Researchers have demonstrated an optical neural computing architecture that uses quantum emitters embedded in inverse-designed nanophotonic structures, according to a peer-reviewed study published in Nature Communications on August 27, 2026. The approach addresses the rising power consumption of deep neural networks by overcoming the shortage of efficient optical nonlinearities in conventional materials.
Analysis published by the research team shows that the quantum activation functions at an intensity measured in nanowatts per square micrometer (nW/μm²). That operating level is seven orders of magnitude below the nonlinearity threshold of conventional optical materials. The system achieves this efficiency through the saturability of quantum emitters, which deliver significantly stronger nonlinear optical responses than traditional media.
The researchers used physics-aware training to numerically demonstrate that the architecture can perform complex tasks, including nonlinear classification and reinforcement learning, entirely within all-optical neural networks. To evaluate performance across hardware types, the authors also introduced a framework that quantitatively links physical nonlinearity to a network's expressive power.
When applied to large language models, the study's estimates indicate that the nonlinearity-limited optical power scales sublinearly with model size. The authors noted that these findings show quantum nanophotonics may provide a practical path toward sustainable artificial intelligence inference.
The study was conducted by researchers at the University of Wisconsin-Madison and Stanford University, with equal contributions from Qingyi Zhou, Jungmin Kim, and Yutian Tao. Funding for the work included support from the National Science Foundation under Grant No. 2016136 for the QLCI center Hybrid Quantum Architectures and Networks.
