At the National Institute of Standards and Technology (NIST), researchers proposed an approach to large-scale artificial intelligence that focuses on integrating photonic components with superconducting electronics as opposed to semiconducting electronics. The work has the potential to boost the function and scale of silicon-based AI chips and optoelectronic advancements as a result. To address problems with the fabrication of silicon chips with electronic and photonic circuits, often used to study cognition, the researchers considered the type of materials used in the process. “We argue that by operating at low temperature and using superconducting electronic circuits, single-photon detectors, and silicon light sources, we will open a path toward rich computational functionality and scalable fabrication,” said NIST researcher and author Jeffrey Shainline. A large neural system with multiple large modules, each containing hundreds to thousands of wafers, enabled by photonic communication and the efficiency of superconducting detectors and electronics. Not shown is the fiber optic white matter that would be woven through the voids between the octagons in this example of hierarchical tiling. Courtesy of Jeffrey Michael Shainline, NIST. The use of light for communication, in conjunction with complex electronic circuits for computation, could enable artificially cognitive systems of scale and function beyond what can be achieved with either light or electronics alone. “What surprised me most was that optoelectronic integration may be much easier when working at low temperatures and using superconductors than when working at room temperatures and using semiconductors,” Shainline said. Photon detectors based on superconductors are able to detect single photons. Those based on semiconductors require about 1000 photons to obtain a reading. Using superconductors rather than semiconductors, however, requires very low temperatures, near 4 K. At this temperature, Shainline noted in a published paper, silicon light sources become available. That indicates that a major impediment optoelectronic very large-scale integration may not be present in the superconducting domain. As computing becomes more distributed, communication becomes a bottleneck. A primary challenge, Shainline’s paper said, lies in the difficulty of achieving a light source on silicon that is robust, efficient, and economical. Operation at 4 K enables the use of single-photon detectors and silicon light sources, both of which increase efficiency and economical scalability. The researchers will explore increasingly complex integration with other superconducting electronic circuits. They also plan to demonstrate all the components that comprise artificial cognitive systems, including synapses and neurons. The ability to manufacture at scale to realize large systems, and at reasonable cost, is also an important step; superconducting optoelectronic integration could also help create scalable quantum technologies based on superconducting or photonic qubits. Such quantum-neural hybrid systems may also lead to new methods of leveraging the strengths of quantum entanglement with spiking neurons. The research was published in Applied Physics Letters (www.doi.org/10.1063/5.0040567).