Employing deep neural networks in the analysis of visual data in ultracold atoms experiments

Deep learning reconstruction models enable the analysis of noisy physical data with unparalleled accuracy. These tools prove to be extremely useful when analyzing absorption imaging signals that suffer from structural structured noise due to slow changes in the illumination. Most of this noise can be removed by taking two successive exposures. Even then, some noise remains. Here we demonstrate the powerful advantage of deep learning in removing structured noise without a need for a second exposure. Specifically, we present an analysis of condensate fraction measurements taken with a superfluid Fermi gas, which is especially sensitive to background noise. The denoising of the spatial distribution obtained by absorption imaging uncovers an intriguing dynamical behavior of a periodically driven gas. To distill the data from the noise, we have developed a method that relies on a single exposure and an image-completion autoencoder neural network. The autoencoder is trained to reconstruct the noise from the information in the area surrounding the data, thus generating an ideal reference frame. Subtraction of this reference image provides a clean signal and thus higher accuracy, which enables deeper insight into the underlying physics.