masonlw

The increasing ability to perform high throughput electron microscopy has created a need for robust, automated analysis that appears addressable by machine learning (ML) tools. Approaches such as convolutional neural nets (CNNs) are finding increased application in scientific data analysis tasks, including analysis of electron microscopy imaging data. However, electron microscopy data varies significantly from natural images. To provide robust ML analysis for increasingly large and complex electron microscopy datasets, we have performed experimental and simulated studies to examine how experimental variation should be included in a training dataset for robust performance and maximum accuracy across workflows. We have also analyzed the role that network architecture plays, and present methods to increase performance of lightweight networks. Ultimately, our tools advance performance of ML workflows that can shed light on structural variation and correlation of large populations of nanomaterials from imaging data.

Many Microscopy techniques have been developed to explore atomic structures of material and biological specimens at a nano-scale. Coherent Diffractive Imaging (CDI) and Atomic Electron Tomography (AET) are the most powerful among these techniques. Mathematics, especially variational optimization, also follows and supports microscopy, helping to solve various image processing problems such as deblurring and denoising. In our latest research, variational methods can help obtain super-resolution ptychography images, marking a substantial improvement in computational microscopy and bypassing the resolution limit. While sub-pixel shifting and structured probes are two crucial keys to the super-resolution problem, total variation and l1 regularization play a significant role in reconstruction. In another research project, we explore the magnetic vector fields created in the vacancy between atoms of a magnetized material. Our analysis shows that the vector tomography needs in-plane rotation and constraint support to guarantee reconstruction. The sparsity inducing l1 minimization is the right tool to generate highly accurate support. Our variational methods have successfully solved a wide range of image processing problems. We will continue to utilize these techniques again in our upcoming research.

Magnetic systems exhibiting chirality have gained significant attention over the past decade. From enabling the efficient movement of magnetic domain walls to stabilizing the topological magnetic quasiparticles known as skyrmions, chirality and chiral interactions may play a central role in the next generation of magnetic memory and computing schemes. The possible responses of chiral magnetic systems to stimuli such as magnetic fields, electrical currents, light, and thermal gradients is at the forefront of thin-film magnetismresearch [1]. I will briefly introduce the underlying interactions that give rise to chiral magnetism, the emergence of topology and discuss some of the surprising systems where chirality and topology are emerging.  I will focus on the various techniques that can be combined to image and understand the three-dimensional magnetic structures that emerge.
[1]. Hellman et al., Reviews of Modern Physics 89, 025006 (2017).

The discovery and analysis of X-ray diffraction from crystals by Max von Laue, William Henry Bragg and William Lawrence Bragg in 1912 marked the birth of crystallography. Over the last century, crystallography has been fundamental to the development of many fields of science. However, many samples in physics, chemistry, materials science, nanoscience, geology, and biology are non-crystalline, and thus their 3D structures are not accessible by traditional crystallography. Overcoming this hurdle has required the development of new structure determination methods. In this talk, I will present two methods that can go beyond crystallography: coherent diffractive imaging (CDI) and atomic electron tomography (AET). In CDI, the diffraction pattern of a non-crystalline sample or a nanocrystal is first measured and then directly phased to obtain an image. The well-known phase problem is solved by combining the oversampling method with iterative algorithms. In the first part of the talk, I will briefly discuss the principle of CDI and highlight its capability of direct observation of 3D topological spin textures and their interactions in a ferromagnetic superlattice. In the second part of the talk, I will present a general tomographic method, termed AET, for 3D structure determination of crystal defects and disordered materials at the single atomic level. By combining advanced electron microscopes with powerful computational algorithms, AET has been used to reveal the 3D atomic structure of crystal defects and chemical order/disorder and to precisely localize the 3D coordinates of individual atoms in materials without assuming crystallinity. The experimentally measured coordinates can then be used as direct input for quantum mechanical calculations of material properties such as atomic spin and orbital magnetic moments and local magnetocrystalline anisotropy. As coherent X-ray sources and powerful electron microscopes are under rapid development around the world, we expect that CDI and AET will find broad applications in both the physical and biological sciences.