pith. machine review for the scientific record. sign in

arxiv: 2604.18506 · v1 · submitted 2026-04-20 · 🪐 quant-ph · physics.comp-ph

Recognition: unknown

Physics-Informed Neural Networks for Maximizing Quantum Fisher Information in Time-Dependent Many-Body Systems

Antonio Ferrer-S\'anchez , Yolanda Vives-Gilabert , Yue Ban , Xi Chen , Jos\'e D. Mart\'in-Guerrero
Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:16 UTC · model grok-4.3

classification 🪐 quant-ph physics.comp-ph
keywords quantum fisher informationphysics-informed neural networkscounter-diabatic dynamicsquantum metrologymany-body systemsoptimal controlMagnus expansionspin Hamiltonians
0
0 comments X

The pith

Physics-informed neural networks learn counter-diabatic protocols that maximize quantum Fisher information beyond Euler-Lagrange references in driven many-body systems.

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

The paper develops a framework that uses physics-informed neural networks to find time-dependent control protocols maximizing the quantum Fisher information, the fundamental limit on parameter estimation precision in quantum systems. The method combines a variational loss enforcing the Euler-Lagrange structure with a Magnus expansion to approximate the time-ordered evolution operator, allowing the network to infer both the adiabatic gauge potential and the control scheduling function directly from the physics. This addresses the difficulty of optimal control in interacting systems where non-commutativity and Hilbert-space growth make direct optimization intractable. Numerical tests on nearest-neighbor, dipolar, and trapped-ion spin Hamiltonians with up to six qubits show higher normalized QFI, improved fidelity and balance metrics, and small residuals compared with simpler reference protocols. The results also indicate that explicitly learning the scheduling function improves outcomes in most cases and expose size-dependent variations in performance.

Core claim

The PINN framework, by variationally learning the counter-diabatic driving term together with the scheduling function while enforcing the Euler-Lagrange condition via a Magnus-expanded time-evolution operator, produces protocols that achieve systematically higher normalized quantum Fisher information, favorable fidelity, and low physical residuals than reference solutions based solely on the Euler-Lagrange condition for families of driven spin Hamiltonians up to six qubits.

What carries the argument

A variational physics-informed neural network loss that incorporates the Magnus expansion of the time-ordered exponential to infer the adiabatic gauge potential and scheduling function while satisfying the Euler-Lagrange structure of the counter-diabatic protocol.

If this is right

  • The framework yields high normalized QFI together with favorable fidelity and extremal-balance metrics.
  • Small physical residuals are preserved while performance exceeds that of reference Euler-Lagrange solutions.
  • Explicitly learning the scheduling function provides a performance advantage in most tested cases.
  • Non-trivial finite-size effects appear, with the q=3 regime emerging as particularly challenging.

Where Pith is reading between the lines

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

  • The exponential growth of the operator space noted in the paper implies that hybrid approximations such as tensor-network representations of the Magnus terms would be needed to reach larger qubit numbers.
  • The same variational-PINN structure could be reused for other inverse quantum-control problems that require enforcing adiabatic or counter-diabatic conditions.
  • Finite-size effects highlighted for q=3 suggest that the method may reveal scaling laws for metrological advantage in small interacting systems before the large-N limit is taken.

Load-bearing premise

The Magnus-expansion truncation combined with the variational PINN loss accurately captures the true time-ordered dynamics and the learned protocols remain optimal when the Hilbert-space dimension grows beyond six qubits.

What would settle it

Applying the trained networks to a seven-qubit instance of the same Hamiltonians and observing that the achieved normalized QFI falls below the reference Euler-Lagrange value or that the physical residuals become large would falsify the claim of systematic improvement.

Figures

Figures reproduced from arXiv: 2604.18506 by Antonio Ferrer-S\'anchez, Jos\'e D. Mart\'in-Guerrero, Xi Chen, Yolanda Vives-Gilabert, Yue Ban.

