Abstract

“We use global, regional, and local models to evaluate the chemical composition of the air we breathe and explore questions motivated by air quality and climate change. These models require 2D input values for variables such as concentrations of chemicals at the sea surface. However, it is neither desirable nor possible to observe all species in all locations for reasons including technical challenges and available resources. Therefore, we need ways to translate the sparse sea-surface observations into spatial fields for models. Currently, this is often done by using simplistic parameterisations through spatially resolved proxies, such as sea-surface temperature or ocean colour data retrieved by satellites. This approach’s accuracy suffers when there is a paucity of data or a lack of full understanding of chemical or physical processes that dictate concentrations. Machine learning techniques offer strong potential to improve and refine parameterisations based on spatially resolved proxies. An example is sea-surface halogen species; due to their potential to destroy ozone, a climate and air-quality gas, improving their parameterisation is important. Here I present examples of machine learning techniques that can be applied to sea-surface concentrations of species, including iodide (I-), bromoform (CH3Br) and dibromomethane (CH2Br2). I show this approach has skill for these species and is transferable to other species and problems in environmental chemistry. “