Linearity improvement of HfO_x-based memristor with multilayer structure

Abstract

The limitation of traditional Von Neumann architecture could be resolved by machine learning training in neuromorphic computing. However, the nonlinearity characteristic during conductance modulation in memristor severely restricts its application in neuromorphic computing. To improve the analog switching including linearity of the hafnium oxide memristors, Ti metal layer has been inserted on hafnium oxide by using the redox reaction on the interface to achieve more gradual switching. Furthermore, a new multilayer structure device is fabricated utilizing the Ti/HfO_x interface characteristic, which enhanced the electrical characteristic and the conductance modulation linearity to 98.4%, amplitude and symmetry have also been improved. The conductive mechanism of segmented growth of conductive filaments by adjusting the oxygen vacancy concentration gradient was analyzed and further characterized by XPS and TEM. The results of this paper will exalt conductance modulation linearity of hafnium oxide memristors for application to neuromorphic computing systems.

Introduction

Traditional computing system has suffered memory wall problems due to the physical separation of storage and computing [1]. Under such circumstances, memristor defined as the fourth basic element of the circuit is proposed to break through the limit of Von Neumann-architecture [2]. The resistance of memristor can switch between two stable resistance states or among less than two resistance states under the action of a specific external electrical signal. This resistance switching characteristic makes memristor attract attention as a new type of non-volatile memory used in the field of information storage which called resistance access memory (RRAM). Moreover, RRAM employed as memristor is generally considered to have a promising application in emulating synapses due to its high compatibility with CMOS technology and low energy consumption [[3], [4], [5], [6]]. The traditional RRAM comprised of a sandwich structure can store data during the formation and rupture of conductive filaments (CF) under the drive of an electric field [[7], [8], [9], [10]]. This characteristic of the two resistance states swtiching can be regarded as digital switching. However, for neuromorphic computing, memristor can also act as synapse to work in high-density neural networks [11]. Synaptic device can be optimized much continuous tunable resistance states equivalent to analog switching, which is widely used in kinds of training to mimic human brain, such as short-term plasticity (STP) to long-term plasticity (LTP), spike-rate-dependent plasticity (SRDP), and spike-timing-dependent plasticity (STDP) [[12], [13], [14]]. The linearity of incremental switching in memristor describes the uniformity of conductance change upon the adjusting of identical electric pulses [15]. When the memristor is applied with the same pulse, the linear increase in conductance (weight) highlights better predictability of synaptic modulation. It is worth noting that the resistance can be continuously decreased or increased between multiple intermediate states without going back to the original state [16,17].

According to the review of Wu et al. [18], the growth of the CF increases the local electric field or temperature, which in turn promotes V_O (O vacancy) generation and migration and CF growth. This process forms a single strong CF, which hinders gradual control of SET. To avoid abrupt SET in filamentary RRAM, thermally enhanced layer [19], conductive filament lateral extension [20], channel composition modulation [21], interface engineering [22] are four dominant methods that have a relatively good effect. Among these four methods, interface engineering is considered to acquire more gradually switching. Hence, titanium serves as an inserting layer to recompose the interface state of hafnium oxide is a feasible way to achieve gradual change characteristic. H. Y. Lee et al. earlier put forward a thin Reactive Ti Buffer Layer in robust HfO₂ based RRAM, the reaction between the Ti electrode and the HfO₂ layer has been demonstrated [23]. Sk. ZiaurRahaman et al. further proposed an effortless Ti thickness modulation-based methodology to control the oxygen vacancy concentrations in the HfO_x layer,thus indicates the drift of charged oxygen vacancies are crucial for filament formation and rupture [24]. So it is a effective way to improve linearity by inserting titanium layer.

In this work, based on TiN/HfO_x/Ta device, titanium is inserted as a buffer layer to form a non-stoichiometric hafnium oxide at the interface, forming an oxygen vacancy concentration gradient from Ti/HfO_x interface to the position away from the interface. The analog switching displays through Ti insert layer. Furthermore, a novel type HfO_x-based multilayer structure device have been demonstrated to enhance the linearity by generating a roughly oxygen vacancy concentration gradient along the filament growth direction. The electrical characteristic more in line with the memristor are obtained, and linearity during pulse modulation has been raised to 98.4%, amplitude and symmetry have also been improved. Moreover, there exists potential to apply such devices to synaptic mimic measurement and neuromorphic computing.