LAUSR

While Machine Learning (ML) has made numerous advances in the analysis of "big data", as this type of data is initially generated for other purposes, ML tends to uncover new insights by serendipity, i.e. valuable previously unknown trends in the data are found, but which specific areas these trends will relate to, cannot be determined a priori. Scientific data used to characterize materials or processes on the other hand is generated with a specific project in mind from highly specialized and expensive equipment, making the generation and analysis of a diverse "big dataset" impossible. However, we can use the way that ML analyses "big data" to learn which parts of any given dataset are most useful for interpretation, and then design our experiments to generate only that "targeted data" which is most useful. For state-ofthe-art electron microscopy, compressively sensing (CS) and reconstructing images / spectra using an ML framework is particularly valuable for obtaining quantitative multi-scale hyperspectral data under extremely low dose and/or dose rate conditions with significantly accelerated framerates. The dose / speed / resolution optimization that is possible using these methods creates wide-ranging new opportunities to avoid major electron beam damage and enables quantified observations of metastable materials and materials dynamics to be made.