SkyBlink

Methods

A blink comparator is only as good as its discipline. These are the exact decisions SkyBlink makes between the archive and your screen — including the two mistakes we caught and fixed during the build.

Validation, with numbers

A blink tool is only as good as its measured recovery — so we measured it. We injected 540 synthetic movers (Gaussian, σ ≈ 1.1 px) into real Barnard's Star and R Draconis epoch pairs and ran the product's own difference-dipole test (positive lobe > 6σ at the new position, negative lobe > 4σ at the old):

  1. 100% recovered at brightness ≥ 15σ, for shifts of 1, 2 and 4 pixels.
  2. At 8σ: 97–98% recovered for shifts ≥ 2 px, and 60% for a 1 px shift — a sub-PSF move partially cancels inside a difference image, which is exactly why the blink itself, not the difference alone, is the primary view.
  3. Control: the same detector on the six uninjected pairs finds 439 peaks above 6σ — genuine residuals from bright-star saturation and cross-detector distortion. That number is why the Two-Epoch Check exists: a difference peak is a candidate, never a discovery. (Reproduce it: scripts/validate.py in the source archive.)

One limit, stated plainly: the injected sources are idealised Gaussians placed on real frames and tested with our own detection rule. This measures the sensitivity of the difference test, not performance on real moving objects, whose shapes and trails are messier.

The pipeline

  1. Discover. IRSA's SIA service is queried for every SPHEREx observation intersecting the field, restricted to the current uniform releases: DR1 and QR3 (collections spherex_dr1, spherex_dr1_deep, spherex_qr3, spherex_qr3_deep). Images are DR1/QR3 only; the one QR2-era item anywhere in the project is IRSA's published 3I/ATLAS position table, used for the comet's coordinates in Case 1.
  2. Cluster & choose. Observations are clustered into epochs (gaps over 20 days). The archive returns frames whose footprints merely intersect the search circle, so for each epoch we keep the frame whose centre is closest to the target — across all six detectors, because the detectors image offset strips of sky and forcing one detector loses whole epochs.
  3. Cut out. Each chosen exposure is fetched as an archive cutout centred on the target, following the IRSA tutorial notebooks' recipe.
  4. Align. Cutouts are reprojected (Astropy + reproject) onto one shared north-up tangent-plane grid at 3.1″/pixel for the target.
  5. One stretch. A single percentile + asinh brightness stretch is computed across all epochs of a target and applied identically — so brightness differences between frames are the sky changing, not the scaling.
  6. Date honestly. Epoch dates come from the archive query's observation times, never from file-processing timestamps in FITS headers.

Two mistakes we caught — and how

  1. The upside-down sky. Reprojection arrays run from low to high declination, so our first renders were south-up. We caught it by checking a rendered frame against the Gaia-propagated position of Barnard's Star: the star sat mirrored. Frames are now flipped to true north-up, verified both numerically (centroid within ~1 px of the WCS prediction) and visually (the difference dipole points along the known proper-motion direction).
  2. The 4.4-arcminute stale centre. Barnard's Star moves so fast that its catalogue J2000 position is ~4.4′ away from where it is now — outside a 5′ frame's centre. SkyBlink centres it on its Gaia DR3 position propagated to the survey era; the same propagation powers the Gaia overlay, so reference stars land on their epoch-correct positions in every frame.

The Two-Epoch Check

A candidate mover is promoted only if it (1) appears in both epochs, shifted consistently; (2) is not a known Gaia star at the new position, with Gaia positions proper-motion propagated to the epoch; and (3) survives the difference view as a clean dipole. Cosmic rays, ghosts and saturation artefacts fail at least one test by construction. Full case treatment on the Case Files page.

Relation to WiseView

The challenge points builders at WiseView, the established WISE blink tool, and SkyBlink deliberately follows its proven interaction grammar — blink interval, epoch stepping, difference imaging, percentile stretch — while differing where the data differs: SPHEREx instead of WISE, 102 bands across six detectors instead of two, DR1 + QR3 multi-epoch sets up to a year apart, and a Gaia overlay propagated per epoch. Where WiseView helped us, we say so.

