Automated Optimization of Laser Fields for Quantum State Manipulation

Edgar Pogosyan; Emil A. Gazazyan; Karen Arzumanyan; Roman Sahakyan; Romik Sargsyan

REVIEW 3 major objections 4 minor 32 references

Automated Optimization of Laser Fields for Quantum State Manipulation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper shows that automatic differentiation plus bound-constrained gradient optimization can design smooth Gaussian pulses that transfer population across a five-level M-type system while keeping lossy intermediate states nearly…

desk verdict A competent but thin numerical methods paper: the temporal-ordering penalty and optimized M-system pulse tables are the only genuinely new pieces, and the absence of quantitative fidelities and robustness tests makes the broader claims unsubstantiated. read the letter →

arxiv 2506.08485 v2 pith:LMOG74NS submitted 2025-06-10 quant-ph

Roman Sahakyan , Romik Sargsyan , Edgar Pogosyan , Karen Arzumanyan , Emil A. Gazazyan This is my paper

classification quant-ph

keywords quantumcontrolpopulationtransferSTIRAPautomaticdifferentiationL-BFGS-BLindbladmasterequationfive-levelM-typesystemlaserpulseoptimization

verification ladder T0 review T1 audit T2 compute T3 formal

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 aims to establish a fully numerical route to quantum control: instead of relying on analytic schemes like STIRAP, an optimizer tunes the amplitudes, timings, widths, and detunings of four Gaussian pulses to drive a dissipative five-level M-type system from state $|1\rangle$ to state $|5\rangle$. The loss function penalizes premature departure from the initial state, integrated occupation of the lossy intermediate states, and any terminal deviation from the target population. In simulations of the Lindblad master equation with spontaneous emission, the method finds smooth, bounded pulses that achieve near-unit transfer while keeping $\rho_{22}$ and $\rho_{44}$ close to zero, passing through these levels virtually. The authors present the framework as a universal and experimentally applicable tool for automated control-pulse design, particularly where analytic methods or manual tuning are inefficient.

What carries the argument

The load-bearing mechanism is a scalar loss functional that sums three terms: an integral of the initial-state population $\rho_{11}(t)$ over the latter half of the protocol, integrals of the intermediate-state populations $\rho_{kk}(t)$ over the whole interval, and a squared terminal-error term $(\rho_{55}(T)-1)^2$. The minimization runs over the four Gaussian-pulse parameters per transition (center time, width, peak amplitude, and detuning) under box constraints, with an optional smooth sigmoid product that softly enforces temporal ordering of the pulse centers. Gradients through the ODE integration are obtained by reverse-mode automatic differentiation, and the bound-constrained L-BFGS optimizer performs the minimization; the same pipeline can in principle be reused for other level configurations and higher-dimensional parameter spaces.

What would settle it

A concrete test would be a Monte Carlo sensitivity study over the optimized pulse parameters: if the final population in $\rho_{55}(T)$ drops markedly under small perturbations in timing, amplitude, width, or detuning, or if including Doppler averaging over atomic velocities degrades the transfer, the paper's claim of experimental applicability would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that a differentiable loss functional over the parameters of four Gaussian pulses, minimized with L-BFGS-B using gradients from automatic differentiation, suffices to solve the inverse quantum control problem in a dissipative five-level M-type system. For representative decay channels $|2\rangle\to|1\rangle$, $|2\rangle\to|3\rangle$, $|4\rangle\to|3\rangle$, and $|4\rangle\to|5\rangle$, the optimized pulses drive the system from $|1\rangle$ to $|5\rangle$ with near-unit final population and minimal transient occupation of the excited states, acting through virtual rather than real excitation. A differentiable temporal-ordering penalty lets the optimizer discover counterintuitive pulse sequences in the spirit of STIRAP, and all optimized parameters respect physically motivated bounds.

Load-bearing premise

The load-bearing premise is that the idealized Lindblad model with no atomic motion or Doppler shifts is an adequate proxy for a real experiment, and that the optimized pulses remain effective under parameter fluctuations and technical noise.

