Scanning transmission electron microscopy (STEM) has emerged as a uniquely powerful tool for structural and functional imaging of materials on the atomic level. Driven by advances in aberration correction, STEM now allows the routine imaging of structures with single-digit picometre-level precision for localization of atomic units. This Primer focuses on the opportunities emerging at the interface between STEM and machine learning (ML) methods. We review the primary STEM imaging methods, including structural imaging, electron energy loss spectroscopy and its momentum-resolved modalities and 4D-STEM. We discuss the quantification of STEM structural data as a necessary step towards meaningful ML applications and its analysis in terms of the relevant physics and chemistry. We show examples of the opportunities offered by structural STEM imaging in elucidating the chemistry and physics of complex materials and how the latter connect to first-principles and phase-field models to yield consistent interpretation of generative physics. We present the critical infrastructural needs for the broad adoption of ML methods in the STEM community, including the storage of data and metadata to allow the reproduction of experiments. Finally, we discuss the application of ML to automating experiments and novel scanning modes.
Machine learning in scanning transmission electron microscopy
Grillo, Vincenzo;
2022
Abstract
Scanning transmission electron microscopy (STEM) has emerged as a uniquely powerful tool for structural and functional imaging of materials on the atomic level. Driven by advances in aberration correction, STEM now allows the routine imaging of structures with single-digit picometre-level precision for localization of atomic units. This Primer focuses on the opportunities emerging at the interface between STEM and machine learning (ML) methods. We review the primary STEM imaging methods, including structural imaging, electron energy loss spectroscopy and its momentum-resolved modalities and 4D-STEM. We discuss the quantification of STEM structural data as a necessary step towards meaningful ML applications and its analysis in terms of the relevant physics and chemistry. We show examples of the opportunities offered by structural STEM imaging in elucidating the chemistry and physics of complex materials and how the latter connect to first-principles and phase-field models to yield consistent interpretation of generative physics. We present the critical infrastructural needs for the broad adoption of ML methods in the STEM community, including the storage of data and metadata to allow the reproduction of experiments. Finally, we discuss the application of ML to automating experiments and novel scanning modes.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.