Liang Feng (right) and Tianwei Wu demonstrate some of the apparatus used to develop the light-powered chip [Image: Sylvia Zhang/University of Pennsylvania]
Researchers at the University of Pennsylvania, USA, have developed what they believe to be the first photonic chip that can be programmed to deliver the nonlinear functionality that enables deep neural networks to tackle complex learning tasks (Nat. Photonics, doi: 10.1038/s41566-025-01660-x). The team showed that nonlinear networks configured on the device can outperform those with only linear connections, while the all-optical architecture processes information more quickly and consumes less energy than electronic chips.
A nonlinear response
Previous work has demonstrated optical chips that can be programmed to perform linear calculations, but in this new study the Penn team has engineered a reconfigurable platform that manipulates light to generate a nonlinear response. “This is a true proof-of-concept for a field-programmable photonic computer,” says team leader Liang Feng. “It’s a step toward a future where we can train AI at the speed of light.”
The nonlinear functionality is achieved by controlling the behavior of charge carriers within a light-sensitive semiconductor, in this case InGaAsP. When a signal beam carrying the input data passes through a thin layer of the material, a second pump beam is used to manipulate both the carrier dynamics and the spatial distribution of excited carriers within the film. Changing the shape and intensity of the pump beam alters the nonlinear interactions between the signal light and the semiconductor, allowing the signal beam to be reshaped into a wide range of polynomial functions.
When trained with a standard dataset describing 150 different iris flowers, a polynomial network ... could identify the specific species with an accuracy of almost 96%, compared with 86% for an equivalent linear network.
In particular, patterning the pump beam into areas of low and high intensity causes the signal power to oscillate along the propagation direction, which produces the high-order polynomials that are desirable for data-intensive applications. The form of the polynomial can also be reconfigured in real time by changing the beam pattern, while more complex nonlinear functions can be realized by combining multiple polynomials together.
Testing with iris flowers
To test their photonic processor, the researchers configured nonlinear optical networks that use these polynomial functions to connect multiple input and output nodes. When trained with a standard dataset describing 150 different iris flowers, a polynomial network composed of four input nodes and three outputs could identify the specific species with an accuracy of almost 96%, compared with 86% for an equivalent linear network. A larger network that was trained to recognize four distinct voice commands also identified the correct word in 92% of cases.
While this particular study focused on polynomial functions that are widely used in machine learning, the team believes that its approach could be extended to other nonlinear operations such as exponential or inverse functions. Combining the photonic processor with optical multiplexing technologies could also yield parallel processing architectures that would support more powerful learning models while also improving the energy efficiency.