Learning the shape of and what shapes dark structure
Abstract
LCDM shows some tension with observations in the non-linear regime of structure formation. This includes abundance, density profiles and substructure amount for dark matter haloes at a range of masses and redshifts. Several recipes have been suggested to remedy such tensions including more detailed modelling of baryonic physics, modifications to gravity or variations to the collisionless, cold dark matter paradigm. However, current observational signatures to distinguish such models are elusive due to the vast information compression in the characterisation of observed dark matter structure via e.g. two-point correlation functions or 1D functional forms of density profiles. I will present a full morphological analysis of dark matter structure in the sky based on optimal mass mapping, computer vision and machine learning that alleviates this shortcoming. More specifically, I will focus on two characterisation schemes. The first one derives from classical computer vision and machine learning and extracts up to 3500 characterising elements directly from dark matter mass or shear maps, several of their transformations (Fourier, Chebyshev, Edge, Wavelet, ...) and transformation of transformations. The second approach is based on deep learning and applies a multi-layered convolutional neural network, including state-of-the-art methods such as inception layers and region detection, to learn the main characterising features in the dark matter mass and shear maps.We apply these techniques to several sets of numerical simulations, all of which explore different aspects of the underlying model of structure formation. This includes several models of modified gravity, the degeneracy between modified gravity and the presence of massive neutrino, baryonic feedback and models of self-interacting dark matter. I will close with the application of this technique to classify the underlying structure formation model of real data coming from the KiDS and CLASH surveys and will give an outlook on the potential of future space and ground-based experiments such as Euclid, WFIRST and LSST.
- Publication:
-
42nd COSPAR Scientific Assembly
- Pub Date:
- July 2018
- Bibcode:
- 2018cosp...42E2264M