Efficient Exploration of High-Dimensional Configuration Spaces for the Generation of Chemical Datasets

In this work, we introduce an automated methodology for the efficient and relatively inexpensive exploration of large high-dimensional chemical spaces, with particular focus on number-of-atoms-conserving processes, such as in mechanochemical reactions. Our approach combines: (1) a physically motivated stochastic global-landscape exploration phase (mechanochemical distortion), which efficiently overcomes entropic barriers, and (2) a local exploration phase in which the previously determined local basins are sampled with molecular dynamics and graph theory. Specifically, this last phase makes use of the (vdW-) transition state search using a chemical dynamical simulations algorithm. Our methodology requires minimal input from the user. As a case study, we have explored the conformational landscape, including transition states and minimum energy paths, of the C60H10 hydrogen-carbon clusters owing to their astrochemical relevance as potential carriers of the aromatic infrared bands. From a single initial seed (geometry), we have obtained a series of 212 mechanochemically relevant conformers, and from just 3 of them, we have obtained a set of >13 000 minima spanning the domain of our interest. The underlying chemical network has been fully characterized and rationalized using statistical analysis tools. Our case study perfectly illustrates the potential of our approach in the automatic generation of chemical databases, in other words, annotated data for the training of data-hungry deep learning models in chemistry.