Extendable Machine Learning Model for the Stability of Single Atom Alloys

In this work, we aim to update the understanding of how impurity or promoter metals segregate on metal surfaces, particularly in the application of single-atom alloys (SAA) for catalysis. Using density functional theory, we calculated the relative stability of the idealized SAA relative to subsurface, dimer, and adatom configurations to determine the tendency of the promoter atom to diffuse into the bulk, form surface clusters, or avoid alloying with the host, respectively. We selected 26 d-block metals augmented with Al and Pb to create a 28 × 28 database that indicates a total of 250 combinations for which the SAA configuration is most stable, and an additional 358 systems for which the SAA geometry is within 0.5 eV of the most stable configuration. We classified the data using decision tree, support vector machine, and neural network machine learning algorithms with tabulated atomic properties as the input vector. These black box approaches are unable to extrapolate to other possible geometries, which was circumvented by redefining the stability problem as a regression. We propose a physical bond counting model to formulate intuitive criteria for the formation of stable SAAs. The accuracy is then improved by using the bonding configuration and tabulated atomic properties with a kernel ridge regression (KRR) algorithm. The hybrid KRR model correctly identifies 190 SAAs with 85 false positives. Importantly, its physical basis allows the hybrid model to extend to similar geometries not included in the training data, thereby expanding the domain where the model is useful.