Start with the scale gap
A global climate model divides the atmosphere into grid cells. The cell is useful for planetary patterns, but it is too coarse to explicitly represent many clouds and turbulent motions.
Research explained
Global climate models cannot resolve every cloud and turbulent motion. My research explores how super-parameterisation can provide that missing detail, and how a machine learning model might eventually reproduce it at a fraction of the cost.
A GCM provides the large-scale state; an embedded CRM supplies detail that the grid cannot resolve.
A global climate model divides the atmosphere into grid cells. The cell is useful for planetary patterns, but it is too coarse to explicitly represent many clouds and turbulent motions.
In super-parameterisation, a small cloud-resolving model (CRM) is embedded inside each GCM grid column. It resolves more of the physics and returns the subgrid tendencies to the larger model, but running all of those CRMs is computationally demanding.
The CRM becomes a high-fidelity teacher. A machine learning emulator learns the relationship between the GCM state and the CRM's tendencies, with the goal of replacing the costly inner loop while retaining the useful small-scale effects.
A model that works in one climate model is not automatically useful elsewhere. The final test is transferability: can the learned emulator work with other GCMs, grids, resolutions, and climates without losing physical credibility?
The question
This is not only a question of prediction accuracy. It is also about stability, physical behaviour, and whether one learned representation can travel between the climate models that scientists actually use.
Talk to me about the research