Editorial extensions

If this is right

For a five-level M-type system with Lindblad dissipation, the method finds control pulses that achieve near-complete population transfer $|1\rangle\to|5\rangle$ while keeping lossy intermediate-state populations near zero.
The same loss-and-solver pipeline applies to other multi-level configurations and high-dimensional parameter spaces without redesigning the control scheme.
The differentiable temporal-ordering penalty allows the optimizer to discover counterintuitive STIRAP-like pulse orderings rather than requiring a hand-specified sequence.
Because the optimized pulses are smooth and bounded, they are presented as physically implementable candidates for experimental realization.
The approach is offered as a practical alternative when analytic schemes such as STIRAP are inefficient or inapplicable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next step, not reported in the paper, is a Monte Carlo sensitivity study over pulse parameters: the claimed experimental applicability rests on robustness to timing, amplitude, width, and detuning fluctuations, which the paper does not test.
The same differentiable loss structure could be reused with non-Gaussian pulse parameterizations such as splines or piecewise-constant segments, extending the method to regimes where Gaussian envelopes are too restrictive.
Because the framework only requires simulated trajectories and their gradients, it could be paired with measured populations in a closed loop for automated recalibration on a real apparatus, although the paper stops at offline simulation.
The method's apparent success in suppressing intermediate-state occupation suggests it may also work for quantum memory writing or gate design in M-type and related systems, but those applications are only mentioned as future directions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

This is a competent but small numerical methods paper. The genuinely new pieces are the differentiable temporal-ordering penalty in Eq. (11) and the three optimized pulse parameter sets in Tables I-III for a dissipative five-level M-type system. Everything else—gradient-based optimization with automatic differentiation, Gaussian pulse ansatz, Lindblad dynamics, L-BFGS-B—is assembled from established tools, and the authors mostly cite that prior work.

What the paper does well: the optimization setup is clean. The loss is sensible (final target population plus suppression of intermediate-state occupation), the box constraints are physically motivated, and the use of Zygote plus DifferentialEquations.jl is a reasonable, modern pipeline. They also tested other solvers and kept finite differences as a cross-check. If the reported pulses really do what the figures suggest, the method works as a practical way to tune Gaussian pulses for this kind of system.

The soft spots are real and mostly in the reporting. The central numerical claim is not actually quantified: no final ρ55(T), no loss values, no fidelities. “Close to unity” from a plot is not enough, especially when the final fidelity is part of the loss function (Eq. 8) and therefore fitted, not predicted. The paper also calls the scheme robust and experimentally applicable, but Section IV explicitly says Doppler shifts and atomic motion are neglected, and there is no sensitivity analysis or Monte Carlo to back “robust.” That is the biggest overreach. There are also no comparisons against STIRAP, Krotov, or GRAPE, so the implied advantage over existing control methods is not demonstrated. The 16-parameter example is too small to support the scalability claims. No code or data is provided, which makes the tables hard to verify independently. The citation pattern leans on the authors’ own prior M-system work, but that is not a problem by itself.

This is not a paper with a load-bearing flaw. The optimization loop and Lindblad model are standard, and the reported dynamics are consistent with successful transfer. It is just a modest contribution that needs stronger quantitative reporting before the broader claims can be taken seriously.

Who it is for: people doing numerical quantum control on multilevel systems, especially M-type and STIRAP-like setups. A serious referee should see it, but the review should insist on fidelity numbers, baseline comparisons, a basic robustness study (at least parameter jitter and Doppler sensitivity), and ideally code release.

I would not cite it yet, but I would take a revised version seriously.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a gradient-based numerical optimization framework, implemented in Julia with automatic differentiation and L-BFGS-B, for designing Gaussian laser pulses that transfer population from state |1> to state |5> in a five-level M-type system with Lindblad dissipation. The loss functional in Eq. (8) penalizes early population of the initial state, intermediate-level occupation, and deviation of rho55(T) from unity; an additional differentiable ordering penalty in Eq. (11) encodes temporal pulse sequencing. The authors report three sets of optimized pulse parameters (Tables I-III) and population/Rabi-frequency curves (Fig. 2), claiming nearly complete population transfer with suppressed intermediate-state occupation, and describe the framework as universal and experimentally applicable.

