Machine learning models for aerosol particle size hygroscopic growth factor

Based on the hourly observational data of nephelometer, aethalometer and GRIMM180 environment particle monitor from October to December 2017 in Chengdu, as well as the simultaneous data of atmospheric visibility (V), relative humidity (RH) and nitrogen dioxide (NO2), aerosol hygroscopic growth factor (Gf) was retrieved by the aid of Mie scattering theory and immune evolutionary algorithm. Firstly, RH, CBC, CBC/CPM1, CPM1/CPM2.5 and CPM2.5/CPM10 were used as explanatory variables set, three machine learning models for aerosol particle size hygroscopic growth factors were constructed (XGBoost model, CatBoost model, and LightGBM model), and the corresponding judgment coefficients (R2) were 0.869, 0.893 and 0.898, root mean square error (RMSE) were 0.108, 0.097 and 0.090, mean absolute error (MAE) were 0.061, 0.054 and 0.052, respectively. Tests of three models showed that, machine learning models for aerosol particle size hygroscopic growth significantly reduced the simulation bias of traditional univariate aerosol particle size hygroscopic growth models under high humidity conditions, and it also improved the calculation accuracy of multivariate GAM model for aerosol particle size hygroscopic growth. Finally, the effects on different explanatory variables of the simulation results of machine learning models were analyzed, black carbon was confirmed as the main control variable in the aerosol hygroscopic growth model. The above study further explained the complexity of the multifactorial influences on aerosol particle size hygroscopic growth factors, and provided a new approach to scientifically characterisation of Gf models. © 2024 Chinese Society for Environmental Sciences. All rights reserved.