Weather and climate forecasts are produced using complex computer simulators of the Earth System. The process of developing weather and climate simulators involves representing the physics equations describing the system in a piece of computer code. Doing so involves selecting a discretisation scale in space and time, which is in effect the smallest scale represented in our solution. However in weather and climate models this effective pixel-size is on the order of 5-50km, so there are many processes occurring at scales smaller than this which are important and must be represented in some way. We do this using so-called ‘parametrisation’ schemes. This parametrisation process introduces uncertainty into forecasts, due to the simplifications and approximations that must be made. And even though such schemes are only an approximation to the small scales, they are very computationally expensive.
An exciting new area of research is using ideas from the machine learning and artificial intelligence communities to improve our weather and climate simulators. One potential application is to replace costly parametrisation schemes with a computationally cheap neural network. This project would investigate the use of machine learning to efficiently process large volumes of observational or high resolution model data, with the goal of building a neural network that accounts for the uncertainty in parametrisation schemes as well as their mean effect. In this way, neural networks could not only significantly speed up our weather and climate simulators, but also improve their fidelity.