In ACS nano ; h5-index 203.0
Scanning transmission electron microscopy-based electron energy loss spectroscopy spectral imaging (STEM-EELS-SI) has been widely used in material research to capture a wealth of information, including elemental, electron density, and bonding state distributions. However, its exploitation still faces many challenges due to the difficulty of extracting information from noisy and overlapping edges in the convoluted spatial and spectroscopic data set. A traditional EELS spectral imaging analysis lacks the capability to isolate noise and deconvolute such overlapping edges, which either limits the resolution or the signal-to-noise ratio of the maps generated by EELS-SI. Existing machine learning (ML) algorithms can achieve denoising and deconvolution to a certain extent, but the extracted spectra lack physical meaning. To address these challenges, we have developed a ML method tailored to a spectral imaging analysis system and based on a non-negative robust principal component analysis. This approach offers an effective way to analyze EELS spectral images with improved space-time resolution, signal-to-noise ratio, and the capability to separate subtle differences in the spectrum. We apply this algorithm to 13 nanomaterial systems to show that ML can greatly improve image quality compared to a traditional approach, especially for more challenging systems. This will expand the type of nanomaterial systems that can be characterized by EELS-SI, and aid the analysis of structural, chemical, and electronic properties that are otherwise difficult to obtain.
Jia Haili, Wang Canhui, Wang Chao, Clancy Paulette
2022-Dec-20
complex systems, electron energy loss spectroscopy, machine learning, nanomaterial characterization, signal separation, spectral imaging analysis
