The development of a machine learning diagnostic package to enhance chemical mechanism implementations within atmospheric chemistry models

The chemical solver in air quality models (AQMs) is often the most time-consuming component. Emulation via machine learning offers a computationally inexpensive alternative method to simulate atmospheric chemistry. Here we are developing a model agnostic user-friendly machine learning package using regression trees to run on multiple different AQMs including GEOS-Chem and CMAQ. The developed package has the potential to support running large model ensembles and enable the inclusion of complex situation specific surrogate chemistry schemes (e.g., biomass burning, urban, biogenic). Additionally, the package will include a diagnostic toolset that will be developed to facilitate the study of the chemical mechanisms, aid scientific insight and to support future development of AQMs Due to the large number of target variables (chemical species) coupled with a high degree of spacial and temporal chemical variation. Efficiently processing the very large training chemical datasets represents a substantial technical hurdle.