(659b) Catalyst Characterization from Complex Infrared Spectroscopy: A Machine Learning Approach
Adsorbate vibrational excitations are a unique descriptor of adsorbate/surface interactions with the benefit of being relatively independent of temperature and pressure effects. The infrared spectra (IR) associated with activating adsorbate vibrational modes is very accurate, captures details of most vibrational modes, and can be obtained operando. Despite the possibility of obtaining detailed catalyst structural information and adsorption site preference throughout the course of a reaction from IR spectroscopy, the ability to interpret the resulting complex spectra is lacking. Current techniques depend on heuristic peak assignments and can only be done for relatively simple spectra. Quantitative assignment of adsorption sites can only be done on simple surfaces where the structure is known a-priori and requires concurrent measurements from mass spectrometry (MS) and low energy electron diffraction (LEED) studies at ultra-high-vacuum (UHV). We present a machine learning approach to determine adsorption site preference and local surface structure where complexity in the IR spectra is preferred. We combine first-principles calculations of carbon monoxide on platinum nanoparticles and a forward surrogate model to simulate IR spectra for training an inverse surrogate machine learning model. We test the machine learning model on both synthesized spectra generated from our forward surrogate model and on experimental spectra. This work demonstrates the ability to gain relevant information from experimental data using theory and computation.