GCMs are fairly coarse models, run with grid boxes that are tens or hundreds of kilometers across. To represent features like clouds and updrafts that are smaller than this, they use parameterizations that predict these small-scale features based on larger, resolved features. These parameterizations are tested against output from high-resolution LES models. Machine learning is a promising tool to improve these parameterizations.
Machine learning is applied in statistical downscaling of coarse model output, and neural networks have been used to emulate existing atmospheric parameterizations with 50-80x speedups. Problems of physical parameterization in atmospheric models can be seen as supervised learning problems. I am interested in potential applications of machine learning as a driver of GCM parameterizations, with the use of high-resolution models and analysis products to create training data.
Model code is often unclear, tightly coupled instead of modular, poorly documented, and without unit tests, causing long delays in research. I helped develop a modelling framework called Sympl to address these issues. It uses Python for flexible, readable, self-documenting overhead code while allowing compiled languages for computation, and can also be used for developing pure Fortran models. You can find the documentation here and the code here.
Jeremy is a machine learning research scientist for climate model development at Vulcan in Seattle, Washington. He received his B. Sc. in Physics from the University of Toronto in 2014, and graduated with his doctorate in Atmospheric Sciences from the University of Washington in 2019.
He is currently researching how machine learning can be applied in parameterizations for global climate models.