Leveraging\nMachine Learning for Advanced Nanoscale\nX‑ray Analysis: Unmixing Multicomponent Signals and Enhancing\nChemical Quantification

Energy dispersive X-ray (EDX) spectroscopy in the transmission\nelectron microscope is a key tool for nanomaterials analysis, providing\na direct link between spatial and chemical information. However, using\nit for precisely determining chemical compositions presents challenges\nof noisy data from low X-ray yields and mixed signals from phases\nthat overlap along the electron beam trajectory. Here, we introduce\na novel method, non-negative matrix factorization based pan-sharpening\n(PSNMF), to address these limitations. Leveraging the Poisson nature\nof EDX spectral noise and binning operations, PSNMF retrieves high-quality\nphase spectral and spatial signatures via consecutive factorizations.\nAfter validating PSNMF with synthetic data sets of different noise\nlevels, we illustrate its effectiveness on two distinct experimental\ncases: a nanomineralogical lamella, and supported catalytic nanoparticles.\nNot only does PSNMF obtain accurate phase signatures, but data sets\nreconstructed from the outputs have demonstrably lower noise and better\nfidelity than from the benchmark denoising method of principle component\nanalysis.