Neural network emulation of the formation of organic aerosols based on the explicit GECKO-A chemistry model

Abstract

Secondary organic aerosols (SOA) have been an active area of research over the past decades with the goal of improving their representation in air quality and climate models (Hodzic et al., 2016; Tsigaridis et al., 2014), which is essential for predicting their effect on human health (Mauderly & Chow, 2008) and their contribution to radiative forcing in the climate system (Boucher et al., 2013). The misrepresentation of SOA formation pathways in 3D models has led to a long-standing discrepancy between observed and modeled organic aerosol concentrations that has been reported from urban to remote regions (de Gouw, 2005;Hodzic et al., 2020). Unlike sulfate and other inorganic aerosols, which are made from a few dominant chemical pathways, SOAs result from the condensation of a very large number of partly oxidized gases. These gases are generated from the multi-generational oxidation of volatile organic compounds (VOCs) emitted from anthropogenic and natural sources. This complexity is not included in current 3D models that rely on simplified SOA parameterizations that have been developed and

An Online‐Learned Neural Network Chemical Solver for Stable Long‐Term Global Simulations of Atmospheric Chemistry

“…We recognize that parameter tuning has the potential to improve the accuracy of our results. Furthermore, due to computational expense, we were not able to generate an ensemble of ML solvers for each season which may also provide a source of improved accuracy and stability (Schreck et al., 2022). These are directions for future research.…”

“…Moreover, stability of the solution is required over the full range of tropospheric conditions from polluted to remote and from the surface to the upper troposphere. Operator splitting between chemistry and transport resets initial conditions after each transport time step, meaning that one cannot easily project the solution along long-term chemical trajectories as with dedicated ML time series algorithms such as Recurrent Neural Networks (RNNs) (Rumelhart et al, 1986) and Long-Short-Term Memory networks (LSTMs) (Hochreiter & Schmidhuber, 1997) without introducing additional complexity (Schreck et al, 2022). Success in applying ML solvers to box models, such as in Kelp et al (2020), may not translate to a global CTM.…”