Memristive devices based on 2D materials, such as WSe 2 and MoS 2 , have been demonstrated to exhibit analogue switching characteristics and enable emulation of ionic interactions involved in synaptic activities. These attractive features hold great potential for construction of energy-efficient artificial neural networks. However, the memristors made from pristine 2D materials typically exhibit a small dynamic range and poor linearity of switching characteristics. The neural network simulated in the basis of such switching characteristics has a poor learning accuracy of ~43%. In this work, we find that Ar plasma treatment can greatly improve both the dynamic range and linearity of analogue switching characteristics of few-layer MoS 2 memristors. The neural network consisting of such plasma-treated memristors is simulated to be able to result in a significantly improved learning accuracy of 94.3% for the MNIST handwritten digits dataset. Our addition...

中文翻译：

---

通过等离子体处理改善MoS 2忆阻器的模拟开关特性

已经证明基于2D材料的忆阻器件（例如WSe 2和MoS 2）表现出类似的开关特性，并能够模拟涉及突触活动的离子相互作用。这些诱人的功能为构建节能型人工神经网络具有巨大潜力。但是，由原始2D材料制成的忆阻器通常具有较小的动态范围和较差的开关特性线性度。基于这样的开关特性模拟的神经网络的学习精度很低，约为43％。在这项工作中，我们发现Ar等离子体处理可以极大地改善几层MoS 2忆阻器的模拟开关特性的动态范围和线性。模拟由这种经过等离子体处理的忆阻器组成的神经网络，能够为MNIST手写数字数据集带来94.3％的显着改善的学习准确性。我们的补充...