Quantifying Black Carbon Mixing State Heterogeneity Using a Machine Learning Model

Abstract The climate impact of black carbon (BC) is strongly sensitive to the mass ratio of non‐BC coatings to BC (R BC ). However, current global climate models (GCMs) typically assume a uniform composition within individual BC‐containing particle populations, neglecting the significant particle‐to‐particle R BC heterogeneity. To address this, we train a machine learning (ML) emulator using high‐fidelity particle‐resolved model (PartMC‐MOSAIC) simulations, and then integrate the ML emulator with the Community Atmosphere Model version 6 (CAM6) to quantify global R BC distributions. By integrating k‐means clustering with multi‐metric cross‐validation on PartMC‐MOSAIC data, we identify three distinct patterns corresponding to progressive aging: long‐tailed (fresh BC), bimodal (transitional BC), and unimodal (aged BC), accurately fitted using Gamma, Bi‐Gaussian, and Gaussian functions, respectively. Our trained ML emulator predicts both the R BC pattern categories (81% overall accuracy) and their specific function parameters with coefficients of determination ( R 2 ) above 0.62 for all the parameters. We utilize offline CAM6 outputs to drive the ML emulator for global predictions, yielding R BC distributions highly consistent with multi‐site field observations ( R 2 > 0.74). Ultimately, global quantification reveals three distinct regimes with the following occurrence frequencies: the bimodal regime (16%) with the strongest heterogeneity (coefficient of variation, CV > 1), the unimodal distribution (70%) with the highest homogeneity (CV < 0.4), and the long‐tailed regime (14%, CV < 0.6). This study establishes a robust ML framework for BC mixing state heterogeneity, providing a critical tool for improving BC parameterizations and radiative forcing assessments within GCMs.