Compact aerosol size-composition representations preserving CCN, optical, and ice nucleation properties

Aerosol models rely on reduced state representations to remain computationally tractable, yet commonly used modal and sectional approaches impose strong assumptions about size distribution shape and chemical mixing state. These assumptions can limit fidelity for aerosol processes that depend sensitively on detailed size–composition structure, including cloud droplet activation, aerosol optical properties, and ice nucleation. Here, we present a data-driven framework for learning compact aerosol state representations directly from particle-resolved simulations. Using a large library of aerosol populations generated with the PartMC–MOSAIC model, we compress speciated mass and number size distributions comprising hundreds of variables into a ten-dimensional reduced representation. We evaluate the fidelity of this representation using climate-relevant aerosol diagnostics, including cloud condensation nuclei (CCN) spectra, wavelength-dependent scattering and absorption coefficients, and immersion freezing frozen-fraction spectra. We find that CCN spectra are reconstructed with the highest accuracy, optical properties with moderate accuracy, and ice nucleation properties with the greatest difficulty, reflecting the differing information content and nonlinearity of these processes. To improve robustness, we introduce a noise-resilient preprocessing strategy and a realism metric based on the sliced Wasserstein distance to guide model regularization. These results demonstrate that detailed aerosol size–composition distributions can be represented compactly while preserving key aerosol–cloud–radiation diagnostics, providing a foundation for reduced aerosol state representations that complement existing aerosol modeling approaches.