CAREER: Entropy Oxide Memristors for Software-equivalent Neuromorphic Computing

人工智能(AI)的训练预计将在未来十年消耗指数级数量的计算资源，并付出巨大的环境和财务成本。记忆技术已经显示出巨大的潜力，可以成为高密度下一代神经形态处理器的一部分，以支持节能的人工智能算法。记忆开关的物理学是很有前途的，因为在理论上，在非常低的电压下，只有少数离子的运动才能改变器件的状态。然而，由于设备的非理想性，由于性能低于软件解决方案，因此无法使用忆阻器芯片进行神经网络训练。忆阻器和其他新兴设备通过需要紧凑和高效计算的关键应用，如医疗植入物、机器人、物联网等，有望产生重大的社会影响。然而，它们的商业应用需要在性能、产量、可靠性和劳动力培训方面实现学术飞跃。该项目将开发下一代忆阻器，基于具有工程开关动力学的熵稳定氧化物。这项工作将使发现用于忆阻器设备的新材料，并支持高效计算技术的开发，与国家对半导体竞争力和下一代微电子的关注保持一致。此外，该项目将研究和教育目标无缝地结合在一起，以支持该领域必要的劳动力发展。这些努力包括对一门纳米电子学课程进行全面改革，使其更容易获得，为不同的学生提供跨学科的研究培训，开发一个全面的虚拟现实记忆阻器模拟器，向社区提供GitHub文档，以及与当地高中进行接触。开发的模型和实验平台将为学生提供关于这些器件、它们的制造以及它们在神经形态硬件中的新应用的实践教育。这个职业项目旨在通过利用熵稳定氧化物的潜在材料属性和物理现象来创造超低变异性模拟忆阻器。通过与现有的尝试使用相分离在介观(微米)尺度上进行结构灯丝约束的方法进行比较，该项目提出在纳米尺度上进行灯丝约束以获得更好的开关一致性。将探索熵稳定的氧化物，因为它们提供了广阔的组成空间和潜力，可以在合适的温度下沉积高质量的薄膜，在这种温度下可以满足所需的开关动力学要求。这些复合氧化物将被用来研究纳米尺度的细丝动力学控制，允许氧空位优先沿着能量有利的轨迹移动，同时晶格是熵稳定的。为此，这项工作将在四个相互交织的研究和教育目标上进行创新。目标#1将集中于识别合适的熵稳定膜以及它们在工程记忆阻器器件中的超低变异性应用。目标2将使用合适的缓冲层在晶体管芯片上集成和表征这些熵稳定的忆阻器。目标#3将使用实验数据集开发用于神经网络模拟的多维器件模型，并使用制造的芯片以软件同等精度制作感知器和变压器神经网络的原型。目标4将以发展和纳入体验式学习活动为目标，以支持在实地培训不同的学生。该项目将在了解超低变异性忆阻器的基本材料和物理要求方面取得重大进展，并调查其在下一代人工智能硬件中的潜在性能。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Training of artificial intelligence (AI) is foreseen to consume exponential amounts of computing resources in the next decade, at large environmental and financial costs. Memristive technology has shown significant potential to be part of high-density next generation neuromorphic processors to support energy efficient AI algorithms. The physics of memristive switching is promising since in theory, the device state can change through the movement of only few ions under very low voltage. However, device non-idealities have prevented using memristor chips for neural network training due to performance sub-par to software solutions. Memristors and other emerging devices promise major societal impact via key applications e.g. medical implants, robotics, Internet of Things, etc. that need compact and efficient computing. However, their commercial adoption requires academic leaps in performance, yield, reliability and workforce training. This project will develop the next generation of memristors, based on entropy-stabilized oxides with engineered switching dynamics. This work will enable the discovery of new materials for memristor devices and support the development of efficient computing technologies, in line with the national focus on semiconductor competitiveness and next generation microelectronics. Moreover, this project seamlessly integrates the research and education objectives to support the necessary workforce development in this field. The efforts include overhaul of a nanoelectronics course to make it more accessible, interdisciplinary research training for diverse students, the development of a comprehensive virtual reality memristor simulator, providing github documentation to the community and outreach to local high schools. The developed models and experimental platforms will give students a hands-on education on these devices, their manufacturing, and their use in neuromorphic hardware for novel applications.This CAREER project aims to create ultra-low variability analog memristors by exploiting the underlying material properties and physical phenomena in entropy-stabilized oxides. By comparison with existing approaches attempting structural filament confinement at mesoscopic (micron) scales using phase separation, this project proposes filament confinement at nanometer scales for better switching uniformity. Entropy-stabilized oxides will be explored because they provide a broad compositional space and potential for high-quality films deposited at CMOS-suitable temperatures where the requirements for desired switching dynamics can be met. These complex oxides will be used to investigate the nanoscale control of the filament dynamics by allowing the oxygen vacancies to preferentially move along energetically-favorable trajectories, while the lattice is entropy-stabilized. To this effect, the work will innovate across four intertwined research and educational objectives. Objective #1 will focus on the identification of suitable entropy-stabilized films and their use in engineering memristor devices with ultra-low variability. Objective #2 will integrate and characterize these entropy-stabilized memristors on transistor chips using suitable buffer layers. Objective #3 will use the experimental dataset to develop multi-dimensional device models useful in neural network simulations and use the fabricated chips to prototype perceptrons and transformer neural networks with software-equivalent accuracy. Objective #4 will target the development and incorporation of experiential learning activities to support the training of diverse students in the field. This project will make critical advances in the understanding of the underlying material and physical requirements for ultra-low variability memristors and investigate their potential performance for next generation artificial intelligence hardware.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.