NTT Research Physis & Informatics (PHI) Lab scientists have completed path-breaking work in the fields of optical amplification and in-memory computing. The results in the first case are discussed in “Efficient parametric down-conversion by gain-trapped solitons,” a paper published in the online journal Optica (Vol. 11, No. 3, March 2024). A second paper, “Hyperspectral In-Memory Computing with Optical Frequency Combs and Programmable Optical Memories,” was presented at the OFC Technical Conference in San Diego, March 24-28, 2024. Another PHI Lab scientist is participated in a panel at OFC, and three NTT researchers also spoke at this premier global event for optical communications and networking professionals.
The paper in Optica, co-authored by 11 scientists representing Stanford University, the University of Neuchâtel and the PHI Lab, demonstrates a new, extremely efficient way to generate broadband coherent light at long wavelengths, which could prove useful in spectroscopy and sensing, enabling further advances in medical diagnostics, drug discovery and environmental monitoring. The most common approach to generating this kind of light is to develop new laser gain media. However, for every new range of colors (or wavelengths), the development time from host medium to mature laser can be a lengthy process. An alternative is to use optical nonlinearities. In a process known as parametric down-conversion (PDC), or optical parametric amplification (OPA), photons from a short-wavelength (high-frequency) pump laser break apart into pairs of photons with less energy, referred to as the signal and idler waves. In other words, the pump amplifies an input wave (either signal or idler) and, in the process, generates all of the desired wavelengths.
“Lower energy photons have longer wavelengths, so we can use this process to convert nice short-wavelength lasers that exist today to essentially any long wavelengths we want,” PHI Lab Scientist Marc Jankowski and co-lead of the project said. A big problem, however, is overhead. These techniques have typically required large, complex and energy-intensive laser systems, which in effect have limited their use to the optical lab.
A new material called thin-film lithium niobate (TFLN) is enabling the latest approaches. PHI Lab Scientist and paper co-author Tim McKenna is leading the PHI Lab’s TFLN team, which is working to scale these technologies beyond the lab. (He spoke about TFLN on Day 2 of Upgrade 2023.) Recent progress in TFLN has enabled the development of degenerate OPAs (which are characterized by identical signal and idler waves) that operate with ultra-low pump powers and realize long interaction lengths. Waves interact as they propagate down a waveguide, and longer interactions lead to large conversion efficiencies (and amplifier gains) with low energy requirements. The development of TFLN photonics enables new design strategies that avert deleterious effects. In particular, by engineering the geometry of a TFLN waveguide, limitations such as temporal walk-off (loss of overlap) between the pump and signal pulses can be eliminated in degenerate OPAs.
But what about the more common, non-degenerate OPAs, where the interaction length is set by the temporal walk-off between three waves? Here the research team opted against aiming for a matched group velocity, which presupposes several practically impossible conditions for the simple waveguide geometries studied here, and instead engineered the group velocities of the signal and idler to walk off from the pump pulse in opposite directions. The result is that the gain localized around the peak of the pump pulse traps the signal and idler pulses to form a three-wave soliton. (Solitons are waves that self-propagate without changing shape; they have a rich history, dating back to 1834, when naval architect John Scott Russell first observed a solitary wave in the Union Canal in Scotland.)
A principal advantage of this technique involving gain-trapped solitons is low power consumption. “The devices we demonstrated reduce the energy requirements needed to achieve efficient down-conversion more than 100-fold over previous approaches, which is the difference between having a pump laser the size of a refrigerator and a pump laser that can fit into a laptop,” Jankowski said. The technique is also extremely robust (highly tolerant of error) and flexible. The research team found that when paired with dispersion engineering, the non-degenerate OPAs could be made to run at any wavelength.
The paper itself advances the field in several ways. “It is the first experimental demonstration of a device that performs optical parametric amplification using confinement in both space and time, the first to provide a simple intuitive picture of the underlying dynamics, and it is the first rigorous study of this process in an engineering context,” Jankowski said. “We show that this approach to optical parametric amplification is very robust and easy to realize, enables extremely high-gain amplification, high conversion efficiency, broad tunability and wide gain bandwidths.” In other words, it resolves the pain points and complicated tradeoffs of traditional approaches to OPA.
The OFC paper on “In-Memory Computing,” co-authored by scientists representing the PHI Lab, the City University of New York (CUNY) and the Korea University-Korea Institute of Science and Technology (KU-KIST), addresses a current dilemma. Namely: The meteoric rise of AI and the limitations of the traditional von Neumann computing architecture. More precisely, this conflict involves the computational demands of machine learning (e.g., matrix-vector multiplication) and the bottleneck implicit in any stored-program computer, wherein instruction “fetching” and data operation cannot be performed simultaneously. In-memory computing systems based on optics are one possible solution.
The idea of optical computing has been around for decades. Yet progress on developing a highly parallel, programmable and scalable optical computer has been elusive. The paper’s four co-authors (PHI Lab Scientist Myoung-Gyun Suh, PHI Lab Director Yoshihisa Yamamoto and PHI Lab Interns Mostafa Honari Latifpur and Byoung Jun Park) suggest a new approach: “We propose a hyperspectral in-memory computing architecture that integrates space multiplexing with frequency multiplexing of optical frequency combs and uses spatial light modulators as a programmable optical memory, thereby boosting the computational throughput and the energy efficiency.”
The results are promising, with demonstrations of the system’s highly precise multiply-accumulate operations indicating a potential application for a variety of ML and optimization tasks. “This system exhibits extraordinary modularity, scalability and programmability, effectively transcending the traditional limitations of optics-based computing architectures,” write the co-authors. “Our approach demonstrates the potential to scale beyond peta operations per second, marking a significant step towards achieving high-throughput energy-efficient optical computing.”
PHI Lab Research Scientist Hiro Onodera participated in a related seven-person panel at OFC on “Photonic Components for In-Physics Computing.” The panel highlighted devices and circuits needed for these new types of computing systems to be competitive with digital hardware. For more on research into non-digital computing that Onodera has conducted, see this write-up and video of a 2022 VentureBeat event on analog ML. For more on Onodera, Jankowski, McKenna and other PHI Lab scientists, visit the PHI Lab video library. Three additional Distinguished Researchers from NTT also spoke at OFC. These include two from the NTT Device Innovation Center: Kenya Suzuki, an expert in optical functional devices, such as optical switches and optical wavelength filters; and Yoshihiro Ogiso, an expert in ultra-high speed optical modulators. Taiji Sakamoto, from the NTT Access Service Systems Laboratories, also spoke at the San Diego event. He specializes in next-generation optical fiber to support the ultra-high capacity, ultra-high-speed networks of the future.
