Hyperspectral data resulting from imaging techniques such nanoARPES, STM, STEM, momentum-resolved EELS, vibrational-EELS, etc. of layered can provide critical information about nanoscale structure including electronic structure. Understanding both novelty/anomaly and conversely normality characterization of multi-modal data of these types would allow for significant acceleration in experimental search for novel materials and preparations and rapid testing. Novel electronic materials (e.g., superconductors) could significantly impact how we transport electricity, power vehicles and transportation, perform medical imaging, and design computer processors and memory. We propose a novel topological data analysis (TDA) approach to finding topological features in hyperspectral data. This algorithm is fast and can be executed on edge devices. We will combine an edge compute capability which generates topological fingerprints with persistent database storage of features and underlying data (a combination of a document database and graph database). These databases will be made consistent in a cloud based and containerized deployment. The cloud-based system will provide additional model generation capability. We will provide an initial demonstration of this capability using nanoARPES data made available by LBNL. During phase I we will use unsupervised and semi-supervised approaches over space of topological fingerprints. We will explore the use of combining topological features with deep learning. At the end of Phase I we will demonstrate our TDA based approach to topological fingerprints on nano ARPES data. We will deploy our algorithm and data pipeline on an edge device such as an NVIDIA jetson and on cloud infrastructure. We compare our approach to existing approaches using clustering and Gaussian process regression in terms of the ability to locate novel/anomalous data in service of automating scans or assessing data. Microscopy and spectroscopy have multiple applications and thus are increasingly used by diverse industries. Nanoscale techniques like those mentioned here, in particular nano ARPES, provide powerful techniques to search for novel electronic materials. In the future, in addition to the discovery of new materials, manufacturing will require the ability to rapidly and successfully test for desirable and undesirable properties. More broadly, the techniques developed in DTFAAST can apply to other forms of molecular spectroscopy such as infrared spectroscopy which is widely applicable to biological systems in plants and animals. Finally, the DTFAAST approach lends itself naturally to the fusion of multimodal data which combine techniques listed above.