Significance. If fully substantiated, the paper would provide a useful demonstration that automatic differentiation combined with a constrained quasi-Newton optimizer can produce smooth, physically plausible pulse parameters for dissipative multilevel population transfer, complementing analytical methods such as STIRAP. However, the current evidence is largely qualitative: no final fidelities, loss values, baseline comparisons, or robustness tests are reported. The central numerical claim is therefore defensible but not yet established quantitatively. Moreover, the explicit neglect of atomic motion and Doppler shifts in Section IV conflicts with the unqualified 'experimentally applicable' and 'robust' claims in the abstract and conclusion. The work is a reasonable starting point but needs substantial additional evidence before it meets the standards of the journal.

major comments (3)

[Section IV, Fig. 2 and Tables I-III] The paper never reports the numerical value of rho55(T), the converged loss L(p), or the final intermediate-level populations. The phrases 'close to unity' and 'nearly complete population transfer' are visual judgments from the figure. Because the loss functional in Eq. (8) explicitly contains the term (rho55(T)-1)^2, reporting the achieved rho55(T) and the loss value is essential to substantiate the abstract's claim that numerical simulations 'confirm the effectiveness' of the method. Without these numbers, the reader cannot distinguish 99.9% transfer from 95% transfer, which matters for the claimed high accuracy.
[Section IV, paragraph 2 and Section V] The manuscript explicitly states that atomic motion and Doppler shifts are neglected, yet the abstract and conclusion describe the framework as 'universal and experimentally applicable' and 'robust' without any sensitivity analysis, Monte Carlo sampling, or experimental comparison. This is a load-bearing gap: the optimized Gaussian pulses could be fragile under Doppler detunings, pulse-amplitude errors, or uncertainty in decay rates, and the paper provides no evidence either way. The authors should either add perturbation studies (e.g., varying detunings, Rabi amplitudes, and Gamma by reasonable percentages and reporting fidelity statistics) or substantially soften the experimental-applicability claims.
[Sections I and V] The paper motivates the method by arguing that STIRAP can be inefficient or inapplicable, but no comparison to STIRAP or to the scanning-based scheme of Ref. [6] is provided under the same model. Without a baseline showing the optimized pulses outperform or match a standard protocol in final fidelity, intermediate-loss suppression, or robustness, the claimed advantage over existing methods is unsupported. A single-figure comparison of rho55(T) and integrated loss for the optimized pulses versus a conventional STIRAP sequence would strengthen the central claim considerably.

minor comments (4)

[Section III.B, Eq. (10) and Tables I-III] The stated bounds sigma_i in [2,4] are violated by the reported optimized parameters: Table I contains sigma values 4.94 and 5.23, and Table II contains 4.02, all above the upper bound. Please correct either the bounds or the reported values, or explain how the constraints were handled.
[Fig. 2 caption and Section IV] The caption states that T is the duration of the shortest pulse and labels panels (a)-(f), while the text refers to 'Figure 2d' and Tables I-III correspond to 'Figure 2d/e/f'. Please clarify which panel corresponds to which table and define T unambiguously, since the time axis extends to 50 units.
[Section III.D.2] The text says L-BFGS-B handles bounds 'as specified in Eq. (13)', but the bounds are actually introduced in Eq. (10). Please correct the cross-reference.
[Sections II and III, notation] The optimization parameters are denoted {t_i, sigma_i, E_i^(0), Delta_i} in Eq. (9) and (10), while the Hamiltonian and later text use Omega_0,j for the pulse amplitude. Please unify the notation for peak Rabi frequency/field amplitude to avoid confusion.

Circularity Check

1 steps flagged · score 4.0 of 10

Target fidelity is included in the loss functional, so the reported population transfer is a fitted outcome; self-citations are present but not load-bearing.

fitted input called prediction [Section III.A, Eq. (8); Section IV (Results), discussion of Fig. 2]
"L(p) = ∫ ... (ρ55(T)−1)^2 ... The final squared-error term drives the population of target state |5⟩ to unity at the terminal time, ensuring high-fidelity transfer. ... At the final time, the curve ρ55(t) reaches a value close to unity, signifying nearly complete population transfer to the target state |5⟩, which is the main indicator of protocol efficiency."

