Characterising dark matter haloes with computer vision

TL;DR

This study demonstrates that computer vision algorithms can effectively distinguish between different gravity models in dark matter halo simulations using surface mass density maps, achieving over 60% accuracy.

Contribution

The paper introduces a novel application of computer vision features to classify dark matter halo models, including modified gravity theories, with high accuracy.

Findings

Over 60% classification success rate among gravity models

Feature importance identified as Zernike moments, Gini coefficient, Haralick, and Tamura textures

Noise reduction via smoothing improves classification accuracy

Abstract

This work explores the ability of computer vision algorithms to characterise dark matter haloes formed in different models of structure formation. We produce surface mass density maps of the most massive haloes in a suite of eight numerical simulations, all based on the same initial conditions, but implementing different models of gravity. This suite includes a standard $\Lambda$CDM model, two variations of $f(R)$-gravity, two variations of Symmetron gravity and three Dvali, Gabadadze and Porrati (DGP) models. We use the publicly available WND-CHARM algorithm to extract 2919 image features from either the raw pixel intensities of the maps, or from a variety of image transformations including Fourier, Wavelet, Chebyshev and Edge transformations. After discarding the most degenerate models, we achieve more than 60% single-image classification success rate in distinguishing the four different models of gravity while using a simple weighted neighbour distance (WND) to define our classification metric. This number can be increased to more than 70% if additional information, such as a rough estimate of the halo mass, is included. We find that the classification success steeply declines when the noise level in the images is increased, but that this trend can be largely reduced by smoothing the noisy data. We find Zernike moments of the Fourier transformation of either the raw image or its Wavelet transformation to be the most descriptive feature, followed by the Gini coefficient of several transformations and the Haralick and Tamura textures of the raw pixel data eventually pre-processed by an Edge transformation. The proposed methodology is general and does not only apply to the characterisation of modified gravity models, but can be used to classify any set of models which show variations in the 2D morphology of their respective structure.