Building a Deep Learning Model for Atmospheric Transport of Gases

Abstract

Obtaining accurate estimates of the distribution of surface emissions of gases like CO2 , NO2 , CO, etc is a technically challenging as well as environmentally relevant problem. Many data assimilation-based algorithms have been developed to tackle this problem over decades. While these algorithms have their own strengths, they also face some prohibitive challenges, like high computational cost, deep mathematical and computational expertise, and errors arising from various numerical simplifications and parameterizations. The rapidly evolving field of deep learning can help address these limitations due to their GPU-powered fast computations and relative ease of application. In this work, we propose a deep learning algorithm to estimate the emission distribution of a given gas, using future information of its concentrations in the atmosphere. The algorithm involves building a deep learning surrogate of an existing numerical transport model and then using the automatic differentiability of neural networks to perform emission estimation. This thesis focuses on the first part, ie building a deep learning model that is trained to predict future concentrations of a gas (CO2 ) given its initial concentrations, the geographical distribution of its emissions/sinks, and the relevant meteorological conditions (eg winds, pressure etc). The numerous model-related choices are described, various alternatives are tested and the model with relatively the best performance is analysed. Our model learns atmospheric transport to a significant degree, having an average error in predicted concentrations of about 0.104 ppm. This thesis is based on the work done at TCS Research, Pune.