The Challenge

Climate models are essential tools for understanding and predicting climate change, but they are limited by their inability to represent physical processes that occur at scales smaller than the model’s grid resolution. This “sub-grid” variability is a major source of uncertainty in climate projections.

My Contribution

I have developed a novel method for representing sub-grid variability in climate models using Gaussian processes. My approach involves:

1. Modeling sub-grid variability: I use Gaussian processes to learn a statistical model of the sub-grid variability from high-resolution data.
2. Stochastic parameterization: I then use this model to generate stochastic perturbations that are injected into the atmospheric fields of a Fortran-based climate model (SPEEDY) at runtime.

This method effectively introduces a more realistic representation of sub-grid processes into the climate model, leading to more accurate and reliable simulations.

Next Steps

This work is now being extended to a much more complex climate model, CESM CAM, in partnership with IIT Delhi. This extension requires the development of more advanced Gaussian process methods using GPyTorch, and their deployment at runtime using FTorch, a tool for calling PyTorch models from Fortran developed by the Alan Turing Institute’s Research Engineering Group.

This project showcases my expertise in combining advanced statistical methods with high-performance computing to address fundamental challenges in climate science.