学术文献库

Single atomic sites often determine the functionality and performance of materials, such as catalysts, semi-conductors or enzymes. Computing and understanding the properties of such sites is therefore a crucial component of the rational materials design process. Because of complex electronic effects at the atomic level, atomic site properties are conventionally derived from computationally expensive first-principle calculations, as this level of theory is required to achieve relevant accuracy. In this study, we present a widely applicable machine learning (ML) approach to compute atomic site properties with high accuracy in real time. The approach works well for complex non-crystalline atomic structures and therefore opens up the possibility for high-throughput screenings of nano-materials, amorphous systems and materials interfaces. Our approach includes a robust featurization scheme to transform atomic structures into features which can be used by common machine learning models. Performing a genetic algorithm (GA) based feature selection, we show how to establish an intuitive physical interpretation of the structure-property relations implied by the ML models. With this approach, we compute atomic site stabilities of metal nanoparticles ranging from 3-55 atoms with mean absolute errors in the range of 0.11-0.14 eV in real time. We also establish the chemical identity of the site as most important factor in determining atomic site stabilities, followed by structural features like bond distances and angles. Both, the featurization and GA feature selection functionality are published in open-source python modules. With this method, we enable the efficient rational design of highly specialized real-world nano-catalysts through data-driven materials screening.