Engineers at the University of Wisconsin-Madison have designed a new type of quantum nanostructure that could enable optical neural networks. This emerging technology has the potential to make artificial intelligence systems, like large language models and image generation, faster and significantly more energy efficient.
The research, led by electrical and computer engineering PhD students Qingyi Zhou and Jungmin Kim, computer science PhD student Yutian Tao, and Zongfu Yu, a professor of electrical and computer engineering, was published in the journal Nature Communications on August 27, 2026.
Many of the most popular AI systems are based on deep neural networks, multilayer systems that mimic the interconnectedness of the human brain. As those systems scale in size and complexity, their energy consumption also increases. That’s one factor in recent concerns about AI energy use and datacenter construction.
This handwringing over AI energy consumption is nothing new; in fact, a decade ago, AI researchers were aware that the energy cost of scaling neural networks was not sustainable. That’s why they proposed something new: optical neural networks.
Instead of relying on traditional computer chips, optical systems use extremely tiny, fast lasers and photodetectors to process information. Theoretically, these systems are orders of magnitude faster and more efficient than the electronics-based GPUs that underpin current AI systems. In other words, they could do more work and use less energy.
But optical systems lack the one key attribute that makes a neural network more than just a fancy calculator. It’s a quality called nonlinearity, which allows AI systems to transcend traditional computing.
Optical systems are great at linear functions, where the output scales proportionally with the input. But, because photons—light particles that carry energy and information in optical systems—don’t like to interact with one another, there are few optical materials that can produce nonlinear functions. Such nonlinear operations allow a network to learn complex patterns rather than just rescaling its inputs.
While a few nonlinear optical materials do exist, the amount of laser energy needed to activate their nonlinear functions is so great that it negates the energy savings of the system.
Zhou, Yu and their colleagues decided to take a fresh look at optical nonlinearity to see if there was another way to approach the problem. “Basically, we asked ourselves, ‘Why is this nonlinearity so important? Why do we need strong nonlinearity in the first place?’” says Zhou. “And, ‘How should we overcome this bottleneck where we don’t have lots of nonlinearity?’ That’s basically the starting point of the entire project.”
The team calculated how much energy it would take to drive nonlinearity in an optical neural network using existing conventional optical materials. The researchers found that the intensity of the laser needed to make the system work was impossible to realize, far exceeding any realistic power budget.
So, if it was not possible to make optical materials nonlinear on their own, the team then began to wonder about combining them with other materials. “We asked, ‘What if we used some not-so-conventional materials, like quantum emitters, which already show very, very strong optical nonlinearity?’” says Zhou. “If you construct a neural network including these materials, hopefully it would produce nonlinearity. That was our intuition; these materials have already demonstrated strong nonlinearity. They just haven’t been used in the optical computing field.”
Using a suite of simulation and computational tools, the team designed a nanostructure surrounding a quantum emitter (technically, a type of structure called a vacancy color center), optimized to produce as much nonlinearity as possible. The detailed simulations showed that incorporating the devices into a full neural network would indeed lead to strong nonlinearity—creating a faster, more powerful system.
Additionally, the model showed the team’s design would reduce power consumption seven orders of magnitude below conventional optical materials. Zhou emphasizes that while the results are a theoretical estimation, the analysis already shows that nonlinearity itself should no longer be the bottleneck for optical computing.
While the current work remains computational, Zhou did scour the literature and analyzed whether it was feasible to build the optical neural network using current materials and techniques. He also says that other types of quantum emitters, like quantum dots or neutral atoms, may also work in the system and improve its performance. “With recent advances in diamond-based quantum photonics, I believe that our proposal is within reach of current technology,” he says.
Zongfu Yu is Grainger Professor in the Department of Electrical and Computer Engineering. Other UW-Madison authors include Guoming Huang and Zewei Shao. Other authors include Ming Zhou of Stanford University.
The authors acknowledge support from National Science Foundation (grant No. 2016136).
Top image caption: PhD student Qingyi Zhou used computational simulations to find a new way to bypass the nonlinearity bottleneck in optical neural networks. Photo: Joel Hallberg