Combining machine learning and satellite observations to predict spatial and temporal variation of surface OH in cities

The hydroxyl radical (OH) is the primary cleansing agent in the atmosphere. The abundance of OH in cities initiates the removal of local pollutants; therefore, it serves as the key species describing the urban chemical environment. We propose a machine learning (ML) approach as an efficient alternative to OH simulation using a computationally expensive chemical transport model. The ML model is trained on the parameters simulated from the WRF-Chem model, and it suggests that six predictive parameters are capable of explaining 76% of the OH variability. The parameters are the tropospheric NO2 column, the tropospheric HCHO column, J(O1D), H2O, temperature, and pressure. We then use observations of the tropospheric NO2 column and HCHO column from OMI as input to the ML model to enable measurement-based prediction of daily near surface OH at 1:30 pm local time across 49 North American cities over the course of 10 years between 2005 and 2014. The result is validated by comparing the OH predictions to measurements of isoprene, which has a source that is uncorrelated with OH and is removed rapidly and almost exclusively by OH in the daytime. We demonstrate that the predicted OH is, as expected, anticorrelated with isoprene. We also show that this ML model is consistent with our understanding of OH chemistry given the solely data-driven nature.