Speakers
Description
The NRL SO82B spectrograph on board of Skylab captured more than 6,000 far-ultraviolet photographic exposures of the Sun between 1973 and 1974. UV flare spectra are still rare with only a few reported during the mission. This work presents a supervised binary image classifier based on a ResNet-18 convolutional neural network (CNN) to identify uncatalogued flares in the SO82B data. The CNN was trained with transfer learning from ImageNet weights to distinguish flare from non-flare SO82B plate images. Training on 1,281 labelled frames (321 flares and 960 non-flares) and evaluating on a held-out test set of 641 independent examples (161 flares and 480 non-flares), the model achieved a ROC-AUC of 0.88 and an area under the precision-recall curve (AUPRC) of 0.75, confirming adequate discrimination capability. At the standard threshold of p > 0.50, the model flagged 400 candidate frames, achieving a held-out flare recall of 72.7% and a false-positive rate of 9.8% on held-out non-flares. Further analysis revealed a strip-level systematic: after plate-level suppression the 253 candidates (p > 0.70) reduce to 136 unique plates, and after strip-level suppression to all 3 film strips. This collapse suggests the model partially responds to strip-level properties, such as emulsion batch or scanner calibration artefacts, rather than exclusively to frame-level flare morphology, which it was expected. To isolate genuine events from this systematic baseline, a within-strip outlier analysis was applied, retaining only plates whose maximum frame score exceeds the strip median by more than 2σ. This conservative approach has led to identification of 68 plates of statistically significant, of which 20, it was from strip 1B (May 30 – Jun 17, 1973) and 48 from strip 2B (Aug 09 – Sep 12, 1973). Cross-referencing the 68 outlier candidates against the NOAA/NGDC flare catalogue for 1973 reveals that 5 candidate plates have capture times that fall within the duration of a catalogued optical flare event, and with 10 more within 30min of a catalogued flare; the strongest temporal match is on 1973 June 15, where four consecutive SO82B plates were acquired during a class 1F optical flare at active region AR 2382. Gradient-weighted class activation maps confirm that the network attends to the chromospheric emission core of the UV spectrum rather than plate artefacts or edges. The resulting candidate list provides a prioritised target list for spectral follow-up and for extending the SO82B flare database.