Figure 1
Figure 1. Figure 1: The proposed neural architecture for the PINN. The main set of stacked layers (below) processes the time coordinate input, 𝑡, and gives the set of {𝑎𝑃(𝑡)} in which the 𝒜𝜆 (𝑡) can be decomposed. On the other hand, the above path produces the 𝑢𝒩𝒩 output used to construct the scheduling 𝜆(𝑡). The main aim of our methodology is to construct a PINN capable of inferring the counter-diabatic contributions, i.e., … view at source ↗
Figure 2
Figure 2. Figure 2: Global (fine-time) truncation error bound for a windowed Magnus propagator as a function of the number of time windows 𝑛𝑤 (with 𝑇 = 1); curves correspond to truncation order 𝑝 and illustrate the predicted scaling 𝜀∕𝐾 ≲ 𝑇(𝑇∕𝑛𝑤) 𝑝 . It is straightforward to see that as the expansion order increases, the approximation error decreases, albeit at the expense of a higher computational cost. As discussed in Appen… view at source ↗
Figure 3
Figure 3. Figure 3: Euler-Lagrange loss ℒE−L versus training epoch for 𝑞 = 2−6 qubits, comparing Hamiltonians (one panel per system size). that point, since the observed improvement is only marginal. By contrast, the cases 𝑞 = 3, and to a lesser extent 𝑞 = 4 for the dipolar and trapped-ion models, remain systematically at higher residual levels throughout the training, in agreement with the be￾havior already discussed in the … view at source ↗
Figure 4
Figure 4. Figure 4: Log-scale error in the normalized QFI between driving the entire time evolution in a sequential way with respect to using the Magnus expansion approximation—up to third order—, versus several number of windows 𝑛𝑤 for 𝑞 ∈ {2, 3, 4, 5, 6} qubits. The Hamiltonian of nearest-neighbors has been considered. 0.0 0.2 0.4 0.6 0.8 1.0 Nearest neighbors (t) 2.4 1.6 0.8 0.0 0.8 1.6 t (t) 0.0 0.2 0.4 0.6 0.8 1.0 Dipola… view at source ↗
Figure 5
Figure 5. Figure 5: Mean 𝜆(𝑡) and 𝜕𝑡𝜆(𝑡) trajectories for different Hamiltonians (rows), averaged over 𝑞 = 2 − 6 qubits (solid). Shaded bands indicate the standard deviation across qubit counts, and the non-trainable scheduling function used as reference is shown for comparison (orange). function 𝜆(𝑡) itself may also be adjusted as an additional degree of freedom for the PINN. As recalled above, once the control Hamil￾tonian … view at source ↗
Figure 6
Figure 6. Figure 6: Training trajectories of the final metrics 𝜂𝜔, ℱ, and ℬ for 𝑞 = 2 − 6 qubits. The Euler-Lagrange loss is also considered for comparison purposes. Curves compare a trainable time-dependent coupling 𝜆𝒩𝒩(𝑡) against the fixed reference 𝜆ref(𝑡). The Hamiltonian of nearest-neighbors has been considered. 0.940 0.950 q=2 0.520 0.540 q=3 0.800 0.900 q=4 0.850 0.900 q=5 0.982 0.985 q=6 0.700 0.710 0.820 0.840 0.900 … view at source ↗
Figure 7
Figure 7. Figure 7: Performance metrics 𝜂𝜔, ℱ, and ℬ as functions of the number of layers considered for the PINN for 𝑞 = 2 − 6 qubits. Each panel compares the two widths considered, showing how neural size can affect the final model performance across system sizes. The Hamiltonian of nearest-neighbors has been considered. 12 Antonio Ferrer-Sanchez et al [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Variation of the final metrics 𝜂𝜔, ℱ, and ℬ as a function of the driving frequency 𝜔, for different system sizes 𝑞. Each panel shows one value of 𝑞, while different curves correspond to different values of the field strength ℎ. The Hamiltonian of nearest-neighbors has been considered. with the loss ℒE−L, across these scenarios. As can be seen, all variables respond favorably, in broad terms, to an increase… view at source ↗
Figure 9
Figure 9. Figure 9: Training and inference computational cost of the PINN model as a function of qubit number 𝑞. Left: mean training time per optimization step (error bars: standard deviation across runs). Middle: mean inference time per forward pass (error bars: standard deviation across runs). Right: peak GPU memory usage during training. It is straightforward to see that the training time remains approx￾imately constant fo… view at source ↗
Figure 11
Figure 11. Figure 11: Scalability of the Schrödinger residual error, 𝜀Schr, and the mean unitarity error, 𝜀uni, as functions of the number of qubits, 𝑞, for the different dynamics considered. The top panels show both the raw and reference results, while the bottom panels display the corresponding rations between them. symmetry properties of the particular dynamics considered. Finally, in terms of the error in the unitarity of … view at source ↗
Figure 12
Figure 12. Figure 12: Memory occupation of the PINN output tensor, 𝑀GiB out , as a function of the locality cutoff 𝑘, for different system sizes 𝑞. The curves are computed from the exact output dimension 𝑁out(𝑞, 𝑘), assuming float32 precision and 𝑁𝑡 = 28 collocation points. 4. Conclusions From a physical standpoint, the problem addressed in this work is the design of time-dependent quantum dynamics capable of encoding a parame… view at source ↗
Figure 13
Figure 13. Figure 13: Time-resolved extremal-subspace population 𝑃ext(𝑡) for the three Hamiltonian families and system sizes 𝑞 = 2, … , 6. This quantity measures the total population of the learned state inside the instantaneous extremal eigenspace of the sensitivity operator [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: reports this quantity for 𝒪(𝑡) = ℋ𝜔(𝑡), 𝜕𝜔ℋ𝜔(𝑡), and ℋtot(𝑡). The results show that, among the sizes 𝑞 ≥ 3, the 𝑞 = 3 case shows the largest mismatch with respect to 𝑆X, consistently across the three Hamiltonian families. This is particularly relevant because the effect is not restricted to the bare Hamiltonian ℋ𝜔(𝑡), but is also present in the sensitivity operator 𝜕𝜔ℋ𝜔(𝑡), being this the operator that de… view at source ↗
read the original abstract

