Faster State Preparation across Quantum Phase Transition Assisted by Reinforcement Learning

Area of Science:

• Quantum physics
• Quantum information science
• Atomic, molecular, and optical physics

Background:

• Adiabatic quantum state preparation is limited by system lifetime and the need to track instantaneous ground states.
• Achieving high-fidelity quantum states in many-body systems is crucial for quantum technologies.
• Previous methods for preparing Dicke states face challenges in balancing speed and accuracy.

Purpose of the Study:

• To develop a faster protocol for preparing quantum states in finite many-body systems.
• To utilize excited level dynamics and deep reinforcement learning (DRL) for enhanced quantum state preparation.
• To improve the fidelity and operational benchmark (interferometric sensitivity) of prepared quantum states.

Main Methods:

• Developed a multistep training scheme for deep reinforcement learning (DRL).
• Trained a DRL agent to control quantum dynamics by optimizing a faster sweeping policy.
• Used interferometric sensitivity as a benchmark for training in the presence of system loss.
• Implemented the protocol in a Bose-Einstein condensate of Rubidium-87 atoms.

Main Results:

• Achieved fidelity ≥99% for the target Dicke state in a fraction of the adiabatic time (without loss).
• Demonstrated improved interferometric sensitivity in approximately half the time compared to previous methods (with loss).
• Prepared a balanced three-mode Dicke state in a Bose-Einstein condensate within 766 ms.
• Observed an improved number squeezing of 13.02±0.20 dB.

Conclusions:

• Deep reinforcement learning offers a powerful approach for controlling quantum dynamics and preparing quantum states in interacting many-body systems.
• The developed DRL-based protocol significantly accelerates quantum state preparation compared to traditional adiabatic methods.
• This work highlights the practical potential of DRL for advancing quantum technologies, particularly in systems with inherent losses.