Machine learning algorithms for single-atom reconstruction in a quantum-gas microscope

Abstract

A quantum-gas microscope is a device that enables the detection of individual ultracold atoms in optical lattices with single-site resolution. In this context, the strontium quantum-gas microscope developed at ICFO is a promising platform for studying strongly correlated many-body states in the Fermi-Hubbard model with high spatial resolution. This work explores the use of a convolutional neural network (CNN) as a non-linear method to reconstruct the single-site occupation of the optical lattice. The neural network is based on unsupervised autoencoders, which learn a compact latent representation associated with lattice occupation while capturing the optical response of the imaging system. The comparison between the experimentally measured point spread function (PSF) and the learned PSF reveals excellent agreement, showing that the network captur the fundamental imaging response. Finally, the measured pinning fidelity, defined from the overlap between two consecutive occupation images, demonstrates that atoms remain trapped at the same site with high probability during imaging, while hopping and loss remain sufficiently suppressed.