Quantum Fisher Information (QFI) sets the ultimate precision limit for parameter estimation and is therefore a central quantity in quantum metrology. In time-dependent many-body systems, however, maximizing QFI is a highly non-trivial task due to the combined effects of non-commutativity, control complexity, and the exponential growth of the Hilbert space. In this work, we present a physics-informed neural network (PINN) framework to address this problem through the learning of counter-diabatic quantum dynamics. Our approach combines a variational PINN formulation with a Magnus-expansion treatment of time-ordered evolution, enabling the adiabatic gauge potential and the scheduling function to be inferred directly from the underlying physics while enforcing the Euler-Lagrange structure of the protocol. The method is applied to several families of driven spin Hamiltonians, including nearest-neighbor, dipolar, and trapped-ion-inspired interactions, for systems of up to six qubits. The numerical results show that the proposed framework systematically improves over reference solutions based only on the Euler-Lagrange condition, yielding high normalized QFI together with favorable fidelity and extremal-balance metrics while preserving small phsical residuals. The analysis further shows that learning the scheduling function provides a clear performance advantage in most cases, and reveals non-trivial finite-size effects, with $q=3$ emerging as a particularly challenging regime. Although scalability remains limited by the exponential growth of the operator space and by automatic-differentiation costs, the results demonstrate that PINNs constitute a viable and physically grounded route for learning metrologically optimal control strategies in interacting quantum systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes a physics-informed neural network (PINN) framework that combines a variational formulation with a Magnus-expansion treatment of time-ordered evolution to learn counter-diabatic protocols maximizing the quantum Fisher information (QFI) in time-dependent many-body spin systems. It is applied to nearest-neighbor, dipolar, and trapped-ion-inspired Hamiltonians for up to six qubits, claiming systematic improvements in normalized QFI, fidelity, and extremal-balance metrics over Euler-Lagrange references while keeping small physical residuals, with additional observations on the benefit of learning the scheduling function and finite-size effects (e.g., q=3 regime).

Significance. If the central claims hold after independent verification, the work offers a concrete, physics-constrained numerical route for optimal quantum control in metrology problems on small interacting systems. It explicitly demonstrates the advantage of learning scheduling functions, reports results across multiple Hamiltonian families, and identifies non-trivial finite-size effects, providing a reproducible benchmark set that could guide future extensions of PINNs to quantum sensing protocols.

major comments (3)
  1. [Abstract and Magnus-expansion treatment] Abstract and Magnus-expansion treatment: the time-evolution operator is obtained from a truncated Magnus series whose truncation order and error relative to the exact time-ordered exponential are neither specified nor validated for the learned controls (up to 6 qubits). Because the variational loss and all reported QFI/fidelity metrics are computed inside this approximation, systematic bias in the Magnus approximant directly affects the optimality assertion and the claimed improvement over references.
  2. [Abstract] Abstract: the central claim that the PINN 'systematically improves over reference solutions based only on the Euler-Lagrange condition' is circular, since both the network loss and the reference derive from the same variational principle; no independent verification (exact QFI via full diagonalization for N=2-4, comparison to GRAPE or other optimal-control baselines, or experimental data) is provided to confirm that the learned protocols are truly optimal beyond the training loss.
  3. [Numerical results] Numerical results (implied in abstract): no error bars, standard deviations, or statistical analysis accompany the reported gains in normalized QFI, fidelity, and extremal-balance metrics across the Hamiltonian families, and there is no explicit post-training check that the protocols satisfy the claimed optimality condition outside the PINN loss.
minor comments (1)
  1. [Abstract] Abstract: typo 'phsical residuals' should be 'physical residuals'.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment point by point below, outlining how we will strengthen the presentation while preserving the core contributions.

read point-by-point responses

  1. Referee: [Abstract and Magnus-expansion treatment] Abstract and Magnus-expansion treatment: the time-evolution operator is obtained from a truncated Magnus series whose truncation order and error relative to the exact time-ordered exponential are neither specified nor validated for the learned controls (up to 6 qubits). Because the variational loss and all reported QFI/fidelity metrics are computed inside this approximation, systematic bias in the Magnus approximant directly affects the optimality assertion and the claimed improvement over references.

    Authors: We appreciate the referee highlighting this important methodological detail. In the revised manuscript we will explicitly state the truncation order of the Magnus expansion used throughout the calculations. We will also add a dedicated validation subsection that compares the Magnus-approximated evolution operator against the exact time-ordered exponential for all systems with N ≤ 4 (where exact computation remains feasible) and reports the norm of the neglected commutator terms for the learned protocols at N = 5 and 6. These additions will quantify any systematic bias introduced by the approximation and confirm that it does not undermine the reported improvements. revision: yes

  2. Referee: [Abstract] Abstract: the central claim that the PINN 'systematically improves over reference solutions based only on the Euler-Lagrange condition' is circular, since both the network loss and the reference derive from the same variational principle; no independent verification (exact QFI via full diagonalization for N=2-4, comparison to GRAPE or other optimal-control baselines, or experimental data) is provided to confirm that the learned protocols are truly optimal beyond the training loss.