The success criterion is the optimized objective: Eq. (8) contains the term (ρ55(T)−1)^2, so the reported 'nearly complete population transfer' is the minimized target, not an independent result. The abstract and Results present this fitted outcome as 'numerical simulations confirm the effectiveness,' but the confirmation merely reflects that the optimizer minimized the loss term that defines success. No quantitative final fidelity, baseline, or out-of-sample test is given, so the effectiveness claim for population transfer is circular in the fitted-input sense. The independent content lies in the optimization machinery (AD, L-BFGS-B, regularization), not in the transfer itself.

full rationale

The central artifact is a numerical optimization: the loss functional of Eq. (8) explicitly includes (ρ55(T)−1)^2, so the 'nearly complete population transfer' reported in the abstract and Section IV is the minimized objective, not an independent prediction. This is a genuine partial circularity of the fitted-input-call-prediction kind. However, the paper's substantive contribution is the optimization framework itself — automatic differentiation, L-BFGS-B, comparator solvers, and temporal-ordering regularization — which has independent content beyond the target fidelity being enforced. The self-citations to prior M-system modeling ([12], [15], [23], [24]) are not load-bearing because the model equations are stated explicitly in the paper. The claims of robustness and experimental applicability are unsupported given that Section IV neglects Doppler shifts and atomic motion, but that is a completeness and correctness gap, not a circularity. Score 4 reflects one partially circular confirmation step while the method has substantial non-circular content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central demonstration depends on the chosen Gaussian ansatz, the dissipative model, and a set of unreported hyperparameters (loss weights, barrier strengths, ordering penalty parameters, simulation time, and optimizer initialization). No new physical entities are introduced. The optimized pulse parameters themselves are outputs, not inputs, and are not counted here as free parameters.

free parameters (5)

Loss function weights
Eq. (8) lists three terms without weighting coefficients; the relative importance of suppressing intermediate populations versus final fidelity is implicitly set to 1, but never stated.
Ordering penalty weight and steepness (lambda, ksharp)
Eq. (11) defines the penalty but the paper never states how it enters the total loss or the values of lambda and ksharp used for Figures 2d-f.
Softplus barrier strength
Section III.B mentions softplus-based barriers to keep parameters in range but does not report the barrier coefficient.
Total simulation time T
Eq. (7) integrates on [0,T], but T is never specified; the figure caption confusingly defines T as the duration of the shortest pulse, while the population dynamics extend to t=50.
Initial conditions for optimizer
The starting parameter vector for L-BFGS-B is not given, so the reported optima cannot be reproduced exactly.

assumptions (5)

domain assumption The rotating-wave approximation and the interaction-picture Hamiltonian in Eq. (3) accurately describe the driven five-level system.
Section II states the Hamiltonian is in the interaction picture and under RWA, which requires the laser detunings to be small compared to optical frequencies; the optimization pushes detunings up to 5 Gamma, still in this regime, but the validity is assumed.
domain assumption The Lindblad master equation with collapse operators in Eq. (5) captures all relevant dissipation.
Section II defines decays from |2> and |4> with rate Gamma and dephasing Gamma/2; no additional dephasing channels or level shifts are considered.
domain assumption Gaussian pulse profiles (Section II) form a sufficient control ansatz for high-fidelity transfer.
The optimization is restricted to Gaussian pulses parameterized by center, width, amplitude, and detuning; this restriction may exclude better solutions, so the claim of 'optimal' is only optimal within this ansatz.
domain assumption Atomic motion and Doppler shifts can be neglected.
Section IV explicitly negates atomic motion and Doppler effects; the optimized pulses are therefore not validated for realistic thermal ensembles.
standard math The Tsitouras 5/4 adaptive solver and L-BFGS-B converge to a sufficiently accurate optimum.
Section III states Tsit5 and L-BFGS-B are used, but convergence tolerances, number of restarts, and gradient accuracy checks are not reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automated Optimization of Laser Fields for Quantum State Manipulation." pith.science (2026). https://pith.science/paper/LMOG74NS

@misc{pith2026250608485,
  author       = {Pith},
  title        = {Pith review of: Automated Optimization of Laser Fields for Quantum State Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMOG74NS}},
  note         = {Machine review of arXiv:2506.08485}
}

read the original abstract

A gradient-based optimization approach combined with automatic differentiation is employed to ensure high accuracy and scalability when working with high-dimensional parameter spaces. Numerical simulations confirm the effectiveness of the proposed method: the population is reliably transferred to the target state with minimal occupation of intermediate levels, while the control pulses remain smooth and physically implementable. The developed framework serves as a universal and experimentally applicable tool for automated control pulse design in quantum systems. It is particularly useful in scenarios where analytical methods or manual parameter tuning--such as standard schemes like STIRAP--prove to be inefficient or inapplicable.

