🌍 Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting

This is the official PyTorch Lightning implementation of the paper:
Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting, accepted at NeurIPS 2025.
[📄 arXiv]

MIGN is implemented using DGL (Deep Graph Library).

# Create and activate a clean conda environment
conda create -n dgl python=3.10
conda activate dgl

# Install PyTorch (CUDA 12.1)
pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu121

# Install DGL (CUDA 12.1 compatible)
pip install dgl -f https://data.dgl.ai/wheels/torch-2.1/cu121/repo.html

# Install additional dependencies
pip install numpy==1.26.0
pip install healpy
pip install lightning
pip install -U 'wandb>=0.12.10'
pip install jupyter

We also provide baseline implementations built upon two widely used spatiotemporal graph learning frameworks:

• PyTorch Geometric Temporal
• Torch Spatiotemporal (TSL)

To install the dependencies required for running these baselines:

# Install PyTorch Geometric and related packages (compatible with CUDA 12.1)
pip install torch_geometric
pip install torch_scatter torch_sparse torch_cluster torch_spline_conv \
    -f https://data.pyg.org/whl/torch-2.1.0+cu121.html
# Install PyTorch Geometric Temporal (local version)
cd baseline/spatial_temporal/pytorch_geometric_temporal/
pip install -e .
# Install Torch Spatiotemporal (TSL)
pip install torch-spatiotemporal

We provide processed datasets for both baseline models and MIGN at

• spherical level 3
• HEALPix level 3

These datasets are designed for one-step input → one-step output prediction tasks.

You can directly download the processed data from our Hugging Face repository:
👉 compasszzn/MIGN

Note:
You can safely skip the Data Preparation sections — the provided datasets are already preprocessed and ready to use.

The metadata are obtained from the Global Summary of the Day (GSOD) dataset provided by NOAA:
🔗 https://www.ncei.noaa.gov/data/global-summary-of-the-day/archive/

Please manually download the data for the years 2017–2024 and place them in the following directory: /realtime/Dataset/

Navigate to the processing directory:

cd data/process_data
# Filter Features; Note: Modify base_dir and output_dir in the script before running.
python data/process_data/step0_filter_feature.py
# Note: Modify folder_path in the script before running.
python data/process_data/step1_get_union_station.py
# Compute Climatology Statistics; Note: Modify base_path and output_path in the script before running.
python data/process_data/step2_climatology.py

# One-step Input → One-step Output
python data/process_data/step3_generate_graph_dgl_step.py --input_day 1 --output_day 1
# Multi-step Input → Multi-step Output
python data/process_data/step3_generate_graph_dgl_multi_step.py --input_day 3 --output_day 4

# One-step Input → One-step Output
python data/process_data/step4_generate_graph_pyg_step.py --input_day 1 --output_day 1
# Multi-step Input → Multi-step Outputt
python data/process_data/step4_generate_graph_pyg_multi_step.py --input_day 3 --output_day 4

We compute the spherical harmonics following the implementation from
👉 MarcCoru/locationencoder

For detailed examples, refer to the notebook:
baseline/locationencoder/test.ipynb

In the default setting, we provide level-3 spherical harmonics embeddings for both station nodes and HEALPix nodes, saved in the following files:

• MAX_embeddings_3.pt
• WDSP_embeddings_3.pt
• ...
• 3_healpix_embeddings_level_3.pt

Place all embedding files in the folder before running the models.

python /home/zinanzheng/project/MIGN/main.py --model MIGN --sh_before --sh_after #with location embedding
python /home/zinanzheng/project/MIGN/main.py --model MIGN #without location embedding
python /home/zinanzheng/project/MIGN/main.py --model tasamp
python /home/zinanzheng/project/MIGN/main.py --model STGCN