    Authors: We respectfully note that the reference solutions are obtained by direct numerical integration of the Euler-Lagrange equations without neural-network parameterization of the scheduling function or adiabatic gauge potential, whereas the PINN optimizes these functions jointly within the same variational structure. The observed gains therefore arise from the additional representational flexibility of the network rather than from a tautological use of the loss. Nevertheless, to provide independent verification we will add exact QFI computations via full diagonalization for N = 2–4 and, where computationally tractable, comparisons against GRAPE baselines. These results will be included in the revised manuscript to demonstrate that the learned protocols outperform both the direct Euler-Lagrange references and standard optimal-control methods. revision: partial

  3. Referee: [Numerical results] Numerical results (implied in abstract): no error bars, standard deviations, or statistical analysis accompany the reported gains in normalized QFI, fidelity, and extremal-balance metrics across the Hamiltonian families, and there is no explicit post-training check that the protocols satisfy the claimed optimality condition outside the PINN loss.

    Authors: We acknowledge the absence of statistical characterization in the current presentation. In the revised version we will report error bars and standard deviations obtained from at least ten independent training runs with different random seeds for each Hamiltonian family and system size. We will also include an explicit post-training diagnostic that evaluates the Euler-Lagrange residual on the learned protocols using finite-difference approximations independent of the PINN loss, thereby confirming that the reported solutions satisfy the optimality condition outside the training objective. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The paper's core chain consists of a variational PINN loss that enforces the Euler-Lagrange structure while additionally learning the scheduling function and gauge potential through a Magnus-expanded time-evolution operator. Numerical results compare the learned protocols against separate reference solutions that satisfy only the Euler-Lagrange condition; these references are independent solutions to the same equations rather than outputs defined by the PINN itself. No equation or claim reduces by construction to a fitted parameter renamed as a prediction, nor does any load-bearing step rely on a self-citation whose content is unverified. The Magnus truncation is an explicit modeling choice whose accuracy is an external assumption, not a definitional loop, and all reported metrics (normalized QFI, fidelity, residuals) are computed consistently inside the chosen model without tautological re-use of inputs as outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on standard variational principles and the Magnus expansion; no new physical entities are postulated. The main free parameters are the neural-network weights, which are fitted to minimize a physics-informed loss.

free parameters (1)
  • neural-network weights
    Fitted during training to minimize the combined loss that includes the Euler-Lagrange residual and the QFI objective.
axioms (2)
  • domain assumption Magnus expansion provides a sufficiently accurate approximation to the time-ordered exponential for the chosen drive durations and strengths.
    Invoked to treat the time-dependent evolution without explicit time-ordering.
  • standard math The Euler-Lagrange condition derived from the variational principle is the correct optimality constraint for the counter-diabatic protocol.
    Used as the physics-informed loss term.

pith-pipeline@v0.9.0 · 5607 in / 1482 out tokens · 24440 ms · 2026-05-10T05:16:44.935493+00:00 · methodology

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

49 extracted references · 42 canonical work pages · 1 internal anchor

  1. [1]

    Bl´ azquez-Salcedo, C

    Samuel L. Braunstein and Carlton M. Caves. “Statistical dis- tanceandthegeometryofquantumstates”.In:Phys.Rev.Lett. 72(22May1994),pp.3439–3443.doi:10.1103/PhysRevLett. 72.3439.url:https://link.aps.org/doi/10.1103/PhysRevLett. 72.3439

    work page doi:10.1103/physrevlett

  2. [2]

    Quantum Estimation for Quantum Technology

    Matteo G. A. Paris. “Quantum Estimation for Quantum Technology”. In:International Journal of Quantum Infor- mation07.supp01 (2009), pp. 125–137.doi: 10 . 1142 / S0219749909004839

    2009

  3. [3]

    Quantum Metrology

    Vittorio Giovannetti, Seth Lloyd, and Lorenzo Maccone. “Quantum Metrology”. In:Phys. Rev. Lett.96 (1 Jan. 2006), p. 010401.doi: 10.1103/PhysRevLett.96.010401.url: https: //link.aps.org/doi/10.1103/PhysRevLett.96.010401

    work page doi:10.1103/physrevlett.96.010401.url: 2006

  4. [4]

    Entanglement,Nonlinear Dynamics,andtheHeisenbergLimit

    LucaPezzéandAugustoSmerzi.“Entanglement,Nonlinear Dynamics,andtheHeisenbergLimit”.In:Phys.Rev.Lett.102 (10 Mar. 2009), p. 100401.doi: 10.1103/PhysRevLett.102. 100401.url: https://link.aps.org/doi/10.1103/PhysRevLett. 102.100401

    work page doi:10.1103/physrevlett.102 2009

  5. [5]

    Truepre- cisionlimitsinquantummetrology

    MarcinJarzynaandRafałDemkowicz-Dobrzański.“Truepre- cisionlimitsinquantummetrology”.In:NewJournalofPhysics 17.1 (Jan. 2015), p. 013010.doi: 10.1088/1367-2630/17/1/ 013010.url:https://doi.org/10.1088/1367-2630/17/1/013010

    work page doi:10.1088/1367-2630/17/1/ 2015

  6. [6]

