SPHEREx triage: find, explore and identify objects in NASA SPHEREx data

An open classifier for every object that NASA's SPHEREx telescope detects in a patch of sky. For each object it gives what it most looks like, how sure it is, and how much it fits nothing known, so unusual objects can be checked first. Built by team Aphelion for NASA Space Apps 2026.

What goes in, what comes out

  • Input: SPHEREx only. Brightness in the D4/D5/D6 infrared bands across every visit, its colours, how much it changes between visits, and how it moves (17 numbers, listed in features.json).
  • Output: one of star, ir-only (seen by WISE but not Gaia: galaxies, dusty or cool objects), asteroid, unmatched (a one-off detection, an artefact, or something new), plus an oddness score from an Isolation Forest.

How it was trained

  • Data: 9427 objects from 43 sky zones (10β€² each), measured by our open engine straight from NASA's public SPHEREx photos on AWS. Dataset: YasenDzhenkov/spherex-triage-data.
  • Labels: catalogue matches only (Gaia DR3 within 2β€³ = star; AllWISE within 3β€³ and no Gaia = ir-only; SkyBoT = asteroid; nothing = unmatched). Catalogues are used as the answer key, never as inputs, so the model cannot copy the answer.
  • Model: scikit-learn HistGradientBoostingClassifier (balanced class weights) + IsolationForest fitted on catalogued objects.

How it was tested

Whole zones held out (GroupKFold by zone (held-out sky)): the model is always scored on sky it never saw.

value
accuracy on held-out sky 93.1%
balanced accuracy (mean of per-class recall) 68.2%
baseline: always guess unmatched 65.7%
oddness: share of unmatched among the 5 % oddest 100.0% (base rate 65.7%)
the app's 210 objects around 6 named targets (never trained on) 92.4%
class precision recall objects
star 0.86 0.96 2916
ir-only 0.23 0.13 316
unmatched 0.99 0.96 6195
truth \ predicted star ir-only unmatched
star 2786 95 35
ir-only 236 41 39
unmatched 206 43 5946

Limits (read before using)

  • unmatched objects are mostly single-visit detections, which are easy to separate; the hard, useful split is among objects seen on several visits.
  • ir-only has few examples, and is the weakest class.
  • No brown-dwarf, variable-star or young-star classes yet: their labels need SIMBAD, which was down while we built this.
  • Photos up to Feb 2026 only (our own S3 header index; NASA's IRSA search was returning errors).
  • A label is a catalogue match, not ground truth: faint stars missing from Gaia are labelled ir-only or unmatched.

Use

python predict.py report.json      # a zone report from the app's engine (POST /scan β†’ report)

Credits

SPHEREx data: NASA/JPL-Caltech/IPAC (IRSA). Gaia DR3: ESA/Gaia/DPAC. AllWISE and 2MASS: NASA/IPAC. Catalogue access via the VizieR service (CDS, Strasbourg; Harvard mirror). Asteroids: SkyBoT (IMCCE).

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train YasenDzhenkov/spherex-triage