Machine learning based catalytic property prediction of metal nanoparticles

Abstract

Metal nanoparticles have higher specific area and quantum effect because of its small size, and is used in various applications, especially for catalysts. However, DFT calculations to study metal nanoparticles for catalysts requires high computational costs, and current machine learning study for materials science could not be easily applicable to nanoparticle structures. To overcome this, we applied graph based convolutional neural networks to develop general machine learning model for metal nanoparticles to predict catalytic properties. In chapter 1, we developed PCA-GCNN machine learning models to predict electronic density of states of metal nanoparticles, not only pristine but also bimetallic core-shell and alloy nanoparticles. This model also can predict DOS of bigger nanoparticles that are not used in training. In chapter 3, to build a surface Pourbaix diagram of Pt NPs, we predicted adsorption energies on Pt NPs with various number of adsorbates using GCNN machine learning models. This model could predict adsorption energies of bigger nanoparticles as well and the surface Pourbaix diagram of it could be built from the predicted value by machine learning model.