EXAFS spectroscopy is one of the very few experimental techniques that are able to decipher atomic
structure of small (1 – 2 nm) nanoparticles (NPs). The conventional approach to EXAFS data analysis,
based on the fitting of theoretical EXAFS equation to the experimental data, does not work well for small
NPs, due to (i) intra-particle heterogeneity of atoms and charges due to the interactions with adsorbates
and support, (ii) inter-cluster heterogeneity, due to the coexistence of multiple species in real catalysts,
with different sizes, structures, morphologies and degrees of reduction, and (iii) the large number of
structural parameters, required for the complete description of 3D structure of NPs.
As possible remedies for the first limitation, I will mention the applications of high energy
resolution X-ray absorption and emission spectroscopies (XAS and XES) for resolving competing
interactions (metal-metal, metal-support, metal-adsorbate) in working nanocatalysts. For solving the
second challenge, we developed and tested a multi-modal investigation scheme, relying on the
combination of average (EXAFS and XANES) and local, statistical (STEM) methods.
In this talk I will focus on the last challenge (iii). It was demonstrated recently that XANES
features in mono- and bimetallic nanoparticles can be interpreted in terms of architectural and
compositional information using theoretical simulations. The problem is that there is no direct method of
obtaining the unknown structure from its XANES spectrum. The use of machine learning methods
allowed us to train the artificial neural network to recognize the features originating from different sizes,
shapes and compositions, and, in turn, reliably reconstruct them from the experimental spectra. This
method has potential applications to the broad range of materials and reaction regimes and couples nicely
with the rapid data collection during operando investigation of chemical transformations by high
throughput synchrotron spectroscopies.