Spiking neural networks (SNNs) promise energy-efficient computing by mimicking biological spike-based processing, but hardware constraints like limited quantization in neuromorphic chips hinder deployment. Here, we characterize quantized SNNs inspired by magnetic tunnel junctions (MTJs), discretizing weights and activations to 2–18 levels using straight-through estimators, on the neuromorphic DVS128 Gesture dataset (1,176 recordings, 11 classes, preprocessed into 4–16 Poisson-encoded frames) and static CIFAR10. Employing a DVSGestureNet with Leaky Integrate-and-Fire neurons, trained via MSE(mean squared error) loss, Adam optimization, and cosine annealing over 100 epochs on UCLA’s own Hoffman GPU cluster, we find that 14-bit quantization yields >95% accuracy on DVS128 by epoch 100, with performance degrading below that threshold due to noise disrupting temporal dynamics. CIFAR10 plateaus at 10 bits, reflecting the simpler non-temporal nature of this dataset. These results highlight quantization thresholds for temporal vs. static tasks, advocating domain-wall MTJs due to their ability to achieve high quantization levels that enhance precision and reduce noise. Addressing the scalability of neuromorphic hardware could enable the implementation of large-scale spiking algorithms, thereby advancing neuromorphic computing.

Low Earth Orbit (LEO) satellite constellations are transforming global high-speed internet access, yet performance characteristics vary widely between networks. OneWeb, in particular, remains underexplored, with limited prior public measurement work. We present the first outside-in measurement campaign of OneWeb across multiple geographic locations, identifying publicly exposed terminals via Censys scans and measuring them from globally distributed probes and a Google Cloud VM using ICMP Paris-Traceroute. Across hundreds of traceroutes, we characterize OneWeb’s routing paths, orbital patterns, and satellite handover behavior, and compare results to Starlink in similar locations. Our Netherlands-based measurements show Starlink achieving consistently lower latencies than OneWeb. While most of this difference is expected from Starlink’s lower orbital altitude, the remaining gap, after accounting for propagation delay, likely reflects unique network conditions. We also observe a repeating orbital latency pattern in OneWeb, suggesting satellite and SNP selections remain consistent across successive orbits. Further analysis shows that OneWeb’s average satellite connection durations are substantially shorter than the time satellites remain in view, indicating frequent handovers. These results reveal key architectural distinctions between OneWeb and Starlink and demonstrate the value of multi-location, multi-orbit observations for capturing LEO network behavior. By combining scalable internet scanning with targeted traceroute probing, our approach offers a framework for evaluating emerging satellite networks.

This work presents a framework for enhancing read margin and recall reliability in nvSRAM through per-transistor threshold-voltage (Vt) mismatch compensation using Charge Trapping Transistors (CTTs). The method determines each transistor’s contribution to read static noise margin (RSNM) degradation and maps it to a corresponding Vt deviation. CTTs, with post-fabrication Vt programmability enabled by charge trapping in high-k dielectrics, can then be tuned to compensate for these mismatches. However, quantifying how PFET programming affects RSNM and improves non-volatile recall accuracy remains a key challenge. RSNM—defined as the maximum tolerable DC noise at an internal storage node without altering stored data—is extracted from Cadence Virtuoso simulations by sweeping injected Vt perturbations. Equal-RSNM contours are reconstructed via a Python pipeline, and under a zero-mean Gaussian assumption, the minimal-radius tangent from the origin identifies the maximum-likelihood mismatch vector, decomposed into individual ΔVt values. Compensation is applied by introducing a Vt shift in the pull-up PFET acting as a CTT, and ΔRSNM is evaluated for logic “1” and “0” states. Linear interpolation determines the PFET Vt shift that minimizes ΔRSNM; the resulting (mismatch, PFET shift) pairs are fit using monotone piecewise-cubic interpolation to create a continuous runtime calibration map. Results show PFET Vt compensation can reduce ΔRSNM to 0.03% within a 2-σ range of the ideal. Ongoing work will further examine PFET programming’s impact on recall accuracy and extend the method to full 6T mismatch analysis, with implications for SRAM Vmin scaling at the 22-nm SOI node and robust non-volatile memory design. 

Kelp forests are important coastal ecosystems that store carbon, provide habitat, and protect shorelines. In California, giant kelp (Macrocystis pyrifera) populations have declined over the past decade, and understanding how spores move and settle is important for restoration planning. This project examines the physical transport mechanisms of water movement inside and around a kelp forest and how they impact residence time and transport pathways of spores, represented by a dye tracer.  Three Rhodamine-WT dye releases, each diluted to  ~21% concentration and deployed 1 meter above the bottom, were conducted inside the Arroyo Quemado giant kelp forest in Santa Barbara, California. To investigate canopy-to-open water gradients inside and outside the kelp forest, an RBR Concerto CTD-fluorometer was used for vertical casts to measure pressure, temperature, salinity, and dye concentrations.   The results from these casts show that dye residence times were notably longer within the kelp forest in comparison to adjacent open waters, as well as a prominent retention zone along the southern edge of the forest. These findings indicate that localized flow patterns and canopy structure can create spatially heterogeneous retention zones, potentially enhancing spore settlement. This work provides insight into kelp forest transport dynamics and can inform site selection and design for future restoration projects.     

As nickel demand rapidly increases due to its integral role in the development of electric vehicles and steel production, conventional nickel reserves are depleting on a global scale. Laterite ores are a promising alternative source; however, the leachates they produce are concentrated in magnesium (II) ions and deficient in nickel (II) ions, making the separation of ions with similar size and charge both difficult and rarely addressed in previous literature. In this project, we developed a modified anion-exchange membrane with embedded Chevrel particles, which preferentially bind with transition metals over alkaline earth metals to enable nickel-selective transport through membrane intercalation. The modified membrane's selectivity was evaluated within an electrodialysis (ED) system positioned between a cation- and an anion-exchange membrane under an applied electric field. Key performance parameters, including three different feed compositions which are nickel-rich, equimolar, and magnesium-rich; limiting current densities from 0.1 to 0.5 A/m^2; and acidic and neutral pH conditions, were tested to identify optimal operating conditions for selective nickel transport. It was found that operating the ED system at a current density of 0.2 A/m^2 and using a nickel-enriched feed solution yielded maximum nickel transport, while pH had little influence. These findings provide a foundation for a more targeted and efficient industrial-scale electrodialysis separation of nickel and magnesium from laterite ore leachates, while introducing Chevrel particles as a new ion-selective medium that enables future work in the selective recovery of other critical metals from industrial streams.