    Geometryandnon-adiabaticre- sponseinquantumandclassicalsystems

    MichaelKolodrubetzetal.“Geometryandnon-adiabaticre- sponseinquantumandclassicalsystems”.In:PhysicsReports 697(2017).Geometryandnon-adiabaticresponseinquantum and classical systems, pp. 1–87.issn: 0370-1573.doi: https: //doi.org/10.1016/j.physrep.2017.07.001.url:https://www. sciencedirect.com/science/article/pii/S0370157317301989

    work page doi:10.1016/j.physrep.2017.07.001.url:https://www 2017

  7. [7]

    Quantumspeedlimitsandthemaximal rate of information production

    SebastianDeffner.“Quantumspeedlimitsandthemaximal rate of information production”. In:Phys. Rev. Res.2 (1 Feb. 2020),p.013161.doi:10.1103/PhysRevResearch.2.013161

    work page doi:10.1103/physrevresearch.2.013161 2020

  8. [8]

    Optimaladaptivecon- trolforquantummetrologywithtime-dependentHamiltoni- ans

    ShengshiPangandAndrewN.Jordan.“Optimaladaptivecon- trolforquantummetrologywithtime-dependentHamiltoni- ans”. In:Nature Communications8.1 (2017), p. 14695.issn: 2041-1723.doi:10.1038/ncomms14695.url:https://doi.org/ 10.1038/ncomms14695

    work page doi:10.1038/ncomms14695.url:https://doi.org/ 2017

  9. [9]

    Fowler, Matteo Mariantoni, John M

    PaoloZanardi,MatteoG.A.Paris,andLorenzoCamposVenuti. “Quantumcriticalityasaresourceforquantumestimation”.In: Phys.Rev.A78(4Oct.2008),p.042105.doi:10.1103/PhysRevA. 78.042105.url:https://link.aps.org/doi/10.1103/PhysRevA. 78.042105

    work page doi:10.1103/physreva 2008

  10. [10]

    Fisher information and multiparticle entanglement

    Philipp Hyllus et al. “Fisher information and multiparticle entanglement”.In:Phys.Rev.A85(2Feb.2012),p.022321.doi: 10.1103/PhysRevA.85.022321.url:https://link.aps.org/doi/ 10.1103/PhysRevA.85.022321

    work page doi:10.1103/physreva.85.022321.url:https://link.aps.org/doi/ 2012

  11. [11]

    Scheie, P

    ShaoyingYinetal.“QuantumFisherinformationinquantum critical systems with topological characterization”. In:Phys. Rev.B100(18Nov.2019),p.184417.doi:10.1103/PhysRevB. 100.184417

    work page doi:10.1103/physrevb 2019

  12. [12]

    Measuring multipartite entanglement throughdynamicsusceptibilities

    Philipp Hauke et al. “Measuring multipartite entanglement throughdynamicsusceptibilities”.In:NaturePhysics12.8(Aug. 2016),pp.778–782.doi:10.1038/nphys3700

    work page doi:10.1038/nphys3700 2016

  13. [13]

    AttheLimitsofCriticality-BasedQuan- tumMetrology:ApparentSuper-HeisenbergScalingRevisited

    MarekM.Ramsetal.“AttheLimitsofCriticality-BasedQuan- tumMetrology:ApparentSuper-HeisenbergScalingRevisited”. In:Phys. Rev. X8 (2 Apr. 2018), p. 021022.doi: 10.1103/ PhysRevX.8.021022

    2018

  14. [14]

    Neural-network-basedparameterestimation forquantumdetection

    YueBanetal.“Neural-network-basedparameterestimation forquantumdetection”.In:QuantumScienceandTechnology 6.4 (Aug. 2021), p. 045012.doi: 10.1088/2058-9565/ac16ed. url:https://doi.org/10.1088/2058-9565/ac16ed

    work page doi:10.1088/2058-9565/ac16ed 2021

  15. [15]

    Stochastic properties of the frequency dynamics in real and synthetic power grids,

    Giacomo Torlai et al. “Precise measurement of quantum ob- servableswithneural-networkestimators”.In:Phys.Rev.Res. 2 (2 June 2020), p. 022060.doi: 10.1103/PhysRevResearch. 2 . 022060.url: https : / / link . aps . org / doi / 10 . 1103 / PhysRevResearch.2.022060

    work page doi:10.1103/physrevresearch 2020

  16. [16]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. “Physics- informedneuralnetworks:Adeeplearningframeworkforsolv- ingforwardandinverseproblemsinvolvingnonlinearpartial differentialequations”.In:JournalofComputationalPhysics 378(2019),pp.686–707.issn:0021-9991.doi:https://doi.org/ 10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  17. [18]

    Molecular ion-pair states in ungerade h2.Phys

    TommasoCaneva,TommasoCalarco,andSimoneMontangero. “Choppedrandom-basisquantumoptimization”.In:Phys.Rev. A84 (2 Aug. 2011), p. 022326.doi: 10.1103/PhysRevA.84. 022326.url:https://link.aps.org/doi/10.1103/PhysRevA.84. 022326

    work page doi:10.1103/physreva.84 2011

  18. [19]

