GIGA-Lens: Fast Bayesian Inference for Strong Gravitational Lens Modeling

Gu, A.; Huang, Xiaosheng; Sheu, W.; Aldering, G.; Bolton, A.; Boone, K.; Dey, Arjun; Filipp, A.; Jullo, Eric; Perlmutter, S.; Rubin, D.; Schlafly, Edward F.; Schlegel, David J.; Shu, Y.; Suyu, S. H.

doi:10.3847/1538-4357/ac6de4

Cited by 17 publications

(12 citation statements)

References 96 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The latter is the best match to our networks, and is comparable in performance, as mentioned in S21a. There is also currently a lot of work going into automated modeling without machine learning (e.g., Nightingale et al 2018Nightingale et al , 2021Rojas et al 2022;Savary et al 2022;Ertl et al 2023;Etherington et al 2022;Gu et al 2022;Schmidt et al 2023), which typically performs better than neural networks but requires significantly longer run times of hours to days. We refer to Schuldt et al (2022) for a direct comparison between the network presented here and traditionally obtained models for real HSC lenses.…”

Section: Network Results and Performancementioning

confidence: 99%

Holismokes

Schuldt

Cañameras

Shu

et al. 2023

A&A

Self Cite

View full text Add to dashboard Cite

Modeling of strong gravitational lenses is a necessity for further applications in astrophysics and cosmology. With the large number of detections in current and upcoming surveys, such as the Rubin Legacy Survey of Space and Time (LSST), it is pertinent to investigate automated and fast analysis techniques beyond the traditional and time-consuming Markov chain Monte Carlo sampling methods. Building upon our (simple) convolutional neural network (CNN), we present here another CNN, specifically a residual neural network (ResNet), that predicts the five mass parameters of a singular isothermal ellipsoid (SIE) profile (lens center x and y, ellipticity ex and ey, Einstein radius θE) and the external shear (γext, 1, γext, 2) from ground-based imaging data. In contrast to our previous CNN, this ResNet further predicts the 1σ uncertainty for each parameter. To train our network, we use our improved pipeline to simulate lens images using real images of galaxies from the Hyper Suprime-Cam Survey (HSC) and from the Hubble Ultra Deep Field as lens galaxies and background sources, respectively. We find very good recoveries overall for the SIE parameters, especially for the lens center in comparison to our previous CNN, while significant differences remain in predicting the external shear. From our multiple tests, it appears that most likely the low ground-based image resolution is the limiting factor in predicting the external shear. Given the run time of milli-seconds per system, our network is perfectly suited to quickly predict the next appearing image and time delays of lensed transients. Therefore, we use the network-predicted mass model to estimate these quantities and compare to those values obtained from our simulations. Unfortunately, the achieved precision allows only a first-order estimate of time delays on real lens systems and requires further refinement through follow-up modeling. Nonetheless, our ResNet is able to predict the SIE and shear parameter values in fractions of a second on a single CPU, meaning that we are able to efficiently process the huge amount of galaxy-scale lenses expected in the near future.

show abstract

Section: Network Results and Performancementioning

confidence: 99%

Holismokes

Schuldt

Cañameras

Shu

et al. 2023

A&A

Self Cite

View full text Add to dashboard Cite

show abstract

“…Various studies are investigating this approach for groundbased images and also future Euclid-like and LSST-like images (e.g., Pearson et al 2019;Schuldt et al 2021;Pearson et al 2021), and for its compatibility with the hierarchical inference framework introduced above (Wagner-Carena et al 2021;Park et al 2021). In addition, the use of Graphics Processing Units (GPU) can save substantial computational time (Gu et al 2022). The challenges are to validate these new modeling methods on real lenses since these methods have so far been demonstrated only on mock lenses.…”

Section: Deflector-modeling Challengesmentioning

confidence: 99%

Strong lensing time-delay cosmography in the 2020s

Treu

Suyu

Marshall

2022

Astron Astrophys Rev

View full text Add to dashboard Cite

Multiply imaged time-variable sources can be used to measure absolute distances as a function of redshifts and thus determine cosmological parameters, chiefly the Hubble Constant H$$_0$$ 0 . In the two decades up to 2020, through a number of observational and conceptual breakthroughs, this so-called time-delay cosmography has reached a precision sufficient to be an important independent voice in the current “Hubble tension” debate between early- and late-universe determinations of H$$_0$$ 0 . The 2020s promise to deliver major advances in time-delay cosmography, owing to the large number of lenses to be discovered by new and upcoming surveys and the vastly improved capabilities for follow-up and analysis. In this review, after a brief summary of the foundations of the method and recent advances, we outline the opportunities for the decade and the challenges that will need to be overcome in order to meet the goal of the determination of H$$_0$$ 0 from time-delay cosmography with 1% precision and accuracy.

