A machine learning approach for refining ECHAM-HAMMOZ-derived PM2.5

In this study, we present our findings for correcting global model-derived surface fine particulate mass (PM2.5) concentrations with a machine learning approach. We simulated the PM2.5 concentrations with an aerosol-climate model ECHAM-HAMMOZ, and trained a machine learning model to downscale the PM2.5 concentrations modelled for an Indian mega city, New Delhi. This way, we are able to utilize a global atmospheric model for analyzing aerosol emission mitigation effects on both Earth's energy budget and local air quality.

Similarly as with many other global-scale models, ECHAM-HAMMOZ underestimates surface PM2.5 at several urban locations. One apparent explanation for this is the coarse grid resolution of global climate models, which results in averaged aerosol concentrations over a much larger area than what urban cities typically cover. Therefore, due to averaging over a large grid box, the very high peak concentrations from urban areas can be evened out. Furthermore, the input fields describing aerosol emissions might lack information of some local emission sources, which can as well affect the simulated surface air pollution levels.

We used the random forest (RF) regression algorithm in order to downscale ECHAM-HAMMOZ-derived surface PM2.5 concentrations towards measured PM2.5 values from the New Delhi capital region in India. In addition, we applied the trained RF model to additional simulations where we had future anthropogenic aerosol emissions according to a business as usual scenario and a mitigation scenario. This allowed us to evaluate the effects of aerosol emission reductions on both global radiative balance, and on local air quality in New Delhi.

The obtained results indicate that surface PM2.5 concentrations from RF prediction correlate with the measured PM2.5 concentrations much better than the original ECHAM-HAMMOZ particulate concentrations for New Delhi region. However, with the current setup and input variables, the PM2.5 concentrations produced by the RF model seems to be lacking some of the short-term variations and very low and high values.

All in all, the downscaling method used in this project shows very promising potential, but requires further adjustment with the selection of input variables and the RF hyperparameters.