[Paper Review] Hunting for Dark Matter Subhalos in Strong Gravitational Lensing with Neural Networks

Joshua Yao-Yu Lin, Hang Yu+3 more|arXiv (Cornell University)|Oct 24, 2020

Galaxies: Formation, Evolution, Phenomena12 references4 citations

TL;DR

This paper proposes using deep neural networks to detect dark matter subhalos in strong gravitational lensing images without traditional lens modeling. Trained on simulated lensing data with subhalo probability maps, the network detects multiple subhalos accurately and surprisingly learns to reject false positives on smooth lensing arcs, achieving inference in under a second per image.

ABSTRACT

Dark matter substructures are interesting since they can reveal the properties of dark matter. Collisionless N-body simulations of cold dark matter show more substructures compared with the population of dwarf galaxy satellites observed in our local group. Therefore, understanding the population and property of subhalos at cosmological scale would be an interesting test for cold dark matter. In recent years, it has become possible to detect individual dark matter subhalos near images of strongly lensed extended background galaxies. In this work, we discuss the possibility of using deep neural networks to detect dark matter subhalos, and showing some preliminary results with simulated data. We found that neural networks not only show promising results on detecting multiple dark matter subhalos, but also learn to reject the subhalos on the lensing arc of a smooth lens where there is no subhalo.

Motivation & Objective

• To develop an automated, fast method for detecting dark matter subhalos in strong gravitational lensing images, bypassing computationally expensive lens modeling.
• To investigate whether deep learning can detect multiple subhalos simultaneously in complex lens systems with realistic substructure distributions.
• To explore whether neural networks can implicitly learn to reject spurious subhalo signals on lensing arcs caused by smooth lens potentials.
• To enable scalable analysis of upcoming large-scale lensing surveys (e.g., LSST, Euclid) with ~170,000 expected strong lens systems.

Proposed method

• Trained convolutional neural networks (DenseNet and ResNet) on simulated strong lensing images with subhalo probability maps as targets.
• Used raytracing under thin-lens approximation to generate 224×224 lensing images with smooth SIE or power-law macro models and embedded subhalos.
• Constructed subhalo probability maps using Gaussian distributions with cutoffs to represent subhalo position likelihoods, enabling regression via binary cross-entropy loss.
• Modified the final layers of pre-trained CNNs to output 56×56×1 probability maps instead of classification logits, preserving spatial resolution.
• Normalized input images to [0,1] and duplicated channels to match RGB input requirements of ResNet and DenseNet.
• Trained the network to minimize binary cross-entropy loss between predicted and ground-truth subhalo probability maps.

Experimental results

Research questions

• RQ1 Can deep neural networks detect multiple dark matter subhalos in strong lensing images without explicit lens modeling?
• RQ2 Can a neural network trained solely on subhalo probability maps learn to reject false positive detections on smooth lensing arcs where no subhalos exist?
• RQ3 How does the performance of neural networks compare to traditional likelihood-based methods in terms of speed and accuracy for subhalo detection?
• RQ4 Can the network generalize to complex lensing systems with cumulative substructure effects on arcs, even when not explicitly trained on such features?

Key findings

• The neural network successfully detected multiple subhalos in simulated lensing images, accurately predicting their positions in the probability map.
• When no subhalos were present, the network predicted zero detections, confirming proper generalization to subhalo-free systems.
• The network learned to assign lower probability to regions along the lensing arc in smooth models, effectively rejecting false positives without explicit supervision.
• Inference time was less than one second per image using a single GPU, representing a dramatic speedup over traditional modeling methods that take days.
• The network’s ability to reject subhalos on smooth arcs emerged as an emergent property, not explicitly programmed, suggesting robust feature learning.
• The method shows strong potential for scaling to large surveys like LSST and Euclid, where automated, fast subhalo detection is essential.