Use of machine learning to reduce uncertainty in anthropogenic radiative forcing associated with aerosol–cloud interactions

Abstract:

The effective radiative forcing (ERF) of anthropogenic aerosols associated with aerosol–cloud interactions (ERFaci) remains the largest source of uncertainty in climate prediction. The calculation of particle number concentration (PNC), one of the key parameters affecting ERFaci, is generally simplified in climate models to avoid expensive computing cost and thus contributes to the uncertainty in ERFaci. Here we employ outputs from long-term (30-years) simulations of a global size-resolved (sectional) aerosol microphysics model (GEOS-Chem-APM) and a machine-learning tool to develop a Random Forest Regression Model (RFRM) for PNC. The GEOS-Chem-APM contains several features of relevance towards the accurate simulation of PNC, including 40 bins to represent secondary particles, and the state-of-the-art ternary ion mediated nucleation mechanism and nucleation involving highly oxidized organics. GEOS-Chem-APM PNC simulations have been validated with a large number of in-situ measurements. The RFRM, trained on GEOS-Chem-APM outputs, captures the complex dependence of PNC not only on aerosol mass concentrations but also on other key variables representing meteorological and chemical conditions. We have implemented the PNC RFRM in the version of GISS-ModelE2.1 with a mass-based One-Moment Aerosol (OMA) module, which is one of the models participating in the Coupled Model Intercomparison Projects 6 (CMIP6). Here we show that the implementation of the PNC RFRM in GISS ModelE2.1-OMA improves significantly the agreement of its predicted PNC with measurements across the globe (both spatial distributions and temporal variations), reduces the relative changes of cloud droplet number concentration associated with changes of emissions from pre-industry to present-day, and decreases the ERFaci from −1.46 W⋅m⁻² to −1.11 W⋅m⁻². The PNC RFRM can be used to reduce uncertainties of climate models in predicting PNC and ERFaci without compromising their computing efficiency, which is critical for long-term climate simulations.