FuSe-REG: Enabling Photonic Computing Engines through Hetero-integration

由于人工智能和机器学习算法的快速发展，基于数字电子逻辑的传统计算面临挑战，这些算法需要大量的处理能力，超过了计算机历史上的发展速度（根据摩尔定律）。光子学技术提供了一个有吸引力的替代方案，因为它有几个优点，包括高速、能源效率和信息处理大规模并行化的潜力。该项目旨在解决在下一代计算平台中使用集成光子深度神经网络的障碍。尽管之前的努力，这些网络仍然面临着挑战，使它们在现实世界的应用中不切实际。特别是以往的光子神经网络大多是不可扩展的，需要电子电路来实现非线性效果，因此不可避免地失去了光子学的高速优势。需要一种跨学科的方法来解决几个问题，包括设计新的神经网络架构来有效处理光编码的信息，以及集成不同的材料来实现所需的功能，例如可重构性和非线性效果，这对机器学习至关重要。该项目专注于开发一种新的深度学习架构，该架构与光子学和光电子技术兼容，与之前的尝试相比，可以显着减少光神经网络的大小。该平台旨在创建一个可扩展的深度神经网络，实现从电信、成像和生物医学应用到经典和量子信息处理系统等广泛应用的实时光信号处理。提出的努力将为加速实现功能光子深度神经网络的基础研究和应用技术开发制定路线图。这一努力还提供了一个独特的机会，通过参与高级研究活动和实习、研讨会和会议，培养和教育学生超越他们的实验室研究。参加的研究人员将为两个技术讲习班编写短期课程，其中将涉及基础和高级研究课题。此外，该团队将组织旨在确定其他合作者的会议。最后，将为本科生创造实习和研究轮岗机会。该项目汇集了来自光学计算、集成光子学和异质集成的专业知识，围绕集成光子电路中光传播的基本物理定律，开发了一种新的深度学习架构。与异构集成相结合，这种结构可以显著减小尺寸，从而实现大规模光子深度神经网络。所提出的架构利用线性和非线性操作的交错，并具有独特的参数化，以促进光子实现中数十个网络层的集成。该项目的重点包括(i)为光子深度学习开发理论/计算基础设施；（ii）半导体和其他光学材料（包括铌酸锂）的异质集成，以在光子芯片上实现强而灵活的非线性效应；(3)基于相变材料的超紧凑型移相器的实现与优化；（四）全光子神经网络的设计、制造和实验演示。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Conventional computing based on digital electronic logic faces challenges due to the rapid growth of artificial intelligence and machine learning algorithms, which need massive amounts of processing power, outpacing the rate at which computers have historically progressed (according to Moore's Law). Photonics technologies offer an attractive alternative due to several advantages, including high speed, energy efficiency, and the potential for massive parallelization of information processing. This project aims to tackle the obstacles in using integrated photonic deep neural networks for the next generation of computing platforms. Despite previous efforts, these networks still face challenges that make them impractical for real-world applications. In particular, most previous photonic neural networks are not scalable and require electronic circuits for achieving nonlinear effects, thus inevitably losing the high-speed advantages of photonics. An interdisciplinary approach is needed to tackle several problems, including designing novel neural network architectures for efficient processing of information encoded in light and integrating different materials for achieving desired functionalities, such as reconfigurability and nonlinear effects, which are essential for machine learning. This project focuses on developing a new deep learning architecture that is compatible with photonics and optoelectronics technologies and significantly reduces the size of optical neural networks compared to previous attempts. The proposed platform aims to create a scalable deep neural network, enabling real-time optical signal processing for a wide range of applications, from telecommunications, imaging, and biomedical applications to classical and quantum information processing systems. The proposed effort will create a roadmap for accelerating fundamental research and applied technology development for realizing functional photonic deep neural networks. This effort also provides a unique opportunity to train and educate students beyond their laboratory research through engagement in advanced research activities and through internships, workshops, and conferences. The participating investigators will develop short courses for two technical workshops that will involve fundamental and advanced research subjects. In addition, the team will organize conferences aiming to identify additional collaborators. Finally, internship and research rotation opportunities will be created for undergraduate students.This project brings together expertise from optical computing, integrated photonics, and hetero-integration to develop a novel deep-learning architecture that is built around the fundamental physical laws of light propagation in integrated photonic circuits. In conjunction with hetero-integration, this architecture leads to dramatic size reduction to enable the realization of large-scale photonic deep neural networks. The proposed architecture utilizes an interlacing of linear and nonlinear operations and is uniquely parametrized to facilitate integration of tens of network layers in a photonic implementation. The thrusts of this project involve (i) developing a theoretical/computational infrastructure for photonic deep learning; (ii) hetero-integration of semiconductor and other optical materials, including lithium niobate, to achieve strong and flexible nonlinear effects on a photonic chip; (iii) realizing and optimizing ultracompact phase shifters based on phase change materials; and (iv) design, fabrication, and experimental demonstration of all-photonic neural networks.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.