Optical Neural Engine brings together light computing and neural networks

Computing

Technological Innovation Website Editorial Team - 06/23/2025

As a wave encoded with a partial differential equation passes through the series of components, its properties gradually change and morph, until it finally represents the solution to that equation. [Image: Gao Lab/University of Utah]

Optical computing

Photonic computing , or computing with light instead of electricity, is no joke — light processors already do all the calculations needed for AI , for example, and they do so with 100% accuracy .

But it's all very new, and we don't yet know exactly which light platforms or photonic processors will dominate. One alternative is based on convolutional neural networks , written in multiple light-sensitive layers, that instantly solve sets of very complex equations, such as partial differential equations. And these are equations that are important for a number of practical problems, but they are very computationally intensive when using digital computers.

Yingheng Tang and colleagues at the University of Utah in the US have now created what they call an "optical neural engine", an architecture that combines diffractive optical neural networks and optical matrix multipliers.

Instead of representing the partial differential equations digitally, they are represented as variations in the optical properties of a material, a plate constructed using metamaterial techniques. Several of these plates are then placed in series, making up what the researchers call a metatronic network - each network is tailored to solve a specific equation.

The variables, in turn, are represented by the various properties of a light wave, such as its intensity and phase. As a wave passes through the series of optical components of the partial differential equation, these properties gradually change and modify, until they finally emerge on the other side of the structure representing the solution to the equation given at the input.

Use of optical device in solving Navier-Stokes and Maxwell equations. [Image: Yingheng Tang et al. - 10.1038/s41467-025-59847-3]

Solving equations with light

Machine learning and digital neural network techniques currently used to solve partial differential equations involve passing the equation through a network of computational nodes, each of which weights its output as it passes it to the next node. As the signal travels through the network, the correct solution becomes weighted more heavily and eventually becomes the output.

The difference now is that it's all done with the photonic devices, the different layers that deal differently with the light passing through them, which the team calls the ONE ( Optical Neural Engine ), doing what is essentially analog computation .

"ONE uses the spatiotemporal data of an input physical quantity, which is a function of positions and time, to predict the spatiotemporal data of an output physical quantity as a function of positions and time," explained Professor Weilu Gao.

The team demonstrated ONE's capabilities on several partial differential equations, including the Darcy flux equation, the magnetostatic Poisson equation in demagnetization, and the Navier-Stokes equation in incompressible fluid.

“The Darcy flow equation, for example, describes the flow of a fluid through a porous medium,” Gao explained. “Given data about the permeability and pressure fields within a given medium, the ONE architecture essentially learns the mapping between these qualities and can predict flow properties without having to perform experiments.”

The device is expected to have immediate applications in a range of areas, from basic research to engineering applications. "This research provides a versatile and powerful platform for large-scale scientific and engineering computing, such as geology and chip design," Gao said.

Other news about:

More topics