The Library
Advances in electron microscopy with deep learning
Tools
Ede, Jeffrey M. (2021) Advances in electron microscopy with deep learning. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Ede_2021_Redacted.pdf - Submitted Version - Requires a PDF viewer. Download (676Kb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3598140~S15
Abstract
Following decades of exponential increases in computational capability and widespread data availability, deep learning is readily enabling new science and technology. This thesis starts with a review of deep learning in electron microscopy, which offers a practical perspective aimed at developers with limited familiarity. To help electron microscopists get started with started with deep learning, large new electron microscopy datasets are introduced for machine learning. Further, new approaches to variational autoencoding are introduced to embed datasets in low-dimensional latent spaces, which are used as the basis of electron microscopy search engines. Encodings are also used to investigate electron microscopy data visualization by t-distributed stochastic neighbour embedding. Neural networks that process large electron microscopy images may need to be trained with small batch sizes to fit them into computer memory. Consequently, adaptive learning rate clipping is introduced to prevent learning being destabilized by loss spikes associated with small batch sizes.
This thesis presents three applications of deep learning to electron microscopy. Firstly, electron beam exposure can damage some specimens, so generative adversarial networks were developed to complete realistic images from sparse spiral, gridlike, and uniformly spaced scans. Further, recurrent neural networks were trained by reinforcement learning to dynamically adapt sparse scans to specimens. Sparse scans can decrease electron beam exposure and scan time by 10-100× with minimal information loss. Secondly, a large encoder-decoder was developed to improve transmission electron micrograph signal-to-noise. Thirdly, conditional generative adversarial networks were developed to recover exit wavefunction phases from single images. Phase recovery with deep learning overcomes existing limitations as it is suitable for live applications and does not require microscope modification. To encourage further investigation, scientific publications and their source files, source code, pretrained models, datasets, and other research outputs covered by this thesis are openly accessible.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > Q Science (General) Q Science > QC Physics |
||||
Library of Congress Subject Headings (LCSH): | Electron microscopy -- Data processing, Machine learning -- Mathematical models, Data sets | ||||
Official Date: | February 2021 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Physics | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Sloan, Jeremy ; Beanland, R. | ||||
Sponsors: | Engineering and Physical Sciences Research Council | ||||
Format of File: | |||||
Extent: | 1 volume (various pagings) | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year