Source-linked AI summary
Fast Automated Analysis of Strong Gravitational Lenses with Convolutional Neural Networks
Yashar D. Hezaveh, Laurence Perreault Levasseur, Philip J. Marshall
TL;DR
The paper develops neural-network methods that map image data to lensing parameters using simulated and real galaxy images. It also applies ICA to separate lensed arcs from lens-galaxy images, while robustness decreases for lensing conditions outside training coverage.
Problem
Estimating lensing parameters from image data requires a mapping from pixel inputs to predicted physical parameters.
Method
The paper trains feed-forward convolutional networks on simulated lensed images, applies observational effects during training, and uses ICA to unmix multi-filter lens images before inference.
Results
Realistic non-Gaussian HST PSFs did not significantly reduce the networks’ parameter accuracies.
Takeaways & Limitations
Freely available simulation, parameter-estimation code, and trained network weights support reuse of the analysis pipeline.
Takeaways & Limitations
Performance decreases when lensing parameters fall outside the training range or when untrained nuisance features such as external shear are present.
Abstract
from arXiv · showhide
Quantifying image distortions caused by strong gravitational lensing and estimating the corresponding matter distribution in lensing galaxies has been primarily performed by maximum likelihood modeling of observations. This is typically a time and resource-consuming procedure, requiring sophisticated lensing codes, several data preparation steps, and finding the maximum likelihood model parameters in a computationally expensive process with downhill optimizers. Accurate analysis of a single lens can take up to a few weeks and requires the attention of dedicated experts. Tens of thousands of new lenses are expected to be discovered with the upcoming generation of ground and space surveys, the analysis of which can be a challenging task. Here we report the use of deep convolutional neural networks to accurately estimate lensing parameters in an extremely fast and automated way, circumventing the difficulties faced by maximum likelihood methods. We also show that lens removal can be made fast and automated using Independent Component Analysis of multi-filter imaging data. Our networks can recover the parameters of the Singular Isothermal Ellipsoid density profile, commonly used to model strong lensing systems, with an accuracy comparable to the uncertainties of sophisticated models, but about ten million times faster: 100 systems in approximately 1s on a single graphics processing unit. These networks can provide a way for non-experts to obtain lensing parameter estimates for large samples of data. Our results suggest that neural networks can be a powerful and fast alternative to maximum likelihood procedures commonly used in astrophysics, radically transforming the traditional methods of data reduction and analysis.
METHODS
The paper develops neural-network and ICA methods for automated strong-lensing analysis, using simulated and observationally varied data to estimate lens parameters and remove lens light.
- METHODS: Training uses simulated real and clumpy background galaxies with randomized lensing configurations and realistic observational effects.The effects include PSF convolution, shot and Gaussian noise, cosmic rays, hot pixels, zero bias, and circular masks.
- METHODS: Independent Component Analysis separates lens light from source arcs using multi-wavelength images whose foreground and background morphologies are treated as statistically independent.The separated arc images are then passed to the networks; the observational pipeline uses at least two HST WFC3 filters and masks bright contaminants.
- METHODS: The network pipeline is evaluated on independently seeded validation and test sets, including background galaxy images excluded from training.The methods also test realistic non-Gaussian HST PSFs and find no significant reduction in parameter accuracy.
- METHODS: Performance decreases for lens parameters or nuisance features outside the training distribution, although adding external shear to training data could reduce the resulting errors.With external shear above 0.15, ellipticity RMS errors rise to 0.23, compared with 0.11 below shear 0.01; lens-center predictions remain robust.