Improving WRF-Chem aerosol optical depth prediction over India using artificial neural network

Accurate prediction of aerosol optical depth (AOD) is critically important in atmospheric science research, as aerosols significantly influence climate forcing, air quality, and human health. In this study, AOD for the years 2010 to 2012 is predicted using the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem). Furthermore, the WRF-Chem AOD predictions are corrected using an artificial neural network (ANN) to reduce systematic errors. The corrected AOD values are then compared with observations from the AErosol RObotic NETwork (AERONET), retrievals from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite, and final analysis from the Monitoring Atmospheric Composition and Climate (MACC) and the Navy Aerosol Analysis and Prediction System (NAAPS) products. Results suggest that the default WRF-Chem predictions (WRFO runs) improve significantly with the use of machine learning (here, a feed-forward neural network) in the WRF-Chem model predictions (WRFN runs). Moreover, the spatial and temporal distributions of AOD in the WRFN runs show notable improvement compared with independent analyses from MACC and NAAPS, MODIS satellite measurements, and AERONET observations. A statistical t-test shows significant improvement in error statistics for WRFN throughout the study period over the Indian region. The WRFN runs show a higher correlation of 0.58 with AERONET observations compared to 0.13 for WRFO runs, which is similar to correlations from the MACC reanalysis and NAAPS analysis. The bias in WRFN is reduced from 0.40 to zero, and the RMSD decreases from 0.50 to 0.25. These results suggest that the quality of WRF-Chem predictions can be further enhanced using machine learning techniques.