Deep Probabilistic Forecasting of Global Temperature Fields

Description

In this challenge, students will use ERA5 reanalysis data to develop deep learning models that forecast monthly global temperature maps. The focus is on quantifying and analyzing predictive uncertainty, including aleatoric (data-related) and epistemic (model-related) components. Students will implement CNN/ConvLSTM architectures in PyTorch, explore probabilistic outputs, and evaluate their models using metrics such as RMSE, NLL, CRPS, and calibration diagrams, providing insights into spatial, temporal, and seasonal patterns of uncertainty.

Recommended reading material

• https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5
• https://doi.org/10.1146/annurev-statistics-062713-085831
• https://arxiv.org/abs/1612.01474
• https://arxiv.org/abs/1506.02142
• https://arxiv.org/pdf/1703.04977
• https://romainstrock.com/blog/modeling-uncertainty-with-pytorch.html