Determining Strong Lens Mass Models using Convolutional Neural

Overview Strong Gravitational Lensing Theory and Surveys Detection & Characterisation - What’s been done

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs

Slides: 19

Download presentation

Determining Strong Lens Mass Models using Convolutional Neural Networks James Pearson Simon Dye, Nan Li

Overview Strong Gravitational Lensing Theory and Surveys Detection & Characterisation - What’s been done already? Simulating Lensing Machine Learning and Convolutional Neural Networks Results Summary and Future Work 2

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Strong Gravitational Lensing Image Credit: Kavan Ratnatunga (Carnegie Mellon Univ. ) and NASA/ESA Image Credit: NASA, ESA, A. Bolton (Harvard-Smithsonian Cf. A) and the SLACS Team Cosmic Horseshoe, LRG 3 -757. Image Credit: ESA/Hubble & NASA 4

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Strong Gravitational Lensing Sloan Lens ACS (SLACS) Survey CFHTLS Strong Lensing Legacy Survey Dark Energy Survey (DES) Credit: Gunn et al. 2006 Credit: CFHT website, www. cfht. hawaii. edu/en/about / Large Synoptic Survey Telescope (LSST) Credit: LSST Project/NSF/AURA Credit: Reidar Hahn, Fermilab Euclid Telescope Credit: ESA 5

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Detection & Characterisation - What’s been done already? Lens Detection: ● Geometrical quantification (Bom et al. 2017; Seidel & Bartelmann 2007) ● Analysis of colour bands & spectroscopy (Maturi et al. 2014; Baron & Poznanski 2016) 6

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Detection & Characterisation - What’s been done already? Lens Detection: ● Non-CNN Machine Learning (Joseph et al. 2014 (Principle Component Analysis); Avestruz et al. 2017 (Histogram of Oriented Gradients)) ● Convolutional Neural Networks (Petrillo et al. 2017; Lanusse et al. 2017; Jacobs et al. 2017; Schaefer et al. 2017) 7

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Detection & Characterisation - What’s been done already? Lens Parameter Estimation: ● Parameter Fitting techniques (Vegetti & Koopmans 2009; Warren & Dye 2003, Nightingale, Dye & Massey 2017) ● CNNs (Hezaveh et al. 2017) 8

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Machine Learning and Convolutional Neural Networks Credit: Cornell University Blogs https: //blogs. cornell. edu/info 2040/2015/09/ 08/neural-networks-and-machine-learning/ IMAGE Input Convolution 1 Pooling 1 Convolution 2 Pooling 2 Fully. Output Connected 11

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work Summary ● Can use lensing to study distant galaxies, with upcoming surveys expected to produce many thousands of lensing images. ● Created my own CNN to estimate the mass model parameters of lenses. ● Simulated my own images for network training, using SLACS and LSST data. ● In testing the CNN, I found that: 1. increasing the width of the network and changing the optimizer greatly increased accuracy, 2. larger parameter errors can arise depending on the correlation between lens light and mass profiles, 3. the CNN achieves higher accuracy for images without their lens light, 4. the CNN works extremely well for low-noise, low-PSF images, but is less accurate when both are increased to LSST levels. 17

Lensing Theory & Surveys | What’s Been Done? | Simulations | ML & CNNs | Results | Summary & Future Work ● Compare simulated LSST and Euclid images ● Improve simulations, including external shear, other galaxies, using multi-band images and larger catalogues. ● Test my CNN on EAGLE 1 hydrodynamical simulations ● Compare my CNN to Simon Dye’s parameter-fitting Semi-Linear Inversion technique 2 ● Ultimately connect my CNN to Auto. Lens 3 source reconstructor => highly efficient automated software for studying the properties of both lenses and sources. Any Questions? 1(Crain et al. 2015; Schaye et al. 2015), 2(Warren & Dye 2003), 3(Nightingale, Dye & Massey 2017) 18