Machine Learning Derivation of Composition-Resolved Particle Number Size Distribution and Aerosol Properties
ARSHAD ARJUNAN NAIR, Fangqun Yu, The State University of New York at Albany
Abstract Number: 391
Working Group: Advancing Aerosol Science through Data Analysis Tools
Abstract
Aerosol microphysics plays a critical role in atmospheric composition, air quality, and climate. However, the computational cost of representing these microphysical processes in chemistry and climate models remains a challenge. Building on our previous work, in which we developed, optimized, and probed the inner workings of machine learning (ML) models to quantify cloud condensation nuclei (CCN0.4) and particle number concentrations (PNC), this work presents an implicitly physics-informed ML parameterization to emulate complex aerosol microphysical processes, such as nucleation, condensation, and coagulation, within a computationally efficient random forests framework. This parameterization uses inputs such as atmospheric state variables, precursor gas concentrations, and bulk aerosol properties. By integrating physical constraints with scalable ML techniques, the parameterization maintains interpretability and fidelity while accelerating simulations. The approach is validated within the framework of a chemistry transport model (GEOS-Chem), showcasing substantial improvements over the default bulk aerosol treatment in quantifying aerosol abundance and optical properties, enhancing accuracy with trivial increase in computational expense. The presented approach is particularly suited for global models lacking detailed size-resolved aerosol microphysics and is readily applicable to regional models (WRF-Chem, WRF-GC, CMAQ, etc.). This development can enable more detailed, computationally feasible simulations of aerosol-cloud-climate interactions, offering reliable and actionable insights.