(595f) Resolving 3D Structures of Metallic Nanoparticles from X-Ray Absorption Data Using Artificial Neural Network

Many applications of metallic clusters and nanoparticles (NPs) rely on the observation that properties (for example, magnetic, optical, chemical, catalytic properties) of such materials are different from those of bulk metals. These unique properties are a result of interplay between NPs core sites and undercoordinated surface sites, the quantum confinement effects that modify the electronic properties of metal atoms, and the interactions between metal sites and support and/or adsorbates. Understanding NPs atomic structure and its relation to the NPs properties is thus important to enable new potential applications of nanostructured materials. More importantly, the atomic structure of NPs can change in-situ, e.g., during the chemical reactions, where NPs are used as catalysts.

One of the factors that hinder such understanding is that structural information, provided by experimental techniques, is scarce in case of small (with sizes ca 1 – 2 nm) NPs, especially under operando conditions. Conventional crystallographical approaches that commonly are used to learn the structure of bulk materials are not applicable in this case, and the sensitivity of high resolution electron microscopy methods is not sufficient to resolve in details the atomic structure of such small NPs. X-ray absorption spectroscopy (XAS) is one of a few methods that are useful in this case, due to its sensitivity to the chemical state of absorbing atoms and to the types and arrangements of atoms around the absorber. Importantly, XAS method also allows in-situ and in-operando investigations, thus allowing one to directly correlate NPs structure and properties.

While the extended X-ray absorption fine structure (EXAFS) is widely used in NPs structure studies, much less attention has been paid to the information encoded in X-ray absorption near edge structure (XANES). At the same time, there are number of reasons, why analysis of XANES data can be more beneficial for studies of small NPs. First of all, XANES is less sensitive to disorder effects, which affect severely EXAFS spectra quality and complicate the interpretation of EXAFS data. Secondly, XANES is more sensitive to the 3D geometry of the environment around absorbing atoms. The main challenge that hinders the usage of XANES for quantitative analysis is the lack of methodology that would allow one to extract such 3D structural information from experimental data.

Recent advances in data-enabled discovery science methods in chemical research provide a key to this problem. In this study for the first time we employ machine learning methods and ab-initio XANES calculations to correlate XANES features with the descriptors of 3D local structure of metallic nanoparticles. Here we demonstrate the potentiality of this method on the example of XANES study of small supported platinum NPs. We use theoretical site-specific XANES spectra, calculated by ab-initio codes, to train an artificial neural network (NN). The trained NN is then used to extract structural information from the experimental XANES data, and to determine the size and shape of investigated metallic NPs.