Luis Saldana successfully defends thesis: “Novel Signal Reconstruction Techniques in Cyclotron Radiation Emission Spectroscopy for Neutrino Mass Measurement”

August 9, 2021

On August 6, 2021 Luis Saldana successfully defended the thesis: “Novel Signal Reconstruction Techniques in Cyclotron Radiation Emission Spectroscopy for Neutrino Mass Measurement”. (Advisor: Karsten Heeger)

Saldana explained, “The Project 8 experiment aims to measure the absolute neutrino mass scale and/or determine the neutrino mass hierarchy using a variation of the tritium endpoint method via Cyclotron Radiation Emissions Spectroscopy (CRES). Project 8 has developed the CRES technique over the last six years, which involves magnetically trapping tritium beta-decay electrons to faithfully sample the emitted cyclotron radiation as the electron performs cyclotron motion around the field lines. Since these electrons are mildly-relativistic, the frequency of emission is inversely proportional to the kinetic energy of the emitted electron, thus, an energy spectrum may be reconstructed by first reconstructing the signal in frequency-time space and converting it to energy using a calibrated value of the average magnetic field. For my thesis work, I developed machine-learning (ML) methods which help us reconstruct the signal (amongst the overwhelming RF background) and classify its topology (to discriminate it from Doppler-shifted signals) with great accuracy and robustness against variance. Project 8 plans to reach a sensitivity to the effective electron neutrino mass m_β of 40 meV/c², pushing the current sensitivity past the inverted mass ordering allowed by oscillation experiments. In order to reach this limit, Project 8 will have to move the waveguide-confined experiment to a very large volume O(10² m) to capture enough statistics to be sensitive to the tiny effect of a massive neutrino at the β-endpoint. Efficiency of signal reconstruction is then paramount in reaching such a limit, which will dominate the uncertainty on the extracted m²_β for the next few years. The ML methods I have developed for this task involve the use of Support Vector Machines and Convolutional Neural Networks (CNN) alongside the appropriate simulations required to train these algorithms.

For the next several months, Saldana will continue to work at Yale with Prof. Heeger and the Project 8 collaboration in order to prepare a paper on the CNN model he has developed for signal reconstruction. He will also be supervising and working with students Chris Xu and Harper Cho on developing these methods even further for related Project 8 applications.

Thesis Abstract: The Project 8 experiment is developing Cyclotron Radiation Emission Spectroscopy (CRES) on the β-decay spectrum of tritium for the measurement of the absolute neutrino mass scale. CRES is a frequency-based technique which aims to probe the endpoint of the tritium energy spectrum with a final target sensitivity of 0.04 eV, pushing the limits beyond the inverted mass hierarchy. A phased-approach experiment, both Phase I and Phase II efforts use a combination of 83mKr and molecular tritium T2 as source gases. The technique relies on an accurate, precise, and well-understood reconstructed β-spectrum whose endpoint and spectral shape near the endpoint may be constrained by a kinematical model which uses the neutrino mass mβ as a free parameter. Since the decays in the last eV of the tritium spectrum encompass O(10-13) of all decays and the precise variation of the spectrum, distorted by the presence of a massive neutrino, is fundamental to the measurement, reconstruction techniques which yield accurate measurements of the frequency (and therefore energy) of the signal and correctly classify signal from background are necessary. In this work, we discuss the open-problem of the absolute neutrino mass scale, the fundamentals of measurements tailored to resolve this, the underpinning and details of the CRES technology, and the measurement of the first-ever CRES tritium β-spectrum. Finally, we focus on novel reconstruction techniques at both the signal and event levels using machine learning algorithms that allow us to adapt our technique to the complex dynamics of the electron inside our detector. We will show that such methods can separate true events from backgrounds at > 94% accuracy and are able to improve the efficiency of reconstruction when compared to traditional reconstruction methods by > 23%.