1
Two-Terminal Neuromorphic Memristors

Hui-Kai He, He-Ming Huang, and Rui Yang

Huazhong University of Science and Technology, School of Materials Science and Engineering, Wuhan 430074, P. R. China

1.1 Memristive Devices

A memristive device is a resistive device with an inherent memory; its theory was creatively conceived by Prof. Chua in 1971 [1] and was connected to the physical devices in 2008 by HP [2]. Since then, memristive devices have been extensively studied over the past decade due to their prominent advantages, such as simple structure, high operation speed, and low power consumption in applications of data storage, logic operation, and neuromorphic computation [3]. In this section, we will introduce traditional two-terminal memristive devices, mainly focusing on device structure and memristive materials.

1.1.1 Memristive Device Structure and Materials

1.1.1.1 Memristive Device Structure

Typically, a memristive device has a metal/insulator/metal (MIM) structure, composed of a switching layer sandwiched between two metal electrodes (possibly different), as shown in Figure 1.1a. Because of its simple structure, highly scalable cross-point and multilevel stacking memory structures have been proposed (Figure 1.1b), which is promising for the construction of huge neural networks and neuromorphic computing systems [3]. It is well known that electrodes play a crucial role in the resistive switching behavior of memristive devices. To date, in addition to metals (such as Ag [5], Cu [6], Pt [7], Au [8], Al [9], and W [10]), a variety of conductive materials have been explored as electrodes for memristors, including nitrides such as TiN [11], carbon materials such as graphene [12] and carbon nanotubes [13], conductive oxides such as ITO [14] and SrRuO3 (SRO) [15], p- and n-type Si [16], and so on. Among these metals, Ag and Cu are the most popular ones due to their ability to dissolve in thin film electrolyte at low electric field and their high ionic mobility [17]. In addition to the electrodes, the switching layer where the resistive switching takes place is the key layer in memristive devices and has a great impact on the device performance. Typically, the switching layer is an insulator or a semiconductor. Also, it is normally in the form of thin film, which is compatible to large-scale integration in the semiconductor industry. Recently, other forms of the switching layer are also intensively investigated, including nanoparticles [18], nanowires [19], two-dimensional (2D) materials [20], three-dimensional nanoarrays, etc. Note that we mainly discuss memristors in the form of thin film in all below sections.

Figure 1.1 (a) Diagram of a memristive device with a capacitor-like structure in which a switching layer is sandwiched between two metal electrodes. (b) Diagram of a cross-point memory structure. Word and bit lines are used for selecting a memristive device and writing/reading data, respectively.

Source: Sawa [4].

1.1.1.2 Memristive Materials

As mentioned above, the materials involved in memristors include switching materials and electrode materials. Here, we mainly focus on switching materials, which is also termed as memristive materials. Up to now, a great number of memristive materials have been explored for memristive devices used in neuromorphic computing. In this chapter, the memristive materials are subdivided into inorganic and organic materials. Generally speaking, inorganic materials have significant advantages over organic ones in switching stability and manufacturing technology, while organic ones stand out in terms of high-mechanical flexibility and low cost.

Inorganic materials for memristors can be loosely divided into binary oxides (e.g. TiOx [21], TaOx [22], HfOx [23], WOx [24], and ZnO [14]), perovskite oxides (e.g. SrTiO3 [25] and BiFeO3 [26]), and 2D materials (e.g. graphene [27], hexagonal boron nitride (h-BN) [28], and molybdenum disulfide (MoS2) [20]). Among these inorganic materials, binary oxides have been intensively studied since they are the most abundant and show superior switching characteristics including ultrahigh ON/OFF ratio, sub-ns operation speed, and extreme endurance. In addition, their simple composition enables them to be easily fabricated by various film deposition technologies, mainly including magnetron sputtering [14, 24], atomic layer deposition (ALD) [29], thermal oxidation [30], and plasma oxidation [11]. Magnetron sputtering is a high-rate, high-efficient film deposition technology and is becoming increasingly popular owing its high-yield and low-cost production of uniform films over large areas. Recently, ALD has also attracted increasing attention for the deposition of binary oxides due to its ability to accurately control the thickness and uniformity of the films. Furthermore, binary oxides have good compatibility with conventional complementary metal oxide semiconductor (CMOS) process and good thermal stability. Thus, binary oxides have been the focus of both academic and industrial communities over the past decade. In particular, research interest in HfOx and TaOx has been extremely high in the past few years since they exhibit both sub-ns operation speed and extreme endurance of >1010 cycles and may be the most promising memristive materials for practical applications in the near future.

In addition to binary oxide, perovskite oxides such as SrTiO3, SmNiO3, BiFeO3, and SrRuO3 have also been researched for memristors over the past few years. Among these perovskite oxides, SrTiO3 receives the most attention for the implementation of memristive synapses because of its superior memristive properties and rich switching dynamics [25, 31, 32]. It has been found that perovskite oxides have advantages of excellent localized accumulation of oxygen ions and can be easily converted into a defective structure. However, it should be mentioned that they are generally obtained through pulsed laser deposition (PLD) under high temperature. Although this is an advanced deposition method that can obtain high-quality thin films with accurate stoichiometry, it is not widely used in the semiconductor industry due to its high-cost and the small uniform area of the deposited film, greatly hindering the development and application of perovskite oxides in memristors.

In recent years, 2D materials have become a new focus in memristors for the realization of artificial synapses and neurons due to their superior physical, chemical, and mechanical properties, including graphene, MoS2, and h-BN. Graphene is one of the highly desirable materials for memristive bioinspired devices owing to its excellent properties of low cost, tunability, nontoxicity, flexibility, and biocompatibility [33, 34]. However, graphene is inherently a semimetallic material and needs to be oxidized or doped when it is used as a switching layer. In contrast to graphene, transition metal chalcogenides (TMDs), such as MoS2 and tungsten selenide (WSe2), are semiconductors with proper bandgaps from 1 to 2 eV [35]. Therefore, TMDs are considered as ideal substitutes for graphene. MoS2, the common member of the TMDs family, has been intensively investigated and shows superior performance including ultrahigh ON/OFF ratio [36], ultralow operating voltage [37], and excellent thermal stability [38]. Reliable production of 2D materials with uniform properties is essential for translating their new electronic and optical properties into applications. Until now, various fabrication methods including mechanical exfoliation [39], liquid-phase exfoliation [40], and chemical vapor deposition (CVD)[41] have been employed to obtain atomically thin flakes of 2D materials. First discovered by Novoselov et al. in 2004, ultrathin 2D materials are peeled from their parent bulk crystals by mechanical exfoliation using adhesive tape. This method produces single-crystal flakes of high purity and cleanliness that are suitable for fundamental characterization. Liquid-phase exfoliation method is also a feasible way to prepare atomically thin 2D material. It can produce gram quantities of submicrometer-sized monolayers, but the resulting exfoliated material differs structurally and electronically from the bulk material [42]. To obtain large-area and uniform layers, CVD method is very promising. Such methods give reasonably good-quality material with typical flake sizes of hundreds of nanometers to a few centimeters, although the flake thickness is not conclusively shown to be monolayers.

Compared with inorganic materials, organic materials have the advantages of low cost, ease of fabrication, and, especially, high-mechanical flexibility. In addition, it is easy to modulate the electrical performance of organic materials by a designed molecular synthesis [43]. Accordingly, organic materials are attracting more and more attention that enable them to be promising for future flexible electronics, although most switching...