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.

All sections

Edition