Optimizing measurement-based cooling by reinforcement learning

Abstract: Conditional cooling-by-measurement holds a significant advantage over its unconditional (nonselective) counterpart in the average-population-reduction rate. However, it has a clear weakness with respect to the limited success probability of finding the detector in the measured state. In this work, we propose an optimized architecture to cool down a target resonator, which is initialized as a thermal state, using an interpolation of conditional and unconditional measurement strategies. An optimal measurement-interval $\tau_{\rm opt}^u$ for unconditional measurement is analytically derived for the first time, which is inversely proportional to the collective dominant Rabi frequency $\Omega_d$ as a function of the resonator's population in the end of the last round. A cooling algorithm under global optimization by the reinforcement learning results in the maximum value for the cooperative cooling performance, an indicator to measure the comprehensive cooling efficiency for arbitrary cooling-by-measurement architecture. In particular, the average population of the target resonator under only $16$ rounds of measurements can be reduced by four orders in magnitude with a success probability about $30\%$.