Simulation of chemical transport model estimates by means of a neural network using meteorological data

Abstract

Chemical substances of either anthropogenic or natural origin affect air quality and, as a consequence, also the health of the population. Therefore, there is a high demand for reliable air quality scenarios that can support possible management decisions. However, generating long term assessments of air quality assuming different emission scenarios is still a great challenge when using detailed atmospheric chemistry models. In this study, we test machine learning technique based on neural networks (NN) to emulate process-oriented modeling outcomes. A successfully calibrated NN might estimate concentrations of chemical substances in the air several orders faster than the original model and with reasonably small errors. We designed a simple recurrent 3-layer NN to reproduce daily mean concentrations of NO2, SO2 and C2H6 over Europe as simulated by the Community Multiscale Air Quality model (CMAQ). The general structure of the NN can be shown to approximate a continuity equation. Inputs of the network are daily mean meteorological state variables, taken from the climate model COSMO-CLM. The proposed NN emulates CMAQ outputs with an error not exceeding the difference between CMAQ and other known chemical transport models.