Deep Neural Networks Hunting Ultra-Light Dark Matter
Pavel Kůs 1,2, Diana López Nacir 3,4, Federico R. Urban 1 · 2025
Read Deep Neural Networks Hunting Ultra-Light Dark Matter by Pavel Kůs 1,2, Diana López Nacir 3,4, Federico R. Urban 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-04100.txt.
Introduction
Pavel Kůs, Diana López Nacir, and Federico Urban test machine learning as a way to hunt ultra-light dark matter in this 2025 paper. Such matter, if it touches ordinary matter, should leave a faint pattern in pulsar timing data, since it tugs the orbits of pulsars in pairs. The team built three kinds of neural network and trained them on made-up data.
The networks were an autoencoder, a two-class sorter, and a multi-class sorter. They were tried on four models of the dark matter: two scalar fields, a vector field, and a tensor field. The machine results matched the reach of a standard statistical method used before. The multi-class sorter could also tell the four models apart. The authors present the work as groundwork for searches in real pulsar timing records.
Catalog record
| CREATOR | Pavel Kůs 1,2, Diana López Nacir 3,4, Federico R. Urban 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 119 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-04100
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep Neural Networks Hunting Ultra-Light Dark Matter Pavel Kůs 1,2, Diana López Nacir 3,4, Federico R. Urban 1 Abstract Ultra-light dark matter (ULDM) is a compelling candidate for cosmological dark matter. If ULDM interacts with ordinary matter, it can induce measurable, characteristic signals in pulsar-timing data because it causes the orbits of pulsars in binary systems to osculate. In this work, we investigate the potential of machine learning (ML) techniques to detect such ULDM signals. To this end, we construct three types of neural networks: an autoencoder, a binary classifier, and a multiclass classifier. We apply these methods to four theoretically well-motivated ULDM models: a linearly coupled scalar field, a quadratically coupled scalar field, a vector field and a tensor field. We show that the sensitivity achieved using ML methods is comparable to that of a semi-analytical Bayesian approach, which to date has only been applied to the linear scalar case. The ML approach is readily applicable to all four ULDM models and, in the case of the multiclass classifier, can distinguish between them. Our results, derived from simulated data, lay the foundation for future applications to real pulsar-timing observations. keywords ultra-light dark matter, pulsar timing, binary pulsars, deep learning, autoencoder, convolutional neural networks, binary classifier, multiclassifier 1 Introduction Dark matter constitutes about 27% of the energy density of the Universe and is about five times more abundant than baryonic matter, yet its fundamental nature remains elusive. A promising solution to this problem is a class of dark matter candidates known as ultra-light dark matter (ULDM). This type of dark matter consists of light bosons with masses in the range 10^{-23}\lesssim m\lesssim 1 eV. If these particles account for the entirety of the dark matter, their number density must be enormous, allowing the dark matter to be described by a coherent classical field oscillating at a frequency determined by the particle’s mass. When the mass approaches the lower end of the interval, the de Broglie wavelength of ULDM stretches across astrophysical scales of order of pc to kpc. Consequently, ULDM exhibits wave-like behaviour on galactic scales, leading to a rich phenomenology while still reproducing the behaviour of cold dark matter on cosmological scales. The dynamics of non-relativistic ULDM is governed by the Schrödinger–Poisson system of equations; numerical simulations predict that ULDM halos form with a central solitonic core, while the outer regions exhibit a granular structure due to wave interference effects ( Schive et al. 2014 ) . 1 1 1 For studies on soliton merger simulations for spin-1 and spin-2 ULDM see Amin et al. 2022 ; López-Sánchez et al. . For a comprehensive review of ULDM properties and tests we refer to Ferreira 2021 . Binary pulsars—binary star systems where one star is a pulsar—have proven to be promising ULDM detectors. Studies in Blas et al. 2017 ; Blas et al. 2020 ; López Nacir and Urban 2018 ; Armaleo et al. 2020 ; Kůs et al. 2024 have shown that both direct and gravitational interactions between ULDM and the components of a binary pulsar lead to perturbations in the dynamics of the binary—see Fig. 1 for an artistic illustration—which leave their imprints on pulsar timing data. If this effect is not incorporated into the timing model describing the times of arrival of the pulses on Earth, it results in a discrepancy between the theoretically predicted values and the actual measurements. This discrepancy manifests as time residuals with a characteristic pattern—a unique fingerprint of each ULDM model—which we refer to as a signal . Figure 1: Illustration of a binary pulsar inside a dark matter halo that follows a non-Keplerian orbit caused by interaction with ULDM. Although the gravitational imprint on pulsar timing data appears too weak to be detectable ( Blas et al. 2020 ) , a signal from direct ULDM-matter interaction could be observed. If such a signal is not detected, it would place stringent constraints on the ULDM parameter space. Initial studies were conducted for models with spins 0, 1 and 2, within the mass range of 10^{-23} to 10^{-18} eV, beyond which the binary pulsar’s backreaction to dark matter can not be ignored ( Blas et al. 2020 ) . Recently, a new two-step Bayesian method to determine the sensitivity of binary systems to the linear spin 0 ULDM interaction constant \alpha (which characterises a direct universal interaction between binary pulsars and ULDM) was given in Kůs et al. 2024 , inspired by a similar method developed for gravitational wave (GW) signals in pulsar timing ( Moore et al. 2015 ) . The sensitivity is described by the curve \alpha(m) , above which the ULDM parameter space is excluded by signal non-detection. The Bayesian method significantly extends previous studies in that it applies to the whole mass range and not only when the ULDM and the binary system are in resonance, that is, when m=Q\omega_{b} where \omega_{b} is the orbital angular frequency of the binary system and Q is the integer resonance number. Despite the success of this Bayesian method, it requires handling and solving complex algebraic equations that make its generalisation to higher spins a challenging task. In order to surmount these difficulties, we look again at GW research, which also seeks to detect coherent signals in very noisy data. Recently, researchers have been exploring the use of machine learning (ML) tools for GW signal detection. These methods leverage deep neural networks (DNNs) with various architectures, such as convolutional neural networks (CNNs) and residual neural networks, see for instance Baltus et al. 2021 ; Baltus et al. 2022 and references therein. Another architecture that is naturally suited for coherent signal detection is the autoencoder, a specialised type of DNN that can be used as an anomaly detector ( Erdmann et al. 2021 ) . Inspired by advancements in the application of ML algorithms for GW searches, we aim to adapt these tools to address two key objectives. 1. Our primary goal is to determine the sensitivity of ML techniques to ULDM signals for the models discussed in Blas et al. 2017 ; Blas et al. 2020 ; López Nacir and Urban 2018 ; Armaleo et al. 2020 . This sensitivity sets limits beyond which signal non-detection excludes a given region of the ULDM parameter space. To achieve this, we employ two independent approaches: an autoencoder, which is an unsupervised ML technique, and a binary classifier, which instead is a supervised ML technique. In order to assess the power of the ML method, once we obtain the sensitivity across the full mass range of 10^{-23} eV to 10^{-18} eV for all ULDM spins, we compare them to the Bayesian sensitivity limits for the spin-0 case presented in Kůs et al. 2024 . 2. Our second goal is to develop a multiclass classifier capable of distinguishing between five types of inputs: pure noise and four different ULDM signal types embedded in the noise. We demonstrate that, given a sufficiently strong ULDM signal, this classification can be performed with high accuracy. Our aim is to provide a proof of concept that ML tools can be used to detect and classify ULDM signals of various nature and, in the absence of a detected signal, to impose constraints