Predicting Planetary Atmospheric Chemical Reaction Rates Using Machine Learning

Abstract:

Atmospheres are important potentially observable features of exoplanets. Their composition can provide evidence for biological activity (i.e., a biosignature) and provide information about other planetary processes and events. Models of exoplanetary atmospheres will assist in the understanding of observations, but these are highly dependent on input parameters such as reaction rate constants. Unfortunately, databases of such rate constants are incomplete and only contain information for a limited set of reactions that have been studied in isolation in laboratories. These known reaction rates may not be sufficient to accurately model planetary atmospheres hosting reactions with unknown rate constants or that are impractical to measure in the lab. To address this problem, we applied a series of machine learning techniques to STAND-2019, an atmospheric chemical network detailed in Rimmer et. al, 2019, to explore how well machine learning can predict reaction rate constants[1]. After creating features for the dataset such as reaction mass, number of species involved, type of species involved, and SMILES string representation, we examined how different properties of the reactions affected the accuracy of the algorithm’s predictions. Preliminary results of this approach to predicting atmospheric chemical reaction rates are presented and discussed.

References

[1] P. B. Rimmer and S. Rugheimer, “Hydrogen cyanide in nitrogen-rich atmospheres of rocky exoplanets,” Icarus, vol. 329, pp. 124–131, Sep. 2019.