Multi-directional beam steering using diffractive neural networks.

Area of Science:

• Optics
• Computational Physics
• Machine Learning

Background:

• Machine learning resurgence prompts novel approaches to physical system simulation.
• Deep neural networks show promise in mimicking complex physical phenomena like wave propagation.
• Diffractive deep neural networks (D2NN) offer new insights into wave propagation and diffractive element design.

Purpose of the Study:

• To derive an efficient, GPU-friendly diffractive deep neural network (D2NN) methodology.
• To utilize the D2NN implementation for virtually creating optical phase mask cascades.
• To investigate the beam steering capabilities of D2NNs under various conditions.

Main Methods:

• Developed a D2NN methodology based on Rayleigh-Sommerfeld diffraction.
• Employed GPU acceleration for efficient computation.
• Trained D2NN instances using input/output conditions from electro-optic modulated waveguide systems.
• Analyzed beam steering efficacy across 27 D2NN instances with varied parameters.

Main Results:

• Successfully implemented an efficient GPU-friendly D2NN for optical simulations.
• Virtually generated cascades of optical phase masks capable of beam steering.
• Demonstrated the D2NN's capability to mimic wave propagation and design diffractive elements.
• Analyzed the performance of 27 D2NN configurations for beam steering.

Conclusions:

• The derived D2NN methodology provides an efficient tool for simulating wave propagation.
• D2NNs can be used to discover and design multi-functional diffractive elements.
• The approach facilitates the virtual design and analysis of optical systems for beam steering.