Direct Detection of Dark Matter Substructure in Strong Lens Images with Convolutional Neural Networks

TL;DR

This paper explores using convolutional neural networks to detect dark matter substructure in strong lens images, aiming to identify subhalos efficiently and improve constraints on dark matter models.

Contribution

It demonstrates a CNN-based method for detecting dark matter substructure directly from lens images without prior modeling, with sensitivity to subhalo masses above 5×10^9 solar masses.

Findings

CNN detects subhalos with >75% accuracy for masses ≥5×10^9 M☉ within the lens galaxy.

Adding lower-mass subhalos does not significantly improve detection accuracy.

The method can quickly identify candidate images for detailed analysis in upcoming surveys.

Abstract

Strong gravitational lensing is a promising way of uncovering the nature of dark matter, by finding perturbations to images that cannot be well accounted for by modeling the lens galaxy without additional structure, be it subhalos (smaller halos within the smooth lens) or line-of-sight (LOS) halos. We present results attempting to infer the presence of substructure from images without requiring an intermediate step in which a smooth model has to be subtracted, using a simple convolutional neural network (CNN). We find that the network is only able to infer the presence of subhalos with $>75\%$ accuracy when they have masses of $\geq 5 \times 10^9$M$_{\odot}$ if they lie within the main lens galaxy. Since less massive foreground LOS halos can have the same effect as higher mass subhalos, the CNN can probe lower masses in the halo mass function. The accuracy does not improve significantly if we add a population of less massive subhalos. With the expectation of experiments such as HST and Euclid yielding thousands of high-quality strong lensing images in the next years, having a way of analyzing images quickly to identify candidates that merit further analysis to determine individual subhalo properties while preventing extensive resources being used for images that would yield null detections could be very useful. By understanding the sensitivity as a function of substructure mass, non-detections could be combined with the information from images with substructure to constrain the cold dark matter scenario, in particular if the sensitivity can be pushed to lower masses.