2D Semiconductor Memristors towards Neuromorphic Hardware Applications

这笔赠款支持为制造未来人工智能系统而创造电子设备的关键知识和技术的研究，以提高美国的技术竞争力和国家繁荣。目前的人工智能系统，如人工神经网络，仍然基于传统的计算原理，与生物神经元过程不匹配，导致巨大的计算复杂性和不可接受的功耗，难以实现规模放大。为了应对这一挑战，该提案支持基础研究，以探索关键的器件物理知识，以实现基于生物相似性高的2D纳米材料的新型记忆开关器件(或忆阻器)，从而潜在地实现生物神经元功能的仿真。由这种设备构建的基于硬件的人工神经网络系统预计能够执行新兴的类脑神经形态计算算法，并能够实现卓越的推理能力以及与生物同行相当的功率效率。如果开发成功，这种神经网络系统可以实现广泛的应用，如无人驾驶车辆的控制，复杂计算机视觉数据的处理，以及基于机器学习的疾病快速诊断，从而大大提高系统的数据处理能力。此外，这项工作的科学和技术成果还将促进开发先进计算和机器人系统的能力。这项研究还促进了来自未被充分代表的群体的学生和教育工作者对与电子学、集成电路芯片、先进控制和计算技术相关的教育活动的参与。新提出的2D半导体忆阻器预计将显示出与基于块材的最先进的忆阻器相比的几个优点，包括无悬挂键表面，潜在地能够以更高的器件集成密度、更低的阈值电压和开关状态的能量、更高水平的器件之间的互连以及更多的可用的器件状态来实现经济高效地生产器件结构。可以进一步利用这些期望的特性来解决与基于硬件的神经网络相关的上述挑战。尽管有这些预期的优点，基于2D半导体忆阻器的神经网络系统的最终实现需要研究努力来解决几个重要的面向器件的挑战。具体地说，二维半导体忆阻器的突触权重更新特性需要改进为响应脉冲式编码信号的线性和对称性，并且需要新的器件掺杂/集成技术来形成不同的突触区域来模拟生物现实功能。此外，还需要进行更多的实验尝试，以构建由2D半导体忆阻器组成的小规模网络，寻求探索神经形态计算算法，以充分利用2D半导体忆阻器在处理动态时空信号方面的上述优势。为了应对这些挑战，PI将执行一系列研究任务，以生产适用于实际网络实施的可靠的2D半导体忆阻器，并初步展示用于神经形态控制应用的小型网络。具体的子目标包括：(1)在微观层面深入了解二维半导体忆阻器的记忆开关方案，生产具有改进的突触权重更新特性的忆阻器；(2)实现具有确定性和统一突触属性的2D忆阻器的可扩展集成；(3)初步展示由2D半导体忆阻器组成的用于时间数据分析的小规模网络。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。

This grant supports research that advances key knowledge and techniques for creating electronic devices for fabrication of future artificial intelligence systems to enhance U.S. technological competitiveness and national prosperity. The current artificial intelligence systems, such as artificial neural networks are still based on conventional computing principles, which do not match biological neuronal processes and result in a formidable computing complexity and unacceptable power consumption for scale-up implementation. To address this challenge, this proposal supports fundamental research to explore critical device physics knowledge for the realization of new memristive switching devices (or memristors) based on 2D nanomaterials which have a high biological similarity, potentially enabling emulation of biological neuronal functions. The hardware-based artificial neural network systems constructed from such devices are anticipated to be capable of executing emerging brain-like neuromorphic computing algorithms and enable superior inference capability as well as power efficiency comparable to those of biological counterparts. Such neural network systems, if successfully developed could be implemented to a broad range of applications, such as controlling of unmanned vehicles, processing of complicated computer vision data, and rapid diagnosis of illness based on machine learning, thereby greatly improving the data processing capability of the systems. In addition, the scientific and technical results from this work will also promote capability in developing advanced computing and robotic systems. This research also enhances participation of students and educators from underrepresented groups in the education activities related to electronics, integrated circuit chips, advanced controlling and computing techniques.The newly proposed 2D semiconductor memristors are anticipated to exhibit several advantageous properties in comparison with state-of-the-art memristors based on bulk materials, including dangling-bond-free surfaces that potentially enable cost-efficient production of device structures with the higher device integration density, the lower threshold voltages and energies for switching states, the higher level of interconnectivity among devices, and the larger number of available device states. These desirable properties could be further leveraged for addressing the aforementioned challenge related to hardware-based neural networks. In spite of such anticipated advantages, the ultimate realization of the neural network systems based on 2D semiconductor memristors demands the research efforts to address several important device-oriented challenges. Specifically, the synaptic weight update characteristics of 2D semiconductor memristors need to be improved to be linear and symmetric in response to pulse-like encoding signals, and new device doping/integration techniques are needed to form different synaptic regions for emulating bio-realistic functions. In addition, more experimental attempts for constructing small-scale networks consisting of 2D semiconductor memristors need to be performed, seeking to exploring the neuromorphic computing algorithms that can fully harvest the aforementioned advantages of 2D semiconductor based memristive devices in processing dynamic spatiotemporal signals. To address these challenges, the PI will perform a series of research tasks to produce reliable 2D semiconductor memristors suitable for practical network implementation and also preliminarily demonstrate small-scale networks for neuromorphic control applications. The specific sub-aims include: (1) Obtain an in-depth understanding of the memristive switching schemes of 2D semiconductor memristors at the microscopic level and produce memristors with improved synaptic weight update characteristics; (2) Realize scalable integration of 2D memristors with deterministic and uniform synaptic properties; (3) Preliminarily demonstrate a small-scale network consisting of 2D semiconductor memristors for temporal data analysis.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.