A neural-network model to find correlations in astrochemical space
The detection of new interstellar molecules is central to astrochemistry, and arguably the most transformative development in the field in past decade has been the increase in both number and diversity of detected species. Using natural language processing models, we develop a machine-learning pipeline, similar to other approaches reported in the literature, to connect molecular structures, encoded as chemical strings, and molecular column densities in TMC-1. The workflow generates astrochemically informed embeddings that can be employed in subsequent supervised learning tasks. These representations are used within a neural-network framework to perform column-density regression on both the original molecular inventory and out-of-distribution species. The resulting model achieves mean squared error values of 0.17/0.58 and mean absolute error values of 0.29/0.59, and coefficients of determination of 0.88/0.51 for the train+validation and validation datasets. The latent vectors learned by the model are rich mathematical objects that capture the underlying chemical relations and also the potential interstellar detectability of the molecule. Using these vectors in molecules obtained from systematic permutations of the TMC-1 inventory, we identify several promising candidates for future detection in this cloud, together with a large number of apolar species. By training the model on datasets reflecting different historical epochs of molecular detections in TMC-1, we show that recent molecular detections have substantially reshaped our understanding of interstellar chemistry, in a transformative, rather than incremental, way. Beyond accurate prediction of column-density trends for both in-distribution and out-of-distribution molecules, the latent-space representations produced by the model provide a stepping stone for developing new forms of data-driven astrochemical intuition and exploration.