Figures

Figures reproduced from arXiv: 2506.08485 by the authors.

**Figure 2.** FIG. 2: Time evolution of populations [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 31 canonical work pages

[6]

Per- spective: Stimulated raman adiabatic passage: The sta- tus after 25 years,

K. Bergmann, N. V. Vitanov, and B. W. Shore, “Per- spective: Stimulated raman adiabatic passage: The sta- tus after 25 years,” The Journal of Chemical Physics , vol. 142, p. 170901, 05 2015

work page 2015
[1]

s = +1: enforce tj > tk (j occurs last),

work page
[2]

chain of order

s = −1: enforce tj < tk (j occurs first). Each term σ s ksharp(tj − tk) satisfies 5 σ s ksharp(tj − tk) ≈ ( 1, if the desired ordering holds (e.g. tj ≫ tk), 0, if it is strongly violated (e.g. tj ≪ tk). Because σ(x) is smooth, this penalty remains differen- tiable, yet becomes sharply selective as ksharp grows. By multiplying over all k ̸= j, P softly—but...

work page
[3]

Automatic Differentiation We initially implemented gradient computations via forward-mode finite differences using AutoFiniteDiff (backed by FiniteDiff.jl), which provided a robust base- line and useful cross-checks. We then transitioned to source-to-source reverse-mode automatic differentiation with Zygote.jl [20] (via the AutoZygote interface) to ob- ta...

work page
[4]

Optimization Algorithms After benchmarking a variety of solvers, we selected the Limited-memory BFGS (L-BFGS) family for its blend of performance and convergence quality. The stan- dard L-BFGS [21] algorithm (unconstrained) approxi- mates the inverse Hessian by storing only the most recent m vector pairs {si, yi}, reducing both storage and per- iteration ...

work page
[5]

Implementation and Diagnostics All solvers and differentiation backends are accessed via the Optimization.jl ecosystem, with callback hooks configured for real-time monitoring. These diagnos- tics—implemented through user callbacks—provide in- teractive loss and gradient visualizations, enabling rapid identification of convergence issues and on-the-fly pa...

work page
[7]

Roadmap on stirap applications,

K. Bergmann, H.-C. Nagerl, C. Panda, G. Gabrielse, E. Miloglyadov, M. Quack, G. Seyfang, G. Wichmann, S. Ospelkaus, A. Kuhn, S. Longhi, A. Szameit, P. Pirro, B. Hillebrands, X.-F. Zhu, J. Zhu, M. Drewsen, W. K. Hensinger, S. Weidt, T. Halfmann, H. Wang, G. S. Paraoanu, N. V. Vitanov, J. Mompart, T. Busch, T. J. Barnum, D. D. Grimes, R. W. Field, M. G. Rai...

work page 2019
[8]

Stimulated raman adiabatic passage in physics, chemistry and beyond,

N. V. Vitanov, A. A. Rangelov, B. W. Shore, and K. Bergmann, “Stimulated raman adiabatic passage in physics, chemistry and beyond,” Reviews of Modern Physics, vol. 89, no. 1, p. 015006, 2017

work page 2017

Show all 32 references

[9]

Picturing stimulated raman adiabatic passage: a stirap tutorial,

B. W. Shore, “Picturing stimulated raman adiabatic passage: a stirap tutorial,” Adv. Opt. Photon. , vol. 9, pp. 563–719, Sep 2017

2017
[10]

Chirped frac- tional stimulated raman adiabatic passage,

J. Chathanathil, A. Ramaswamy, V. S. Malinovsky, D. Budker, and S. A. Malinovskaya, “Chirped frac- tional stimulated raman adiabatic passage,” Phys. Rev. A, vol. 108, p. 043710, Oct 2023

2023
[11]

Effective full population transfer in M-System using scanning technique,

A. Y. Aleksanyan, “Effective full population transfer in M-System using scanning technique,” Journal of Con- temporary Physics (Armenian Academy of Sciences) , vol. 56, pp. 6–12, Jan. 2021

2021
[12]

Optimal pulse sequences for population transfer in multilevel sys- tems,

I. R. Sol´ a, V. S. Malinovsky, and D. J. Tannor, “Optimal pulse sequences for population transfer in multilevel sys- tems,” Phys. Rev. A , vol. 60, pp. 3081–3090, Oct 1999

