“Inverting” X-ray Absorption Spectra of Catalysts by Machine Learning in Search for Activity Descriptors

Abstract

The rapid growth of methods emerging in the past decade for synthesis of “designer” catalystsranging from the size and shape-selected nanoparticles to mass-selected clusters, to precisely engineered bimetallic surfaces, to single site and pair site catalystshas opened opportunities for tailoring the catalyst structure for the desired activity and selectivity. It has also sharpened the need for developing approaches to the operando characterization, ones that identify the catalytic active sites and follow their evolutions in reaction conditions. Commonly used methods for determination of the activity descriptors in the nanocatalysts, based on the correlation between the changes in catalyst performance and evolution of its structural and electronic properties, are hampered by the paucity of experimental techniques that can detect such properties with high accuracy and in reaction conditions. Out of many such techniques, X-ray absorption spectroscopy (XAS) stands out as an element-specific method that is very sensitive to the local geometric and electronic properties of the metal atoms and their surroundings and, therefore, is able to track catalyst structure modifications in operando conditions. Despite the vast amount of structure-specific information (such as, e.g., the charge states and radial distribution function of neighbors of selected atomic species) stored in the XAS data of catalysts, extracting it from the spectra is challenging, especially in the conditions of low metal weight loading, nanoscale dimensions, heterogeneous size and composition distributions, and harsh reaction environment. In this Perspective, we discuss the recent developments in XAS data analysis achieved by employing supervised and unsupervised machine learning (ML) methods for structural characterization of catalysts. By benefiting from the sensitivity of ML methods to subtle variations in experimental data, a previously “hidden” relationship between the X-ray absorption spectrum and descriptors of material’s structure and/or composition can be found, as illustrated on representative examples of mono-, hetero-, and nonmetallic catalysts. In the case of supervised ML, the experimental spectra can be rapidly “inverted”, and the structure of the catalyst can be tracked in real time and in reaction conditions. Emerging opportunities for catalysis research that the ML methods enable, such as high-throughput data analysis, and their applications to other experimental probes of catalyst structure are discussed.

A Critical Review of Machine Learning of Energy Materials

“…Finally, an application of ML models that has received far less attention in other reviews is their potential to enhance experimental characterization. The interpretation and labeling of experimental images (e.g., scanning transmission electron microscopy (STEM)) and spectra (e.g., X‐ray diffraction (XRD), X‐ray absorption near‐edge structure (XANES), nuclear magnetic resonance (NMR), etc.) are today still mostly painstakingly carried out by humans.…”

Linking the evolution of catalytic properties and structural changes in copper–zinc nanocatalysts using operando EXAFS and neural-networks

“…[30][31][32][33][34][35][36][37][38][39] To address this problem, we have recently proposed a machine learning-based approach for the interpretation of experimental XAS data, where we employed articial neural networks (NNs) to extract the structural information. [40][41][42][43][44] In the case of EXAFS data interpretation, by exposing NN to thousands of training examples (theoretical EXAFS spectra, for which the corresponding structure was known) we could establish a relationship between spectral features and structure descriptors. The trained NN can then be used to directly invert experimental data, and quickly and accurately extract relevant structural information, such as partial radial distribution functions (RDFs).…”

“…We have recently shown that the catalytic selectivity of this system can be tuned by varying the copper to zinc ratio, and that signicant structural changes take place under reaction conditions. 13,47 Here we take advantage of the sensitivity of our NN-EXAFS method 44 to monitor subtle chemical and structural changes taking place during the CO 2 RR as a function of the Cu-Zn NP composition and time. In particular, we will show that the reduction of copper and zinc species takes place on different time-scales, and that gradual alloying and a transition from a close-packed to a more disordered structure take place.…”

Progress in Computational and Machine‐Learning Methods for Heterogeneous Small‐Molecule Activation

“…While the data are highly convoluted, a machine learning model's exceptional flexibility could enable atomistic-level characterization of the surface. There has been some exciting progress in this direction, [242][243][244] and is expected to significantly expand the power of spectroscopy.…”