Overlays, spectra and staying current

  1. Two overlays, two epochs. The Gaia overlay (rings) propagates every Gaia DR3 star to the frame's own date with its measured proper motion. The WISE overlay (diamonds) draws NASA AllWISE catalogue positions exactly as catalogued — the catalogue extract carries no proper motions, so its epoch is ~2010 and a fast mover's diamond can sit visibly off its SPHEREx position. That offset is real information, not an error: it is the same displacement WiseView hunters measure.
  2. Spectra are sampled, not interpolated. A dossier spectrum is built exposure by exposure: every archive visit that covered the position contributes one point, at the wavelength that visit's spectral WCS assigns to the target's pixel. SPHEREx has 102 channels; a single position has only been sampled in the channels its visits happened to place on it, so the chart shows the sampled channels and states how many — the count grows as the survey revisits each field. The live backend runs the same measurement on demand for searched coordinates.
  3. Links carry the view. Every explorer view has a shareable URL: a pre-loaded target by name, or live coordinates with their field of view (?ra&dec&fov). The FoV control offers 3 or 5 arcmin. For a pre-loaded target it switches instantly on the client: the 3-arcmin view is the centre crop of the pre-rendered 5-arcmin frame, same pixels and same scale. For live coordinates, the backend renders at the chosen field, and the search stops with a clear message if the archive does not answer within about 20 seconds.
  4. Staying current. A scheduled job re-queries the archive for all five targets every week. When a new visit epoch appears, that target's frames, quality figures, overlay lists and spectrum are rebuilt and the site is re-published; when the archive has nothing new, the check changes nothing. Live coordinate searches always render from the archive as it stands at search time.

Stated limits

  1. WISE 0855 is intrinsically near-invisible below ~3 µm; its short-band frames are shown as honest non-detections, not enhanced into a “signal”.
  2. Barnard's Star saturates its core in SPHEREx exposures; centroiding uses the unsaturated wings, which is why Case 3 quotes arcsecond-level tolerances.
  3. Band fluxes on the dossier charts are aperture sums on rendered frames — indicative. The Barnard's Star spectrum on its dossier page is the calibrated route: per-exposure wavelengths from each frame's spectral WCS with aperture photometry on the Level-2 data (74 channels, 0.80–4.89 µm).
  4. The pre-loaded site covers the five official targets plus the 3I/ATLAS case; arbitrary-coordinate rendering runs on the same pipeline in the project repository's API service.
  5. Colour vs motion (band mixing). Epochs are imaged by whichever detector's strip covers the target, and the six detectors are six different wavelength bands — so a source can brighten or fade between epochs purely because the band changed, with no motion at all. That is why SkyBlink's motion evidence is never brightness alone: a real mover shifts position — a centroid offset against Gaia, and in Difference mode a dipole (dark at the old position, bright at the new). A pure band change stays in one place and produces a single-signed blob, which the Two-Epoch Check rejects. Every frame's band is labelled in the viewer, the “Same band only” toggle isolates same-detector epochs where the archive provides them, and Case 3 publishes the measured centroid scatter in full (1.3–7.4″).
  6. Zodiacal light and flagged pixels. Frames show the archive's IMAGE plane as delivered: the modelled zodiacal-light plane (a separate FITS extension) is not subtracted, and flagged pixels are not masked out, so a slow background gradient or an occasional hot pixel can appear in a difference view without anything in the sky having changed. The frame-info “Clean pixels” figure reports the share of cutout pixels whose FLAGS value is 0 (no flag bits) or the SOURCE bit alone; bit meanings are the archive's own header definitions, and the OVERFLOW bit does not count as clean.
  7. Epoch counts are the archive's, not a choice. The 3–5 epochs per target are every distinct visit epoch the released DR1 + QR3 data contains for those fields — SPHEREx launched in 2025 and re-surveys the sky roughly every six months. The pipeline is epoch-agnostic: new releases drop in as new epochs without code changes.