1999
[13]

Reverse-engineered exact control of pop- ulation transfer in lossy nonlinear three-state systems,

A. Ishkhanyan, “Reverse-engineered exact control of pop- ulation transfer in lossy nonlinear three-state systems,” Photonics, vol. 11, no. 11, p. 1007, 2024

2024
[14]

Op- timal stirap shortcuts using the spin-to-spring mapping,

V. Evangelakos, E. Paspalakis, and D. Stefanatos, “Op- timal stirap shortcuts using the spin-to-spring mapping,” Phys. Rev. A , vol. 107, p. 052606, May 2023

2023
[15]

Robust two-state swap by stimulated ra- man adiabatic passage,

G. T. Genov, S. Rochester, M. Auzinsh, F. Jelezko, and D. Budker, “Robust two-state swap by stimulated ra- man adiabatic passage,” Journal of Physics B: Atomic, Molecular and Optical Physics , vol. 56, no. 5, p. 054001, 2023

2023
[16]

Robust gates inspired by stimulated raman adiabatic passage for a superconducting dual-rail qubit,

U. Singhal, H. V. Upadhyay, I. Ahmad, and V. Singh, “Robust gates inspired by stimulated raman adiabatic passage for a superconducting dual-rail qubit,” Physical Review Applied, vol. 23, Jan. 2025

2025
[17]

All-optical four-bit Toffoli gate with pos- sible implementation in solids,

G. Grigoryan, V. Chaltykyan, E. Gazazyan, and O. Tikhova, “All-optical four-bit Toffoli gate with pos- sible implementation in solids,” in Nonlinear Optics and Applications VII (M. Bertolotti, J. Haus, and A. M. Zheltikov, eds.), vol. 8772, p. 87721N, International So- ciety f...

2013
[18]

Blum, Density matrix theory and applications

K. Blum, Density matrix theory and applications . New York : Plenum Press, 2nd ed ed., 1996

1996
[19]

A short introduction to the lindblad mas- ter equation,

D. Manzano, “A short introduction to the lindblad mas- ter equation,” AIP Advances, vol. 10, p. 025106, 02 2020

2020
[20]

Programmable quantum gate operations using qutrit quantum dots,

E. A. Pogosyan and E. A. Gazazyan, “Programmable quantum gate operations using qutrit quantum dots,” Optical Memory and Neural Networks , vol. 32, pp. S396– S401, Dec 2023

2023
[21]

Quantum and atom optics,

D. A. Steck, “Quantum and atom optics,” 2007

2007
[22]

L. G. Ixaru and G. Vanden Berghe, Runge-Kutta Solvers for Ordinary Differential Equations , pp. 223–304. Dor- drecht: Springer Netherlands, 2004

2004
[23]

Optimization.jl: A uni- fied optimization package,

V. K. Dixit and C. Rackauckas, “Optimization.jl: A uni- fied optimization package,” Mar. 2023

2023
[24]

Differential equations.jl–a performant and feature-rich ecosystem for solving differ- ential equations in Julia,

C. Rackauckas and Q. Nie, “Differential equations.jl–a performant and feature-rich ecosystem for solving differ- ential equations in Julia,” Journal of Open Research Soft- ware, vol. 5, no. 1, 2017

2017
[25]

Don’t unroll adjoint: Differentiating ssa-form programs,

M. Innes, “Don’t unroll adjoint: Differentiating ssa-form programs,” CoRR, vol. abs/1810.07951, 2018

2018 arXiv
[26]

Algorithm 778: L-bfgs-b: Fortran subroutines for large-scale bound- constrained optimization,

C. Zhu, R. H. Byrd, P. Lu, and J. Nocedal, “Algorithm 778: L-bfgs-b: Fortran subroutines for large-scale bound- constrained optimization,” ACM Trans. Math. Softw. , vol. 23, p. 550–560, Dec. 1997

1997
[27]

On the limited memory bfgs method for large scale optimization,

D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large scale optimization,” Mathematical Pro- gramming, vol. 45, pp. 503–528, Aug 1989

1989
[28]

Influence of multi- photon detunings from resonance on adiabatic processes in a five-level system,

