Aerosol Iron Solubility Specification in the Global Marine Atmosphere with Machine Learning

Aerosol iron (Fe) solubility is a key factor for the assessment of atmospheric nutrients input to the ocean but poorly specified in models because the mechanism of determining the solubility is unclear. We develop a deep learning model to project the solubility based on the data that we observed in a coastal city of China. The model has five variables: the size range of particles, relative humidity, and the ratios of sulfate, nitrate and oxalate to total Fe (TFe) contents in aerosol particles. Results show excellent statistical agreements with the solubility in the literature over most worldwide seas and margin areas with the Pearson correlation coefficients (r) as large as 0.73–0.97. The exception is the Atlantic Ocean, where good agreement is obtained with the model trained using local data (r: 0.34–0.66). The model further uncovers that the ratio of oxalate/TFe is the most important variable influencing the solubility. These results indicate the feasibility of treating the solubility as a function of the six factors in deep learning models with careful training and validation. Our model and projected solubility provide innovative options for better quantification of air-to-sea input of aerosol soluble Fe in observational and model studies in the global marine atmosphere.