V12 Ensemble Discovery Scan: Unseen Euclid Q1 Strong Lens Candidates

Fuente: Zenodo
Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: Bhabhrawala, Aliasger
Format: Recurso digital
Sprache:Englisch
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866902142896308224
author Bhabhrawala, Aliasger
author_facet Bhabhrawala, Aliasger
contents <p>Inference results from the V12 ensemble model (V7×0.3 + V10×0.2 + V11×0.5, <br>weighted logit averaging, AUROC=0.8871 on sacred test set) applied to 1,415 <br>unseen Euclid Q1 ERO images — all files outside the locked training, <br>validation, and test split. Contamination verified at zero overlap via <br>skycoord crossmatch <2 arcsec against all 1,000 sacred split files.</p> <p>Scanned data: 247 Grade B probable-lens candidates + 1,168 Grade C holdout <br>files from the ESA SLDE Q1 catalog (VIS band, 300×300px FITS cutouts, <br>0.1 arcsec/px). Inference used 4-angle TTA (0°/90°/180°/270° rotations, <br>averaged softmax). Classification threshold: P≥0.70.</p> <p>Key results:<br>- 33/86 pure Grade B candidates (no NEG prefix) flagged as lens at P≥0.70 (38.4%)<br>- 228 files with all three component models independently scoring P>0.90<br>- 625 total files flagged; 790 not flagged<br>- Grade C false positive rate reflects documented out-of-distribution <br>  generalization limit consistent with ESA arXiv:2512.05899</p> <p>Package contents:<br>- v12_unseen_q1_scan_full.csv: All 1,415 files with per-model TTA and <br>  no-TTA scores (p_v7, p_v10, p_v11, p_ensemble), RA, Dec, grade, <br>  classification<br>- v12_unseen_lens_candidates.csv: 625 flagged candidates only (P≥0.70)<br>- v12_unseen_q1_scan_full.xlsx: Same as CSV in Excel format<br>- v12_discovery_gallery_final.pdf: 32-page visual gallery of top-ranked <br>  candidates, 4 per page, sorted by ensemble TTA score descending, <br>  annotated with individual model scores</p> <p>These are lens candidates for follow-up visual inspection, not confirmed <br>lens discoveries. Real lens prevalence in photometric surveys is ~0.1–1%.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_20037150
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle V12 Ensemble Discovery Scan: Unseen Euclid Q1 Strong Lens Candidates
Bhabhrawala, Aliasger
gravitational lensing
Euclid
Q1
strong lensing
machine learning
lens detection
galaxy survey
<p>Inference results from the V12 ensemble model (V7×0.3 + V10×0.2 + V11×0.5, <br>weighted logit averaging, AUROC=0.8871 on sacred test set) applied to 1,415 <br>unseen Euclid Q1 ERO images — all files outside the locked training, <br>validation, and test split. Contamination verified at zero overlap via <br>skycoord crossmatch <2 arcsec against all 1,000 sacred split files.</p> <p>Scanned data: 247 Grade B probable-lens candidates + 1,168 Grade C holdout <br>files from the ESA SLDE Q1 catalog (VIS band, 300×300px FITS cutouts, <br>0.1 arcsec/px). Inference used 4-angle TTA (0°/90°/180°/270° rotations, <br>averaged softmax). Classification threshold: P≥0.70.</p> <p>Key results:<br>- 33/86 pure Grade B candidates (no NEG prefix) flagged as lens at P≥0.70 (38.4%)<br>- 228 files with all three component models independently scoring P>0.90<br>- 625 total files flagged; 790 not flagged<br>- Grade C false positive rate reflects documented out-of-distribution <br>  generalization limit consistent with ESA arXiv:2512.05899</p> <p>Package contents:<br>- v12_unseen_q1_scan_full.csv: All 1,415 files with per-model TTA and <br>  no-TTA scores (p_v7, p_v10, p_v11, p_ensemble), RA, Dec, grade, <br>  classification<br>- v12_unseen_lens_candidates.csv: 625 flagged candidates only (P≥0.70)<br>- v12_unseen_q1_scan_full.xlsx: Same as CSV in Excel format<br>- v12_discovery_gallery_final.pdf: 32-page visual gallery of top-ranked <br>  candidates, 4 per page, sorted by ensemble TTA score descending, <br>  annotated with individual model scores</p> <p>These are lens candidates for follow-up visual inspection, not confirmed <br>lens discoveries. Real lens prevalence in photometric surveys is ~0.1–1%.</p>
title V12 Ensemble Discovery Scan: Unseen Euclid Q1 Strong Lens Candidates
topic gravitational lensing
Euclid
Q1
strong lensing
machine learning
lens detection
galaxy survey
url https://doi.org/10.5281/zenodo.20037150