High-performance mode decomposition using physics- and data-driven deep learning.

Area of Science:

• Optical Fiber Communications
• Computational Physics
• Machine Learning Applications

Background:

• Mode decomposition (MD) is crucial for analyzing few-mode fibers (FMFs).
• Traditional deep learning (DL) methods for MD face challenges with generalization and error fluctuation.
• Physics-guided learning is needed to improve neural network (NN) performance in optical systems.

Purpose of the Study:

• To introduce a novel physics- and data-driven deep-learning (PDDL) method for complete mode decomposition in FMFs.
• To enhance the robustness, adaptability, and generalization ability of NNs in MD tasks.
• To validate the PDDL method theoretically and experimentally for accurate modal analysis.

Main Methods:

• Developed a PDDL scheme integrating FMF beam propagation models into NNs.
• Guided NNs to learn essential physical features and discard unphysical ones.
• Investigated the method for obtaining real modal weights (ρ2) and relative phases (θ) in theory and experiment.

Main Results:

• The PDDL method overcomes generalization defects and reduces error fluctuations compared to traditional DL-based MD.
• Achieved 12x and 100x error reduction in ρ2 and θ for 8-mode FMFs with dissimilar beam patterns.
• Maintained high accuracy (error < 0.43% for ρ2, < 2.08% for θ) with noise and fiber parameter variations without retraining.
• Experimental validation showed >98% correlation between real and reconstructed beam patterns.

Conclusions:

• The PDDL method offers superior accuracy and adaptability for mode decomposition in FMFs.
• Demonstrated significant improvements over existing DL methods, especially for complex scenarios.
• The PDDL scheme holds substantial potential for practical modal coupling characterization and laser beam quality analysis.