A machine learning-based parameterization of OH

A practical limitation of Earth System Models (ESMs) is the computational expense associated with the numerical simulation of key processes of the Earth system, such as radiation or atmospheric chemistry. This is particularly true as ESMs become more coupled and the representation of Earth system processes becomes more comprehensive. The development of ESMs requires computationally-efficient, yet accurate, modules to capture the feedbacks between elements of the Earth system. Within the NASA GEOS ESM, the computationally Efficient CH4–CO–OH (ECCOH; pronounced “echo”) module interactively simulates the atmospheric chemistry of the CH4–CO–OH cycle. Developed ~20 years ago, the parameterization of OH within the ECCOH module represents OH as a set of high-order polynomials that describe the functional relationship between the concentration of OH and meteorological, solar irradiance, and chemical variables.

However, an opportunity for updating and improving this parameterization is afforded through methods of machine learning. These methods are computationally efficient and well suited for simulating non-linear, multi-dimensional systems. Here we present an updated version of ECCOH based on machine learning. Using the NASA MERRA2-GMI full chemistry replay simulation (driven by MERRA-2 meteorology and available at ~50km horizontal resolution from 1980-2016) as a training dataset, we have developed parameterizations for OH using neural networks and extreme gradient boosting (XGBoost). We will highlight the capabilities and limitations of these approaches, and discuss how they can be further improved through targeted parameter selection and training dataset balancing.