Data clustering and machine learning in electron microscopy and spectroscopy

The progress in data clustering algorithms permits nowadays relatively fast analysis, nearly on-the-fly, especially useful in the progressively growing size of microscopy and spectroscopy data. For instance, electron energy-loss (EELS) datasets are suitable 3D data to explore the effectiveness of unsupervised algorithms, with the aim of a precise chemical characterization of nanoparticles and thin films.
In this Colloquium, I will discuss some of the ‘classical’ unsupervised clustering algorithms of machine learning, such has k-means (KM), gaussian mixture models (GMM), and agglomerative clustering (AGC), and they comparison with supervised algorithms, all supported by principal components analysis (PCA), to explore the classification effectiveness in the latent space of the components with the highest variance. I will show some results of their applications on hyperspectral data from nanostructured films. Finally, I will show the application of deep convolution neural networks to the diagnosis of aberration of the probe in scanning transmission electron microscopy (STEM) for high spatial resolution.
Both shallow and deep learning algorithms may contribute to progressively automate the analysis and experiments, and to reduce resulting biases when different users perform the analyses.

Seminar realized in the framework of the funded projects:
-IMPRESS. HORIZON-INFRA-2022-TECH-01; GA 101094299;
-iENTRANCE@ENL. Italian Ministry of Research and Next generation EU (IR0000027);
-AI-TEM. PRIN 2022 (2022249HSF).