A world first: The NRC and McGill build a pure AI-based virtual fabrication model to accelerate silicon photonics production from design to market
As we generate and process more data, global demand is soaring for better, faster and cheaper telecommunications, data connections and sensing technologies. The pressure is on for researchers to quickly come up with innovative solutions to meet those needs. One driving force in this race is silicon photonics.
Integration and miniaturization in silicon photonics is one of the most promising technologies for carrying and processing large-capacity data at lower cost while consuming less power. This technology is also attractive because it is compatible with existing semiconductor infrastructure.
We're surrounded by photonics, the science and technology of light manipulation. It's in cellphones, CT scans, LED lights, driverless cars and optical communications that power the internet. It transports light through fibres to send broadband communications into homes and businesses in the farthest corners of the world at unprecedented speeds, all while using less energy.
The next generation of photonic integrated circuits must be even smaller, as communication systems grow exponentially. But shrinking these computer chips introduces new challenges to the manufacturing process, even in today's world of state-of-the-art semiconductor fabrication. During production, imperfections and fluctuations of even a few nanometres can affect optical performance. Perfecting the fabricated components requires significant trial and error.
Recently, the National Research Council of Canada ( NRC ) in collaboration with McGill University has devised an ingenious digital solution to reduce the impact of fabrication imperfections on component's optical performance.
"Our novel approach is a machine learning-based model to predict fabrication outcomes of design layouts," says the NRC 's Dan‑Xia Xu, Principal Research Officer at the Advanced Electronics and Photonics (AEP) Research Centre, adjunct professor at Carleton University, and a Fellow of the Royal Society of Canada. "Designers can conduct virtual trials on the model, ensure that clients'expectations are met and collect useful data before sending their designs to the manufacturer." She points out that the goal is "first time right" design and fabrication.
In addition to significantly reducing the design-to-fabrication cycle, the model cuts costs dramatically. "In the physical trial-and-error process, each iteration could run into tens of thousands of dollars. On an industrial scale this becomes hundreds of thousands—or millions," she adds.
Virtual fabrication model: predict and correct
Currently, photonics researchers and industrial designers are all faced with the process of trial-and-error in fabrication. Many system-makers that design and sell hardware do not fabricate the silicon wafers or chips used in their products. These "fabless" facilities do the designing but outsource the manufacturing to specialty foundries. Before the product can be finalized, manufacturers create "drafts" of components that will go back and forth between designers and manufacturers for months.
"This process is time-consuming, challenging and expensive," says Yuri Grinberg, Associate Research Officer, Digital Technologies (DT) at the NRC and adjunct professor at the University of Ottawa. "Companies that send out designs often work on several iterations with foundries before chips have been completed, so they don't know how the previous one has turned out before they have to start another."
The virtual fabrication model comes with benefits that provide competitive advantages to all. Prototyping foundries can demonstrate product feasibility before they start manufacturing chips or can sell the technology to customers. In addition, the model is easy to use and makes it possible for designers and researchers to predict how components will perform in the end. It also allows designers and researchers to correct their design layouts virtually, so circuits turn out as intended.
"We use a fairly standard AI tool that takes multiple images, learns the patterns and predicts the precise structure of a silicon photonic component once it's fabricated," adds Grinberg, pointing out that the innovation is in the concept. "This is the first experimentally validated model in the world of photonics that is purely based on AI without any knowledge of the complexity of the manufacturing process."
He adds that this ability to integrate AI methods into photonic design and fabrication gives Canada a unique technological advantage in a very competitive market.
Collaborations advance invention
Photonics brings together physics, electrical engineering and materials science knowledge about photons and electrons.
In this project, AEP 's knowledge of integrated photonics was critical in not only designing components for chips but also prototyping them. The expertise in photonic circuits and systems of McGill University's Professor Odile Liboiron-Ladouceur was harnessed to identify new optical functionalities that future circuits will require and to carry out system-level demonstrations. The talent and experience at DT helped researchers identify design problems that could benefit from AI and create models to address those. Dusan Gostimirovic, a postdoctoral fellow at McGill dedicated to the project, contributed his talent, enthusiasm and broad range of knowledge to driving the technology development.
"This is a good example of a successful collaboration where a multidisciplinary group works closely together," adds Grinberg. "We are now in the process of fine-tuning and patenting the technology."
The research is supported by the NRC 's High-throughput and Secure Networks Challenge program and Artificial Intelligence for Design Challenge program. The product should be ready for commercialization sometime in 2023.
Co-founders of PreFab AI Photonics, Professor Odile Liboiron-Ladouceur and Dr. Dusan Gostimirovic, were recently announced as the recipients of the 2022-2023 William and Rhea Seath Awards Competition and the 2022 TechAcceIR Grants. They have been recognized for their work developing ML-based corrector tools for photonics – based on joint IP with the NRC from this project.
Learn more about other collaborative R&D programs and initiatives that have been funded through the NRC .
For a more in depth look at the research behind the story, read the team's publication on Deep learning-based prediction of fabrication-process-induced structural variations in nanophotonic devices.
Contact us
Media Relations, National Research Council of Canada
1-855-282-1637 (toll-free in Canada only)
1-613-991-1431 (elsewhere in North America)
001-613-991-1431 (international)
media@nrc-cnrc.gc.ca
Follow us on X: @NRC_CNRC