Programmable Photonics Advanced with Lithium Niobate-based Waveguide
SUNNYVALE, Calif., Dec. 29, 2025 — A research team comprising of NTT, Cornell University, and Stanford University demonstrated a photonic processor whose refractive index can be rewritten across a two-dimensional waveguide to exercise virtually arbitrary control over the propagation of light waves and perform machine-learning computations. The processor is built around a lithium niobate slab waveguide and provides about 10,000 programmable spatial degrees of freedom. The researchers implemented all-optical neural-network inference on benchmark tests with up to 49-dimensional vectors in a single pass.
By enabling complex optical computations within a single, reprogrammable waveguide, the approach could help overcome growing energy and scaling limitations faced by processors used for AI. Optical neural networks are particularly well-suited for matrix-vector multiplications, which are a core operation in machine-learning workloads, and the ability to perform these operations in a single optical pass could significantly reduce latency and power consumption in future AI accelerators and data-center systems.
Large-scale photonic devices are fabricated by combining many discrete components, which must be isolated from each other, then connected by waveguides and individually controlled by many electrodes. That design creates overhead in the unused chip area and control systems. The waveguide, based on lithium niobate, is simple to fabricate and still permits more spatial complexity.
Unlike traditional photonic processors that rely on large numbers of discrete components such as phase shifters and directional couplers, the two-dimensional programmable waveguide puts functionality into a single slab. This reduces unused chip area and eliminates the need for wiring and electrode networks, while still supporting a high degree of spatial complexity and reconfigurability.
At fabrication, the device is a uniform slab and is programmed by projecting illumination patterns that act as virtual electrodes, locally controlling the refractive index in the chip without any physical wiring. Bright regions of the pattern create a strong bias voltage across the waveguide and a larger electro-optic response, while dark regions leave the bias voltage unchanged. This arrangement lets a single chip realize distinct and highly complex linear optical transformations by changing only the illumination pattern.
“The level of control offered by this new waveguide means that we can make linear optical devices of unprecedented spatial complexity,” said Martin Stein, a postdoctoral fellow at NTT and a co-lead author of the study. “The device is a first of its kind, allowing us to essentially paint any optical circuitry we want and then redraw it in the blink of a second.”
The team used the chip to perform neural-network inference by training the refractive-index distribution and consequently the multimode wave propagation through the chip. They developed a physics-based model of the chip’s behavior, along with a data-driven refinement, allowing the model to be sufficiently accurate to support a backpropagation algorithm. In the tests, the chip performed vowel classification with 96% test accuracy and handwritten-digit recognition with 86% test accuracy, each in a single optical pass.
The demonstration significantly increases the number of controllable optical modes compared with prior integrated photonic systems; however, the researchers note that further scaling will be required for optical computing to compete directly with state-of-the-art electronic processors. Future work will focus on extending the approach to control hundreds or thousands of optical modes, a key threshold for achieving energy-efficiency advantages in large-scale machine-learning applications.
This research was published in Nature Physics (www.doi.org/10.1038/s41567-025-03094-2).
/Buyers-Guide/NTT-Advanced-Technology-Corp/c31803