全忆阻器超低能耗单层感知器研究

The core of constructing artificial neural networks (ANNs) is the hardware implementation of emulation of synaptic and neural functions. Memristor is ideal for the construction of ANNs since it is good synaptic and neural emulator and has simple structure. At present, it is still a big challenge to hardware implement ANNs with high device density, especially considering the energy dissipation of synaptic and neural devices. In the human brain, the energy dissipation of synapses and neurons is as low as several fJ per spike. However, the energy dissipation of most memristive neuromorphic devices is far higher than that of biological ones. Then, a large amount of Joule heat will severely deteriorate the device performance. Our previous study demonstrates that the lowest energy dissipation of neuromorphic devices based on sulfide memristors can reach several fJ. In this project, in order to obtain stable memristive synapses and neurons with ultra-low energy dissipation (1-10 fJ), the microstructure of sulfide materials and the structure of memristive devices will be improved. Next, crosstalk-free crossbar structured synaptic and neural arrays will be constructed. Finally, a simple all-memristive artificial neural network, i.e., single-layer perceptron, will be built. The performance of as-constructed perceptron including effectiveness, stability and reliability will be tested by simple pattern recognition.

构建人工神经网络（ANN）的核心是硬件实现人脑突触和神经元功能的模拟。忆阻器能很好模拟突触和神经元的功能，并且易于高密度集成，因此是构建ANN的理想元器件。目前，高密度ANN的硬件实现依然是一个挑战，尤其是考虑到器件的能耗。人脑突触和神经元单次放电能耗低至fJ量级，但是绝大部分忆阻型类脑器件，单脉冲能耗远高于fJ量级，大量焦耳热将严重影响器件正常工作。在我们前期工作中发现，基于硫化物忆阻器的类脑器件，其单脉冲能耗最低可以达到fJ量级。本项目将进一步优化硫化物材料种类、改善材料微结构和忆阻器件结构，获得性能稳定的超低能耗（1-10 fJ）突触和神经元器件。在此基础上，制备无串扰交叉矩阵结构突触和神经元阵列，结合外围电路，构建全忆阻器超低能耗单层感知器，并且通过简单图像识别，测试单层感知器的有效性、稳定性和可靠性。

硬件实现人脑突触和神经元功能的模拟是构建类脑神经形态芯片的基础。忆阻器结构简单，工作能耗低，因此是构建人工突触和神经元的较理想元器件。本项目聚焦离子迁移能力强的硫化物忆阻器，分别研发出适合模拟突触功能的非易失性忆阻器和适合模拟神经元发射行为的易失性忆阻器。1）通过调控硫化物薄膜沉积温度、器件结构、光电信号激励方案，获得了超低工作电压和可见光透明忆阻器，可用于模拟低能耗人工突触；进一步研发出了能模拟人眼明暗适应能力的光电突触。此外，通过选择导电硫化物作为电极，可有效抑制金属离子向硫化物中的迁移，从而大幅提高了硫化物忆阻器的稳定性。发现并系统研究了忆阻器低阻态自恢复特性，有望拓展离子型忆阻器的应用场景。2）通过优化沉积温度，获得了无需电形成过程的阈值阻变忆阻器，并且创新性实现了双极性阈值阻变，器件在正向电压下表现出常规阈值阻变行为，而施加负向电压可调控后续正向阈值电压。利用双极性阈值阻变特性，构建了可单器件整合兴奋性和抑制性突触信号的人工神经元，实现了多种生物神经元功能，包括泄漏-累积发射、不应期、随机发射、激活函数模拟、发射抑制和输出频率可调等。该新型神经元可大幅降低神经网络中器件数量，减少芯片尺寸和能耗。3）通过仿真模拟，实现了全忆阻器单层感知器及简单类脑功能。本项目工作为忆阻器类脑芯片的研发奠定了一定的器件基础。