Copper-Based 3-Terminal Synaptic Cell with Multiple Resistance Levels

Abstract

Synaptic elements capable of reversible multilevel analog switching have the potential to become key components in future neuromorphic computing technologies and exceed the practical limits of current digital technology. Artificial neural networks based on analog synapses could potentially deliver performance one step closer to the human brain which far surpasses the efficiency of current computing capabilities by many orders of magnitude. For example, this could be achieved by mapping the synaptic weights of deep neural net on the multi-bit device and performing the multiply-accumulate operation in-memory during learning or inferencing cycles [1]. While different types of 2-terminal emerging non-volatile memory, such as phase change memory (PCM) or resistive RAM (ReRAM), permit some degree of reversible analog tuneability, they exhibit significant stochasticity, and asymmetry which are detrimental to learning accuracy. In addition to precise and predictable control of the resistance state, decoupling the write and read cycle is a major challenge for the practical application of such devices. Memistor-based three-terminal devices allow decoupling the read and write steps, potentially permitting drastic increase in device operational precision and reliability. Electrochemical copper plating devices of this type for the purpose of “adaptive switching circuits” were conceived as early as the1960s in the ADALINE neuron element [2] and more recently as “atomic transistors” [3], however these devices utilized liquid electrolytes which are a big challenge for practical fabrication of an integrated chip. Other systems, such as lithium-based intercalation materials have been used for solid-state three-terminal synaptic devices and scaled to sub-micron dimensions [4], [5]. However. Li+ based systems are chemically reactive and sensitive to atmospheric exposure while also possessing a residual open circuit potential which opposes continued potentiation and creates challenges for voltage-driven neural networks. Here, we report a solid-state, lithium-free, Cu+-based 3-terminal synaptic device capable of reversible multilevel switching. Cu+ is driven in or out of a channel to create a conductance change and Cu+ ion transport is achieved through a solid electrolyte by writing pulses of only 100mV with near-zero open circuit potential. We analyze device performance, physical, and chemical characteristics, and provide insight into operation mechanisms and limitations. Exceptionally high ionic conductivity of 0.36S/cm (significantly higher than any reported solid-state lithium-based system) was measured in the electrolyte by impedance spectroscopy which is critical for high speed device operation. The first device prototypes were over 100 microns in size with very low dynamic range and channel resistance of only several hundred ohms, which is too low for practical applications. We demonstrate a path for increasing the dynamic range and channel resistance of our devices, as well as first steps towards transfer to sub-micron scale fabrication. Figure 1 Impedance spectrum of solid copper electrolyte with inset micrograph of completed device in measurement setup (a). Schematics of device structure and operation (b). Channel resistivity change during cycling measured between writing pulses (two consecutive cycles of 5, +100mV, and 5,-100mV pulses each). Reference: [1] T. Gokmen, Y Vlasov, Frontiers in Neuroscience 2016, 10, 333 [2] B. Wirdow. Technical Report No1553-2, Office of Naval Research 1960 [3] Xie et al. Beilstein J. Nanotechnol. 2017, 8, 530–538 [4] Fuller et al. Li-Ion Synaptic Transistor for Low Power Analog Computing. Adv. Mater. 2017, 29 [5] J.Tang et al, presented at IEEE IEDM 2018, 13.1 Figure 1