Physics-Informed Machine Learning for Aerosol Microphysics: Efficient Parameterization for Chemistry and Climate Models

Atmospheric aerosols influence air quality, human health, cloud formation, and climate by altering radiation and cloud microphysics. Accurately representing aerosol size and composition is key to predicting pollution episodes and climate forcing and consequent effects. Yet, resolving aerosol microphysical processes (e.g., nucleation, condensation, coagulation) remains computationally expensive in chemistry & climate models, particularly at global, multi-decadal scales.

State-of-the-science models (here GEOS-Chem with advanced particle microphysics) can explicitly simulate particle number size distributions (PNSDs), enabling improved quantifications of cloud condensation nuclei (CCN) and aerosol radiative effects. Our recent studies have demonstrated that machine learning (ML) can recover key aerosol abundance properties such as CCN0.4 and PNC from commonly available atmospheric variables. These efforts highlight ML’s potential to reduce computational burden while preserving fidelity to detailed microphysics.

However, most climate and chemical transport models still rely on bulk or simplified aerosol schemes because fully size- & composition-resolved microphysics is computationally prohibitive. It is important for ML approaches to be physicochemically grounded and extensively evaluated for reproducing composition-resolved PNSDs, particularly in the size ranges relevant to cloud activation. A scalable, interpretable, physics-informed ML framework that emulates complex aerosol microphysics within operational models remains an unmet need.

Here, we present a physics-informed random forest parameterization that emulates aerosol microphysical evolution within a chemistry transport modeling framework. Using atmospheric state variables, precursor gases, and bulk aerosol properties as inputs, the model reproduces composition-resolved PNSDs with strong agreement relative to detailed microphysics. Implemented within GEOS-Chem, the approach substantially improves aerosol abundance and optical property predictions with negligible additional computational cost. This framework enables more accurate yet efficient simulations of aerosol—cloud—climate interactions and offers a pathway for integrating physically grounded ML parameterizations into global and regional atmospheric models.