Researchers test photonic AI diagnostics

Researchers behind two recent photonics papers are testing whether light-based hardware can handle parts of AI workloads now run on conventional chips. A University of Pennsylvania group reported an ultra-low-energy optical switch in May, while a Shenzhen University-led team described a prototype photonic neural network for medical-image diagnosis in a paper indexed in recent weeks. Both results were amplified in science posts on X on July 8 and July 9, but the underlying work comes from a peer-reviewed physics paper, a university release and an optics journal article. ### Which research papers are people actually talking about? Bo Zhen’s group at the University of Pennsylvania published a Physical Review Letters paper in 2026 on “strongly nonlinear nanocavity exciton-polaritons in gate-tunable monolayer semiconductors,” according to the lab’s publication list and the journal’s accepted-paper page. Penn said the work used exciton-polaritons — hybrid light-matter quasiparticles — to perform all-light switching at about 4 quadrillionths of a joule, or roughly 4 femtojoules. (penntoday.upenn.edu) Han Zhang and co-authors at Shenzhen University reported a separate result in Optical and Electronic Advances. The article describes a phosphorene-based electro-optic modulator and a prototype all-fiber photonic neural network built on a time-division multiplexing architecture. ### What did the Penn group say it demonstrated? The Penn team said the main technical hurdle is that photons are efficient for carrying information but typically interact too weakly to do computing tasks such as switching. (web.sas.upenn.edu) Penn’s release said the researchers used a nanoscale cavity and an atomically thin semiconductor to create exciton-polaritons that interact strongly enough to switch optical signals. (oejournal.org) Physical Review Letters said the device enabled all-optical switching of the cavity spectrum with excitation energies as low as about 4 fJ and on picosecond timescales. Penn said that, if scaled, the approach could help photonic chips process light directly from cameras and reduce the power demands of large AI systems. ### What did the Shenzhen medical-diagnostics paper actually do? (penntoday.upenn.edu) The Optical and Electronic Advances paper said the prototype photonic neural network was tested on handwritten digit classification, retinal B-scan analysis and multiphase liver CT image diagnosis. The authors said the system used a black phosphorus-based van der Waals heterostructure with a microring or microfiber-knot-resonator design to achieve low-power modulation. (journals.aps.org) EurekAlert’s summary of the same work said the platform was developed by Shenzhen University researchers with industry partners and was aimed at low-power, high-speed optical computing for medical diagnosis. The paper itself described the result as a foundation for more highly integrated, low-power, fiber-based photonic neural networks. ### Are these clinical products or lab prototypes? (oejournal.org) The Shenzhen paper describes experimental validation on diagnostic tasks, not a hospital-deployed product. The Penn work is a physics demonstration of an optical switching mechanism, not a finished AI processor. Nature and other recent photonics papers show the broader field is also working on programmable 3D photonic neural network chips and photonic mixture-of-experts systems, but those are separate efforts from the two results highlighted in the social-media posts. (sciencesources.eurekalert.org) ### Why are these two results being grouped together? The July 8 X post grouped them because both address the same hardware problem from different angles: one tries to lower the energy cost of optical switching, and the other tries to run inference-style image tasks on photonic hardware. (oejournal.org) That pairing is consistent with the source papers, which both frame photonics as a way to ease energy and data-movement bottlenecks in AI workloads. (nature.com) Physical Review Letters and Optical and Electronic Advances are the next places to watch for follow-up citations, replication and device-scale updates. Penn’s paper is already listed in the Zhen lab’s 2026 publications, and the Shenzhen article includes supplementary materials tied to the diagnostic prototype. (web.sas.upenn.edu) (oejournal.org)