    Blanes, F

    S.Blanesetal.“TheMagnusexpansionandsomeofitsappli- cations”. In:Physics Reports470.5 (2009), pp. 151–238.issn: 0370-1573.doi:https://doi.org/10.1016/j.physrep.2008.11.001. url: https://www.sciencedirect.com/science/article/pii/ S0370157308004092

    work page doi:10.1016/j.physrep.2008.11.001 2009

  19. [20]

    Quantummetrologyfora general Hamiltonian parameter

    ShengshiPangandToddA.Brun.“Quantummetrologyfora general Hamiltonian parameter”. In:Phys. Rev. A90 (2 Aug. 2014), p. 022117.doi: 10.1103/PhysRevA.90.022117.url: https://link.aps.org/doi/10.1103/PhysRevA.90.022117

    work page doi:10.1103/physreva.90.022117.url: 2014

  20. [21]

    Shortcuts to adiabaticity: Concepts, methods,andapplications

    D. Guéry-Odelin et al. “Shortcuts to adiabaticity: Concepts, methods,andapplications”.In:Rev.Mod.Phys.91(4Oct.2019), p.045001.doi:10.1103/RevModPhys.91.045001.url:https: //link.aps.org/doi/10.1103/RevModPhys.91.045001

    work page doi:10.1103/revmodphys.91.045001.url:https: 2019

  21. [22]

    Beweis des Adiabatensatzes

    M. Born and V. Fock. “Beweis des Adiabatensatzes”. In: ZeitschriftfürPhysik51.3(Mar.1928),pp.165–180.issn:0044- 3328.doi:10.1007/BF01343193.url:https://doi.org/10.1007/ BF01343193

    work page doi:10.1007/bf01343193.url:https://doi.org/10.1007/ 1928

  22. [23]

    Adiabaticquantumtrans- portinmultiplyconnectedsystems

    J.E.Avron,A.Raveh,andB.Zur.“Adiabaticquantumtrans- portinmultiplyconnectedsystems”.In:Rev.Mod.Phys.60(4 Oct. 1988), pp. 873–915.doi: 10.1103/RevModPhys.60.873. url:https://link.aps.org/doi/10.1103/RevModPhys.60.873

    work page doi:10.1103/revmodphys.60.873 1988

  23. [24]

    Minimizingirreversible lossesinquantumsystemsbylocalcounterdiabaticdriving

    DriesSelsandAnatoliPolkovnikov.“Minimizingirreversible lossesinquantumsystemsbylocalcounterdiabaticdriving”.In: ProceedingsoftheNationalAcademyofSciences114.20(2017), E3909–E3916.doi: 10.1073/pnas.1619826114. eprint: https: //www.pnas.org/doi/pdf/10.1073/pnas.1619826114.url: https://www.pnas.org/doi/abs/10.1073/pnas.1619826114. Antonio Ferrer-Sanchez et a...

    work page doi:10.1073/pnas.1619826114 2017

  24. [25]

    Fast and power efficient GPU-based explicit elastic wave propagation analysis by low- ordered orthogonal voxel finite element with INT8 tensor cores

    J.R.Johansson,P.D.Nation,andFrancoNori.“QuTiP:Anopen- sourcePythonframeworkforthedynamicsofopenquantum systems”.In:ComputerPhysicsCommunications183.8(2012), pp.1760–1772.issn:0010-4655.doi:https://doi.org/10.1016/j. cpc.2012.02.021.url:https://www.sciencedirect.com/science/ article/pii/S0010465512000835

    work page doi:10.1016/j 2012

  25. [26]

    PennyLane: Automatic differentiation of hybrid quantum-classical computations

    Ville Bergholm et al.PennyLane: Automatic differentiation ofhybridquantum-classicalcomputations.2022.arXiv:1811. 04968[quant-ph].url:https://arxiv.org/abs/1811.04968

    work page internal anchor Pith review arXiv 2022

  26. [27]

    Mul- tilayer feedforward networks are universal approximators

    KurtHornik,MaxwellStinchcombe,andHalbertWhite.“Mul- tilayer feedforward networks are universal approximators”. In:NeuralNetworks2.5(1989),pp.359–366.issn:0893-6080. doi: https://doi.org/10.1016/0893- 6080(89)90020- 8. url: https://www.sciencedirect.com/science/article/pii/ 0893608089900208

    work page doi:10.1016/0893- 1989

  27. [28]

    Locally adaptive activation functions with slope re- covery for deep and physics-informed neural networks

    Ameya D. Jagtap, Kenji Kawaguchi, and George Em Karni- adakis. “Locally adaptive activation functions with slope re- covery for deep and physics-informed neural networks”. In: ProceedingsoftheRoyalSocietyA:Mathematical,Physicaland EngineeringSciences476.2239(July2020),p.20200334.issn: 1364-5021.doi: 10.1098/rspa.2020.0334. eprint: https:// royalsocietypub...

    work page doi:10.1098/rspa.2020.0334 2020

  28. [29]

    MaximalquantumFisherinformation forgeneralsu(2)parametrizationprocesses

    Xiao-XingJingetal.“MaximalquantumFisherinformation forgeneralsu(2)parametrizationprocesses”.In:Phys.Rev.A 92(1July2015),p.012312.doi:10.1103/PhysRevA.92.012312. url:https://link.aps.org/doi/10.1103/PhysRevA.92.012312

    work page doi:10.1103/physreva.92.012312

  29. [30]

