Programmable and nonvolatile computing with composition tuning in thin film lithium niobate

TL;DR

This paper introduces a nonvolatile, programmable photonic computing approach using thin film lithium niobate, enabling energy-efficient matrix-vector multiplications through composition tuning and electrochemical lithiation.

Contribution

It demonstrates a novel nonvolatile tuning mechanism in TFLN for programmable photonic matrix operations, including design and validation of Mach-Zehnder interferometers and microring resonators.

Findings

Achieved matrix-vector multiplication with 1.5% average relative error.

Demonstrated composition-dependent refractive index control in TFLN.

Validated programmable photonic operations with realistic material loss constraints.

Abstract

Matrix-vector multiplications are fundamental operations in artificial intelligence and high-throughput computations, and are executed repeatedly during training and inference. Their high energy cost in electronic processors motivate scalable photonic computing approaches that reduce the energy required per operation. Thin film lithium niobate (TFLN) is a dominant photonic platform due to its large electro-optic effect. However, it lacks nonvolatile index tuning mechanisms, which promise to pave the way for energy-efficient photonic computing. Here, we explore electrochemical lithiation as a route to nonvolatile matrix-vector multiplications in TFLN. The LiNbO3 phase is stable at room temperature over a 2% Li composition window with an associated composition-dependent refractive index. We computationally demonstrate this as a programmable, low-loss approach to perform matrix-vector multiplications by using composition to control matrix weights. We design Mach-Zehnder interferometers to perform image processing tasks under realistic material loss constraints. We also design microring resonators for iterative weight updates, using gradient descent training to program target matrix operations with matrix-vector multiplication accuracy validated at 1.5% average relative error. These demonstrations show a facile route towards nonvolatile photonic computing in TFLN, addressing a critical requirement for energy-efficient photonic matrix operations at scale.