AUTHOR=Ghosh Rajib Ratan , Liu Danqing , Reniers Sander , Jiao Yuqing , Yao Weiming 

TITLE=Reconfigurable optical synaptic weighting engine using a liquid crystal-based multimode interference coupler

JOURNAL=Frontiers in Nanotechnology

VOLUME=Volume 7 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/nanotechnology/articles/10.3389/fnano.2025.1593347

DOI=10.3389/fnano.2025.1593347

ISSN=2673-3013

ABSTRACT=Artificial neural networks (ANNs) have become ubiquitous in high-performance information processing. However, conventional electronic hardware, based on the sequential Von Neumann architecture, struggles to efficiently support ANN computations due to their inherently massive parallelism. Additionally, electrical parasitics further limit energy efficiency and processing speed, pushing traditional architectures toward their fundamental constraints. To overcome these limitations, researchers are exploring integrated photonics, leveraging the inherent parallelism of optical devices for more efficient computation. Despite these efforts, most existing optical computing schemes encounter scalability challenges, given that the number of optical elements typically grows quadratically with the computational matrix size. In this work, a compact programmable multimode interference (MMI) coupler on an indium phosphide membrane platform is proposed for realizing a photonic feedforward neural network. MMIs present a unique opportunity to accelerate matrix multiplication processes by exploiting the interference properties of light modes, promising advancements in both speed and energy efficiency. The programmable MMI coupler, comprising four input and three output (4 × 3 MMI) InP waveguides, makes use of hybrid integration of liquid crystals as cladding material, which offers reconfigurability to the MMI structure. Three electrically tunable sections are made to perform parallel multiplication operations. A novel modeling technique is introduced to facilitate effective training and inference operations. Finite-Difference Time-Domain (FDTD) simulations are employed for calculating the optical mode propagation process within the programmable MMI structure. Based on the FDTD results, a compact optical neural network is implemented and assessed on the Iris flower dataset, demonstrating a testing accuracy of 86.67%. This novel MMI device concept offers a promising pathway toward energy-efficient, scalable optical computing systems, contributing to the advancement of next-generation artificial intelligence hardware.