Euclid Quick Data Release (Q1). From simulations to sky: Advancing machine-learning lens detection with real Euclid data

TL;DR

The study quantifies the domain gap between simulated training data and real Euclid observations for strong-lens detection and demonstrates that a purely simulation-trained model achieves $AUC ≈ 0.9991$ on simulations but only $AUC ≈ 0.941$ on real Q1 data, with $92\%$ completeness at $100\%$ purity in simulations versus $50\%$ completeness at $24\%$ purity on real data.$+0$ Incorporating real Q1 lenses and non-lenses into training yields substantial improvements, boosting expected lens discoveries in DR1 and the full EWS by $25$–$30\%$ and reducing the number of inspected images by about a factor of ten; most gains arise from real lenses, with additional gains from more diverse non-lens contaminants.$+0$ Embedding analyses (UMAP) reveal real lenses occupy an intermediate region between simulated and non-lens instances, supporting a hybrid approach that blends high-fidelity real data with broad simulations to capture lens diversity.$+0$ The findings establish a practical pathway for maximizing lens yields in Euclid and inform similar strategies for LSST, while highlighting the need to quantify selection effects and mitigate domain biases as data volumes grow.$+0$

Abstract

In the era of large-scale surveys like Euclid, machine learning has become an essential tool for identifying rare yet scientifically valuable objects, such as strong gravitational lenses. However, supervised machine-learning approaches require large quantities of labelled examples to train on, and the limited number of known strong lenses has lead to a reliance on simulations for training. A well-known challenge is that machine-learning models trained on one data domain often underperform when applied to a different domain: in the context of lens finding, this means that strong performance on simulated lenses does not necessarily translate into equally good performance on real observations. In Euclid's Quick Data Release 1 (Q1), covering 63 deg2, 500 strong lens candidates were discovered through a synergy of machine learning, citizen science, and expert visual inspection. These discoveries now allow us to quantify this performance gap and investigate the impact of training on real data. We find that a network trained only on simulations recovers up to 92% of simulated lenses with 100% purity, but only achieves 50% completeness with 24% purity on real Euclid data. By augmenting training data with real Euclid lenses and non-lenses, completeness improves by 25-30% in terms of the expected yield of discoverable lenses in Euclid DR1 and the full Euclid Wide Survey. Roughly 20% of this improvement comes from the inclusion of real lenses in the training data, while 5-10% comes from exposure to a more diverse set of non-lenses and false-positives from Q1. We show that the most effective lens-finding strategy for real-world performance combines the diversity of simulations with the fidelity of real lenses. This hybrid approach establishes a clear methodology for maximising lens discoveries in future data releases from Euclid, and will likely also be applicable to other surveys such as LSST.

Figures (9)

• Figure 1: Top: simulated lenses from Q1-SP052. Bottom: real lenses found in Q1.
• Figure 2: ROC curve of the fine-tuned Zoobot ML model that is trained only on pre-Q1 data (simulated lenses), as evaluated on both pre-Q1 and Q1 data.
• Figure 3: Projected number of lenses discoverable in DR1 and EWS as a function of number of images to inspect, for the network trained with and without the Q1 data.
• Figure 4: Impact of augmenting the training data with available Q1 lenses and non-lenses, in terms of performance across a range of metrics. These include F1 score and AUC, as well as the projected fraction of lenses that would be discoverable in a data set the size of EWS, assuming that one million images can be visually inspected.
• Figure 5: Same as in Fig. \ref{['fig:adding-Q1-data']}, but isolating the impact of adding Q1 lenses and non-lenses separately. The $x$-axis corresponds to adding the Q1 lenses to the pre-existing training data, and the two lines show the change in performance with and without the addition of the Q1 non-lenses.