Diagnosing the Model Bias in Simulating Daily Surface Ozone Variability Using a Machine Learning Method: The Effects of Dry Deposition and Cloud Optical Depth

Machine learning methods are increasingly used in air quality studies to predict air pollution levels, while few applied them to diagnose and improve the underlying mechanisms controlling air pollution represented in chemical transport models (CTMs). Here, we use the random forest (RF) method to diagnose high biases of surface daily maximum 8 h average (MDA8) ozone concentrations in the GEOS-Chem CTM evaluated against measurements from the nationwide monitoring network in summer 2018 over China. The feature importance results show that cloud optical depth (COD), relative humidity, and precipitation are the top three factors affecting CTM high biases. Such results indicate that the high ozone biases in summer over China mainly occur on wet/cloudy days (∼40% biased high), while biases on dry/clear days are small (within 5%). We link the important features with model parameterizations and variables, identifying model underestimates in the dry deposition velocity and COD on wet/cloudy days. By accounting for the enhanced dry deposition on wet plant cuticles and using satellite observation constrained COD, we find that CTM high ozone biases can be halved with an improved agreement in the temporal variability, highlighting the effects of dry deposition and COD on ozone, as suggested by the RF outcomes.