Neural network picks bright exploding stars from sky survey alerts and books telescope time

The work

Technical · from the record

BTSbot is a multi-modal convolutional neural network that assigns each Zwicky Transient Facility alert packet a score for whether its source is, or will become, a bright extragalactic transient, using three 63x63-pixel image cutouts and 25 extracted features. It was trained on 608,943 alerts from 19,558 catalogued ZTF sources and reached 94.9% accuracy and ROCAUC of 0.985 on a concealed test split, recovering 100% of the bright transients in that split, with 93.0% purity under the stricter of two follow-up policies and 84.6% under the looser one. Running in production inside ZTF's alert broker and marshal during October 2023, it saved 296 sources and sent automatic spectroscopic follow-up requests; 92.6% of those sources were confirmed as extragalactic transients. Median differences in saving and triggering time relative to human scanners on the test split were -0.0381 and -0.0147 days, and one selected source, SN 2023tyk, was detected, identified, spectroscopically classified and publicly reported with no human action.

How AI was used

The authors assembled a labelled training set of ZTF alert packets by querying internal Bright Transient Survey catalogues for four source classes (spectroscopically confirmed bright transients; AGN, CVs, VarStars and QSOs; dim transient-like sources; and candidates scanners declined to save), retrieving all alerts for those identifiers from the Kowalski broker. Science, reference and difference cutouts were Euclidean-normalised and stacked into 63x63x3 triplets; six custom light-curve features (days to peak, days since peak, age, peakmag so far, maxmag so far, nnondet) were computed from each alert's own perspective and combined with alert-packet metadata to give 25 features, and all alerts were re-scored with the d6 m9 version of the braai real/bogus classifier. Cleaning cuts removed corrupted cutouts, ambiguous labels, missing Pan-STARRS1 cross-matches, i-band alerts, negative difference images and sources with transients in the reference image; sources were split 81/9/10 into train, validation and test, and alerts per source were thinned to Nmax = 100. The multi-modal network passes images through a VGG-like convolutional branch and metadata through batch normalisation and two dense layers, concatenates the branches, and emits a unit-interval score; it was implemented in TensorFlow and Keras with the Adam optimiser, binary cross-entropy loss, class weighting, rotation and flip augmentation, learning-rate decay, and Bayesian hyperparameter sweeps on the Weights and Biases platform, with 20 trials trained at the optimal hyperparameters and uni-modal CNN and fully-connected variants trained for comparison. In production, Kowalski computes the custom features, runs BTSbot on incoming alerts, and two policies (bts p1 and bts p2) map sequences of alert scores to source-level decisions that auto-save sources on Fritz and send SEDM spectroscopic follow-up requests at priority 1 or 2, with resulting spectra reduced by pySEDM and classified by SNIascore.