Abstract Precise determination of thermodynamic parameters in ultracold Bose gases remains challenging due to the destructive nature of conventional measurement techniques and inherent experimental uncertainties. We demonstrate a machine learning approach for rapid, minimally destructive estimation of the chemical potential and temperature from a single image of an in situ imaged density profiles of finite-temperature Bose gases. Our convolutional neural network is trained exclusively on quasi-two-dimensional ‘pancake’ condensates in harmonic trap configurations. It achieves parameter extraction within fractions of a second. The model also demonstrates some zero-shot generalisation across both trap geometry and thermalisation dynamics, successfully estimating the temperature (although not the chemical potential) for toroidally trapped condensates with errors of only a few nanokelvin despite no prior exposure to such geometries during training, and maintaining predictive accuracy during dynamic thermalisation processes after a relatively brief evolution without explicit training on non-equilibrium states. These results suggest that supervised learning can overcome traditional limitations in ultracold atom thermometry, with extension to broader geometric configurations, temperature ranges, and additional parameters potentially enabling comprehensive real-time analysis of quantum gas experiments. Such capabilities could significantly streamline experimental workflows whilst improving measurement precision across a range of quantum fluid systems.
Abstract Precise determination of thermodynamic parameters in ultracold Bose gases remains challenging due to the destructive nature of conventional measurement techniques and inherent experimental uncertainties. We demonstrate a machine learning approach for rapid, minimally destructive estimation of the chemical potential and temperature from a single image of an in situ imaged density profiles of finite-temperature Bose gases. Our convolutional neural network is trained exclusively on quasi-two-dimensional ‘pancake’ condensates in harmonic trap configurations. It achieves parameter extraction within fractions of a second. The model also demonstrates some zero-shot generalisation across both trap geometry and thermalisation dynamics, successfully estimating the temperature (although not the chemical potential) for toroidally trapped condensates with errors of only a few nanokelvin despite no prior exposure to such geometries during training, and maintaining predictive accuracy during dynamic thermalisation processes after a relatively brief evolution without explicit training on non-equilibrium states. These results suggest that supervised learning can overcome traditional limitations in ultracold atom thermometry, with extension to broader geometric configurations, temperature ranges, and additional parameters potentially enabling comprehensive real-time analysis of quantum gas experiments. Such capabilities could significantly streamline experimental workflows whilst improving measurement precision across a range of quantum fluid systems.
We design a neural network to extract and process features from absorption images taken of one-dimensional Bose gases in the quasicondensate regime. Specifically, the network is trained to predict both the temperature of single realizations of the system and the uncertainty thereof. For multiple realizations, the individual predictions can be combined in an estimate of the mean temperature, improving precision. We benchmark our model on both simulated and experimentally measured data and compare it to the established method of density ripple thermometry. We find the predictions of the two methods compatible, although the neural network reaches similar precision needing many fewer realizations, thus highlighting the efficiency gain achievable when incorporating neural networks into analysis of data from cold gas experiments. Further, we study feature maps to reveal which local features of the condensate are extracted by the network and how said features correlate with properties of the system. A similar analysis could be employed to uncover physical relations in more complex systems.
This thesis covers two fields of research carried out in the context of ultracold Fermi gases: Non-equilibrium physics in the BEC-BCS crossover and the use of machine learning methods in physics. \n\t \nThe BEC-BCS crossover is of particular interest for the investigation of low-temperature superconductivity and related many-body phenomena. We realise the BEC-BCS crossover experimentally by cooling a mixture of two hyperfine states of fermionic lithium atoms far below degeneracy. To tune the interaction between the constituents, we make use of a broad magnetic Feshbach resonance between the states. We achieve temperatures far in the degenerate regime that can be tuned by a trap release heating procedure. Being able to tune temperature and interaction allows us to access the whole BEC-BCS crossover regime and investigate related phenomena. \n\t \nOne key quantity across the crossover is the critical temperature of the thermal to superfluid phase transition that signals condensation of pairs. The critical temperature can be calculated in the limits of weak interactions, however, for strong interactions theoretical treatment is difficult. In this thesis, we use neural networks to explore the critical temperature in the whole crossover regime with unprecedented accuracy. We show that a supervisedly trained neural network can not only detect the critical temperature, but also distinguish between Cooper pairs and Feshbach molecules, which was not possible before. Moreover, we show that an unsupervisedly trained autoencoder network is able to determine the critical temperature without any external input other than time-of-flight pictures. \n\t \nAnother important quantity in the crossover regime is the superfluid gap parameter. We measure the gap by exciting the Higgs mode through radio-frequency modulation of the atom population. In the BCS limit, the frequency of the Higgs mode is given by twice the gap frequency. By observing the frequency dependent response of the system we can therefore measure the gap parameter. Towards strong interactions, the observed feature broadens, signalling an instability of the mode due to the gradual violation of the required particle-hole symmetry. \n\t \nExploring the non-equilibrium behaviour of a Fermi gas in the BEC-BCS crossover might hold the key to understand the formation dynamics of pairs and how to induce superconductivity in a system that does not exhibit superconductivity in its equilibrium state. In this work, we present two ways to expose the system to a sudden change of interaction, resulting in a non-equilibrium response. Firstly, we use a fast radio-frequency transfer of one of the hyperfine states into a different, third state. We observe dynamics on several timescales in the momentum distribution and condensate fraction. Secondly, we develop a new coil system that is able to change the magnetic field at the atom position by tens of Gauss within a few microseconds. We detail the construction, testing and alignment process and present investigations into the long-time response of the system. \n\t \nMachine learning procedures can not only be used for data analysis, but also to improve upon the control over experimental systems. In this work we present a novel approach to improve the detection and correction of experimental errors exponentially. We use two machine learning approaches, principal component analysis and small, fully connected neural networks to create an empirical model of the experimental setup. Using this model allows us to reduce the required amount of data for compensation of noise sources exponentially. We demonstrate both approaches in the context of micromotion compensation on a single trapped ytterbium ion in a radio-frequency Paul trap.
