Dissertations, Theses, and Capstone Projects

Date of Degree

2-2023

Document Type

Dissertation

Degree Name

Ph.D.

Program

Physics

Advisor

Pouyan Ghaemi

Advisor

Jiadong Zang

Committee Members

Eugene M. Chudnovsky

Viviana Acquaviva

Karl G. Sandeman

Subject Categories

Condensed Matter Physics | Data Science | Physics

Keywords

3D Magnetic Reconstruction, Electron Holography, Machine Learning, Skyrmion, Unet, cycleGAN

Abstract

Revealing three-dimensional (3D) magnetic textures with vector field electron tomography (VFET) is essential in studying novel magnetic materials with topologically protected spin textures potentially being used in the next-generation semiconductor industry. In this dissertation, we use machine learning (ML) models to reconstruct 3D magnetic textures from electron holography (EH) data.

We can feed the EH data, a series of two-dimensional (2D) phasemaps, into a neural network (NN) architecture directly or feed the EH data into a conventional VFET and then feed the reconstructed results into a NN. Thus, perceptive NN, either a simple convolutional neural network (CNN) or Unet architecture, is built and used to reconstruct the 3D magnetic texture. We demonstrate that the magnetic vector potential and magnetic induction field can be successfully reconstructed with an end-to-end Unet-based ML model. Also, reconstruction results of conventional VFET can be significantly enhanced with a plug-and-play Unet attached to it. The scaling law for run time versus dataset size is studied. Reconstruction results of EH data with various defects, such as noise, sparsity, misalignment, and missing wedge, are also discussed in the frame of ML models with Unet architecture.

Furthermore, a generative model is introduced to reconstruct the magnetization to solve the missing information of scalar potential that EH cannot probe. Integrating the cycle consistency and a forward model from magnetization to EH phasemap, we build a cycle consistency generative adversarial network (cycleGAN) based generative model that gives impressive reconstruction results of magnetization. This cycle consistency with a forward model generative model framework is also a promising solution for other inverse problems with an explicit forward model.

Share

COinS