HOLISMOKES

S. Schuldt, R. Cañameras, Y. Shu, S. H. Suyu, S. Taubenberger, Tim Meinhardt, Laura Leal-TaixéAstronomy 2023 · DOI

TL;DR

This paper introduces HOLISMOKES-IX, a Residual Neural Network (ResNet) that can rapidly infer the parameters of strong gravitational lenses and their uncertainties from ground-based astronomical images. This approach significantly speeds up the modeling process, making it feasible to analyze the vast number of lenses expected from upcoming surveys like LSST.

What Problem Does This Paper Address?

Modeling strong gravitational lenses is crucial for various astrophysical and cosmological studies, including understanding dark matter and constraining the Hubble constant. However, traditional modeling methods, like Markov Chain Monte Carlo (MCMC), are computationally intensive and time-consuming, making them impractical for the ever-increasing number of detected lenses from large sky surveys. The authors aim to develop a fast, automated method to predict lens mass models, including uncertainties, from imaging data.

Methods

The core of this work is a Residual Neural Network (ResNet) architecture. Unlike their previous Convolutional Neural Network (CNN), this ResNet is designed to predict not only the five parameters of a Singular Isothermal Ellipsoid (SIE) mass profile (lens center, ellipticity, and Einstein radius) but also the external shear, and importantly, the 1-sigma uncertainty for each of these seven parameters. The network is trained on simulated lens images generated using a sophisticated pipeline that incorporates real galaxy images from the Hyper Suprime-Cam Survey (HSC) and the Hubble Ultra Deep Field (HUDF). This pipeline has been improved to include Poisson noise, realistic external shear, and better handling of line-of-sight objects, making the simulated data more representative of real observations.

Datasets

The primary dataset used is a collection of approximately 165,000 simulated strong gravitational lens images. These simulations are crucial for training and validating the neural network. The simulations leverage:

  • Lens Galaxies: Real galaxy images from the Hyper Suprime-Cam Survey (HSC).
  • Background Sources: Real galaxy images from the Hubble Ultra Deep Field (HUDF).
  • Lens Properties: Velocity dispersion and redshift measurements from the Sloan Digital Sky Survey (SDSS).

The authors note that while they use only simulated data for testing the network's performance, they apply it to a smaller sample of 31 real HSC lenses in a companion paper.

Key Insights and Ideas

  • ResNet Architecture for Complexity: The adoption of a ResNet architecture, with its residual blocks and skip connections, is key to handling the more complex task of predicting both SIE parameters and external shear, especially from lower-resolution ground-based images.
  • Uncertainty Prediction is Crucial: Including the prediction of 1-sigma uncertainties for each parameter significantly enhances the network's power, allowing for more robust parameter estimations and the exclusion of insecure results.
  • Realistic Simulations are Paramount: The improved simulation pipeline, incorporating real galaxy images and realistic noise, is vital for training a network that performs well on actual observational data. The inclusion of line-of-sight objects and random image shifts helps the network generalize better.
  • External Shear Remains a Challenge: Despite advancements, accurately predicting the external shear from ground-based images remains difficult, likely due to limitations in image resolution and the complexity of disentangling it from other image features.

Key Results

  • Accurate SIE Parameter Recovery: The ResNet demonstrates very good recovery of the SIE parameters, particularly the lens center, showing significant improvement over their previous CNN.
  • Fast Inference: The network can predict all seven parameters (SIE + external shear) and their uncertainties in milliseconds on a single CPU, a dramatic speedup compared to traditional MCMC methods.
  • External Shear Limitations: While the network can predict external shear values, the accuracy is significantly lower than for the SIE parameters, with a tendency to predict values closer to zero.
  • Time Delay Estimation: The network-predicted mass models can be used to estimate time delays for lensed transients. While the precision is currently a first-order estimate and requires refinement, it's a significant step towards rapid transient follow-up.
  • Uncertainty Calibration: The predicted uncertainties are scaled to accurately represent the 68.3%, 95.4%, and 99.7% confidence intervals for most parameters, making them interpretable as Gaussian widths.

Limitations and Future Work

  • External Shear Accuracy: The primary limitation acknowledged is the difficulty in accurately predicting the external shear from ground-based images. The authors suggest this is likely due to image resolution and the complexity of generalizing to new systems. Further investigation is needed.
  • Time Delay Precision: The current precision of time delay predictions is sufficient for a first-order estimate, but requires follow-up modeling for more accurate results, especially for systems with larger image separations.
  • LSST Data: The authors plan to generate dedicated mocks and train a new network specifically for LSST data, as slight differences in image characteristics might affect performance.
  • Real Lens Application: While the network is applied to real HSC lenses in a companion paper, this paper focuses on testing with simulated data. Future work will involve more extensive validation on real datasets.

Summary updated

Abstract

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 e x and e y , Einstein radius θ E ) and the external shear ( γ ext, 1 , γ ext, 2 ) from ground-based imaging data. In contrast to our previous CNN, this ResNet further predicts the 1 σ 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.
BibTeX
@inproceedings{schuldt2023holismokes,
  author = {S. Schuldt and R. Cañameras and Y. Shu and S. H. Suyu and S. Taubenberger and Tim Meinhardt and Laura Leal-Taixé},
  title = {{HOLISMOKES}},
  booktitle = {Astronomy},
  year = {2023},
  doi = {10.1051/0004-6361/202244325},
}

Loading related content...

Figure