Deep Neural Networks Hunting Ultra-Light Dark Matter

TL;DR

This paper explores the use of machine learning neural networks to detect ultra-light dark matter signals in pulsar-timing data, demonstrating comparable sensitivity to traditional Bayesian methods across multiple ULDM models.

Contribution

It introduces neural network-based methods for ULDM detection, applicable to various models, and shows they can distinguish between different ULDM types with simulated data.

Findings

ML methods achieve sensitivity comparable to Bayesian approaches

Neural networks can distinguish between different ULDM models

The approach is applicable to multiple ULDM interaction types

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.