How can artificial intelligence help in the building of smaller, faster, and more energy-efficient photonic devices? One such way is through inverse design. 

In traditional forward design, you start with a structure and through understanding and applying the maths, be they Maxwell’s equations or similar, you tune certain parameters until the design goal is met. Inverse design, not surprisingly, takes the opposite approach, with the desired optical response being the starting point and algorithms altering the structure until the goal is met. As an article from the Photonics and Nanostructures – Fundamentals and Applications journal puts it, whereas forward design puts structure first, inverse design generally puts the performance first. 

For Dr. Jobayer Hossain, assistant professor at the University of North Dakota’s school of electrical engineering and computer science, understanding how tools such as AI can support inverse design is one of his key research areas.  

Hossain co-authored no fewer than three papers for the IEEE Photonics Conference back in 2023. One explored inverse design of silicon photonics components from a deep learning perspective. While his work has mostly been at the device level, he notes there is potential to extend the same approach for circuit level design, as well as some aspects of system level work. 

“Forward design is more human-centric,” Hossain explains. “Every time we have a new device, we basically go through the same process. We keep tuning the device structure and parameters until we reach the target characteristics. 

“For inverse design, sometimes we can have non-intuitive design coming directly from the machine learning and deep learning,” he adds. “That can give us a very small footprint, very low-power devices, which was not possible using the forward design method.” 

Yet for all the promise, there is a caveat. The device still needs to be manufacturable. “It needs to be process variation tolerant,” says Hossain. “From this inverse design approach, it can have very sharp corners, like one nanometre or two nanometre feature size, which may not be possible to fabricate repeatedly. So we need to ensure that the device we fabricate… gives us this consistent result.”  

Following stints at AIM Photonics and Phase Sensitive Innovations, Hossain made the move back to academia. He explains that he appreciates the ‘space for free thinking and innovation’ it provides. “[In academic research,] we have the luxury of taking more risk,” he notes. “Certain ideas may fail, and we can take that, but in industry, if that happens, you are out of business. In academia, our failure, our success… everything is learning.” 

One such learning was in a recent paper Hossain and his team published on low-loss, low-crosstalk arrayed waveguide grating (AWG) on a 300mm silicon photonics platform. 

AWGs, Hossain notes, are rather like a prism — just as a prism separates white light into its constituent colors, an AWG separates a combined optical signal into its individual wavelengths. Their integration with semiconductor optical amplifiers, detectors, modulators and switches help create complex photonic integrated circuits (PICs), with applications across LiDAR, to spectral imaging, to neuromorphic computing and AI accelerators. “It’s, of course, one component or one device of a whole system, but it’s a crucial device,” notes Hossain. 

Hossain is keen to avoid the valley of death, the key gap between scientific discovery and practical application, for his cohort of students and researchers. The funding agencies, he notes, emphasise bridging that gap – and this points to another reason why he moved back to academia. 

“I want to use my experience in industry, and bring those insights, and would like to use those to train the next generation of engineers,” says Hossain. “I enjoy following curiosity, and I would like to solve problems, but I want to pursue real-world problems with [an] academic perspective. My mindset is [on] real-world problem solution.” 

Hossain is participating at Microelectronics US on April 22-23 where he will be the one asking the questions, moderating a panel session on the future of the data centre with photonics technologies. So, how does he assess the proposition? 

“Photonics is already transforming the data centres because we had certain limitation from the electrical interconnects,” says Hossain. “Photonic interconnect comes inherently with huge bandwidth, low power consumption, so it is already being used. There are certain bottlenecks, certain challenges, even in photonics. Co-packaged optics… has certain challenges, like how do we couple light [between components] with low loss? How do we pack more and more components within a small area? How do we do thermal isolation? 

“I will be asking all of these questions to the panellists and try to get answers for the audience.”