Abstract Atmospheric deposition of dissolved inorganic phosphorus (DIP) is crucial to nutrient cycling but remains poorly constrained in chemical transport models (CTMs). Large uncertainties in phosphorus emissions and aerosol dissolution processes challenge the reliable estimates of DIP fluxes. Here, we developed a two‐stage random forest residual learning framework integrating observational DIP data sets with CTM simulation outputs. Compared to the original CTM results, our machine learning model substantially narrows the gap of DIP deposition fluxes across the globe between the observations and model estimates, with improved Pearson correlation from 0.10 to 0.85. The optimized global DIP deposition is estimated of 2.02 Tg DIP yr −1 , with a 95% confidence interval of 1.98–2.59 Tg DIP yr −1 . Our estimate is 60%–460% higher than the CTM‐based estimates from previous studies. In particular, 1.35 Tg DIP yr −1 (67%) is deposited over land and 0.67 Tg DIP yr −1 (33%) over ocean. The source appointment analysis suggests dust‐ related predictor variables (including dust aerosols and dust‐sulfate interactions) and black carbon as the major contributors to correcting CTM‐observation bias in DIP deposition. Overall, our approach helps to better understand the atmospheric phosphorus cycling and enhance Earth system models' capabilities by refining ecosystem response simulations.
Jiang et al. (Sat,) studied this question.