Fast generation of meteorological data using machine learning

Abstract

This study aims at speeding up the acquisition of meteorological data for use in the chemical transport model TM5 for calculating methane mixing ratios, by using AI to generate meteorological data. The study starts off with a literature review in which meteorology and AI is discussed, after which it has been decided that GraphCast is the most suitable forecasting model for this project. In order to match GraphCast’s output to TM5’s input, three processing steps have been defined and verified. Based upon this verification it appears that that the conversion of the temperature, humidity and horizontal massfluxes were successful, however the conversion of the vertical massflux was unsuccessful. Subsequently the new pipeline was integrated in TM5, followed by another verification step. From this step it once again appeared that the vertical massflux was incorrect, and was not incorporated. The remainder of the pipeline was incorporated. Finally GraphCast generated meteorological data was ran in TM5, with forecasting lengths ranging from 6 to 72 hours. From this step it appeared the meteorological data provided good results from TM5. The error to the reference simulations increased for longer forecasting times. Peculiarities were found in the stratosphere, where a poor forecast by GraphCast resulted in large errors in the methane mixing ratio of TM5, which would carry over in an error on the South Pole. Similarly a reduction in vertical resolution in the stratosphere resulted in similar errors. The quality of the data close to the surface has generally remained good though. Overall the performance of TM5 was still good, and despite improvements being possible, the results were encouraging.