E. A. Gazazyan and G. G. Grigoryan, “Influence of multi- photon detunings from resonance on adiabatic processes in a five-level system,” Journal of Experimental and The- oretical Physics, vol. 124, pp. 540–545, Apr 2017

2017
[29]

Four-photon adiabatic excitation of rydberg states,

E. A. Gazazyan, G. G. Grigoryan, and D. N. Khacha- tryan, “Four-photon adiabatic excitation of rydberg states,” Journal of Contemporary Physics (Armenian Academy of Sciences), vol. 53, pp. 293–300, Oct 2018

2018
[30]

Realization of the programmable logical atomic gate,

A. Y. Aleksanyan and E. A. Gazazyan, “Realization of the programmable logical atomic gate,” Journal of Contemporary Physics (Armenian Academy of Sciences), vol. 53, pp. 205–211, jul 2018

2018
[31]

Quantum memory for photons in case of many close ly- ing exciton resonances in solids,

A. D. Gazazyan, E. A. Gazazyan, and A. G. Margaryan, “Quantum memory for photons in case of many close ly- ing exciton resonances in solids,” The European Physical Journal D , vol. 53, pp. 243–251, jun 2009

2009
[32]

Armenian National Supercomputing Center (ANSCC),

A. N. S. Center, “Armenian National Supercomputing Center (ANSCC),” 2025. Accessed: 2025-04-12

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [6]

Per- spective: Stimulated raman adiabatic passage: The sta- tus after 25 years,

K. Bergmann, N. V. Vitanov, and B. W. Shore, “Per- spective: Stimulated raman adiabatic passage: The sta- tus after 25 years,” The Journal of Chemical Physics , vol. 142, p. 170901, 05 2015

work page 2015

[2] [1]

s = +1: enforce tj > tk (j occurs last),

work page

[3] [2]

chain of order

s = −1: enforce tj < tk (j occurs first). Each term σ s ksharp(tj − tk) satisfies 5 σ s ksharp(tj − tk) ≈ ( 1, if the desired ordering holds (e.g. tj ≫ tk), 0, if it is strongly violated (e.g. tj ≪ tk). Because σ(x) is smooth, this penalty remains differen- tiable, yet becomes sharply selective as ksharp grows. By multiplying over all k ̸= j, P softly—but...

work page

[4] [3]

Automatic Differentiation We initially implemented gradient computations via forward-mode finite differences using AutoFiniteDiff (backed by FiniteDiff.jl), which provided a robust base- line and useful cross-checks. We then transitioned to source-to-source reverse-mode automatic differentiation with Zygote.jl [20] (via the AutoZygote interface) to ob- ta...

work page

[5] [4]

Optimization Algorithms After benchmarking a variety of solvers, we selected the Limited-memory BFGS (L-BFGS) family for its blend of performance and convergence quality. The stan- dard L-BFGS [21] algorithm (unconstrained) approxi- mates the inverse Hessian by storing only the most recent m vector pairs {si, yi}, reducing both storage and per- iteration ...

work page

[6] [5]

Implementation and Diagnostics All solvers and differentiation backends are accessed via the Optimization.jl ecosystem, with callback hooks configured for real-time monitoring. These diagnos- tics—implemented through user callbacks—provide in- teractive loss and gradient visualizations, enabling rapid identification of convergence issues and on-the-fly pa...

work page

[7] [7]

Roadmap on stirap applications,

K. Bergmann, H.-C. Nagerl, C. Panda, G. Gabrielse, E. Miloglyadov, M. Quack, G. Seyfang, G. Wichmann, S. Ospelkaus, A. Kuhn, S. Longhi, A. Szameit, P. Pirro, B. Hillebrands, X.-F. Zhu, J. Zhu, M. Drewsen, W. K. Hensinger, S. Weidt, T. Halfmann, H. Wang, G. S. Paraoanu, N. V. Vitanov, J. Mompart, T. Busch, T. J. Barnum, D. D. Grimes, R. W. Field, M. G. Rai...

work page 2019

[8] [8]

Stimulated raman adiabatic passage in physics, chemistry and beyond,

N. V. Vitanov, A. A. Rangelov, B. W. Shore, and K. Bergmann, “Stimulated raman adiabatic passage in physics, chemistry and beyond,” Reviews of Modern Physics, vol. 89, no. 1, p. 015006, 2017

work page 2017