Faster State Preparation by Controlled Quantum Dynamics from Reinforcement Learning

Ground state transformation by crossing quantum phase transition can be achieved through adiabatic parameter change taking advantage of the inherent energy gap near quantum critical point from finite size effect of a many-body (MB) system. In real application scenarios, however, the efficacy of such an approach is compromised by the need to balance finite system lifetime with adiabaticity, as exemplified in a recent experiment that prepares three-mode balanced Dicke state near deterministically ^[1]. In this work, we employ reinforcement learning agent to optimize the tuning of quadratic Zeeman shift to control MB quantum dynamics for preparing balanced Dicke state. Benefiting from faster excited level dynamics, the learned non-monotonous control protocol facilitates much shorter sweeping time than previous protocols. Excluding atom loss, a theoretical fidelity ≥ 99% between prepared and the target Dicke state is achieved over a small fraction of the required time for adiabatic tuning. When loss is considered as a one-body stochastic effect, the RL agent remains effective by learning from the approximate dynamics based on the truncated Wigner method, leading to enhanced interferometric sensitivity within about half of the previously reported time. Implemented in a Bose-Einstein condensate of ∼104 87Rb atoms, the balanced three-mode Dicke state exhibiting an improved number squeezing of  dB is prepared within 766 ms, highlighting the potential of RL for quantum dynamics control and quantum state preparation in interacting MB systems ^[2].