Speaker
Description
The molecular composition of exoplanetary atmospheres can be heavily influenced by non-equilibrium and photochemistry. Determining this composition can thus be very computationally demanding, with conventional ODE-solving strategies often taking hours to converge. This makes incorporation into retrieval algorithms difficult. As such, we investigate the application of machine learning for photochemistry models within planetary atmospheres. We present a model capable of determining the mixing ratio of certain species orders of magnitude faster per run, which would make scanning the parameter space for retrievals significantly quicker. A one-dimensional convolutional network is trained to predict the mixing ratio of a single molecule at a time incorporating the data from the VULCAN chemical kinetics code (Tsai et al. 2021), using similar input and output as VULCAN. We find the typical multiplicative error factor of the predicted volume mixing ratios to be around 1.3 (median), ranging from 1.14 for the best performing molecules and 2.9 for the least accurate species. The predicted output is then used to generate a transmission spectrum using the ARCiS framework (Min et al. 2020) to determine the propagation of these errors and thus analyse the feasibility of this approach for atmospheric retrievals.
| Talk category | NOVA Network 2 |
|---|---|
| Second preference | Opportunities of LLMs in Astronomical research, Data science |
