Hunt for Tropical Cyclones pt. 1

[mariusaurus]

Earth2Studio Pull Request

Description

• Adds a new recipes/tc_tracking recipe that generates ensemble forecasts (using FCN3 or AIFS-ENS) and tracks tropical cyclones within the predictions by integrating TempestExtremes as a downstream diagnostic tool.
• Implements an asynchronous CPU/GPU execution mode where TempestExtremes runs cyclone detection on CPU in parallel with GPU inference, resulting in virtually no computational overhead from the tracking process.
• Uses in-memory file handling (/dev/shm) to avoid writing large atmospheric field data to disk, which would otherwise slow down inference significantly and become prohibitive at scale.
• Includes a stability check mechanism for detecting and recovering from numerical instabilities in long-range FCN3 forecasts.
• Includes a comprehensive README with setup instructions (container and uv), configuration reference, and an example workflow. Also serves as a reference for integrating other custom downstream analysis tools into Earth-2 Studio prediction pipelines.
• Includes an end-to-end test (test/test_tc_hunt.sh) that runs a five-member ensemble forecast of Hurricane Helene with FCN3 and verifies that track files are produced correctly.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• Assess and address Greptile feedback (AI code review bot for guidance; use discretion, addressing all feedback is not required).

Dependencies

Licences for all Python dependencies declared in pyproject.toml.
Information sourced from PyPI on 2026-04-09.

Core Dependencies

Package	Licence	PyPI
earth2studio	Apache-2.0	link
huggingface-hub	Apache-2.0	link
hydra-core	MIT	link
importlib-metadata	Apache-2.0	link
more-itertools	MIT	link
omegaconf	BSD	link
ruamel-yaml	MIT	link
scipy	BSD-3-Clause	link
setuptools	MIT	link
tqdm	MPL-2.0 AND MIT	link
zarrdump	BSD-3-Clause	link

Optional – plot

Package	Licence	PyPI
cartopy	BSD-3-Clause	link
ipykernel	BSD-3-Clause	link
matplotlib	PSF-2.0 (matplotlib licence)	link
moviepy	MIT	link

Optional – dev

Package	Licence	PyPI
ninja	Apache-2.0 / BSD	link
pre-commit	MIT	link
pytest	MIT	link

[mariusaurus]

@greptileai

[mariusaurus]

@greptileai

[NickGeneva]

/blossom-ci

[NickGeneva]

LGTM for first PR.
My suggestion for next PR is to make AIFS / FCN potentially different install groups or something to lift the flash attention requirement 😆 . Something to consider maybe.

[NickGeneva]

/blossom-ci

[NickGeneva]

/blossom-ci

[NickGeneva]

/blossom-ci