Searching for Bosonic Dark Matter with Nuclear Magnetic Resonance

Abstract

The nature of dark matter, the invisible substance making up over 80% of the matter in the Universe, is one of the most fundamental mysteries of modern physics. Ultralight bosons such as axions, axion-like particles or dark photons could make up most of the dark matter. Couplings between such bosons and nuclear spins may enable their direct detection via nuclear magnetic resonance (NMR) spectroscopy: as nuclear spins move through the galactic dark-matter halo, they couple to dark-matter and behave as if they were in an oscillating magnetic field, generating a dark-matter-driven NMR signal. In the first chapter of this thesis, we review the predicted couplings of axions and axion-like particles with baryonic matter that enable their detection via NMR. We then describe two measurement schemes being implemented in the Cosmic Axion Spin Precession Experiment (CASPEr), an NMR experiment seeking to detect axion and axion-like particles. The first method, presented in the original CASPEr proposal, consists of a resonant search via continuous-wave NMR spectroscopy. This method offers the highest sensitivity for frequencies ranging from a few Hz to hundreds of MHz, corresponding to masses ma ∼ 10−14–10−6 eV. However, Sub-Hz frequencies are typically difficult to probe with NMR due to the diminishing sensitivity of magnetometers in this region. To circumvent this limitation, we suggest new detection and data processing modalities: a non-resonant frequencymodulation detection scheme, enabling searches from mHz to Hz frequencies (ma ∼ 10−17–10−14 eV).As a second part of this thesis, we apply the above mentioned non-resonant method and use ultralow-field NMR to probe the axionfermion “wind” coupling and dark-photon couplings to nuclear spins. No dark matter signal was detected above background, establishing new experimental bounds for dark-matter bosons with masses ranging from 1.8 × 10−16 to 7.8 × 10−14 eV.In the last chapter of this thesis, we use Deep Neural Networks (DNNs) to disentangle components of oscillating time series, arguably the most common form of signals acquired during dark-matter searches. We show that the regression and denoising performance is similar to those of leastsquare curve fittings (LS-fit). We then explore various applications in which we believe our architecture could prove useful for time-series processing, when prior knowledge is incomplete. Because the Autoencoder needs no prior information about the physical model, the remaining unknown latent parameters can still be captured, thus making use of partial prior knowledge, while leaving space for data exploration and discoveries.