    Critical quantum metrology with fully- connectedmodels:fromHeisenbergtoKibble–Zurekscaling

    Louis Garbe et al. “Critical quantum metrology with fully- connectedmodels:fromHeisenbergtoKibble–Zurekscaling”. In:QuantumScienceandTechnology7.3(May2022),p.035010. doi:10.1088/2058-9565/ac6ca5.url:https://doi.org/10.1088/ 2058-9565/ac6ca5

    work page doi:10.1088/2058-9565/ac6ca5.url:https://doi.org/10.1088/ 2058

  30. [31]

    Quantum Fisher information in one- dimensionaltranslation-invariantquantumsystems:Large-N limitanalysis

    Shu Qu et al. “Quantum Fisher information in one- dimensionaltranslation-invariantquantumsystems:Large-N limitanalysis”.In:PhysicsLettersA529(2025),p.130103.issn: 0375-9601.doi: https://doi.org/10.1016/j.physleta.2024. 130103.url:https://www.sciencedirect.com/science/article/ pii/S0375960124007977

    work page doi:10.1016/j.physleta.2024 2025

  31. [32]

    Generalizablecontrolforquantumparameter estimationthroughreinforcementlearning

    HanXuetal.“Generalizablecontrolforquantumparameter estimationthroughreinforcementlearning”.In:npjQuantum Information5.1(2019),p.82.issn:2056-6387.doi:10.1038/ s41534-019-0198-z.url:https://doi.org/10.1038/s41534-019- 0198-z

    work page doi:10.1038/s41534-019- 2019

  32. [33]

    Improving thedynamicsofquantumsensorswithreinforcementlearn- ing

    JonasSchuff,LukasJFiderer,andDanielBraun.“Improving thedynamicsofquantumsensorswithreinforcementlearn- ing”.In:NewJournalofPhysics22.3(Mar.2020),p.035001.doi: 10.1088/1367-2630/ab6f1f.url:https://doi.org/10.1088/1367- 2630/ab6f1f

    work page doi:10.1088/1367-2630/ab6f1f.url:https://doi.org/10.1088/1367- 2020

  33. [34]

    Parameteres- timation in quantum sensing based on deep reinforcement learning

    TailongXiao,JianpingFan,andGuihuaZeng.“Parameteres- timation in quantum sensing based on deep reinforcement learning”.In:npjQuantumInformation8.1(2022),p.2.issn: 2056-6387.doi: 10.1038/s41534-021-00513-z.url: https: //doi.org/10.1038/s41534-021-00513-z

    work page doi:10.1038/s41534-021-00513-z.url: 2022

  34. [35]

    url:https://arxiv.org/abs/2508.21252

    LaxmishaAshokAttisaraandSathishKumar.QuantumMa- chine Learning for Optimizing Entanglement Distribution in QuantumSensorCircuits.2025.arXiv:2508.21252 [quant-ph]. url:https://arxiv.org/abs/2508.21252

    work page arXiv 2025

  35. [36]

    Physics-InformedNeuralNetworks for Quantum Control

    ArielNorambuenaetal.“Physics-InformedNeuralNetworks for Quantum Control”. In:Phys. Rev. Lett.132 (1 Jan. 2024), p.010801.doi:10.1103/PhysRevLett.132.010801.url:https: //link.aps.org/doi/10.1103/PhysRevLett.132.010801

    work page doi:10.1103/physrevlett.132.010801.url:https: 2024

  36. [37]

    arXiv: 2511.09463 [quant-ph].url: https://arxiv.org/abs/ 2511.09463

    Sofiia Lauten and Matthew Otten.Physics-Informed Neural NetworksforGateDesignusingQuantumOptimalControl.2025. arXiv: 2511.09463 [quant-ph].url: https://arxiv.org/abs/ 2511.09463

    work page arXiv 2025

  37. [38]

    Minimizingirreversible lossesinquantumsystemsbylocalcounterdiabaticdriving

    DriesSelsandAnatoliPolkovnikov.“Minimizingirreversible lossesinquantumsystemsbylocalcounterdiabaticdriving”.In: ProceedingsoftheNationalAcademyofSciences114.20(2017), E3909–E3916.doi: 10.1073/pnas.1619826114.url: https: //doi.org/10.1073/pnas.1619826114

    work page doi:10.1073/pnas.1619826114.url: 2017

  38. [39]

    Meta-learning digitized- counterdiabatic quantum optimization

    Pranav Chandarana et al. “Meta-learning digitized- counterdiabatic quantum optimization”. In:Quantum Science and Technology8.4 (July 2023), p. 045007.doi: 10.1088/2058-9565/ace54a.url:https://doi.org/10.1088/2058- 9565/ace54a

    work page doi:10.1088/2058-9565/ace54a.url:https://doi.org/10.1088/2058- 2023

  39. [40]

