http://arxiv.org/abs/2107.14610
The characterization of interstellar chemical inventories provides valuable insight into the chemical and physical processes in astrophysical sources. The discovery of new interstellar molecules becomes increasingly difficult as the number of viable species grows combinatorially, even when considering only the most thermodynamically stable. In this work, we present a novel approach for understanding and modeling interstellar chemical inventories by combining methodologies from cheminformatics and machine learning. Using multidimensional vector representations of molecules obtained through unsupervised machine learning, we show that identification of candidates for astrochemical study can be achieved through quantitative measures of chemical similarity in this vector space, highlighting molecules that are most similar to those already known in the interstellar medium. Furthermore, we show that simple, supervised learning regressors are capable of reproducing the abundances of entire chemical inventories, and predict the abundance of not yet seen molecules. As a proof-of-concept, we have developed and applied this discovery pipeline to the chemical inventory of a well-known dark molecular cloud, the Taurus Molecular Cloud 1 (TMC-1); one of the most chemically rich regions of space known to date. In this paper, we discuss the implications and new insights machine learning explorations of chemical space can provide in astrochemistry.
K. Lee, J. Patterson, A. Burkhardt, et. al.
Mon, 2 Aug 21
3/82
Comments: 20 pages; 8 figures, 2 tables in the main text. 6 figures, 2 tables in the appendix. Accepted for publication in The Astrophysical Journal Letters. Molecule recommendations for TMC-1 can be found in the Zenodo repository: this https URL