Machine Learning Methods for Determining Nitrogen-Vacancy Center Orientations from ODMR Spectra of Single Crystal Nanoscale Diamonds

Abstract

This thesis investigates new methods to develop an existing idea \cite{dustpatent} that utilizes random spatial positions and orientations of Nitrogen Vacancy (NV) centers in single crystal diamond nanoparticles to create a Physical Unclonable Function (PUF). While the standard techniques in the literature estimate the actual orientation of the NV centers by making an initial estimate of orientation for the $\theta$ and $\phi$ positions of the NV axis with respect to a common frame (e.g. directional cosines) using the locations of local extrema of the spectra, this new method takes advantage of data science and machine learning techniques. Results using engineered data demonstrate a method that makes predictions approximately three orders of magnitude faster than the existing method, with an improvement in ROC plot area-under-curve (AUC) figure of merit ranging from a 4.5\% to a 9.5\% improvement, depending on the magnitude of the noise introduced into the model. Efforts to apply these new methods to real crystals have been met with minor successes, as results suggest a 9\% baseline (AUC) improvement in the detection of true positive results for crystal ensembles, though this has so far been accompanied with a high rate of false positives. In all cases, the new models represent a significant improvement with regards to required computational time.