Advancing Black Carbon Mass Absorption Cross-section Estimation Using Machine Learning Models
UZZAL KUMAR DASH, Andrew A. May, Haikal Rozaidi, Dominic Blackston, Aadit Shah, The Ohio State University
Abstract Number: 433
Working Group: Aerosols, Clouds and Climate
Abstract
Black carbon (BC), a light-absorbing component of atmospheric aerosols, plays a significant role in Earth’s climate system. Its impacts are typically quantified using chemistry-climate models, but predictions of BC mass concentrations often diverge due to inconsistencies in aerosol microphysics among models. Model evaluation using observations is essential, yet most ground-based and remote sensing observations quantify BC indirectly via light absorption measurements rather than direct mass concentrations. Converting absorption measurements to mass concentrations often assumes a constant, wavelength-specific mass absorption cross-section (MAC), which may not capture real-world variability. Alternately, relating light absorption measurements and modeled BC mass concentrations requires assumptions about the applicable light scattering theory and aerosol properties (e.g., mixing state, complex refractive index). To address these limitations, we propose a third approach: a statistical framework that employs multiple regression and machine learning models to estimate the temporal variability of the mass absorption cross-section of black carbon (MACBC). This enables more accurate conversion of aerosol light absorption measurements to BC mass concentrations for both filter-based absorption photometers and remote sensing platforms, with the potential to improve the evaluation of chemistry-climate model predictions. Building on our previous work, which used ground-based observations from a single DOE campaign, we expanded our MACBC prediction tool by constructing comprehensive database incorporating airborne observations from nine NASA and five DOE field campaigns. Multiple regression and machine learning models were trained on this database where we incorporated additional input variables such as the Absorption Ångström Exponent (AAE), Scattering Ångström Exponent (SAE), and Single Scattering Albedo (SSA), along with existing input parameters including multi-wavelength absorption and scattering coefficients and aerosol particle size distributions. The trained models were evaluated using an independent validation dataset, and potential sources of systematic bias in predicted MACBC values were diagnosed.