Collaborative Research: CMOS+X: 3D integration of CMOS spiking neurons with AlBN/GaN-based Ferroelectric HEMT towards artificial somatosensory system

通过先进的封装工艺将联合收割机硅技术与新兴器件相结合的三维异构集成方法可以利用先进电子/光电子的独特半导体组合。特别地，基于Si的人工神经元和人工突触的集成将通过最小化传感器、计算和致动单元之间的数据传输来实现节能的近传感器计算。我们的神经形态阵列将允许对各种传感器采集的数据进行原位处理，并将提供用于致动的必要控制信号，该致动可以普遍用于读取和处理外部刺激并相应地做出响应，例如原位视觉处理和机械响应。具体而言，3D集成神经形态单元将实现高频和高功率操作，为机器人、自动驾驶汽车和医疗设备实现简化的感知到行动系统。因此，我们提出的异构集成系统提供了一个创新的范例，一个紧凑的神经形态边缘计算系统，是从中央处理单元（CPU）和图形处理单元（GPU）分散。为了实现上述目标，该提案旨在设计和演示一种片上人工体感系统，该系统可以通过基于互补金属氧化物半导体（CMOS）的棘波神经元和基于GaN铁电高电子迁移率晶体管（FeHEMT）的人工突触的3D集成来模拟生物体感系统。所设计的神经形态芯片将能够用一维时间序列向量调制小的感觉信号。原始的时间序列的感觉信号可以有效地处理与基于CMOS的尖峰神经网络（SNN）的能源效率和时空编码，以克服冯诺依曼瓶颈。所设计的神经形态芯片提供一次性计算，类似于中枢神经系统（CNS）中的生物计算。此外，Cu-Cu互连将使基于CMOS的SNN与基于宽带隙半导体的铁电晶体管的高密度3D集成成为可能，用于输入刺激的原位处理以触发机械致动。由图像传感器捕获的时间序列数据将通过前端基于CMOS的神经形态芯片在尖峰域中进行编码。编码后的输出信号将直接传输到后端神经形态芯片，基于FeHEMT交叉杆的synchronous阵列编程其权重值。通过AlBN/GaN HEMT交叉杆阵列的解码输出电流可以超过安培的量级，允许其驱动用于系统宏观运动的机械致动，例如机械对象跟踪。我们相信所提出的混合信号神经形态阵列将允许对时间序列感觉数据进行原位处理，从而实现了一种超低功耗的人工体感系统，从传感和数据处理到广泛应用（包括AIoT和机器人）的反应的高效和自发计算。该奖项反映了NSF的法定使命，并通过使用基金会的学术价值和更广泛的影响审查标准。

Three-dimensional heterogeneous integration approaches that combine silicon technology with emerging devices via advanced packaging processes can leverage unique semiconductor combinations for advanced electronics/optoelectronics. In particular, the integration of Si-based artificial neurons and artificial synapses will enable energy-efficient near-sensor computing by minimizing data transfer between sensor, computing, and actuation units. Our neuromorphic array will allow for the in-situ processing of data acquired by various sensors and will provide necessary control signals for actuation that can be universally used to read and process external stimuli and respond accordingly, such as in-situ vision processing and mechanical response. Specifically, 3D integrated neuromorphic unit will enable high-frequency and high-power operation, realizing a simplified sensing-to-action system for robots, autonomous vehicles, and medical devices. Thus, our proposed heterogeneously integrated system provides an innovative paradigm for a compact neuromorphic edge-computing system that is decentralized from central processing units (CPUs) and graphic processing units (GPUs). To achieve the above goal, the proposal aims to design and demonstrate an on-chip artificial somatosensory system that can emulate the biological somatosensory system via 3D integration of complementary metal-oxide-semiconductor (CMOS)-based spike neurons and GaN ferroelectric high electron mobility transistors (FeHEMTs) based artificial synapses. The designed neuromorphic chip will be able to modulate small sensory signals with a one-dimensional time-series vector. The raw time-series sensory signals can be efficiently processed with a CMOS-based Spiking Neural Network (SNN) for energy-efficient and spatiotemporal encoding to overcome the Von Neumann bottleneck. The designed neuromorphic chips provide one-shot computation, analogous to the biological computing in the central nervous system (CNS). Furthermore, Cu-Cu interconnection will enable the high density 3D integration of the CMOS-based SNN with ferroelectric transistors based on wide-bandgap semiconductors for in-situ processing of the input stimulus to trigger mechanical actuation. The time-series data captured by the image sensor will be encoded through the front-end CMOS-based neuromorphic chip in a spiking domain. The encoded output signals will be directly transmitted to the back-end neuromorphic chip based on the FeHEMT crossbar-based synpatic array to program its weight value. The decoded output current through the AlBN/GaN HEMT crossbar array can exceed an order of mangitude of an ampere, allowing it to drive mechanical actuation for system macro-motion, such as mechanical object tracking. We believe the proposed mixed-signal neuromorphic array will allow for the in-situ processing of time-series sensory data, leading to the realization of an ultra-low-power artificial somatosensory system that provides power-efficient and spontaneous computing from sensing and data processing to reaction for widespread applications including AIoT and robotics.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.