Physics-informed neural networks for exponential readout improvement in release-interference cold-atom sensors

We propose a physics-informed neural network (PINN) pipeline that provides readout precision gains for release-interference cold-atom sensors, such as those utilizing the Sagnac effect or ring-trap devices. This pipeline is particularly helpful when the experiment is repeatable and integrated with a supervisory control block that hyper-fine tunes parameters between runs. In the best case, the uncertainty shrinks exponentially with the number of repeats, which means one can achieve a target precision with only logarithmically many shots.

In our case study, we examine the double-target BEC rotation sensor, in which each target combines a central disk BEC and a concentric ring BEC. We initialize one unit of flow in the top ring, raise a potential barrier in the overlap region, then read out whether flow transfers, which places a bound on the frame rotation rate via release interference. We then train a PINN using 2D rotating-frame GPE dynamics and an imaging noise model to infer phase winding, barrier-induced transfer, and confidence scores from readout images. A control block ingests each PINN posterior and selects the next experimental parameters, so the loop remains robust to complex, noisy images while concentrating information across shots. We show exponential precision improvement by pooling PINN posteriors over an array of interferometers to shrink the rotation interval.