Design of a Matter-Wave Gyroscope with Reinforcement Learning

As the complexity for the next generation of quantum sensors increases, it is intriguing to consider a new paradigm in which the design of metrological devices is supported by machine learning. As a demonstration of such philosophy, we apply reinforcement learning to design a shaken-lattice matter-wave gyroscope involving minimal human intuition. That is, the machine is given no instructions to construct the splitting, reflecting, and recombining components used in conventional interferometers. Instead, we assign it the task of optimizing the sensitivity of the gyroscope to rotational signals by shaking the lattice. The machine-learned protocol is completely different from the typical sequence used in a Mach-Zehnder or Bragg matter-wave interferometer, and provides significant improvement in sensitivity compared to the conventional protocol.