Different Machine Learning-Based Algorithms Enabling Inverse Designs Of Photonic Structures: A Physical Approach, A Weakly Adversarial One With Help Inspired By Transfer Learning, And A Distance Metric Learning Based One
Loading...
Author(s)
Bao, Daqian
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Physics-informed Neural Network (PINN) model and its derivatives, altogether,
build a new category of machine learning (ML) problems: scientific machine learning
(SciML). SciML had been active on solving different problems in engineering and applied
science including thermodynamics, fluid mechanics and optics and photonics. The thesis
focuses on the development of SciML methods that enable the inverse design of photonic
devices and eventually, photonic metasurfaces (MS), a very prosperous and prospective
subject in optics and photonics.
This thesis focuses on SciML methods that successfully predict the transmission
pattern of light through photonic devices; in particular, fully connected (FC) simple neural
networks (NN) successfully predicts the transmission pattern of light as planar, timeinvariant wave, through parallel stacked multi-layered thin-film filters by incorporation of
continuity of electric field and magnetic field accompanying the transmission of light. This
thesis also shows the potency of other formulations of derivatives of PINN, particularly,
the weak adversarial network (WAN). The network structure, along with the incorporation
of continuity, sets a new paradigm in photonic inverse design: different from traditional
adjoint optimization (AO) using finite element method (FEM), in particular, finite
difference time domain (FDTD) methods to produce transmission pattern of the light, the
NN based paradigm is better supported by GPU acceleration tool that facilitates the floating
point operations, thus the prediction (or reconstruction, used interchangeably hereafter) of
transmission patterns.
The thesis also covers a ML based method, distance metric learning (DML), that can help the inverse design by modeling different types of responses of different types of
photonic devices. DML can aid the inverse design process by picking the designs that yields peak intensity at desirable wavelength of different types of responses that have different properties, like FWHM.
Altogether, the thesis paves a way for a new inverse design pipeline built with
SciML method that, if work adjointly with DML, can select the proper design of photonic device that produce desirable responses.
Sponsor
Date
2022-12-16
Extent
Resource Type
Text
Resource Subtype
Thesis