The XENONnT experiment is a dark matter detector that strives to perform the most sensitive search for particle dark matter to date. Functioning as a dual-phase time projection chamber, XENONnT relies upon unprecedentedly low background rates, which are achieved through a combination of hardware purification technology and software and analysis techniques. In particular, solar neutrinos pose an irreducible background source to XENONnT, as they interact with the xenon atoms and cannot be physically shielded or vetoed due to their weak interaction strength with matter. This background, however, also presents a unique opportunity for XENONnT to search for neutrino-induced physics beyond the standard model.

The sensitivity of XENONnT is limited by the degree to which background sources dominate signal production in the detector. Analyses aim to reduce background rates through event reconstruction and selection. Background reduction methods can introduce a number of undesirable consequences, including reducing the efficiency of identifying true dark matter or neutrino signals and introducing selection bias into the final data used in analysis. This thesis presents a novel analysis of the first science data obtained with the XENONnT dark matter experiment by employing a waveform-based Bayesian network to perform signal quality selection. This technique was applied to the search for new electronic-recoil interactions in XENONnT and, by extension, was used to quantify a limit on the value of the neutrino magnetic moment.

A Bayesian network was trained on simulated and observed experimental data for classifying scintillation and ionization detector signals. The network showed improved performance for signal classification over the current software method and a baseline neural network method. The Bayesian network outputs were then used in a continuous way to perform signal characterization, identifying those detector signals that were most alike to true physics interactions. This characterization reduced events outside of the analysis region of interest independently of the methods developed in the standard analysis of XENONnT. Compared to the traditional analysis method, this method of data selection improved XENONnT's efficiency of true signals by 3%. The analysis confirmed the background-only hypothesis of electronic recoil data from XENONnT's first science run. A test of a neutrino magnetic moment component in the data yielded a 90% confidence-level upper limit of 1.29 × 10^(−11)μ_B.

The Bayesian network method of signal characterization offers several advantages for the analysis of a dark matter detection experiment. First, the method provides a valid cross-check to an important experimental result. This improves the confidence of the finding reported by the XENONnT experiment and can be utilized in many applications to verify experimental results and check for biases in data analysis. Second, the method reduces potential systematic uncertainties due to the sequential development of highly optimized data selection criteria, where each introduces the possibility of a systematic effect in the analysis dimension(s). The traditional data analysis method relies on an approach of calculating efficiencies and selection criteria definitions that are degenerate (one cannot be used independently of the other). By contrast, in this work, the definition of the Bayesian network-based signal characterization boundary is independent of the calculation of efficiency applied in the analysis. Finally, the result of using fewer cuts optimized on efficiency results in an improved signal acceptance.

Bayesian networks are widely applicable to dark matter and neutrino physics experiments, where signal classification and characterization are central to sensitive measurements. In the future, analysis methods such as this can be used to perform valuable verification of experimental results and for the reduction of backgrounds in cases where waveform-based analysis is beneficial.