ExoDiscover

A classifier for NASA exoplanet transit signals, trained on Kepler and tested on TESS.

This is a rebuilt version of a NASA Space Apps 2025 hackathon submission (A World Away: Hunting for Exoplanets with AI). The original reported a strong accuracy that was measuring the wrong thing: the Kepler table ships the vetting pipeline's own verdict as a column. This version quarantines those columns, holds out whole host stars rather than rows, and reports what the model does on a different mission's data. See docs/LEAKAGE.md for the investigation.

77% accuracy zero-shot on TESS      (majority-class baseline 51%)
ROC-AUC 0.838 · Brier 0.177 · n = 2,562 resolved TESS objects

The app

16,932 catalogued objects at their real right ascension, declination and distance, with Earth at the origin. Drag to look, scroll to travel, click a planet.

Both missions are shown. Kepler observed a single 22°×16° window, so its 9,444 objects form a dense beam in one direction. TESS surveyed the whole sky, so its 7,488 objects lie in every direction, the nearest 21 light years away.

Clicking a planet shows the model's prediction, the archive's disposition, and the SHAP terms behind the score. Both missions are scored by the same 11-feature model, the only features the two catalogues share, so the probabilities are comparable — and it is the same model the 77% above describes, loaded from models/production/transfer.joblib rather than refitted at render time.

Kepler objects with a resolved disposition are in that model's training data, so those are scored out of fold: each gets its probability from a star-grouped fold that never saw it. Nothing on screen is a model's opinion of a row it already learned.

Run it

Everything runs on localhost. The trained model and its metrics are committed, so no catalogue download or training is required.

Prerequisites: Python 3.11 or 3.12, Node 20+.

make install     # pip install -e ".[dev,api]", then npm install in web/

Then, in two terminals:

make serve       # Flask on :8000, interactive docs at /docs
make web         # React UI on :5173

Open http://localhost:5173.

Without make (Windows, or no GNU make installed):

pip install -e ".[dev,api]"
cd web && npm install && cd ..
flask --app api.wsgi run --port 8000 --reload  # terminal 1
cd web && npm run dev                          # terminal 2

Or run the whole stack with docker compose up --build.

Regenerating the artifacts

Only needed to reproduce them rather than use the committed ones:

exo ingest              # NASA archive to data/raw/
exo train --trials 40   # ~25 min, CPU only (exo train --fast, ~6 min)
exo skymap              # joins positions and distances for the scene

Checks

make test        # pytest (offline, against committed fixtures) + vitest
make lint        # ruff, mypy, eslint

Layout

ml/exodiscover/    ingest, leakage firewall, physics features, training, evaluation
api/               Flask: typed prediction, batch CSV, SHAP, metrics, sky map
web/               React and WebGL, reading from the API
docs/              LEAKAGE.md, MODEL_CARD.md, metrics/
models/production/ model.joblib (Kepler triage) and transfer.joblib (cross-mission)
tests/             146 Python tests and 36 web tests, offline against fixtures

docs/LEAKAGE.md covers the leakage investigation. docs/MODEL_CARD.md records intended use and limitations.

Python 3.11, scikit-learn, CatBoost/XGBoost/LightGBM, Optuna, SHAP, Flask, Pydantic, gunicorn, React, TypeScript, Vite, Tailwind, pytest, vitest, ruff, mypy, GitHub Actions. CPU only.