Machine Learning and Photonic Devices

Dr. Keisuke Kojima, Fellow, IEEE, Fellow, Optica

This talk will provide an overview of deep learning applications in nanophotonic device design, focusing on generative neural networks. To achieve inverse design in nanophotonics, optimization of tens of thousands of ‘pixels’ is typically required. The adjoint method, a popular local optimization approach, often necessitates multiple optimization runs. Generative deep learning builds on existing data to generate new designs with specified target specifications such as transmission/reflection spectra. For instance, datasets optimized for discrete wavelengths (e.g., wavelength splitters) or splitting ratios (e.g., power splitters) can be used to generate devices with arbitrary wavelength or splitting ratios. We demonstrate examples using conditional variational autoencoders (CVAE) and denoising diffusion probabilistic models (DDPM) for applications in planar waveguide devices, metasurface gratings, and plasmonic gratings. Additionally, we introduce the concept of latent space optimization and transfer learning.