[Paper Review] Reinforcement-learning based matterwave interferometer in a shaken optical lattice

Liang-Ying Chih, Murray Holland|arXiv (Cornell University)|Jun 21, 2021

Cold Atom Physics and Bose-Einstein Condensates38 references33 citations

TL;DR

This paper proposes a reinforcement learning (RL)-based design for a matterwave interferometer in a shaken optical lattice, using RL to optimize lattice shaking protocols that emulate beam splitters, mirrors, and recombiners. The method achieves higher sensitivity to acceleration than standard Bragg interferometry, demonstrating RL's potential for creating high-performance quantum sensors in complex, multi-path systems.

ABSTRACT

We demonstrate the design of a matterwave interferometer to measure acceleration in one dimension with high precision. The system we base this on consists of ultracold atoms in an optical lattice potential created by interfering laser beams. Our approach uses reinforcement learning, a branch of machine learning, that generates the protocols needed to realize lattice-based analogs of optical components including a beam splitter, a mirror, and a recombiner. The performance of these components is evaluated by comparison with their optical analogs. The interferometer's sensitivity to acceleration is quantitatively evaluated using a Bayesian statistical approach. We find the sensitivity to surpass that of standard Bragg interferometry, demonstrating the future potential for this design methodology.

Motivation & Objective

• To develop a high-precision matterwave interferometer for one-dimensional acceleration sensing using ultracold atoms in an optical lattice.
• To overcome limitations of intuitive or classical optimization in designing complex quantum control protocols for multi-path interferometers.
• To apply model-free reinforcement learning to discover non-intuitive, high-fidelity control protocols for beam splitters, mirrors, and recombiners in a shaken lattice.
• To evaluate the interferometer's performance using Bayesian statistical analysis to quantify sensitivity to acceleration.
• To demonstrate that RL-designed protocols can outperform conventional Bragg interferometry in sensitivity.

Proposed method

• Uses model-free deep reinforcement learning (DDQN) to learn optimal time-dependent phase modulation functions φ(t) for shaking the optical lattice.
• Defines states as matrix elements of the time-evolution operator in the relevant momentum subspaces (e.g., |±4ℏkL⟩).
• Defines actions as discrete amplitudes of sinusoidal phase modulation at the characteristic frequency 12ωr.
• Employs a reward function based on fidelity to target unitary operations (e.g., beam splitter or mirror operations) to guide learning.
• Trains agents using the double deep Q-network (DDQN) algorithm with Adam optimizer and experience replay.
• Validates performance using Bayesian statistical inference to quantify sensitivity to acceleration.

Experimental results

Research questions

• RQ1 Can reinforcement learning discover high-fidelity control protocols for matterwave interferometer components (beam splitter, mirror, recombiner) in a shaken optical lattice?
• RQ2 Does the RL-designed interferometer achieve higher sensitivity to acceleration than standard Bragg interferometry?
• RQ3 Can model-free RL outperform human-intuition-based or classical optimization methods in designing complex, multi-path quantum interferometers?
• RQ4 How does the choice of control frequency (e.g., 12ωr) affect the performance of the mirror and beam splitter components?
• RQ5 What is the achievable sensitivity of the RL-designed interferometer, and how does it compare quantitatively to established benchmarks?

Key findings

• The RL-designed interferometer achieves sensitivity to acceleration that surpasses that of standard Bragg interferometry, demonstrating a key performance advantage.
• The beam splitter and mirror components were successfully learned with channel fidelities reaching up to 0.8 under fixed-amplitude modulation, and further improved via RL optimization.
• The optimal shaking protocols were found to be non-intuitive, highlighting the ability of RL to discover novel solutions beyond human design bias.
• The use of a Bayesian statistical framework confirmed the high precision of the interferometer, with sensitivity quantified in terms of the minimum detectable acceleration.
• The method successfully leveraged the multi-path nature of the shaken lattice by learning control strategies that exploit the rich dynamics of Bloch states.
• Hyperparameter tuning (e.g., learning rate α=0.001, γ=0.99) led to stable training and convergence within 8,000–20,000 episodes, depending on the task.