AI-Guided, Multi-Modal, Multi-Scale (AIM3S) Anomaly Detection in Multi-Dimensional STEM/EELS

C53-14b-271071The increasing data rates in conventional and scanning transmission electron microscopy, afforded by the developments of faster detectors and novel in-situ methods, requires automation of post- acquisition data analysis and, ideally, identification of regions of interest during experimental data acquisition. The overarching objective of this program is to leverage a suite of cutting-edge data- science approaches to decisively augment the ability of scientists to link complex chemical and electronic structure feature across time and space. The Phase I program successfully demonstrated that machine learning (ML) can enable automation of the image data analysis. Sivananthan Laboratories developed an image analysis approach using 2-dimensional neural network (NN)-based ML approaches to detect anomalous regions within STEM HAADF images. Specifically, synthetic (i.e., simulated HAADF images) and experimental atomic-resolution images of SrTiO3 bulk (001), with point defects and grain boundaries were used to train a NN and identify the position of crystal structure anomalies in the test image sets. Sivananthan Laboratories also demonstrated that sub-sampling and compressive sensing approaches can be integrated in the anomaly detection workflow. During the Phase II program, Sivananthan Laboratories will test and expand the current software, and integrate it into existing control suites for various TEM manufacturers. Specifically, during a Phase II program, the team at Sivananthan Laboratories proposes to test the stability of the anomaly detection approach on different crystal structures, including HgCdTe, one of the areas of expertise at Sivananthan Laboratories. Furthermore, the stability of the convolutional neural network (CNN) to various forms of noise and possible solutions to integrating noise reduction methods into the current CNN workflow will be developed. The overall goal of the Phase II program will be the development and integration of a software suite that can either be integrated in existing microscope control platforms, such as the Gatan Microscopy Suite or Protochips Axon, or be available as a stand-alone image acquisition and analysis platform. Ultimately, this anomaly detection software suite will be available for all major TEM manufacturers, but the initial development will be conducted using a JEOL ARM200CF at the University of Illinois at Chicago.