Physical informed neural network improving the WRF-CHEM results of air pollution using satellite-based remote sensing data

Accurate measurement of air gases concentration are crucial for effective air quality management and minimizing the harm caused by pollution events, as well as providing guidance for healthy travel. We developed develop a physical informed deep learning model that combines a traditional atmospheric chemical transport model and a data-based deep learning model. The model uses multisource observation data to constrain and optimize the results of atmospheric pollution gases concentrations. The accuracy and spatial resolution in central and eastern China (20–45°N, 100–125°E) have been improved compared to the original Weather Research and Forecasting model coupled to Chemistry (WRF-CHEM) results by incorporating constraints from ground-based and satellite observations. Compared with the WRF-CHEM results, our model increases the verification Pearson coefficient (R) of NO2 and CNEMC sites from 0.56 to 0.8, and the RMSE decreases from 25.33 to 15.54 μg cm⁻³. For O3 results, the model increased the verification R of the CNEMC from 0.57 to 0.76, and the RMSE decreased from 24.55 to 20.22 cm⁻³. We employed independent MAXDOAS monitoring to authenticate the HCHO results. The R was 0.78, while the validation results for satellite HCHO and MAXDOAS exhibited a R of 0.63 and the validation results for WRF-CHEM and MAXDOAS demonstrated a R of 0.11. Additionally, the RMSE was 6.97E15 mole cm⁻², whereas the validation results for satellite HCHO and MAX-DOAS had an RMSE of 7.16E15 mole cm⁻². By utilizing satellite remote sensing to effectively capture the spatial distribution of pollutants, and by combining satellite data with other sources of monitoring data and optimizing our models, we have improved the accuracy of our results and have overcome the limitation of cloud coverage. As a result, we are now able to better understand the space-time distribution of pollutants in different regions.