Efficient phase identification in coherent beam combination using interpretable deep learning.

Area of Science:

• Optics and Photonics
• Laser Physics
• Artificial Intelligence in Physics

Background:

• Coherent beam combination (CBC) is crucial for high-power fiber lasers, but requires precise phase stabilization.
• Deep learning shows promise for single-step phase retrieval from interference patterns.
• Challenges remain in deep learning model interpretability and optimal imaging system placement for CBC.

Purpose of the Study:

• To investigate the impact of axial imaging system position on phase prediction accuracy in a simulated CBC system.
• To identify optimal regions within interference patterns for enhanced phase sensitivity and retrieval efficiency.
• To develop a lightweight deep learning model for rapid and accurate phase retrieval in CBC.

Main Methods:

• Emulation of a CBC system using a spatial light modulator.
• Systematic evaluation of phase prediction accuracy at varying axial positions.
• Identification of high phase-sensitivity regions in interference patterns.
• Development and application of a lightweight fully connected neural network for phase retrieval.

Main Results:

• Phase retrieval efficiency is significantly improved by focusing on high phase-sensitivity regions.
• A substantial reduction in input data size is achieved.
• A lightweight fully connected neural network attained a phase prediction error of ~λ/60.
• An inference rate of ~35 kHz was achieved for a 7-beamlet hexagonal close-packed array.

Conclusions:

• Optimizing imaging system position and utilizing phase-sensitive regions are key to enhancing deep learning-based CBC phase retrieval.
• Lightweight neural networks can achieve high accuracy and speed for CBC phase stabilization.
• This approach offers a viable solution for power scaling in fiber laser systems through improved phase control.