The core code for understanding the logic and implementing PhyDL-NWP (https://arxiv.org/pdf/2505.14555). Brief video intro can be found in https://www.youtube.com/watch?v=ifV4jPLvlLo.

Get Started

1. Install dependencies

pip install -r requirements.txt

2. Run code to obtain the governing equation with latent force

python main.py

3. Run code to use the learned physics to inform forecasting models

python train_forecasting_models.py

Dataset Description

In this repository, we release the Ningxia dataset as a regional data example to run the code (For other datasets that may require access permission, please refer to the footnote links in our paper). It contains two files:

• NWP_train: Weather forecasts generated by a Numerical Weather Prediction (NWP) model.
• real_train: Corresponding ground-truth weather observations.

Input Features

The input consists of 11 features, including:

• 8 meteorological variables:

  • 100 metre U wind component
  • 100 metre V wind component
  • 10 metre U wind component
  • 10 metre V wind component
  • 2 metre temperature
  • Mean sea level pressure
  • Surface pressure
  • Total precipitation

• 3 geographic variables:

  • latitude
  • longitude
  • dem (digital elevation model, representing elevation information)

Output Features

The output consists of 8 meteorological variables (same as the 8 weather-related input features above).

Poster Presentation

Paper Cover