Machine Learning for Multi-messenger Probes of New Physics and Cosmology: A Review and Perspective

Author(s)

Addazi, Andrea, Belotsky, Konstantin, Beylin, Vitaly, Bikbaev, Timur, Chen, Deen, Fabrocini, Filippo, Giagu, Stefano, Jinklub, Krid, Kharakhashyan, Artem, Khlopov, Maxim, Korchagin, Vladimir, Krasnov, Maxim, Mahajan, Atharv, Marciano, Antonino, Mayorov, Andrey, Morais, Antonio, Pasechnik, Roman, Levi Said, Jackson, Sopin, Danila, Stasenko, Viktor, Trivedi, Oem

Abstract

The multi-messenger exploration of dark matter and physics beyond the Standard Model has emerged as a central direction in modern astro-particle physics, particularly following the discovery of gravitational waves. In this work, we present a comprehensive review and forward-looking perspective on machine-learning-enhanced multi-messenger approaches, combining information from gravitational waves, cosmic rays, gamma rays, neutrinos, and collider experiments. We summarize the current state of the field, discuss recent methodological developments, and outline a coherent research program aimed at integrating heterogeneous datasets within a unified inference framework. Our collaboration proposes here a plan for forthcoming analyses aiming at extracting information on the properties and interactions of dark matter, and finally on its genesis, combining multi-messenger astronomy techniques and inputs from laboratory physics. The main objectives planned in this line of research comprise: i) the multi-messenger analysis of new physics in cosmology, including mainly, but not only, several different models of dark matter; ii) the phenomenology of new physics signatures in ground-based cosmic rays experiments, with cross-correlation to the corresponding physical, astrophysical and cosmological observations; iii) the development of machine learning methods for data analysis in ground-based cosmic rays experiments, in light of the new physics signatures. We note that several groups have explored the use of multi-messenger observations, including gravitational waves, to probe alternative dark matter candidates. The present work builds on these developments by focusing on the role of machine learning in integrating heterogeneous datasets. We foresee that such a cross-fertilizing approach will represent the right path to extract information about the main questions left in fundamental physics.

Figures

Caption

Demonstration of likelihood-free inference (LFI) using a simulation-based approximate posterior in a 1D toy cosmological parameter estimation task. Unlike traditional Bayesian inference, which requires an analytical likelihood, this approach uses distance-weighted simulation outputs to recover the posterior around the true parameter value. Such methods are foundational in machine learning-based cosmological inference pipelines where forward simulations are available but explicit likelihoods are intractable.

Caption

Convolutional Neural Network (CNN) trained to estimate the cosmological parameter $\Omega_m$ directly from synthetic 2D "sky map" images. The left panel shows the training and validation mean squared error (MSE) loss over epochs, indicating successful learning. The right panel compares predicted and true $\Omega_m$ values on held-out data, demonstrating the model’s strong regression performance. This illustrates the potential of CNN-based inference in cosmology.

Caption

Figure 6, from \cite{BENTO2025139690}, shows a schematic representation of the PINN structure. This is relevant to tackle inverse problems while modeling the Boltzmann equation for freeze-in DM particle yield, in alternative cosmology.