XClose

Advanced Research Computing

Home
Menu

TechSocial Series - September 2024

'SpeedyWeather.jl: How to build an atmospheric model towards extensibility and interactivity for', by Milan Klöwer - Schmidt AI in Science Fellow at the University of Oxford

Speaker Bio

Milan Klöwer is a Schmidt AI in Science Fellow at the University of Oxford. After a postdoc at the Massachusetts Institute of Technology, he returned to Oxford where he did his PhD in low-precision climate computing. Milan is climate scientist turned Julia developer with a background in climate physics and studied in Germany, France and Norway. His other research interests include data compression, predictability of weather and climate, forecast and science communication and he has also extensively worked on aviation and decarbonisation.


Abstract

Machine learning is increasingly used in climate science. Most such efforts focus on post processing climate data, so-called offline learning. However, there is a large potential to include machine-learned representation of the more uncertain climate processes (for example to represent clouds or precipitation) directly in climate models. Such neural general circulation models couple conventionally solved dynamical equations (which have relatively small uncertainty) with neural networks to represent the uncertain climate processes. But how to build a climate model, or more specifically an atmospheric model to be easily extensible and ideally also interactive to increase productivity? Here, we are building SpeedyWeather.jl in the Julia programming language for this purpose. Such a model should be modular: Easily plugging in new, potentially machine learning-based, components needs to be easy. It needs to be hackable, allowing for a wide range of future modifications, especially outside of the intended modularity. And it needs to be interactive, productively combining a simulation and its analysis and visualisation. At the same time we want SpeedyWeather to be flexible: Running it on CPUs and GPUs (work in progress) or changing the numerical precision from double to single or even down to 16 bits, with deterministic or stochastic rounding. All towards a productive model for atmospheric and climate research that also keeps every software engineer happy.


Archived Media

Recording |