Single-shot images are the standard readout of experiments with ultracold atoms, the imperfect reflection of their many-body physics. The efficient extraction of observables from single-shot images is thus crucial. Here we demonstrate how artificial neural networks can optimize this extraction. In contrast to standard averaging approaches, machine learning allows both one- and two-particle densities to be accurately obtained from a drastically reduced number of single-shot images. Quantum fluctuations and correlations are directly harnessed to obtain physical observables for bosons in a tilted double-well potential at an extreme accuracy. Strikingly, machine learning also enables a reliable extraction of momentum-space observables from real-space single-shot images and vice versa. With this technique, the reconfiguration of the experimental setup between in situ and time-of-flight imaging is required only once to obtain training data, thus potentially granting an outstanding reduction in resources.
The power of machine learning (ML) provides the possibility of analyzing experimental measurements with a high sensitivity. However, it still remains challenging to probe the subtle effects directly related to physical observables and to understand physics behind from ordinary experimental data using ML. Here, we introduce a heuristic machinery by using machine learning analysis. We use our machinery to guide the thermodynamic studies in the density profile of ultracold fermions interacting within SU(N) spin symmetry prepared in a quantum simulator. Although such spin symmetry should manifest itself in a many-body wavefunction, it is elusive how the momentum distribution of fermions, the most ordinary measurement, reveals the effect of spin symmetry. Using a fully trained convolutional neural network (NN) with a remarkably high accuracy of ~94% for detection of the spin multiplicity, we investigate how the accuracy depends on various less-pronounced effects with filtered experimental images. Guided by our machinery, we directly measure a thermodynamic compressibility from density fluctuations within the single image. Our machine learning framework shows a potential to validate theoretical descriptions of SU(N) Fermi liquids, and to identify less-pronounced effects even for highly complex quantum matter with minimal prior understanding.
Laser cooling and trapping techniques are foundational to modern atomic physics, with applications including atomic clocks and sensors as well as quantum simulators and computers. These applications require stable, well-calibrated laser cooling and accurate determination of atom number and temperature. 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 snapshot 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 × 10 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. These results show that fluorescence images can provide fast, nondestructive diagnostics for cold-atom experiments and may enable real-time experimental feedback.
The power of machine learning (ML) provides the possibility of analyzing experimental measurements with a high sensitivity. However, it still remains challenging to probe the subtle effects directly related to physical observables and to understand physics behind from ordinary experimental data using ML. Here, we introduce a heuristic machinery by using machine learning analysis. We use our machinery to guide the thermodynamic studies in the density profile of ultracold fermions interacting within SU(N) spin symmetry prepared in a quantum simulator. Although such spin symmetry should manifest itself in a many-body wavefunction, it is elusive how the momentum distribution of fermions, the most ordinary measurement, reveals the effect of spin symmetry. Using a fully trained convolutional neural network (NN) with a remarkably high accuracy of ~94% for detection of the spin multiplicity, we investigate how the accuracy depends on various less-pronounced effects with filtered experimental images. Guided by our machinery, we directly measure a thermodynamic compressibility from density fluctuations within the single image. Our machine learning framework shows a potential to validate theoretical descriptions of SU(N) Fermi liquids, and to identify less-pronounced effects even for highly complex quantum matter with minimal prior understanding.