Cavity optomechanics is a rapidly expanding field offering innovative approaches for numerous technologies utilizing sub-wavelength light-matter interactions. Optomechanical detection is based on electromagnetic phenomena and ultralow-noise laser readout, rather than traditional techniques such as capacitive or piezoelectric sensing which have higher Johnson electrical noise, consequently achieving sizable higher sensitivities, even at the quantum backaction limits. This high-sensitivity and precision is pertinent in near-term frontier inertial measurement units, which use gyroscopes and accelerometers to measure position and velocity absent external signals acting as references. Inertial measurement is vital in applications from vehicle navigation to motion sensing in smartphones and wearable devices. The operation of the designed optomechanical accelerometer relies on a slotted photonic crystal, which localizes light at specific frequencies determined by its photonic bandgap. Under acceleration, the slot dimensions shift, causing a corresponding change in the resonant frequency band of trapped light. These shifts show up as measurable changes in the physical oscillation frequency. In this project, I focus on integrating the components of the inertial measurement unit, by modeling suitable housing to be used for navigation in the field. This work compared our optomechanical accelerometer measurement data against leading-edge electronic counterparts. With considerations critical to performance, such as spacing, optical fiber positioning, and heat-weight distributions of required components, I have designed an arrangement with the necessary power supplies, circuits and measurement devices for a near-term flight mission. This assembly transitions from laboratory-scale testing to assessments in the field, effectively increasing the technological readiness level of the system being developed.

This research delves into device fingerprinting using CPU side-channel emanations, capturing the innate information leakage of these channels. Through tools like DeMiCPU, which collects magnetic induction signals from CPUs, we have unlocked a method for deriving distinct device signatures, critical for enhanced security in both software and applications. Central to our approach is a deep learning-based classifier that capitalizes on side-channel data from microcontroller CPUs. Raw IQ data is gathered with an EM probe, and, after undergoing FFT for frequency domain analysis and normalization, is ready for model training. We implemented both a Convolutional Neural Network (CNN) for image-based spectrogram data and a Long Short-Term Memory (LSTM) model for sequence-based IQ data. Our methodology involved recording EM signals from four Arduino devices at several different time instances, resulting in an extensive FFT dataset. Though the LSTM model initially performed well in identifying trained devices, it misidentified an untrained device as a trained device with high confidence, hinting at overfitting. However, by revising the normalization function and introducing an improved validation prediction rule, we achieved 100% accuracy in identifying both trained and untrained devices. In conclusion, this project explores an accurate method of utilizing the CPU's EM emanations for device fingerprinting. By merging FFT-based frequency domain analysis with refined deep learning techniques, we offer a method that suggests potential advancements in device identification, contributing to the field of cybersecurity.

Patients often suffer from multiple co-morbid conditions and may simultaneously take multiple drugs, heightening the risk of side effects. Despite significant patterns in co-morbidity and polypharmacy, patient conditions are treated individually instead of concurrently. An increasingly popular approach for identifying co-occurring diseases involves analyzing the genetic networks linked with these conditions, categorizing them according to shared gene associations. We aimed to understand comorbid diseases using our algorithm, PathFX, which predicts drug-induced phenotypes based on drug binding and downstream proteins at both genetic and network levels. Leveraging the data derived from PathFX for 4264 diseases, we employed the k-means technique to group diseases based on common drug and gene associations. This process created 87 clusters based on shared drug associations and another 87 clusters based on shared gene associations. The diseases within clusters show moderate overlap overall; within drug-based clusters, the average overlap coefficient among diseases' association data was 0.578, whereas within gene-based clusters, this figure was slightly lower at 0.461. This implies that diseases have a greater overlap of proteins within drug networks compared to the overlap of proteins among the diseases themselves. Furthermore, we conducted a comparison between drug and gene-based clusters, finding that the average highest overlap coefficient associated with each cluster was 0.496. This moderate degree of overlap suggests that drug networks associate to diseases through similar network proteins. Future work could focus on analyzing shared diseases in the electronic health record, which would lead to a better understanding of how drugs might predictively treat co-morbid conditions.

LiDAR sensors are heavily adopted by both researchers and industry professionals for addressing localization challenges in autonomous systems. Localization for LiDAR requires aligning two LiDAR measurements obtained from different perspectives, occurring when the robot is in different positions and determining its movement. The alignment of the two scans or point clouds is referred to as the scan-matching or point-cloud registration problem. Existing algorithms for this problem are predominantly heuristic and local in nature, leading to inaccuracies when the initial alignment is suboptimal. Provably Accurate Simple Transformation Alignment (PASTA), designed by the UCLA CyPhy Lab, aims to provide formal error guarantees on localization error without relying on point-to-point correspondences as existing algorithms do. As most scan-matching algorithms utilize 3D point cloud data, we aim to determine the best-performing algorithms by comparing rotation and translation estimations with artificial 3D data generated from existing 2D LiDAR scans. The nominal performance of several variations of such algorithms, including Iterative Closest Point (ICP), Fast Point Feature Histograms (FPFH), and PASTA, on 3D data was successfully validated through their estimated solutions to the manually applied rotation on the point clouds. Validation with ground-truth measurements for the alignment of LiDAR point clouds is in progress to determine the best algorithms for certain environments, in addition to attaching PASTA’s guarantees to supervise alternative heuristic methods. Future work includes supervising machine learning point-registration with PASTA and integration into simultaneous localization and mapping (SLAM) frameworks for mapping of unknown environments and obstacle avoidance.