Electron energy loss spectroscopy has undergone dramatic improvements over the last decades in terms of energy resolution from monochromators, signal to noise from improved detectors and advances in the statistical treatment of the recorded data. Yet, despite many efforts, the technique relies heavily on expert operators and interpretation skills. This hampers the widespread use of EELS and leaves room for experimenters’ bias posing a significant reproducibility risk. In this talk, I will give an overview of recent efforts towards the goal of an entirely autonomous data processing workflow which could significantly improve the quantification of EELS spectra in terms of ease of use, precision and accuracy. The method relies on model-based quantification as was e.g. implemented in EELSMODEL [1] but improves on several important aspects: The physical model is made entirely linear. This results in a single solution without the need for user provided initial parameter estimates. The background modelling process is significantly improved through a combination of a linearized model and constrained quadratic programming methods. The fine structure is described with a polynomial series that naturally expresses that most significant ELNES features occur near the edge onset. A complete Dirac-based database of open generalized oscillator strengths significantly improves the fit with experiments and provides a way to improve accuracy and precision. We validate auto-ID algorithms, ranging from relatively simple feature detection filters to neural network attempts to identify which elements are present in a given dataset. This allows to build a complete model without any user input thereby completely removing the experimenters’ preferences and opinions from the process. We discuss achieved precision approaching the statistical limit and show attempts to evaluate the accuracy with reference samples. The result provides not only a reliable quantification with highest possible precision and excellent accuracy but also seamlessly provides background subtracted and deconvolved ELNES features without any user intervention. This is particularly interesting for high energy K-edges as they provide a direct lab-based alternative for XAS measurements even for edges that are traditionally considered to be far outside the achievable EELS energy range. In summary, we believe that the days of manual processing of EELS data are over and fast and fully unsupervised data interpretation is available. Together with advances in EELS spectrometers, this widens the scope of EELS significantly and confirms its position as an indispensable analytical tool, also for non-expert users. It also satisfies the need for high reproducibility, essential in particular for industry applications [2]. Screenshot of the unsupervised quantification software under development.
No takes yet. Share an insight, caveat, or question.
Verbeeck et al. (2023) studied this question.