Analyzing modal power in multi-mode waveguide via machine learning.

Area of Science:

• Photonics and Optical Engineering
• Machine Learning Applications
• Integrated Optics

Background:

• Modal power distribution is crucial for optical signal analysis in waveguides.
• Existing methods for modal power analysis can be complex and time-consuming.
• Integrated multi-mode waveguides present unique challenges for mode analysis.

Purpose of the Study:

• To develop and validate a machine learning-assisted scheme for analyzing modal power in integrated multi-mode waveguides.
• To correlate far-field diffraction patterns with modal power distribution using convolutional neural networks (CNNs).
• To assess the accuracy and efficiency of specialized and full-scale CNN models for different waveguide configurations.

Main Methods:

• Training convolutional neural networks (CNNs) to analyze far-field diffraction intensity patterns.
• Developing a specialized CNN for thin optical waveguides (single-moded along one axis, multi-moded along the other).
• Cross-validating results with a full-scale CNN and benchmarking prediction accuracy using statistical error metrics.

Main Results:

• The specialized CNN achieved satisfactory accuracy for thin optical waveguides.
• A full-scale CNN showed similar accuracy but doubled training time.
• The approach demonstrated robustness and maintained performance in generalized multi-mode waveguides, even with noise.

Conclusions:

• Machine learning, specifically CNNs, offers an effective method for modal power analysis in integrated optical waveguides.
• The proposed scheme enables high-fidelity information recovery from far-field data.
• This technique holds significant potential for applications in integrated and fiber-based spatial-division demultiplexing.