Machine Learning tools for Analytical Transmission Electron Microscopy


Modern transmission electron microscopes are equipped with aberration correctors and high-resolution spectrometers, which enable the spatially-resolved chemical analysis of an unlimited variety of materials on the nanometric or even atomic  length scales. The ultimate ambition of a researcher is to segment and accurately quantify the analytical data recorded from such the sample, in order to precisely characterize the spatial distribution and nature of each of its phases. However, currently, such an aim can only be realized in a limited fraction of cases, owing to factors including signal convolutions and non-linear backgrounds, detector and signal noise characteristics, and segmentation challenges from the projection effects.

The aim of this project is to address and resolve these deficits, by innovating a radically new approach to the data analysis.

 The aim of our project is to develop an integrated approach where all of these elements are synergistically coupled together: prior information; a directed “dictionary learning” approach using the optimal data science tools; and physical model-based quantification. With interactive feedback between these elements, it will be possible to achieve a step change in the quality of data analysis. A successful project will therefore significantly increase the ability of the researcher to leverage their data for obtaining new scientific insights and technological advances across a wide range of fields.



The spectroscopic data obtain in scanning transmission electron microscopes (STEM) are in most cases not straightforward to analyse because of noise and mixed features (both spectrally and spatially). The current state-of-the-art in this community in terms of data analysis is limited to basic ML algorithms such as PCA or NMF which, in general, do not retrieve the original physical features of the observed sample. In the MLATEM project we aim at designing a physics-guided ML algorithm which will retrieve the physically-correct features of the observed sample. In that perspective, the SDSC will provide an expertise on cutting-edge ML techniques to develop this algorithm while the collaborating team will provide domain knowledge.


Our goal is to develop a new standard algorithm for the electromicroscopy community.

Scheme of the acquisition of an Electron Energy-Loss Spectroscopy (EELS) and/or Energy-Dispersive X-ray Spectroscopy (EDXS) Spectrum Image (SpIm).  The electron beam is scanned over a chosen region of the sample. At each position of the beam, the electrons interact with the sample. Depending on the available equipement in the microscope, the EELS and the High-Angle Annular Dark-Field (HAADF) signals or the EDXS and HAADF signals are acquired simultaneously. Some specialised microscope allow the acquisition of EELS, EDXS and HAADF signals simultaneously. The HAADF signals form an image while the spectroscopy signals form an array of spectra, making a datacube, called a SpIm. An HAADF intensity corresponds to each spectrum of the SpIm. Figure adaptated from ref [1].

[1] Gatan Inc. Spectrum imaging. GATAN. URL: Spectrum Imaging | Gatan, Inc. .