ThoR: A Motion-Dependent Physics-Informed Deep Learning Framework with Constraint-Centric Theory of Functional...

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence
• Computational Physics

Background:

• Accurate precipitation nowcasting is vital for extreme weather event mitigation, especially with climate change.
• Traditional forecasting methods like numerical weather prediction and radar extrapolation have limitations in short-term, high-resolution forecasting.
• Deep learning models for nowcasting often produce blurry results and lack physical consistency.

Purpose of the Study:

• To introduce ThoR, a novel deep learning framework for rainfall nowcasting.
• To improve the accuracy and physical consistency of precipitation nowcasting using deep learning.
• To address limitations of existing deep learning approaches in nowcasting, such as blurry predictions.

Main Methods:

• Developed ThoR, a Motion-Dependent Physics-Informed Deep Learning Framework with Constraint-Centric Theory of Functional Connections (TFC).
• Integrated attention-centric spatio-temporal modeling with physical constraints from partial differential equations (PDEs).
• Employed a cascaded-branch architecture with an attention-driven generator and an unsupervised motion field extraction module, embedding the advection-diffusion equation into the optimization objective via TFC.

Main Results:

• ThoR demonstrated consistent outperformance over existing methods in deterministic metrics on real-world radar datasets.
• The framework showed particular effectiveness at longer lead times and during extreme weather events.
• Physics-informed deep learning, through ThoR, shows significant potential for operational precipitation nowcasting.

Conclusions:

• ThoR offers a significant advancement in rainfall nowcasting by successfully integrating deep learning with physical constraints.
• The physics-informed approach enhances prediction accuracy and consistency, outperforming traditional and current deep learning methods.
• This framework highlights the potential of physics-informed deep learning for operationalizing accurate and reliable precipitation nowcasting, especially for extreme events.