Atom interferometry improved by neural networks

For applications in metrology, it is important to both estimate the parameters of interest from data and characterize the error in those estimates. Here, we present a machine learning-based method for model-free inference of physical parameters from an interferometer data. We consider estimating quantities such as acceleration and rotations from interference patterns generated by an atom interferometer without the need for an exact mathematical model of the device and the error processes affecting it. Imperfections in the model, systematic errors, and noise severely limit the performance. Our method, based on neural network, learns to simultaneously estimate the quantities of interest and the error in those estimates from noisy input images. It extends the applicability of the interferometer when the resolution is limited, and noise and imperfections are present.