RESEARCH PROJECTS

Two sensing systems, University of Virginia

Conducted at the Intelligence Sensing Group, advised by Prof. Kyusang Lee, Department of Electrical and Computer Engineering. Both projects were NSF-supported.

CMOS+X: 3D integration of CMOS spiking neurons with AlBN/GaN-based ferroelectric HEMTs toward an artificial somatosensory system

Dec 2024–Jan 2025 & Dec 2025–Jan 2026 · Hybrid on-site/remote

ABSTRACT

This project explores an energy-efficient artificial somatosensory system built by combining CMOS spiking-neuron circuits with ferroelectric AlBN/GaN HEMTs, converting sensory input into neural-like spike signals processed locally at the device level rather than in a distant processor.

BACKGROUND

Conventional CMOS sensing architectures process raw signals off-chip, which costs power and adds latency. Neuromorphic approaches that mimic biological spiking aim to reduce this cost, but require device platforms — such as ferroelectric HEMTs — capable of stable, repeatable switching behavior at the hardware level.

METHODS

Reviewed more than 20 peer-reviewed papers spanning neuromorphic computing, spiking neural networks, and ferroelectric device physics. Performed polarization–voltage (P-V), positive-up-negative-down (PUND), and current–voltage (I-V) characterization on fabricated FeHEMT devices, and organized short- and long-term plasticity (STP/LTP) and subthreshold-swing measurements for team-wide comparison.

RESULTS

Compiled endurance data showing reliable polarization switching to approximately 10⁷ cycles, and helped confirm plasticity behavior consistent with the requirements of neuromorphic operation.

SIGNIFICANCE

Device-level endurance and plasticity results of this kind are a precondition for any neuromorphic sensing system intended for long-term deployment, supporting further work on 3D-integrated CMOS-ferroelectric sensing architectures.

Integrating a federated split neural network with artificial stereoscopic compound eyes for optical flow sensing in 3D space with precision

Jul–Aug 2026 · 40 hrs/week · Hybrid on-site/remote

ABSTRACT

This project develops a stereoscopic artificial compound eye — an insect-inspired image sensor — capable of tracking 3D object motion with low power draw, by combining in-sensor computing with a federated split neural network (CSE-FSL) that divides computation between the sensor and a background processor.

BACKGROUND

Flat image sensors are constrained in field of view and reaction speed, and require transmitting full raw frames to a central processor. Compound eyes in insects avoid this bottleneck through a curved array of ommatidia that sense and process together — a structure this project attempts to replicate artificially.

METHODS

InGaAs photodiode films were grown via remote epitaxy on graphene and released by 2D layer transfer (2DLT), then paired with HfO2 ReRAM memory to form single "1P-1R" pixels, arrayed on a domed, microlensed substrate. Supported I-V characterization of the photodiode and ReRAM devices, ray-tracing and field-of-view (FOV) measurements on the microlens array, and processing of 16×16 optical-flow output grids recorded across two time points for motion analysis.

RESULTS

Confirmed reliable HRS/LRS switching in the ReRAM devices, measured a per-pixel field of view of approximately 8.5° — consistent with a single ommatidium — and verified that a 1P-1R pixel distinguishes four distinct time-states, (0,0)/(0,1)/(1,0)/(1,1), sufficient to compute motion direction directly in hardware.

SIGNIFICANCE

Hardware-level direction discrimination reduces the volume of data that must reach a background processor, directly supporting the communication and power efficiency goals of the CSE-FSL framework this sensor feeds into.