AI-guided, multi-modal, multi-scale (AIM3S) anomaly detection in multi-dimensional STEM/EELS

The increasing data rates in conventional and scanning transmission electron microscopy provided by the development of faster detectors and novel in-situ methods requires the automation of post-acquisition data analysis and, ideally, the identification of regions of interest during experimental data acquisition. Sivananthan Laboratories will develop a software suite that can analyze multi-dimensional data cubes, such as atomic-resolution electron energy- loss spectrum images combined with high-angle annular dark field images, based on neural networks that have been previously trained on simulated and existing experimental data sets. The neural networks will be further trained during data acquisition using user-identified sample areas. This will enable the detection of structural and chemical anomalies, including point defects such as vacancies, interstitials, and anti-sites. Depending on the user need, the software can be employed to perform post-acquisition data analysis to identify regions of interest within a large field of view dataset, or to automate the data acquisition by enabling the automated identification of anomalous sample areas using low-magnification, low dose approaches, followed by in-depth characterization of specific areas of interest at the highest possible spatial resolution. Preliminary work on training a neural network to identify SrTiO3 defects has been conducted, demonstrating that the proposed methodology has the necessary sensitivity and robustness to identify structural anomalies. This software will be initially tested on a JEOL ARM200CF but can be adapted to all major vendors in this space.