Machine learning techniques in image-based 3D reconstruction of atom clouds for atom interferometry

The next generation of atom interferometers strive for longer interrogation times and improved sensitivities, motivating more precise control and characterization of systematic effects. Accurate knowledge of the atomic density distribution in the interferometer output ports is essential for understanding the spatial dependence of the interference pattern, as well as diagnosing phase errors that arise from imperfections in optical elements and inhomogeneities in the laser wavefront. Imaging these clouds on the beam propagation axis is ideal, but unrealistic in our system due to limited optical access. As a solution, we present computational imaging techniques for reconstructing 3D atom clouds from a set of off-axis images. Our approach combines differentiable ray tracing with neural rendering using Sinusoidal Representation Networks (SIREN), producing volumetric atomic densities from 2D images. We benchmark the reconstruction fidelity and demonstrate reasonable recovery of cloud geometry for a variety of cloud shapes. These techniques provide a framework for imaging atomic ensembles and offer new tools for precision diagnostics in atom interferometry.