath, and Animashree Anand- kumar. FourCastNet: Acceleratingglobalhigh-resolution weather forecasting using adaptive Fourier neural opera- tors, 2022

  59. [64]

    Lin, Julius Berner, Valentin Duruisseaux, David Pitt, Daniel Leibovici, Jean Kossaifi, Kamyar Azizzade- nesheli, and Anima Anandkumar

    Ryan Y. Lin, Julius Berner, Valentin Duruisseaux, David Pitt, Daniel Leibovici, Jean Kossaifi, Kamyar Azizzade- nesheli, and Anima Anandkumar. Enabling automatic differentiation with mollified graph neural operators, 2025

  60. [65]

    Fc- pino: High precision physics-informed neural operators via fourier continuation, 2025

    Adarsh Ganeshram, Haydn Maust, Valentin Duruis- seaux, Zongyi Li, Yixuan Wang, Daniel Leibovici, Os- car Bruno, Thomas Hou, and Anima Anandkumar. Fc- pino: High precision physics-informed neural operators via fourier continuation, 2025

  61. [66]

    A li- brary for learning neural operators, 2025

    Jean Kossaifi, Nikola Kovachki, Zongyi Li, David Pitt, Miguel Liu-Schiaffini, Valentin Duruisseaux, Robert Joseph George, Boris Bonev, Kamyar Azizzade- nesheli, Julius Berner, and Anima Anandkumar. A li- brary for learning neural operators, 2025. URLhttps: //arxiv.org/abs/2412.10354. Appendix A: Hamiltonian

  62. [67]

    We only model the dynamics of the hydronium internal states and the shared motional mode

    Hyperfine Hamiltonian Inthiswork, thespectroscopyionishydronium,H 3O+, and the logic ion isCa+. We only model the dynamics of the hydronium internal states and the shared motional mode. The logic ion enters only through sympathetic cooling and readout. We label each molecular hyperfine state ofH 3O+by J ≡ {J, K, p, mF , ξ},(A1) whereJis the total rotation...

  63. [68]

    Within this frequency range, we can neglect couplings that arise from THz- scale transitions, using the rotating-wave approximation

    Block Diagonal form The control subspace includes a drive-frequency in the ω/2π∈[5050,5300] kHzrange. Within this frequency range, we can neglect couplings that arise from THz- scale transitions, using the rotating-wave approximation. Within the defined MHz frequency range and the selec- tion rules of two-photon transitions, the Hamiltonian be- comes bloc...

  64. [69]

    Among them, standard neural networks map between finite-dimensional vectors, but PDEs define relationships between infinite-dimensional functions

    Neural Operators Machine learning models have been developed to over- comethehighcomputationalcostsandlimitedscalability oftraditionalnumericalintegratorsacrossscientificfields. Among them, standard neural networks map between finite-dimensional vectors, but PDEs define relationships between infinite-dimensional functions. Neural operators a(x) P Fourier ...

  65. [70]

    The FNO Architecture Throughout our numerical experiments, we employ Fourier neural operators to approximate the solution operators of PDEs. AF ourier neural operator (FNO)[37] is a neural operator using Fourier integral operator layers, which are defined via K(ϕ)vt (x) =F −1 Rϕ ·(Fv t) (x).(B3) whereR ϕ is the Fourier transform of a periodic function κpa...

  66. [71]

    The details are summarized in Table II

    FNO Hyperparameters All Hamiltonian blocks mentioned in Appendix A2 were trained using the same FNO architecture and hy- perparameters. The details are summarized in Table II. The checkpoint with the lowest validation loss was re- tained for subsequent surrogate evaluation and inverse- design calculations. Appendix C: Additional Material

  67. [72]

    Reinforcement-learning benchmark details The RL benchmark was adapted from the RL-qMDP method in Ref. [29]. The action library, referred to as TABLE II: FNO training hyperparameters. Hyperparameter Value Fourier modes 60 Hidden channels 256 Fourier layers 4 Lifting/projection ratio 5 Tensor factorization Tucker Tensor rank 0.8 Domain padding 0.1 Batch siz...

  68. [2005]

    doi:DOI: 10.1126/science.1114375

  69. [2021]

    URLhttp://jmlr.org/papers/ v22/21-0806.html

    ISSN 1532-4435. URLhttp://jmlr.org/papers/ v22/21-0806.html

  70. [2023]

    doi:https://doi.org/10.48550/arXiv.2302.06542

  71. [2024]

    doi:https://doi.org/10.1038/s42254-024-00712-5

  72. [2025]

    doi:10.1103/PhysRevResearch.7.L012013

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.