    Counterdiabaticoptimizeddriving inquantumphasesensitivemodels

    FrancescoPioBaroneetal.“Counterdiabaticoptimizeddriving inquantumphasesensitivemodels”.In:NewJournalofPhysics 26.3(Mar.2024),p.033031.doi:10.1088/1367-2630/ad313e. url:https://doi.org/10.1088/1367-2630/ad313e

    work page doi:10.1088/1367-2630/ad313e 2024

  40. [41]

    Physics-informedneuralnet- worksforanoptimalcounterdiabaticquantumcomputation

    AntonioFerrer-Sánchezetal.“Physics-informedneuralnet- worksforanoptimalcounterdiabaticquantumcomputation”. In:MachineLearning:ScienceandTechnology5.2(May2024), p.025035.doi:10.1088/2632-2153/ad450f.url:https://doi. org/10.1088/2632-2153/ad450f

    work page doi:10.1088/2632-2153/ad450f.url:https://doi

  41. [42]

    Speedup for quantum optimal control fromautomaticdifferentiationbasedongraphicsprocessing units

    Nelson Leung et al. “Speedup for quantum optimal control fromautomaticdifferentiationbasedongraphicsprocessing units”. In:Phys. Rev. A95 (4 Apr. 2017), p. 042318.doi: 10. 1103/PhysRevA.95.042318.url:https://link.aps.org/doi/10. 1103/PhysRevA.95.042318

    2017

  42. [43]

    LeveragingStateSparsity forMoreEfficientQuantumSimulations

    SamuelJaquesandThomasHäner.“LeveragingStateSparsity forMoreEfficientQuantumSimulations”.In:ACMTransac- tionsonQuantumComputing3.3(June2022),pp.1–17.doi: 10.1145/3491248

    work page doi:10.1145/3491248

  43. [44]

    Ontheexponentialsolutionofdifferential equationsforalinearoperator

    WilhelmMagnus.“Ontheexponentialsolutionofdifferential equationsforalinearoperator”.In:CommunicationsonPure andAppliedMathematics7.4(1954),pp.649–673.doi:10.1002/ cpa.3160070404

    1954

  44. [45]

    Respect- ingcausalityfortrainingphysics-informedneuralnetworks

    SifanWang,ShyamSankaran,andParisPerdikaris.“Respect- ingcausalityfortrainingphysics-informedneuralnetworks”. In:ComputerMethodsinAppliedMechanicsandEngineering 421(2024),p.116813.issn:0045-7825.doi:https://doi.org/10. 1016/j.cma.2024.116813.url:https://www.sciencedirect.com/ science/article/pii/S0045782524000690

    work page arXiv 2024

  45. [46]

    Optuna:ANext-generationHyperparame- terOptimizationFramework

    TakuyaAkibaetal.“Optuna:ANext-generationHyperparame- terOptimizationFramework”.In:Proceedingsofthe25thACM SIGKDDInternationalConferenceonKnowledgeDiscoveryand DataMining.2019

    2019

  46. [47]

    SOAP:ImprovingandStabilizingShampoo usingAdamforLanguageModeling

    Nikhil Vyas and Depen Morwani and Rosie Zhao and Itai Shapira and David Brandfonbrener and Lucas Janson and ShamM.Kakade.“SOAP:ImprovingandStabilizingShampoo usingAdamforLanguageModeling”.In:TheThirteenthInter- nationalConferenceonLearningRepresentations. 2025.url: https://openreview.net/forum?id=IDxZhXrpNf

    2025

  47. [48]

    Gradient Alignment in Physics-informed NeuralNetworks:ASecond-OrderOptimizationPerspective

    SifanWangandAnanyaeKumarbhartariandBowenLiand Paris Perdikaris. “Gradient Alignment in Physics-informed NeuralNetworks:ASecond-OrderOptimizationPerspective”. In:TheThirty-ninthAnnualConferenceonNeuralInformation ProcessingSystems.2025.url:https://openreview.net/forum? id=iweeVl1RHU. 18 Antonio Ferrer-Sanchez et al. Physics-Informed Neural Networks for Max...

    2025

  48. [49]

    High-ordercommutator-free exponentialtime-propagationofdrivenquantumsystems

    A.AlvermannandH.Fehske.“High-ordercommutator-free exponentialtime-propagationofdrivenquantumsystems”.In: JournalofComputationalPhysics230.15(2011),pp.5930–5956. issn:0021-9991.doi:https://doi.org/10.1016/j.jcp.2011.04.006. url: https://www.sciencedirect.com/science/article/pii/ S0021999111002300

    work page doi:10.1016/j.jcp.2011.04.006 2011

  49. [50]

    SolutionoftheSchrödinger equationbyaspectralmethod

    M.DFeit,J.AFleck,andASteiger.“SolutionoftheSchrödinger equationbyaspectralmethod”.In:JournalofComputational Physics47.3(1982),pp.412–433.issn:0021-9991.doi:https: //doi.org/10.1016/0021-9991(82)90091-2.url:https://www. sciencedirect.com/science/article/pii/0021999182900912. Appendix A. Magnus expansion TheMagnusexpansionprovidesaconvenientframeworkforcomp...

    work page doi:10.1016/0021-9991(82)90091-2.url:https://www 1982