Neural network trained on survey data retuned to spot transients for a liquid mirror telescope

The work

Technical · from the record

The authors trained convolutional neural networks to separate real transient detections from image-subtraction artefacts, and to sub-classify real candidates by host morphology, for the small field-of-view 4-m International Liquid Mirror Telescope. Source models were trained on a large labelled set of ZTF alert stamp triplets obtained through the ALeRCE broker, then retrained by transfer learning on a much smaller ILMT dataset. On a test set built from a different ILMT observation cycle, the transfer-learned real/bogus model reached 97.3% accuracy, against 90.5% for a model trained on the small ILMT dataset without transfer learning and 91.5% for the ZTF source model applied directly; the 3-class and 4-class alert classifiers reached 92.9% and 85.6%. An unpaired t-test over 20 transfer-learned and 20 baseline models gave t-statistics of 8.90, 16.75 and 4.82 for the three model types, with p-values below 0.05, and the models were run inside the PyLMT pipeline on nearly 300 full-frame ILMT images.

How AI was used

Source CNN classifiers were trained from scratch on 63 x 63 x 3 science/reference/difference stamp triplets of ZTF alerts selected using ALeRCE stamp- and lightcurve-classifier labels and scores, with artefact samples drawn from ALeRCE bogus alerts, manually differenced ZTF images and ZTF difference images. The architecture used three convolutional blocks with 5x5, 3x3 and 2x2 filters and 2x2 max pooling, ReLU intermediate activations, dropout, L2 regularisation and batch normalisation, a sigmoid output for real/bogus and softmax outputs of dimension 3 and 4 for the alert classifiers, trained with adam and early stopping on maximum validation accuracy. ILMT target stamps were produced by ILMTDiff image subtraction, downsampled with scipy spline interpolation from 0.323 to 1 arcsec/pixel to match ZTF, rotation-augmented for under-represented extended-host candidates, manually vetted, and split 85:15 into training and validation sets afresh for each model version. Transfer learning then adapted each source model to the ILMT data: fine-tuning at reduced learning rate for the real/bogus and 4-class models, and freezing the convolutional base while retraining only the fully connected layers for the 3-class model, with class weighting for the alert classifiers. The real/bogus decision threshold was set by maximising F1 on the validation set, and alert classes by argmax of the softmax output. Models were evaluated on a test set from an earlier ILMT observation cycle, supplemented with injected simulated Gaussian sources and injected ILMT-PSF sources near galaxies, and compared with models trained without transfer learning using an unpaired t-test over 20 models per group. The trained classifiers were integrated into the PyLMT pipeline, requiring positive classification from two real/bogus models, and applied to full-frame ILMT images with subsequent visual vetting and SkyBoT cross-matching.