Based Synaptic Transistor for Neuromorphic Computing

Abstract

Human brain-inspired neuromorphic computing (NC), which emulates neural activities, is considered for the low-power implementation of artificial intelligence. It gives the benefit of parallel processing, ultralow power consumption, self-learning, and fault tolerance. The realization of hardware elements, such as artificial synaptic devices, which are the basic building blocks of neuromorphic computing systems, is an essential path towards this goal. Two terminal artificial synaptic devices have been studied using resistive random-access memory and memristors and have emulated the synaptic functions. However, in these two-terminal devices, both learning and signal transmission cannot occur concurrently, which leads to an incomplete emulation of a synapse. On the other hand, in a three-terminal synaptic device, signal transmission and learning occur independently through the channel and external gate respectively, leading to complete emulation of a synapse. Much progress has been made in three-terminal devices based on oxides [1], organic materials [2,3], and 2D materials [4]. Electrolyte gating has been widely used for conductance modulation by massive carrier injections at a very low gate voltage. However, studies based on metallic ferromagnetic systems remain highly unexplored. Additionally, a synaptic device based on magnetic metal has a spin degree of freedom that can be tuned along with. In this work, we have demonstrated a three-terminal coplanar synaptic transistor based on permalloy by electrolyte gating.The channel consists of a stack of Ta (3 nm)/NiFe (3 nm)/Ta (1 nm) on the silicon substrate. EMIM-TFSI ionic liquid was used as the gate dielectric. Fig. 1(a) shows the transfer curve measured at different sweeping rates. A giant anticlockwise loop opening in the transfer curve shows a non-volatile and reversible change of channel conductance suitable for emulating synaptic functions. We have obtained the highest room temperature conductance modulation of 11 % for a metal at the slowest sweeping rate, as shown in Fig. 1(b). The transistor's working principle is based on the formation of an electric double layer (EDL) at the interface of ionic liquid and channel and by intercalation/ extraction of oxygen ions.Multilevel non-volatile states are a prerequisite for realizing analog computing. We have obtained gating controlled multilevel, non-volatile, reversible conducting states by applying positive gate pulses of different amplitude shown in Fig. 1(c). The channel conductance remained constant at zero gating showing the non-volatile nature of the state. Furthermore, the channel conductance can be brought back to the initial state by applying negative gate pulses. The retention of conducting states is checked separately, which shows retention of up to a few hours or more.In this permalloy based artificial synapse, the gate pulses act as the external stimuli, and the channel conductance act as the synaptic weight. We have emulated the essential synaptic functions at room temperature. A basic form of short-term plasticity like excitatory postsynaptic conductance (EPSC) and paired-pulse facilitation (PPF) has been emulated, as shown in Fig. 2(a) and Fig. 2(b), respectively.Moreover, we have shown a transition from short-term memory (STM) to long-term memory (LTM) by varying the gate pulse amplitude, duration, and number shown in Fig. 2(c). The change in channel conductance is more significant at higher amplitude, duration, and number due to the intercalation/extraction of more oxygen ions into the channel. Long-term plasticity behavior like long-term potentiation (LTP) and multiple potentiation and depression cycles have been demonstrated in Fig. 2(d)-(f).The writing and reading energy per synaptic event are calculated at different gate amplitude and duration showing a linear escalation. The lowest writing energy consumed in our device is 0.6 nJ, and can be further reduced by reducing the channel dimension and the pulse width. The variation of magnetic transport properties of the device with gating is also studied. Along with the emulation of synaptic functions, the device showed dynamic filtering behavior and can act as a high-pass filter. These results provide an insight into the potential application of ferromagnetic metal-based synaptic transistor for large-scale, low power consuming, silicon compatible neuromorphic computing network. **