The Paper

I am a co-author on the paper “Embedding machine-learnt sub-grid variability improves climate model precipitation patterns”, published in Nature Communications Earth & Environment. You can find the full paper in the publications section.

The Problem

Climate models often struggle to accurately represent precipitation patterns. This is partly due to the difficulty of modeling the complex, small-scale processes that lead to cloud formation and rainfall.

Our Approach

In this work, we developed a novel method for improving precipitation patterns in climate models by using Gaussian processes to represent sub-grid variability. Our approach involves:

• Learning from data: We trained a Gaussian process model on high-resolution data to learn the statistical relationships between large-scale atmospheric variables and small-scale precipitation patterns.
• Stochastic parameterization: We then used this model to generate stochastic perturbations that were injected into the climate model at runtime.

The Impact

Our results show that this method leads to a significant improvement in the accuracy of precipitation forecasts. This work has been presented at several major conferences, including:

• SIAM Conference on Computational Science and Engineering (CSE25)
• SIAM Conference on Uncertainty Quantification (UQ24)
• Climate Informatics 2024
• RMetS Annual Weather and Climate Conference 2024

The code for this project is available on GitHub.