Physics-guided NMF for the analysis of STEM/EDXS data : from raw data to quantified chemical phases.

Energy-dispersive X-ray spectroscopy (EDXS) in the scanning transmission electron microscope (STEM) is a useful tool for the local chemical analysis at the nanoscale. Using the scanning mode it is possible to acquire a so-called spectrum-image which is a 3D datacube with an EDXS spectrum at each position of the scan. In the simplest cases it is straightforward to obtain a chemical map by integrating some part along the energy axis. Often, the spatial overlap between phases with different chemistry or the spectral overlap between X-ray characteristics peaks make the analysis of STEM/EDXS spectrum images a complex task. A few decomposition algorithms, such as Principal Component Analysis (PCA) or Non-Negative Matrix Factorization (NMF), are commonly used to retrieve the pure phases and their distribution in the sample from STEM/EDXS data. However, the outputs of these methods tend to contain artefacts and lack accuracy.
To address these issues we designed an improved version of the NMF that incorporates a physical model of the EDXS into the decomposition. Using this new algorithm, it is possible to obtain an accurate decomposition of a spectrum-image in pure phases as well as a chemical quantification of each phases.
In this presentation we will demonstrate the operating principle of the algorithm. We will also show a comparison of the performances of our algorithm with those in the literature on both simulated and experimental data.