Euclid DR1 lens modelling drops per-image hand tracing: AstroVink-TREX separates lens, source and contaminants, with only 39 of about 3,000 systems failing on segmentation
Synopsis
Euclid DR1 holds more than 15,000 strong-lens candidates, and until now an expert had to mark the lens light, the lensed source light and the foreground contaminants to be masked in every image, which cannot keep pace with the catalogue. AstroVink-TREX replaces that step with a neural network that predicts three flux maps and derives the modelling centre, fitting region and conjugate points from them. Across about 3,000 expert-inspected systems, at a 75% inspector-agreement threshold, only 39 models fail because of a segmentation error.
Interpretation
Mass modelling of strong lenses can now run its image preparation automatically across the whole Euclid DR1 catalogue, with lens light, lensed source light and foreground contaminants separated in one pass instead of being drawn by hand system by system. Within Euclid, identifying and masking the components of strong lens systems had been done entirely by hand, with an expert mapping contaminants and outlining lens and source regions one system at a time, a process that is slow, hard to reproduce and dependent on individual judgement. Across about 3,000 inspected DR1 systems, at a 75% inspector-agreement threshold, 210 systems are marked as failures and 18.6% of those are attributed to segmentation, meaning 39 models fail because of a segmentation error; this ordering holds at 100%, 75% and 50% consensus thresholds.
The network outputs three continuous flux maps, one each for lens, source and contaminant, and derives the modelling centre, fitting radius, conjugate points and binary masks directly from them. The network uses a hybrid architecture that pairs a reused AstroVink vision transformer encoder, which connects arcs far apart in the image, with a parallel convolutional branch that preserves pixel-level detail. Training targets come from 377 Q1 systems already modelled successfully with PyAutoLens, split into 302 for training and 75 for validation; evaluation uses a 750-candidate DR1 subset and the expert-inspected set of about 3,000 systems.
Perspective
The demonstrated scope is Euclid DR1 strong-lens candidates: the network is validated on about 3,000 expert-inspected systems and supplies component separation and modelling setup for every system modelled in DR1, with training drawn from only 377 Q1 systems modelled with PyAutoLens.
Segmentation failures mainly involve confusion between source structures and contaminants, particularly when source images are very faint or the system morphology is complex; such failures can be flagged automatically by comparing the lens and source light reconstructed by the mass-modelling pipeline with the network's predicted maps. The network currently uses a single band, and adding near-infrared bands would require adapting or retraining the encoder.
