Highly efficient machine learning strategy for low-loss eels characterization: nanophotonic resonances as a case study

Abstract Spatially-resolved electron energy loss spectroscopy in scanning transmission electron microscopy for materials characterization provides unparalleled access to nanoscale mapping of physical and chemical properties. In particular, the low-loss energy regime contains key details for understanding the optical (e.g., plasmons), electronic (e.g., band gaps), and structural characteristics of materials. However, overlapping resonances, proximity of relevant features to the intense zero-loss peak, and low signal-to-noise ratio complicate the analysis, requiring advanced methods for effective denoising and identification of complex properties. Here, a novel combination of machine learning algorithms is proposed to improve efficiency and interpretation of the low-loss region of EEL spectra and to automatically identify spatially localized information. This work combines uniform manifold approximation and projection for dimensionality reduction with the unsupervised clustering algorithm hierarchical density-based spatial clustering of applications with noise. In a second step, data classified as outliers are reassigned by the supervised support vector machines algorithm. The later supervised stage naturally extends the unsupervised analysis, providing a trained supervised model that classifies new spectrum images based on the previously discovered clusters at negligible computational cost, enabling near real-time deployment on other datasets. This strategy has been applied to silicon/gold nanopillars, used as hybrid metal-semiconductor nanoantennas, to elucidate their nanophotonic resonances (i.e., plasmon, Mie, and hybrid resonances). The procedure enabled accurate mapping of the different resonances in a straightforward manner. The simplicity and versatility of the approach pave the way for studying more complex types of collective excitations and band transitions.