Low Cost Climate Prediction

Task introduction & significance

Climate prediction models will play a key role in effectively modelling the Earth’s changing climate. The current trade-off between predictions cost and accuracy is an open challenge that needs to be addressed as heterogeneous observational data becomes more readily available.

Simulators are useful for modelling climate and weather patterns, however when using simulators, there is often a trade-off between cost and accuracy. Low-fidelity data can be produced easily using inexpensive and approximate simulation methods, yet it often deviates significantly from reality. In contrast, high-fidelity data may more closely resemble the real-world system. This can be gathered from real-world measurements or computationally expensive simulations.

Multi-fidelity modelling provides a useful framework for combining the accuracy of high-fidelity data with the low cost of low-fidelity data. The goal is to develop computational methods to provide high-fidelity climate predictions from low-fidelity data. Each team is free to explore and propose its own methodologies and solution but a first step towards the goal of this task can be to reproduce the results presented by Hudson et al.[1]. This work provides high-fidelity climate predictions from low-fidelity data for a mountainous region within Peru. The data is available here .

Teams can explore validate their models with other datasets once they have an initial model to provide high-fidelity climate predictions from low fidelity data.

Helpful tools & resources

Initial notebook (you can create your own copy):

Useful resources Other potentially useful datasets:


  1. Ben Hudson, Frederik Nijweide, and Isaac Sebenius. "Computationally-Efficient Climate Predictions using Multi-Fidelity Surrogate Modelling." Preprint, ArXiv (2021).