The growing demand for AI has pushed modern data centers toward unprecedented requirements in computing speed and energy efficiency. Photonic processors, which exploit the massive bandwidth and parallelism of light, have recently demonstrated remarkable capabilities in linear operations such as matrix multiplication. However, neural networks fundamentally rely on nonlinear activation to construct complex decision boundaries. In optical hardware, the lack of practical, high-performance nonlinear elements has therefore remained one of the most critical obstacles to building fully functional photonic neural networks.
In a new paper published in eLight, a team of scientists led by Professors Aaron Danner and Di Zhu from the National University of Singapore reported a passive, ultrafast, and chip-integrable solution to this challenge. By exploiting strong second-order nonlinear interactions inside periodically poled lithium-niobate (PPLN) nanowaveguides, the researchers realized an all-optical activation unit based on a waveguide-only design, eliminating the need for additional material epitaxy. Here, the nonlinearity is not driven by carriers, heaters, or additional control beams—it is generated directly by the data-carrying light itself. As the optical signal propagates through the PPLN nanowaveguide, the intrinsic χ² parametric interaction automatically reshapes the amplitude into a sigmoid-like response, much like how a biological neuron fires once a threshold is reached. The device delivers second-harmonic conversion efficiencies exceeding 80%, and because the effect stems from ultrafast electronic polarization inside the material, the response is essentially instantaneous, in principle supporting data rates beyond 100 GHz.
To show that the activation can work in a realistic computing environment, the team optically cascaded the PPLN chip with a programmable silicon Mach–Zehnder interferometer chip that performs linear transformations. Together, the modules form a complete optical neuron in which weighted summation and nonlinear activation are both executed directly in the photonic domain. Using this architecture, the optical neuron demonstrated competitive performance in tasks ranging from medical image classification to airfoil noise regression, highlighting a viable path toward ultrafast, scalable photonic intelligence. The team highlighted both the working principle and the broader significance of the approach:
“What makes this device distinctive is that the nonlinearity is triggered directly by the signal light itself. No electrical conversion, thermal tuning, or auxiliary control is required.”
“Because the response is governed by ultrafast electronic dynamics, the activation can naturally keep pace with modern photonic processors. This removes the long-standing mismatch between extremely fast linear optics and comparatively slow nonlinear elements.”
“With a passive and fabrication-compatible unit, we can envision cascading many layers on chip without complex routing or electronic overhead. This could accelerate the development of large-scale photonic neural systems for AI computing and ultrafast information processing.”
eLight
Passive all-optical nonlinear neuron activation via PPLN nanophotonic waveguides