Unsupervised machine learning reveals eigen reactivity of metal surfaces.

The reactivity of metal surfaces is a cornerstone concept in chemistry, as metals have long been used as catalysts to accelerate chemical reactions. Although fundamentally important, the reactivity of metal surfaces has hitherto not been explicitly defined. For example, in order to compare the activity of two metal surfaces, a particular probe adsorbate, such as O, H, or CO, has to be specified, as comparisons may vary from probe to probe. Here we report that the metal surfaces actually have their own intrinsic/eigen reactivity, independent of any probe adsorbate. By employing unsupervised machine learning algorithms, specifically, principal component analysis (PCA), two dominant eigenvectors emerged from the binding strength dataset formed by 10 commonly used probes on 48 typical metal surfaces. According to their chemical characteristics revealed by vector decomposition, these two eigenvectors can be defined as the covalent reactivity and the ionic reactivity, respectively. Whereas the ionic reactivity turns out to be related to the work function of the metal surface, the covalent reactivity cannot be indexed by simple physical properties, but appears to be roughly connected with the valence-electron number normalized density of states at the Fermi level. Our findings expose that the metal surface reactivity is essentially a two-dimensional vector rather than a scalar, opening new horizons for understanding interactions at the metal surface.