A new tool for regional meteorological downscaling
The Spherical Physics-informed Neural Operator helps with high-resolution regional climate forecasting and disaster early warning.
The Spherical Physics-informed Neural Operator provides a solution for global-to-regional meteorological downscaling. This framework aims to improve high-resolution regional climate forecasting and disaster early warning.
The method uses a Spherical Laplacian Decomposition. This partitions atmospheric fields into hierarchical frequency components. It maintains exact point-wise correspondence across scales.
A localized spherical integral operator evaluates these representations at arbitrary locations. This operator approximates continuous kernel transforms via geometry-aware attention.
The framework embeds differentiable constraints into the learning process. This enforces dynamical consistency.
Existing frameworks often struggle with spherical manifold geometric distortions. They also face inherent atmospheric multi-scale coupling mismatches and a lack of explicit atmospheric laws.
Key facts
- The Spherical Physics-informed Neural Operator attains superior accuracy and zero-shot generalization.
- The framework works across various meteorological variables and unseen queries.
- The method uses a Spherical Laplacian Decomposition to partition atmospheric fields.