Speakers
Invited Speaker
Electron energy loss spectroscopy (EELS) is a powerful tool for examining elemental composition and local bonding information the scanning transmission electron microscope (STEM). As electron probes have become smaller, with a corresponding decrease in the illuminated volume, the signal to noise ratio (SNR) of atomic resolution EELS becomes an issue. This is due to the low cross sections for core-shell ionization, especially for higher energy ionization events. While in some cases it may be possible to increase the dwell times at each probe position, in general this results in damage to the specimen. This is particularly the case is defects and impurities exist, both features which are of particular interest in many materials systems. In this talk we will examine the denoising of STEM EELS data using convolutional autoencoders.
Perhaps the most common approach to denoising EELS data is principal component analysis (PCA) which represents the 3-dimensional data set as a weighted sum of orthogonal components. It is generally assumed that the components that are most common (i.e. more heavily weighted) represent the signal of interest, while less common components are assume to be related to random noise. By removing these lower weighted components, the aim is to produce “cleaned” spectra which are more easily interpreted. However less common components can also be due to isolated defects or impurities, or even interfaces. It is the EELS signal around these features that are usually of most interest.
An alternative approach for denoising data is the use of convolutional autoencoders (CA) which has been demonstrated in a number of different contexts. In this talk we will demonstrate their use on both simulate and experimental EELS data. An essential step in the use of CAs is the development of a suitable training set. This set must represent all the features observed in the dataset. We will demonstrate the pitfalls of this approach if the training set is poorly chosen. While we will use a simulated data set to generate our training data, other more general approaches will be discussed.
This effort (ML and STEM) is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES), Materials Sciences and Engineering Division (M.P.O., S.V.K.) and was performed and partially supported (M.Z.) at the Oak Ridge National Laboratory’s Center for Nanophase Materials Sciences (CNMS), a U.S. Department of Energy, Office of Science User Facility.
ORNL
Graduated from The University of Melbourne Ph.D. 1999 Theoretical Physics
1999- 2001 Research Fellow Grade 1 (Level A), School of Physics, The University of Melbourne
2002-2004 A.R.C. Research Fellow, School of Physics, The University of Melbourne
2005-2008 Postdoctoral Research Associate, Oak Ridge National Laboratory
2008-2015 Research Associate Professor, Department of Physics and Astronomy, Vanderbilt University, Nashville, TN.
2015-present R&D Staff, Oak Ridge National Laboratory, Oak Ridge, TN
Research Interests: Image simulation in the S/TEM. Phase retrieval. Structure determination. Electron energy loss simulations for core and low loss EELS.