Autonomous alignment of a Mach-Zehnder interferometer via deep reinforcement learning: a foundation for resilient space-based and ground-based optical control

Abstract

This paper presents a Deep Reinforcement Learning (DRL) framework for the autonomous alignment of a Mach-Zehnder Interferometer (MZI). Utilizing a Proximal Policy Optimization (PPO) algorithm and a Convolutional Neural Network (CNN) policy, the agent manages two-degree-of-freedom mirror adjustments to achieve a null-fringe condition. The system employs an observation space of Fourier-domain difference images, allowing the agent to perceive the trajectory of spatial frequency shifts. A reward function based on the negative moment of inertia of Fourier frequency components—identified via K-means clustering—guides the agent from coarse misalignment to high-precision overlap. To ensure robust convergence, the agent was trained using a multiterm reward function and subjected to penalties for beam-loss events. We demonstrate that the learned policy is highly resilient to non-stationarity, successfully recovering alignment following stochastic ”table-kick” perturbations. This work establishes a scalable pathway for autonomous, safety-aware alignment in remote or unstable optical environments.