Revolutionizing Computing With Meta-ONNs
MIE/ECE Professor Yongmin Liu was awarded a $467,930 NSF grant for “Design and Realization of Multiplexed Meta-Optical Neural Networks.” The goal is to design and implement a novel class of metasurface-based ONNs, know as meta-ONNs, which can operate at optical frequencies and realize diverse functions, including all-optical image recognition and pattern generation.
Abstract Source: NSF
Artificial neural networks in machine learning have impacted many fields of science and engineering. Despite tremendous progress and achievements, implementing artificial neural networks on conventional computers is becoming increasingly challenging due to power and speed constraints. Optical neural networks (ONNs), which offer potential advantages in energy efficiency, speed, parallelism, bandwidth, and scalability, stand out as a highly promising solution to this challenge. The goal of this project is to design and implement a novel class of metasurface-based ONNs, termed meta-ONNs, which can operate at optical frequencies and realize diverse functions, including all-optical image recognition and pattern generation. Metasurfaces are composed of artificially engineered structures much smaller than the wavelength of light. They can manipulate light characteristics such as the amplitude, polarization state, and phase in a prescribed manner. By leveraging the unique properties of metasurfaces, this project will demonstrate innovative meta-ONNs capable of encoding multiple functional channels within a single system and achieving functions beyond conventional classification, significantly expanding the capabilities of ONNs by transforming computing, communications, and information processing technologies, thus benefiting the public and the nation. Integrated with the research, the education effort of the project will enhance outreach activities and educate students across different levels. Students will actively participate in the project, gaining frontier knowledge in multiple fields and eventually becoming leaders in the next-generation workforce.
The project aims to unlock the potential of meta-ONNs as a new platform for multifunctional optical computing, complex information processing, and innovative image generation through a software-hardware co-design approach that seamlessly integrates photonics, neural network models, advanced manufacturing, and systems engineering. The project consists of three research thrusts: (1) Design multiplexed meta-ONNs based on artificial intelligence (AI) and optimization techniques to seamlessly integrate multiple wavelength and polarization channels within a single system, greatly enhancing the capacity and versatility of ONNs; (2) Demonstrate generative meta-ONNs that can create distinct images after light propagates through the meta-ONNs, enabling novel optical encryption schemes and serving as pivotal tools for AI-assisted photonic design; (3) Fabricate low-loss, multilayered metasurfaces to implement the designed meta-ONNs, and experimentally characterize the key performance metrics including accuracy, efficiency, and robustness. The precise control of light at the subwavelength meta-neuron level is expected to significantly boost the capacity of ONNs. New manufacturing methodologies and techniques will be developed to realize high-efficiency meta-ONNs. The potential applications of the meta-ONNs include image generation for virtual reality and entertainment, medical imaging and diagnostics, security and surveillance, autonomous driving, and advanced photonic circuits for quantum computing. The research findings will accelerate the interplay between AI and photonics, forming the virtuous AI-photonics-AI circle. The design principles could also serve as inspiration for other physical neural networks and intelligent devices based on mechanical, electrical, and acoustic systems.