06Deep Field guide

Audience: everyone curious about exoplanets, from total beginners to astronomers. Route: /deep-field.

Deep Field is named after the Hubble Deep Field — the famous image where astronomers pointed the telescope at a patch of "empty" sky and found thousands of galaxies. The lesson: point at nothing, find everything. This room points the same loop that forks missions at the night sky.

The big idea (30 seconds)#

When a planet passes in front of its star, it blocks a tiny bit of light and the star dims for a few hours — then brightens again, and repeats every orbit. NASA's TESS and Kepler space telescopes recorded the brightness of millions of stars for years. Hidden in that data are the little repeating dips that betray planets.

Deep Field lets you analyze that real, public NASA data in your browser: find a dip, measure it, check it isn't a fake, and publish a forkable candidate document. On day one you can rediscover π Mensae c — a genuine planet — in about 90 seconds.

If "transit" and "light curve" are new, read Space without the math §9 first. (Layman version: a moth flickering across a distant streetlight on a schedule.)

A note on the units#

This room — and only this room — speaks the astronomer's units instead of pure SI: time in days (BTJD — a TESS-specific day count**)**, brightness dips in parts-per-million (ppm), durations in hours. That's deliberate: it's the language the global exoplanet community uses, and it never touches the orbital physics engine. Don't expect metres and seconds here.

The two halves of the room#

Deep Field has two views you can switch between:

1. THE FIELD — a 3D star map of real candidate host stars you fly through.
2. The Analysis Desk — the six-station pipeline where you actually find and vet a transit.

THE FIELD — the 3D star map#

A WebGL point-cloud of thousands of real TESS Objects of Interest (TOIs) — stars flagged as possibly hosting planets — placed at their true 3D positions (computed from sky coordinates and distance). The committed catalog snapshot holds about 7,600 stars / 7,900 TOIs; the frontier view renders the ~5,000 still- open candidates. (The snapshot's date is stamped in the UI; it refreshes by re-running one query against the NASA Exoplanet Archive.)

What you can do:

• Fly through it. Drag, zoom, and select a star — the camera does a cinematic fly-to, easing through the galaxy to center on it with a halo.
• Filter by disposition — show only confirmed planets, candidates, false positives, etc.
• Color by status or by score — paint the cloud by official disposition or by the ML model's score (see below).
• Selected-star rail — pick a star and a side panel shows its host facts, its candidate signals, real telescope observations of it, and an Analyze button that hands the star off to the Analysis Desk with its data pre-filled.
• Neighbour web — toggle a faint lattice linking each star to its nearest neighbours in 3D space (a sense of the local stellar neighborhood).
• ML vetting desk — train a machine-learning model (scikit-learn) on the known confirmed-vs-false-positive examples, then let it score the open frontier. After training, a worklist shows the top-ranked candidates; click one to fly there and open it. Honest label: the score is a vetting prior (a hint about where to look), never a discovery; the official human dispositions remain ground truth.

Observations & cross-match (on a selected star)#

• Observations — a live search of NASA's MAST archive for every telescope that has looked at this star (TESS, Hubble, Swift, GALEX…), grouped by mission with exposure hours. (A never-seen star may take ~60–75 s to fetch the first time, then it's cached forever.)
• Cross-match — a live query against the NASA Exoplanet Archive that tells you, honestly, whether a candidate period matches a known planet, a known system, an existing TOI, or nothing known — and links you to ExoFOP (the real follow-up hub) to check for yourself. It will say "a rediscovery" when you've re-found a known planet — never "a discovery."

The Analysis Desk — the six-station pipeline#

This is where you turn a star's raw brightness data into a vetted candidate. Each station is a card; the whole sequence is recorded into the document so anyone can reproduce (or challenge) it.

1. TARGET — pick a star. Five curated chips come with textbook ephemerides (known period/epoch) to learn on; or arrive here via Analyze from THE FIELD, or by typing a target name / TIC id.
2. PHOTOMETRY — load and clean the light curve: download from MAST, stitch the sectors together, flatten out slow trends, normalize. The cleaned brightness- vs-time series renders in a fast canvas plot (tens of thousands of points at 60 fps).
3. PERIODOGRAM — search for the repeating dip. A BLS (Box Least Squares) search scans candidate periods and scores how planet-like each is; the peaks are clickable. (Under the hood, careful statistics keep it honest — e.g. a robust outlier clip stops the telescope's own 13.7-day orbit from masquerading as a planet.)
4. PHASE FOLD — fold the data on the best period so every orbit stacks on top of the last, and the transit dip appears cleanly. Zoom in; edit the ephemeris (period / epoch / duration) by hand and watch the fold sharpen or smear.
5. VETTING — run the standard sanity checks, shown as pass / warn / fail cards: is the signal strong enough (SNR vs the community 7.1σ line)? Do odd and even transits match? Is there a secondary eclipse (a sign it's two stars, not a planet)? Could it be a half-period alias? Enough transits? Sensible duration? The verdict is only ever candidate-grade, needs-scrutiny, or false-positive-risk — never "confirmed."
6. PUBLISH — auto-generate a dv-brief and post your candidate to The Loop, where others can fork it, change your detrending window, and either strengthen or kill it.

URL-as-document: the page's permalink encodes the whole analysis (?target=…&period=…&t0=…&dur=…). Send the link and the recipient replays your exact analysis. That link is the forkable candidate document.

The honesty rule (this matters)#

Finding a transit dip makes a candidate, not a confirmed planet. Confirmation needs follow-up the platform can't do — independent telescopes, ruling out look-alikes, sometimes measuring the planet's gravitational tug. Deep Field is scrupulous about this: it says "community-validated candidate," never "confirmed," and it lays out the real path a serious candidate would travel:

Your analysis → ExoFOP cross-match → cTOI submission → TOI team review
   → TFOP follow-up (SG1–4) → refereed paper → NASA Exoplanet Archive

The room even includes the practical rules for submitting a community TESS Object of Interest (cTOI) for real — what parameters you need, what not to claim, and why you don't call it "planet b" before peer review. This is the same fidelity-ladder ethics as the orbital engine: tell the user exactly how much to trust the result.

The day-1 demo (try this)#

1. Open /deep-field, go to the Analysis Desk, pick π Mensae (a curated chip).
2. Run photometry → periodogram. A strong peak appears near 6.27 days.
3. Phase fold on it — the transit dip snaps into view.
4. Vet — it passes all checks (high SNR, 8 transits, no secondary).
5. Cross-match — it flags as π Mensae c, a known planet: a rediscovery. 🎉

You just rediscovered a real exoplanet from real NASA data, end to end, in a browser tab — and the permalink lets anyone replay it.

Where this is going#

Deep Field already does quick-look discovery and publishing. The piece that lands with a later phase (S5) is persisted candidate documents with community Go/No-Go validation polls — turning "I found a dip" into "the community vetted this candidate," with the same review primitive that approves missions. The architecture is built for it; the voting loop is the remaining brick.

Back to the rooms · On to Fidelity & honesty.