Machine Learning Phase Field Reconstruction in a Bose-Einstein Condensate Preprint: APS/123-QED
Jackson Lee, Andrew J Millis · 2026
Read Machine Learning Phase Field Reconstruction in a Bose-Einstein Condensate Preprint: APS/123-QED by Jackson Lee, Andrew J Millis (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09528.txt.
Introduction
Lee and Millis tackle a blind spot in cold-atom imaging. Ordinary snapshots reveal only density, while the phase field stays hidden. That field exposes vortices and confirms superfluidity. The authors train a U-Net model on simulated two-dimensional condensate states to infer phase gradients from density alone. A second model locates vortex cores. Graphical post-processing then reads each vortex's quantized charge. The combination reconstructs the full phase field with high accuracy.
Catalog record
| CREATOR | Jackson Lee, Andrew J Millis |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40339 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09528
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine Learning Phase Field Reconstruction in a Bose-Einstein Condensate Preprint: APS/123-QED Jackson Lee, Andrew J Millis Abstract A basic challenge in experimental physics is the extraction of information related to variables that are not directly measured. The challenge is particularly severe in quantum systems where one may be interested in correlations of operators that are not diagonal in the measurement basis. In this paper we take a step towards addressing this issue in the context of Boson superfluids, where standard in-situ imaging yields only the spatially resolved density, leaving the phase field—crucial for identifying topological defects such as vortices and confirming superfluidity—indirectly encoded. Previous work has shown that the location of vortices in the phase field may be detected, but has not solved the problems of fully reconstructing the phase or identifying the charge (vortex vs. antivortex). This paper shows that a combination of a deep machine learning (ML) model and classical computer vision post-processing steps can address this gap. We use realistic snapshots of the thermal state of a two-dimensional BEC in a harmonic trap using synthetic data obtained from projected Gross-Pitaevskii equation simulations to train a U-Net-based architecture to infer the absolute values of the phase field gradients from an observed density field, and then employ a separate ML model to locate the positions of the vortex cores and a post-processing graphical analysis to determine with high accuracy the phase field, including the quantized charge of each vortex. I Introduction Over the past two decades, advances in cooling and trapping techniques have enabled unprecedented experimental control over quantum systems, enabling the creation of remarkable quantum states of matter 20 ; 1 ; 27 ; 11 ; 6 ; 48 ; 15 ; 12 ; 40 ; 4 ; 10 ; 25 ; 38 ; 17 . However, inferring quantum correlations remains an experimental challenge. Available imaging techniques mainly work by projection of the quantum state onto a particular (typically density) basis 2 ; 24 ; 28 ; 37 . Impressive advances provide very fine-grained information on atom positions, but obtaining information about other correlations, including phase and spin variables, that are conjugate to (non-commuting with) the more straightforwardly measured variables, remains challenging. One possible route is via ’Hamiltonian learning’ 46 ; 32 ; 21 , in which a many-body Hamiltonian is inferred from short time dynamics measurements and the inferred Hamiltonian is then used to compute the desired correlations. However, it is desirable to have a method that learns correlations directly from ”snapshot” data obtained by projection onto one basis at a given time, and does not require measurements of dynamics. An argument that such a reconstruction may be possible in principle follows from the logic underpinning the Hohenberg-Kohn theorem 23 of electronic structure. This theorem says that the ground state wave function of a quantum system is fully determined by the particle density (space dependent one particle density matrix), if the interparticle interaction is specified. If such a relationship exists, then it may be possible to use AI methods to learn the functional relationship and predict other correlations from data. While many papers have attempted to learn the Hohenberg-Kohn density functional of electronic structure theory 42 ; 36 ; 35 ; 3 ; 31 ; 13 , the problem of reconstructing quantum correlations directly from measured densities has received less attention. In this paper we apply this idea in the context of Bose–Einstein condensates (BECs). Formed by cooling a dilute gas of bosons to ultralow temperatures, BECs exhibit striking collective phenomena such as long-range coherence, superfluidity, and quantized circulation 14 , all of which relate to the correlations of the quantum mechanical wave function phase, which is conjugate to the particle density. Typical time-of-flight (TOF) expansion or in-situ imaging experiments extract the density profile with high accuracy, but in these experiments the phase of the wave function is not directly observed 24 ; 9 ; 14 . Experimental efforts to extract phase information rely on interference between different parts of a condensate 22 , or between different condensates 49 ; 43 . Theoretical approaches, typically relying on TOF information, have also been developed. These techniques rely on using many repetitions of an experimental setup 44 ; 33 , several measurements in-situ and in the course of a TOF expansion 45 ; 50 , or TOF measurements with knowledge of in-situ support constraints 29 . These approaches typically require expensive iterative schemes, have limited spatial resolution, and may require several snapshots in a single run, which is difficult when the measurements are destructive. A particularly important feature of the phase field of a BEC is its topological defects— vortices and antivortices—which appear as density depletions around which the phase winds by an integer multiple of 2\pi 14 . The configuration, dynamics, and interactions of these defects play a central role in two-dimensional superfluids, particularly in the context of vortex–antivortex unbinding across the Kosterlitz–Thouless transition 5 ; 30 . As a result, experimental access to the condensate phase is essential for characterizing superfluid order and identifying distinct phases of matter. Recent work has applied machine learning and computer vision techniques to identify vortex cores within noisy density images with high accuracy 47 ; 34 . However, these methods generally stop at locating vortex positions: they do not reconstruct the full phase field, nor distinguish vortices from antivortices (unless given the corresponding phase profile, as in 34 ). In this paper, we present a machine-learning-based approach that successfully reconstructs the phase field of a superfluid from measurements of the density. Attempts to use a neural network to learn the full (signed) phase field were not successful. The method proposed here is a two-stage approach in which a U-Net convolutional neural network architecture, known for its power in image-to-image translation, is used to determine the magnitude of the gradients of the phase field. The signed phase gradients are then determined using a classical post-processing method. The U-Net was trained on synthetic data obtained from numerical solutions of the Projected Gross Pitaevskii Equation (PGPE) (representing the long-wavelength modes of the system) combined with a uniformly distributed random density (modelling the incoherent, short-wavelength modes). We note that in the absence of the incoherent background the vortex cores are points of vanishing density, so vortices may easily be identified; the presence of the incoherent background means the vortex cores are simply regions of lower than average density, making the vortex identification problem significantly more challenging. The U-Net architecture identifies lines where \partial_{x}\phi=0 and another set of lines where \partial_{y}\phi=0 , where \phi is the phase; these lines bound domains where a component of \nabla\phi has a definite sign. The remaining problem is to identify the signs by ”coloring” the domains consistently; this problem is complicated by the fact that the ML-predicted borders are returned as probabilities, and are not guaranteed to be always correct. We introduce a classical post-processing that finds a configuration of the sign of \partial_{x}\phi that maximally respects the ML-predicted borders. The rest of this paper is organized as follows. Section II presents the needed background, including the theoretical definition of the Bose condensate problem and the PGPE approach to generating synthetic data, as well as general features of the machine