OCELOT

Author: Azadeh Gholoubi

OCELOT is an end-to-end graph neural network workflow for direct observation prediction in weather applications. The repository includes data preparation, GNN training/inference, mesh-grid prediction, and evaluation utilities for OCELOT, persistence, truth, and GFS comparisons.

The ocelot-v1.0 tag is the reference code snapshot for the OCELOT v1 AIES submission.

Repository Layout

• data_prep/: readers, mappings, and Zarr/Parquet data preparation tools.
• gnn_model/: PyTorch Lightning/PyTorch Geometric model, training scripts, prediction workflows, and evaluation scripts.
• gnn_model/evaluation/: OCELOT-vs-truth, persistence, and GFS comparison documentation and plotting utilities.

v1.0 Highlights

• Heterogeneous graph neural network for multi-instrument observation prediction.
• Obs-space and mesh-space prediction outputs.
• Persistence-baseline columns (persist_*) for obs-space and mesh-grid CSVs.
• Mesh-grid GFS forecast and optional GFS analysis interpolation (gfs_*, anl_*).
• 3-panel and 6-panel mesh plotting for OCELOT/GFS workflows and GFS analysis diagnostics.
• 2025 and seasonal Slurm workflows for prediction/evaluation campaigns.

Quick Start

For model details and training commands, see:

• gnn_model/README.md

For evaluation workflows, see:

• gnn_model/evaluation/README_EVALUATION.md