Abstract

Understanding composition and phase distribution at the nanoscale is essential for advancing materials research—from optimizing catalytic nanostructures to unraveling complex phase behavior in geological and functional materials. Transmission electron microscopy (TEM) offers unparalleled spatial resolution, enabling direct imaging and analysis at the atomic scale. When combined with energy-dispersive X-ray spectroscopy (EDX) in scanning mode (STEM-EDX), it allows for spatially resolved chemical mapping with sub-nanometer precision.

Despite its widespread use, STEM-EDX faces two critical challenges that limit its effectiveness for accurate chemical and phase analysis: spectral mixing due to spatially overlapping phases along the electron beam path, which obscures phase-specific signals; and low signal-to-noise ratio (SNR), particularly in beam-sensitive materials or when detecting trace elements at concentrations below 100 ppm.

In this talk, I will present two self-developed machine learning strategies that directly address these limitations. The first employs non-negative matrix factorization (NMF) with integrated physical constraints to effectively unmix overlapping spectral signals. The second approach leverages the Poisson statistics of X-ray counting and a data fusion strategy to reconstruct high-quality datasets from a single noisy acquisition. These methods are demonstrated on complex experimental systems, including nanoscale mineral assemblages and heterogeneous catalysts. These represent a significant advancement in quantitative chemical analysis in electron microscopy, expanding its applicability to materials with complex phase relationships, low SNR, beam sensitivity, and critical trace constituents.