Strong Lensing Source Reconstruction
Interdisciplinary Research Achievement �The aim of this project is to use continuous neural fields (also known as implicit neural representations) to reconstruct the complex morphology of a background source galaxy non-parametrically while simultaneously inferring a distribution over possible lens configurations in observations of strongly-lensed systems. The efficacy of the method is demonstrated through experiments targeting high-resolution lensing images similar to those anticipated in near-future astrophysical surveys like Euclid and the Extremely Large Telescope (ELT). The first paper [1] was presented as a Spotlight Oral (8% spotlight rate) at the ICML 2022 Machine Learning for Astrophysics Workshop.
Impact on Artificial Intelligence�Several innovations on the AI methodological side are introduced. Rather than predicting a point estimate for the source galaxy image, we obtain a distribution over possible sources using variational inference. Additionally, we include the continuous neural representation of the source as part of a larger probabilistic model that simultaneously infers a posterior over possible lens galaxy configuration. A fully probabilistic treatment is crucial for downstream applications in searching for dark matter substructure. A probabilistic treatment of continuous neural fields could be more widely applicable in the field of computer vision, where this AI method is typically used in methods like neural radiance fields.
Impact on Fundamental Interactions
Domain knowledge is introduced through a fully differentiable forward model (simulator) that renders pixelized strong lensing observations given a continuous representation of the unlensed source. We have implemented a Jax-based, fully differentiable simulator for rendering strong lensing observations of background galaxies with complex non-parameteric morphologies and completed the initial project goal of developing a pipeline for strong lensing source and lens reconstruction using continuous neural field representations. We have also tested and validated the pipeline on a suite of mock strong lensing images to make sure trustworthy posteriors on all quantities of interest (characterizing both the lens and source galaxies) can be recovered.
Outlook
The end goal of this project is to be able to efficiently model future high-resolution strong lensing observations at their full complexity. The reconstruction model can then be used for a variety of downstream analyses, for example targeting the effect of dark matter substructure (correlated with the dark matter particle physics) on the lensing image using likelihood-based as well as likelihood-free methods.
Siddharth Mishra-Sharma, Ge Yang (IAIFI Fellows)
Figure 1: A schematic overview of the method used in this work. Figure 2: Results of reconstructing the source through our coordinate-based neural network pipeline using a mock lensed image of galaxy NGC2906. The (a) true source, (b) reconstructed source mean, (c) reconstructed mean minus true source residuals, (d) true lensed image, (e) reconstructed lensed image, and (f) reconstructed mean minus true lensed image residuals are shown. All images are normalized by the observation noise.
[1] arXiv:2206.14820
The NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) is �supported by National Science Foundation under Cooperative Agreement PHY-2019786
Figure 1
Figure 2