show abstract

Section: Introductionmentioning

confidence: 99%

“…One possibility is to automate the modeling procedure while still relying on Bayesian inference such as MCMC sampling (e.g., Nightingale et al 2018;Rojas et al 2022;Savary et al 2022;Ertl et al 2022;Etherington et al 2022;Gu et al 2022;Schmidt et al 2023), which reduces the user input dramatically, resulting in an overall runtime on the order of days. A further speed-up can be achieved by using Graphical Processing Units (GPUs, e.g., Gu et al 2022). Another option is to use machine learning (Hezaveh et al 2017;Perreault Levasseur et al 2017;Morningstar et al 2018Morningstar et al , 2019Pearson et al 2019Pearson et al , 2021Schuldt et al 2021b, 2023, hereafter S21b andS23, respectively).…”

Section: Introductionmentioning

confidence: 99%

“…they had an additional a slope parameter to constrain. In an attempt to accelerate the modeling,Gu et al (2022) presented Giga-Lens, a modeling code tested on simulated HST-like lenses that still relies on a Bayesian framework but is much faster through the use of GPUs, high parallelization, and implementations in TensorFlow(Abadi et al 2015) and Jax(Bradbury et al 2018). In contrast to those codes for lensed galaxies,Schmidt et al (2023) andErtl et al (2022) have developed both modeling pipelines that are dedicated to HST images of strongly lensed quasars to derive mass models more rapidly which is necessary to predict time delays and give a good initial model with minimal user input required for cosmological analysis.Besides these codes for high-resolution images, Rojas et al (2022) andSavary et al (…”

mentioning

confidence: 99%

See 1 more Smart Citation

Holismokes

Schuldt¹,

Suyu²,

Cañameras³

et al. 2023

A&A

View full text Add to dashboard Cite

Modeling of strongly gravitationally lensed galaxies is often required in order to use them as astrophysical or cosmological probes. With current and upcoming wide-field imaging surveys, the number of detected lenses is increasing significantly such that automated and fast modeling procedures for ground-based data are urgently needed. This is especially pertinent to short-lived lensed transients in order to plan follow-up observations. Therefore, we present in a companion paper a neural network predicting the parameter values with corresponding uncertainties of a singular isothermal ellipsoid (SIE) mass profile with external shear. In this work, we also present a newly developed pipeline glee auto.py that can be used to model any galaxy-scale lensing system consistently. In contrast to previous automated modeling pipelines that require high-resolution space-based images, glee auto.py is optimized to work well on ground-based images such as those from the Hyper-Suprime-Cam (HSC) Subaru Strategic Program or the upcoming Rubin Observatory Legacy Survey of Space and Time. We further present glee tools.py, a flexible automation code for individual modeling that has no direct decisions and assumptions implemented on the lens system setup or image resolution. Both pipelines, in addition to our modeling network, minimize the user input time drastically and thus are important for future modeling efforts. We applied the network to 31 real galaxy-scale lenses of HSC and compare the results to traditional, Markov-Chain Monte-Carlo sampling-based models obtained from our semi-autonomous pipelines. In the direct comparison, we find a very good match for the Einstein radius. The lens mass center and ellipticity show reasonable agreement. The main discrepancies pretrain to the external shear, as is expected from our tests on mock systems where the neural network always predicts values close to zero for the complex components of the shear. In general, our study demonstrates that neural networks are a viable and ultra fast approach for measuring the lens-galaxy masses from ground-based data in the upcoming era with ∼ 10 5 lenses expected.

show abstract

GIGA-Lens: Fast Bayesian Inference for Strong Gravitational Lens Modeling

Cited by 17 publications

References 96 publications

Holismokes

Holismokes

Strong lensing time-delay cosmography in the 2020s

Holismokes

Contact Info

Product

Resources

About