Presentation #301.03 in the session Machine Learning Applications.
With the launch of the Euclid space telescope in 2023, its imaging survey covering up to 15,000 deg2 is expected to reveal ~100,000 strong gravitational lensing events. Strong lenses are valuable for a variety of astrophysical and cosmological applications, but only ~1,000 strong lenses have been identified so far. Detecting the strong lenses in Euclid’s survey will thus provide vital resources for new scientific discoveries. However, these events are rare, occurring once in every ~10,000 observed massive galaxies, making visual inspection of the survey data unfeasible. We address this issue by using Convolutional Neural Networks (CNNs) to automate this detection process.
To train the CNN for this task, we must first generate mock lensing images. We sampled pairs of galaxies from HST’s CANDELS survey at different redshifts, where the foreground galaxy acts as the lens and the background galaxy as the source. The lensing potential of a Singular Isothermal Ellipsoid (SIE) was used to artificially lens the background galaxy, and the image of the lensed background was overlaid onto a CANDELS cutout of the foreground galaxy. This produces an image of strong lensing events in Hubble’s resolution, which was then downgraded to Euclid’s resolution. We simulated a total of 60,000 training images for both VIS and NISP (Euclid’s visual and IR band imagers), with lenses simulated in a wide range of configurations to fully cover the parameter space of resolvable lenses.
We used a CNN architecture derived from VGG as it was found to be most suitable for this task. Two CNNs were trained separately on the VIS and NISP images. The performance of the CNNs was evaluated on a separate set of test images using the Receiver Operating Curve (ROC), and the Area Under the Curve (AUC) was measured. The VIS classifier achieved an AUC of 0.987, and the NISP classifier achieved 0.797. We also characterized the CNN’s bias towards different morphological/physical properties of the lensing configuration, this will be important for any future statistics work with the detected lenses.
In application, the CNN will go through Euclid’s imaging data and propose candidates of strong lenses for visual inspection. By choosing an appropriate classification threshold, the CNN have a precision of 99% while retrieving most of the strong lenses. This will increase the purity of lenses from 1 in ~10,000 galaxies to 1 in ~100 proposed candidates. Further reduction can be achieved by feeding only elliptical galaxies to the CNN or constraining other physical properties of the fed galaxies.