Dark from light (DfL): Inferring halo properties from luminous tracers with machine learning trained on cosmological simulations. I. Method, proof of concept & preliminary testing

Abstract

We present Dark from Light (DfL) - a novel method to infer the dark sector in wide-field galaxy surveys, leveraging a machine learning approach trained on contemporary cosmological simulations. The aim of this algorithm is to provide a fast, straightforward, and accurate route to estimating dark matter halo masses and group membership in wide-field spectroscopic galaxy surveys. This approach requires a highly limited number of input parameters and yields full probability distribution functions for the output halo masses. To achieve this, we train a series of Random Forest (RF) regression models on the IllustrisTNG and EAGLE simulations at z=0-3, which provide model-dependent mappings from luminous tracers to dark matter halo properties. We incorporate the individual regression models into a virial group-finding algorithm (DfL), which outputs halo properties for observational-like input data. We test the method at z=0-2 for both the EAGLE and IllustrisTNG models, as well as in a cross-validation mode. We demonstrate that known halo masses can be recovered with a mean systematic bias of $\langle b \rangle = \pm 0.10\,$dex (resulting from simulation choice), a mean statistical uncertainty of $\langle \sigma \rangle = 0.12 \,$dex across epochs, and a central - (core) satellite classification accuracy of 96%. We establish that this approach yields superior halo mass recovery to standard abundance matching applied to groups identified through a friends-of-friends algorithm. Additionally, we compare the outputs of DfL to observational constraints on the $M_* - M_{\rm Halo}$ relation from strong gravitational lensing at $z \sim 0$, demonstrating the promise of this novel approach. Finally, we systematically quantify how DfL performs on observational-like input data with varying stellar mass uncertainty and spectroscopic incompleteness, enabling robust error calibration.