Multimode optical fiber transmission with a deep learning network.

Area of Science:

• Optics
• Machine Learning
• Fiber Optics

Background:

• Multimode fibers (MMFs) are highly scattering media that scramble light.
• Applications like imaging and projection through MMFs require understanding input-output relationships.
• Scrambled light patterns in MMFs typically lose phase information.

Purpose of the Study:

• To demonstrate that a deep neural network can learn the input-output relationship in a multimode fiber.
• To show that a deep convolutional neural network (CNN) can learn nonlinear relationships between input and output light patterns.
• To achieve high-fidelity image reconstruction and projection through MMFs using a trained neural network.

Main Methods:

• Utilizing a deep convolutional neural network (CNN) to learn the nonlinear mapping between input light and output speckle patterns.
• Training the CNN on amplitude information from the output speckle pattern, effectively performing a nonlinear inversion task.
• Evaluating image reconstruction and projection performance by comparing results with those obtained using a fully characterized system transmission matrix.

Main Results:

• Achieved image fidelities as high as ~98% for reconstruction and ~94% for image projection through a 0.75m MMF.
• Demonstrated the CNN's ability to learn the nonlinear input-output relationship, even with lost phase information.
• Showcased successful transfer learning, enabling the network to transmit images from classes not used during training.

Conclusions:

• Deep neural networks, specifically CNNs, can effectively learn and invert the complex light propagation dynamics in multimode fibers.
• High-fidelity image reconstruction and projection are achievable through scattering media like MMFs using learned nonlinear mappings.
• The developed approach offers a powerful method for optical wavefront shaping and image transmission in complex optical systems.