03 May 2023
University of Pennsylvania approach promises accuracy and flexibility for AI applications
Lithography processes have traditionally been used to manufacture photonic chips, analogous to the operations used to create electron-guiding paths in conventional chips, but a project at the University of Pennsylvania (UPenn) has now developed an alternative route.
Reported in Nature Photonics, the project's approach maintains the programmable on-chip behavior without the use of lithography, potentially improving speed and flexibility for AI applications in particular, according to the project.
"Photonic chips intended for machine learning applications face the obstacles of an intricate fabrication process where lithographic patterning is fixed, limited in reprogrammability, subject to error or damage and expensive," said Liang Feng from UPenn.
"By removing the need for lithography, we are creating a new paradigm. Our chip overcomes those obstacles and offers improved accuracy and ultimate reconfigurability given the elimination of all kinds of constraints from predefined features."
The project's approach hinges on the way that spatial patterning of optical gain can be created on an unpatterned area of III-V semiconductor material with no predefined features using the "imaginary" attenuation component of the materials refractive index, as compared to normal lithography methods which rely on modulating the "real" conventional dimension of the same index property.
"An imaginary-index-driven methodology can tailor optical-gain distributions to execute prescribed optical responses, and configure desired functionality to route and switch optical signals," noted the project in its published paper.
Taking a vowel
"Conventional photonic chips are technologies based on passive material, meaning its material scatters light, bouncing it back and forth," explained UPenn's Marco Menarini. "Our material is active. A beam of pumping light modifies the material such that when the signal beam arrives, it can release energy and increase the amplitude of signals."
This active nature is the key aspect, used to reroute optical signals and program optical information processing on-chip, according to the project.
To test its approach with a data-intensive machine learning task, the project applied the chip to a data set of audio files containing four different vowel sounds. Distinctive frequencies of the vowels were extracted and encoded as input signals able to trigger a set of micro-ring lasers. Once trained, the ability of the chip to correctly "recognize" a vowel in a new audio file by outputting the correct optical response was measured at 94 percent.
Because patterns are not pre-defined or etched in, as they would be in lithography, the device is intrinsically free of defects and also reprogrammable, able to tailor its laser-cast patterns for optimal performance in tasks with small or large datasets, according to the UPenn project.
"What we have here is something incredibly simple,” said UPenn's Tianwei Wu. "We can build and use it very quickly, integrate it easily with classical electronics, and we can reprogram it, changing the laser patterns on the fly to achieve real-time reconfigurable computing for on-chip training of an AI network."