Improved Prediction of Adsorption Energies and Vibrational Frequencies of Molecules Adsorbed on Monometal Catalyst Surfaces by Machine Learning using Structural Descriptors

Designing surfaces of metallic catalyst is essential to improve the efficiency of reaction. Recently prediction of intermediate states of adsorbates by using both first-principles calculation and machine learning has been attracted much attention. First-principles calculations for adsorbates on surfaces have been performed, and various databases of adsorption energies have been reported[1]. In order to improve the accuracy of the prediction by machine learning, various descriptors have been developed[2]. Structures of both metal surfaces and adsorbates could affect the adsorption energies. We thus focus on the structural descriptors for machine learning to predict the adsorption energies and vibration properties of adsorbates on metals. We used a surface slab model which has (3 or 5)-layer mono metal (Ti, V, Cu, etc) fcc surface and evaluated the adsorption energies and vibrational frequencies of adsorbates (C, H, N, CO, etc) by first-principles calculation.<br/>We calculated from first-principles more than 7000 surface structures which were employed for prediction of adsorption energies and vibrational frequencies. We evaluated the accuracy of prediction based on several machine learning techniques and found that structural descriptors such as bulk lattice parameter, <i>d</i>-band center and width of the surface model are effective to improve the accuracy compared to the prediction using only composition descriptors. The prediction accuracies of the mean squared error were improved from 0.161 to 0.136. The importance of structural descriptors was examined using the “permutational importance”. The lattice parameter and <i>d</i>-band center were turned out to contribute highly to the prediction of adsorption energy.<br/><br/>[1] https://www.catalysis-hub.org<br/>[2] H.-J. Peng, <i>et</i>. <i>al</i>., Nature Communications <b>13</b>, 1 (2022).