Thèses

Vendredi 30 Septembre 2022 à 14h00.

Quantitative measurement of chemical composition in small alloy nanoparticles by STEM-EDS and Machine Learning

''
Directeur de thèse / thesis director :

Matthias HILLENKAMP

Membres du jury / jury members :

Wolfgang ERNST (U Graz, Austria)
Antonius VAN HELVOORT (U Trondheim, Norway)
Giorgio DIVITINI (U Genova, Italy)
David AMANS (UCBL)
Antonio RIUL jr (U Campinas, Brazil)
Daniela ZANCHET (U Campinas, Brazil)
Matthias HILLENKAMP (UCBL)

Résumé / Abstract :

EDS-STEM has gained significant enhancement in its acquisition systems, making it possible to perform chemical analysis in nanometric objects, such as bimetallic nanoparticles (BNPs). These advances turn possible to extract quantitative information from individual small BNPs (Diam. 10 nm), opening the path for the understanding of chemical composition vs. size properties and elemental distribution in these systems. This is a significant improvement concerning purely qualitative chemical mapping, as it is widely used in various communities. Alloy NPs can be produced in various ways, with typical options being chemical synthesis in colloidal suspension and physical synthesis methods such as gas phase aggregation. In this context, developing tools with specific and well-adapted capabilities for studying BNPs becomes essential. The complete characterization of nanometric systems requires detailed knowledge of how the different chemical elements of the materials are distributed at the interfaces/surfaces (roughness, interdiffusion, etc.); and, finally, how these factors modify the electronic properties of the system. Henceforth, we propose quantifying the chemical composition of individual AuAg BNPs produced by gas-phase aggregation. We aim to use EDS-STEM technique and hyperspectral image (HSI) data treatment using machine learning to improve the quantitative chemical composition information data extraction. Therefore, we deeply investigated the effects of machine learning procedures such as Principal Component Analysis (PCA) applied to analyze HSIs. PCA is used as a standard procedure for HSI denoising through reconstructing the data set on a more representative basis, reducing Poisson noise. However, our studies suggest that the statistical tool must be used carefully, mainly in quantitative analysis. When reconstructed by PCA, a low signal-to- noise ratio (SNR) data set is biased, and artifacts appear in the results, such as a systematic error in the chemical composition quantification “averaging” all the voxels from the BNPs. We quantitatively evaluate the bias of HSI reconstruction with different SNR levels, comparing both experiments and simulations. We increased X-ray counts and obtained a bias reduction in the reconstruction. Based on statistical models, we proposed an information loss estimator that allows us to analyze the quality of the reconstruction according to bias. We also analyzed the interplay effects of counts, pixels, and channels on the HSI denoising. Proper quantification requires confidence intervals; therefore, we show a methodology to estimate uncertainty in chemical composition analysis after PCA denoising. With that, we can quantify the chemical composition of AuAg BNP’s and identify, for example, size-dependent composition effects hidden by Poisson noise. Finally, we propose in this thesis the use of PCA and NMF for the study of unmixing signals in HSIs, aiming at the knowledge of the elemental distribution inside these BNPs. We show that we can measure a chemical gradient of 0.45±0.02 to 0.60±0.02 (Ag at. fraction) from the projected center of the BNP to its surface. The Ag enrichment towards the surface characterizes a core rich in Au and a shell rich in Ag. Consequently, to explain the radial Ag enrichment, we studied the effect of reactivity towards oxidation by analyzing carbon-protected and unprotected BNPs. Hence, due to the quantitative aspect of our analysis, we can know the number of atoms in different regions of the BNPs. This allowed us to identify that carbon-protect and unannealed show a partially segregate Ag pattern with an increased level of alloying in contrast to the oxidized situation where the Ag atoms tend to the BNP’s surface. Finally, with the established quantitative analysis methodologies employed in this thesis, we expect that the chemical composition characterization of small volumes can be improved to support theory and simulations of modeling in nanoscale physics and chemistry for both fundamental and applied studies.

''