Explainability, uncertainty & physics
Predictability Limit of the 2021 Pacific Northwest Heatwave from Deep-Learning Sensitivity Analysis
In one sentence. Backpropagation through GraphCast optimizes initial conditions, cutting 10-day Pacific Northwest heatwave errors over 90% and extending predictability to about 23 days.
Abstract
The traditional method for estimating weather forecast sensitivity to initial conditions uses adjoint models, which are limited to short lead times due to linearization around a control forecast. The advent of deep-learning frameworks enables a new approach using backpropagation and gradient descent to iteratively optimize initial conditions to minimize forecast errors. We apply this approach to forecasts of the June 2021 Pacific Northwest heatwave using the GraphCast model, yielding over 90% reduction in 10-day forecast errors over the Pacific Northwest. Similar improvements are found for Pangu-Weather model forecasts initialized with the GraphCast-derived optimal, suggesting that model error is not an important part of the initial perturbations. Eliminating small scales from the initial perturbations also yields similar forecast improvements. Extending the length of the optimization window, we find forecast improvement to about 23 days, suggesting atmospheric predictability at the upper end of recent estimates.
Labels
- Topic
- Explainability, uncertainty & physics
- Type
- Application study
- Method
- Graph neural network, Differentiable physical model
- Time scale
- 3–15 days
- Domain
- Regional
- Data
- ERA5 / reanalysis
- Status
- Research
- Identifiers
- arXiv:2406.05019 DOI:10.1029/2024gl110651 OpenAlex:W4403085937
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.