Simple Heuristics for Advanced Sampling of Reactive Species on Surfaces

Understanding interactions between reactive fragments and surfaces is a central objective in functional materials design and discovery. No available computational methods can reliably generate such structures for an arbitrary choice of fragment and surface while explicitly describing the binding mode. As a result, human curation is accepted as the norm, introducing bias and hampering efforts toward computational materials screening and database engineering, limiting the reach and impact of machine learning in the field. By implementing two complementary concepts, surrogate-SMILES (*SMILES) and an atomic-position-only algorithm for surface site identification, we achieve simple parametrization of all fragment-surface configurations. We harness this fully hands-off approach to revisit CO2 reduction on copper, discovering diversity in the adsorption energy spectra, even for this benchmark system. We generate a holistic picture of fragment-surface energy features and compare them across systems and facets. Efficient parametrization and string-based expression provide a strong bridge of our approach to generative, optimization, and language-driven machine learning frameworks. With no domain-specific assertions, this approach extends to all domains in which surface reactivity is key for rational design.