Machine learning for earth system observation and forecasting

In this thesis we explore the idea of replacing complex, expensive systems for earth system observation and forecasting with simple, streamlined machine learning models that achieve superior performance at a fraction of the cost. Three separate applications of this idea are investigated. We begin by developing Aardvark Weather, the first end-to-end data-driven weather forecasting system. We demonstrate that Aardvark not only produces skilful forecasts for both global and local forecasting, but outperforms fully operational systems for multiple variables and lead-times. We next explore an earth observation task, developing the first systems to automatically detect methane super-emissions in multi-spectral satellite imagery. The resulting model, MARS-S2L, is now operational at the United Nations Environment Programme and is capable of detecting emissions globally with formal notifications already issued to governments and stakeholders with the ability to act in 20 countries. We present a case study of how these notifications were used in the mitigation of a super-emitter in Algeria, an equivalent annual climate impact to removing 480,000 cars from US roads or entirely mitigating the emissions of several countries. Finally, we turn to foundation modelling and build Aurora, the first large-scale foundation model for the earth system. We show that Aurora outperforms current operational dynamical models for tropical cyclone track, atmospheric chemistry and ocean wave forecasting at a fraction of the cost of existing systems at inference time. We conclude with a discussion of how these three works can be combined into a next-generation earth system observation and forecasting model.