V12 Ensemble Discovery Scan: Unseen Euclid Q1 Strong Lens Candidates

Inference results from the V12 ensemble model (V7×0.3 + V10×0.2 + V11×0.5, weighted logit averaging, AUROC=0.8871 on sacred test set) applied to 1,415 unseen Euclid Q1 ERO images — all files outside the locked training, validation, and test split. Contamination verified at zero overlap via skycoord crossmatch <2 arcsec against all 1,000 sacred split files. Scanned data: 247 Grade B probable-lens candidates + 1,168 Grade C holdout files from the ESA SLDE Q1 catalog (VIS band, 300×300px FITS cutouts, 0.1 arcsec/px). Inference used 4-angle TTA (0°/90°/180°/270° rotations, averaged softmax). Classification threshold: P≥0.70. Key results:- 33/86 pure Grade B candidates (no NEG prefix) flagged as lens at P≥0.70 (38.4%)- 228 files with all three component models independently scoring P>0.90- 625 total files flagged; 790 not flagged- Grade C false positive rate reflects documented out-of-distribution generalization limit consistent with ESA arXiv:2512.05899 Package contents:- v12_unseen_q1_scan_full.csv: All 1,415 files with per-model TTA and no-TTA scores (p_v7, p_v10, p_v11, p_ensemble), RA, Dec, grade, classification- v12_unseen_lens_candidates.csv: 625 flagged candidates only (P≥0.70)- v12_unseen_q1_scan_full.xlsx: Same as CSV in Excel format- v12_discovery_gallery_final.pdf: 32-page visual gallery of top-ranked candidates, 4 per page, sorted by ensemble TTA score descending, annotated with individual model scores These are lens candidates for follow-up visual inspection, not confirmed lens discoveries. Real lens prevalence in photometric surveys is ~0.1–1%.本数据集包含V12集成模型（V7×0.3 + V10×0.2 + V11×0.5，采用加权对数几率平均（weighted logit averaging），在sacred测试集上的受试者工作特征曲线下面积（Area Under the Receiver Operating Characteristic Curve, AUROC）为0.8871）对1415张未纳入锁定的训练、验证与测试划分集的欧几里得（Euclid）Q1 ERO图像的推理结果。通过与全部1000份sacred划分文件进行天球坐标（skycoord）交叉匹配（匹配半径<2角秒），验证了本数据集无重叠污染。 本次扫描所用数据来自欧洲空间局（European Space Agency, ESA）SLDE Q1星表中的247个B级疑似透镜候选体与1168个C级留存文件（VIS波段，300×300像素FITS（Flexible Image Transport System）截帧，空间分辨率0.1角秒/像素）。推理过程采用四角度测试时数据增强（Test-Time Augmentation, TTA）：对图像分别进行0°、90°、180°、270°旋转后取平均softmax得分。分类阈值设置为P≥0.70。 关键实验结果： - 86个无NEG前缀的纯B级候选体中，有33个在P≥0.70时被标记为透镜候选体，占比38.4% - 228份文件的三个独立组件模型均给出P>0.90的得分 - 总计625份文件被标记为候选体，790份未被标记 - C级样本的假阳性率符合已发表的分布外泛化极限，与ESA arXiv:2512.05899文献一致 数据集包含文件如下： - v12_unseen_q1_scan_full.csv：包含全部1415份文件的各模型TTA增强与非增强得分（p_v7、p_v10、p_v11、p_ensemble）、赤经（RA）、赤纬（Dec）、等级、分类结果 - v12_unseen_lens_candidates.csv：仅包含625份符合P≥0.70阈值的候选体文件 - v12_unseen_q1_scan_full.xlsx：与CSV格式一致的Excel版本文件 - v12_discovery_gallery_final.pdf：32页的候选体可视化图集，每页展示4张候选图像，按集成模型TTA得分降序排列，并标注了各独立模型的得分 本数据集仅为供后续目视检查的透镜候选体，并非已确认的透镜发现。测光巡天中真实透镜的占比约为0.1%~1%。