Abstract

Photonic neural networks use integrated photonic circuits to perform neural network operations with high bandwidth and the potential for energy-efficient computation. In these systems, optical device responses can represent weights, and wavelength division multiplexing enables multiple wavelength channels to be processed in parallel within the same waveguide. However, scalable photonic neural network hardware needs weights that can be programmed, stored, and read reliably. Many current approaches require continuous electrical or thermal tuning to hold the optical state. This dissertation addresses this challenge by developing a hybrid photonic-electronic approach to non-volatile electro-optic memory. Here, we design and demonstrate a non-volatile electro-optic circuit, referred to as the NoVo-EOC memory cell. The cell combines a ferroelectric field-effect transistor (FeFET), a source-follower buffer, and a silicon PIN microring modulator. The FeFET stores the memory state electronically through a retained threshold-voltage shift after program and erase pulses. The stored FeFET state is then read as an electrical response, transferred through the source follower, and applied to the PIN microring modulator. The microring converts this electrical memory state into a change in optical transmission, allowing the stored electronic state to be read optically. We first characterize the individual devices needed to build the NoVo-EOC memory cell. The FeFET demonstrates distinct program and erase states through shifted drain current curves and a measurable memory window. The PIN microring modulator demonstrates passive resonance behavior, forward-bias electrical tuning through the PIN junction, and thermal tuning through integrated heaters. These measurements establish the two main functions required by the proposed cell. The first is non-volatile electronic storage, and the second is electrical-to-optical conversion. We then experimentally demonstrate the complete hybrid NoVo-EOC memory cell at the circuit level. The programmed FeFET state is read electrically, buffered through the source-follower stage, and converted into distinct optical spectral responses in the microring. We demonstrate fixed-wavelength optical hysteresis, program/erase readout repeatability over multiple write–read cycles, and a read-window selection method that identifies the wavelength and read voltage with the strongest optical state separation. By varying the write-pulse amplitude, we generate multiple optical transmission states and estimate the optical memory precision. The NoVo-EOC achieves up to six distinguishable optical levels, corresponding to two bits of optical memory precision. We also measure retention of the programmed state over seven days and show that the memory state remains readable both electrically and optically. Finally, we demonstrate the use of the NoVo-EOC in a wavelength-division-multiplexed photonic neural-network circuit. The fabricated circuit uses multiple wavelength channels to perform photonic linear-neuron operations and demonstrates how the stored NoVo-EOC states can be mapped to optical weight settings. We then use the retained memory state of the NoVo-EOC to directly set a photonic weight and demonstrate stored-weight operation within the circuit. Together, these results demonstrate a proof-of-concept approach toward WDM photonic neural-network hardware with non-volatile stored weights, where electronic memory can be programmed, retained, and used to control optical computation.

Publication Date

8-11-2026

Document Type

Dissertation

Student Type

Graduate

Degree Name

Microsystems Engineering (Ph.D.)

Department, Program, or Center

Microsystems Engineering

College

Kate Gleason College of Engineering

Advisor

Stefan Preble

Advisor/Committee Member

Karl Hirschman

Advisor/Committee Member

Tejasvi Das

Comments

This thesis has been embargoed. The full-text will be available on or around 2/13/2027.

Campus

RIT – Main Campus

Download

Share

COinS