Machine Learning Physics Parametrisation: Impact of pre-processing and architecture

Abstract:

Machine learning promises to deliver big gains in efficiency and accuracy in the weather and climate models and increasing number of atmospheric science researchers are adopting machine learning methods to improve their models. But just how easy is it to develop these emulators? In this presentation I will go through some of the issues and successes of developing a neural network based emulator for moisture and heat increments from a cloud resolving model (CRM). We have used Met Office high resolution nested model as the ground truth in physics parametrisation and then coarse grained to global model resolution. We haved aimed for NWP-like set-up for the CRM, which creates challenges for the emulator to generalise and leads to instabilities in prognostic validations. I will outline some of the key data processing steps and neural network architectures required to have a high diagnostic accuracy as well as a relatively stable prognostic emulator. Finally, I will also discuss if all the data processing and transformations tricks mean that linear statistical models can also perform as well as neural networks.