We want to produce high-fidelity temperature predictions for a mountainous, costal region of Peru.


Region of Interest (ROI) for high-fidelity temperature predictions

We have access to a Regional Climate Model (RCM) which can compute these predictions, but it is prohibitively expensive to run for the whole region of interest. We also have access to high-fidelity altitude data and a Global Climate Model (GCM) which can compute inexpensive, low-fidelity temperature and wind speed predictions.


Low-fidelity temperature and wind speed (GCM)

High-fidelity altitude

High-fidelity temperature (RCM, target)

We use the RCM to produce high-fidelity temperature predictions over small regions within the region of interest. These are much less expensive to produce. Then, we use Gaussian Processes combine these predictions with the low-fidelity temperature predictions from the GCM and the high-fidelity altitude data to infer the high-fidelity temperature predictions over the entire region of interest.

For example:


Low-fidelity temperature training data

High-fidelity temperature training data

Inferred high-fidelity temperature predictions

True high-fidelity temperature predictions

This work done during my master’s degree at University of Cambridge.

Read on arXiv.