Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Nondestructive characterization of laser-cooled atoms using machine learning

G. De Sousa, M. Doris, D. D'Amato, B. Egleston, J. P. Zwolak, I. B. Spielman

TL;DR

The paper tackles non-destructive extraction of internal MOT properties from fluorescence images of laser-cooled $^{39}$K atoms. It trains regression models, culminating in a CNN that predicts atom number $N$ and temperature $T$ from two fluorescence images, using TOF-based labels for supervision and reliability metrics to weight losses. The results show substantial gains in accuracy, with a typical $N$-uncertainty of $4\times10^6$ atoms and a fractional $T$-uncertainty of about $0.2$, especially when employing reflection/translation data augmentation to exploit spatial structure. The approach enables rapid, real-time diagnostics and paves the way for ML-enabled feedback control in MOT-based quantum technologies, while providing publicly available data for further research.

Abstract

We develop machine learning techniques for estimating physical properties of laser-cooled potassium-39 atoms in a magneto-optical trap using only the scattered light -- i.e., fluorescence -- that is intrinsic to the cooling process. In-situ snap-shot images of fluorescing atomic ensembles directly reveal the spatial structure of these millimeter-scale objects but contain no obvious information regarding internal properties such as the temperature. We first assembled and labeled a balanced dataset sampling $8\times10^3$ different experimental parameters that includes examples with: large and dense atomic ensembles, a complete absence of atoms, and everything in between. We describe a range of models trained to predict atom number and temperature solely from fluorescence images. These run the gamut from a poorly performing linear regression model based only on integrated fluorescence to deep neural networks that give number and temperature with fractional uncertainties of $0.1$ and $0.2$ respectively.

Nondestructive characterization of laser-cooled atoms using machine learning

TL;DR

The paper tackles non-destructive extraction of internal MOT properties from fluorescence images of laser-cooled K atoms. It trains regression models, culminating in a CNN that predicts atom number and temperature from two fluorescence images, using TOF-based labels for supervision and reliability metrics to weight losses. The results show substantial gains in accuracy, with a typical -uncertainty of atoms and a fractional -uncertainty of about , especially when employing reflection/translation data augmentation to exploit spatial structure. The approach enables rapid, real-time diagnostics and paves the way for ML-enabled feedback control in MOT-based quantum technologies, while providing publicly available data for further research.

Abstract

We develop machine learning techniques for estimating physical properties of laser-cooled potassium-39 atoms in a magneto-optical trap using only the scattered light -- i.e., fluorescence -- that is intrinsic to the cooling process. In-situ snap-shot images of fluorescing atomic ensembles directly reveal the spatial structure of these millimeter-scale objects but contain no obvious information regarding internal properties such as the temperature. We first assembled and labeled a balanced dataset sampling different experimental parameters that includes examples with: large and dense atomic ensembles, a complete absence of atoms, and everything in between. We describe a range of models trained to predict atom number and temperature solely from fluorescence images. These run the gamut from a poorly performing linear regression model based only on integrated fluorescence to deep neural networks that give number and temperature with fractional uncertainties of and respectively.

Paper Structure

This paper contains 17 sections, 7 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Experiment
  3. Experimental hardware
  4. Experimental sequence
  5. Data
  6. Labeling strategy
  7. Data organization
  8. Machine learning toolbox
  9. Regression
  10. Data augmentation
  11. Training
  12. Results and discussion
  13. Overview of results
  14. Impact of model architecture and augmentation
  15. Out-of-distribution data
  16. ...and 2 more sections

Figures (7)

  • Figure 1: Workflow of classification and regression system in operation. (a) Acquisition. $^{39}\mathrm{K}$ atoms are laser cooled and then florescence-imaged along $\mathbf{e}_x$ and $\mathbf{e}_z$ as described Sec. \ref{['sec:experiment']} and \ref{['ssec:experimental_sequence']}. Images for optimal experimental parameters are shown. (b) Classification. After acquisition data is classified. (c) Regression. The data is then passed into a regression model which: first potentially processes the images independently; fuses the data; and predicts atom number $N$ and temperature $T$ along with corresponding signal-to-noise ratios ${\rm SNR}_N$ and ${\rm SNR}_T$. Section \ref{['sec:toolbox']} discusses data pre-processing, classification, and regression.
  • Figure 2: (a) Schematic of experimental geometry including top (left) and side (right) views. (b) Relevant energy levels for laser cooling and trapping $^{39}\mathrm{K}$ using only the D2 line. (c) Experimental sequence for MOT operation.
  • Figure 3: Comparison between compact and diffuse clouds. (a) and (b) correspond to ( BATCH 00, SET 150) and ( BATCH 03, SET 995) respectively. Note that the signal in (b) is scaled by $4\times$ to make this lower quality MOT visible.
  • Figure 4: Normalized fluorescence histograms. Panels (a) and (b) derive from the $\mathbf{e}_x$- and $\mathbf{e}_z$-cameras respectively. Horizontal axes are arcsinh-scaled, a symmetric scaling which interpolates between linear for small arguments and logarithmic for large arguments. Solid black curves, obtained with no atoms present, are plotted along with Gaussian fits (dashed black curves). Red curves describe the complete dataset, while the pink curves are the difference between the complete and the no-atoms histograms. The vertical gray lines represents the threshold below which data is assigned MOT = False. Insets: $F2$ metric as a function of threshold with operating point marked in red.
  • Figure 5: Inference for the CON (dashed purple), LIN (solid blue), MM (fine-dotted red), MLP (dot-dotted green) and CNN (dash-dotted orange) models. These models are trained using RT augmentation, and compared to the in-distribution test set with RT augmentation (configured to give a ten-fold increase in test data). To focus the effect of the test distribution's statistics, we evaluate a randomly selected models from those trained in the 10-fold cross validation. (a) marginal distributions of residual vectors $\bm{\ell}$. (b) and (c) number and temperature inference respectively. Points are shaded according to their SNR (color bars) and lines of slope one indicate the ideal behavior.
  • ...and 2 more figures