Highly parallel simulation and optimization of photonic circuits in time and frequency domain based on the...

Area of Science:

• Photonics
• Computational Electromagnetics
• Machine Learning

Background:

• Photonic circuit simulations are crucial for designing integrated optical devices.
• Current simulation methods can be computationally intensive, limiting scalability.
• Optimization of photonic components often requires iterative and time-consuming processes.

Purpose of the Study:

• To develop a novel, efficient, and scalable method for simulating photonic circuits.
• To leverage deep learning for reimagining photonic circuit simulation.
• To enable straightforward optimization of photonic component parameters.

Main Methods:

• Utilizing the scatter matrix formalism for photonic circuit representation.
• Employing the PyTorch deep-learning framework to model circuits as neural networks.
• Implementing complex-valued, sparsely connected neural networks for simulations.

Main Results:

• Achieved highly parallelized simulations of large photonic circuits.
• Enabled simulations in both time and frequency domains.
• Facilitated easy optimization of individual component parameters using backpropagation.

Conclusions:

• The proposed deep-learning-based scatter matrix method offers a powerful new paradigm for photonic circuit simulation.
• This approach significantly enhances simulation speed and scalability.
• It integrates seamlessly with machine learning algorithms for efficient component design and optimization.