Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Review and experimental benchmarking of machine learning algorithms for efficient optimization of cold atom experiments

Oliver Anton, Victoria A. Henderson, Elisa Da Ros, Ivan Sekulic, Sven Burger, Philipp-Immanuel Schneider, Markus Krutzik

TL;DR

This work addresses the challenge of efficiently optimizing noisy, high-dimensional control parameters in cold-atom experiments by benchmarking a range of heuristic and machine-learning–driven optimizers on $d=10$ and $d=18$ parameter spaces. It compares Bayesian optimization variants (FMFN, AX, JCM) with noise-adaptive capabilities against classical heuristics such as CMA-ES, PSO, DE, and Nelder–Mead, using the atom number $N_{ m atoms}$ as the objective (minimizing $f=-N_{ m atoms}$). The results show that a noise-aware BO implementation (JCM with noise detection) delivers the fastest convergence and highest achieved atom numbers in 10D, with CMA-ES offering strong final performance in both 10D and 18D; 18D optimization is notably harder for all methods. The findings guide optimizer choice for mobile or autonomous cold-atom setups, highlighting the trade-offs between speed and ultimate fitness and demonstrating the value of explicit noise handling in BO for real-world experiments. The study also provides practical implications for autonomous operation of quantum devices where repeated, rapid optimization is essential, such as mobile gravimeters or space-based clocks.

Abstract

The generation of cold atom clouds is a complex process which involves the optimization of noisy data in high dimensional parameter spaces. Optimization can be challenging both in and especially outside of the lab due to lack of time, expertise, or access for lengthy manual optimization. In recent years, it was demonstrated that machine learning offers a solution since it can optimize high dimensional problems quickly, without knowledge of the experiment itself. In this paper we present results showing the benchmarking of nine different optimization techniques and implementations, alongside their ability to optimize a Rubidium (Rb) cold atom experiment. The investigations are performed on a 3D $^{87}$Rb molasses with 10 and 18 adjustable parameters, respectively, where the atom number obtained by absorption imaging was chosen as the test problem. We further compare the best performing optimizers under different effective noise conditions by reducing the Signal-to-Noise ratio of the images via adapting the atomic vapor pressure in the 2D+ MOT and the detection laser frequency stability.

Review and experimental benchmarking of machine learning algorithms for efficient optimization of cold atom experiments

TL;DR

This work addresses the challenge of efficiently optimizing noisy, high-dimensional control parameters in cold-atom experiments by benchmarking a range of heuristic and machine-learning–driven optimizers on and parameter spaces. It compares Bayesian optimization variants (FMFN, AX, JCM) with noise-adaptive capabilities against classical heuristics such as CMA-ES, PSO, DE, and Nelder–Mead, using the atom number as the objective (minimizing ). The results show that a noise-aware BO implementation (JCM with noise detection) delivers the fastest convergence and highest achieved atom numbers in 10D, with CMA-ES offering strong final performance in both 10D and 18D; 18D optimization is notably harder for all methods. The findings guide optimizer choice for mobile or autonomous cold-atom setups, highlighting the trade-offs between speed and ultimate fitness and demonstrating the value of explicit noise handling in BO for real-world experiments. The study also provides practical implications for autonomous operation of quantum devices where repeated, rapid optimization is essential, such as mobile gravimeters or space-based clocks.

Abstract

The generation of cold atom clouds is a complex process which involves the optimization of noisy data in high dimensional parameter spaces. Optimization can be challenging both in and especially outside of the lab due to lack of time, expertise, or access for lengthy manual optimization. In recent years, it was demonstrated that machine learning offers a solution since it can optimize high dimensional problems quickly, without knowledge of the experiment itself. In this paper we present results showing the benchmarking of nine different optimization techniques and implementations, alongside their ability to optimize a Rubidium (Rb) cold atom experiment. The investigations are performed on a 3D Rb molasses with 10 and 18 adjustable parameters, respectively, where the atom number obtained by absorption imaging was chosen as the test problem. We further compare the best performing optimizers under different effective noise conditions by reducing the Signal-to-Noise ratio of the images via adapting the atomic vapor pressure in the 2D+ MOT and the detection laser frequency stability.
Paper Structure (17 sections, 4 equations, 4 figures, 3 tables)

This paper contains 17 sections, 4 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Optimization algorithms for tuning experimental parameters
  3. Heuristic minimization methods
  4. Bayesian optimization
  5. Bayesian optimization of noisy functions
  6. Different implementations of Bayesian optimization
  7. Experimental set-up
  8. Results
  9. 10-dimensional optimization
  10. 18-dimensional optimization
  11. Influence of noise on the experiment optimization
  12. Outlook
  13. Acknowledgements
  14. Experimental parameters
  15. Best parameters reached in the 10 dimensional case
  16. ...and 2 more sections

Figures (4)

  • Figure 1: A block diagram showing the relevant parts of the experimental setup used within this paper, alongside a rendering of the apparatus (on the right). Dashed lines indicate electrical and solid lines optical signals, similarly green shading represents electronics components and yellow shading optical components.
  • Figure 2: Schematic (block diagram) of the control software showing the server's workflow running the experiment with its sub-processes. Using the client, a sequence is created to be used by the server for compilation and later used by the sequence player. Feedback for the optimizer is provided using the acquired data via the data evaluation software. An adapted sequence is then fed into the compiler.
  • Figure 3: Development of the fitness function (maximization of atom number) over time for different optimizers for (a) a 10-dimensional parameter space and (b) a 18-dimensional parameter space. Each algorithm runs for 400 iterations. The lines represent the mean value, while the shaded areas show the standard deviation over 10 repetitions.
  • Figure 4: Development of the fitness function (maximization of atom number) over time for the highest performing optimizers with different noise levels. Figure (a) shows optimization with an atom number noise level of ±5.6%, while (b) shows the same for a noise level of ±19.5%. The JCM optimizer is tested with and without noise detection activated, with this comparison highlighted in the insets. The reduced atom number reached in (b) is due to the experimental conditions choosen to increase the noise level. Each algorithm runs for 400 iterations. The lines represent the mean value, while the shaded area shows the standard deviation over 10 repetitions.