1 Introduction

Resistive switching (RS) devices are being extensively studied due to their potential to overcome the von Neumann bottleneck for facilitating in-memory computing [1, 2]. Besides that machine learning and other neural network-based in-memory computing systems are also widely exploiting the RS devices to mimic the basic bio-synaptic functionalities [3, 4], wherein the memristors work as artificial synapses. As a future technology requirement, high-density integration of such artificial electronic synapses necessitates scalability of the devices down to nanometers [5]. To date, RS devices are mostly investigated at bulk level, inferences drawn from which may not be appropriate to nanoscale devices [6]. For example, it has been observed that requirement of forming voltage increases with a decrease in device size [7]. Nevertheless, the phenomena regulating their RS behaviour such as filament formation/annihilation [8], local phase variation [9] or ferroelectric domain switching [10] are confined at nanoscale, which therefore, makes it necessary to perform nanoscale analysis to fully comprehend the underlying mechanism. Interestingly, by exploiting atomic force microscopy (AFM), wherein the nano-sized AFM tip (~ 25 nm radius of curvature) works as one of the electrodes, one can investigate the device scalability and local electrical properties at nanoscale without the complexity of fabricating several small devices with ultra-high packing density [11, 12].

In this paper, we demonstrate synaptic functionalities in transition metal oxide (TMO)-based filamentary memristors using different versatile modes of AFM. In this regard, most of the TMOs, such as ZnO, TiO2, NiO, and HfO2, are preferred candidates due to their adaptability to the deviation in their local intrinsic/extrinsic defect landscape that promotes the formation of conducting channels/filaments, which largely govern their memristive/synaptic functionalities [13,14,15,16]. We investigate the influence of defect modification on memristive and synaptic properties in two types of metal oxide [viz. ZnO:Cu (2 wt%) (CZO) and TiOx (x < 2)]-based memristors at nanoscale. The CZO films are modified in situ by incorporating O2 gas of varying flow rates during the film growth to compensate for the excess amount of oxygen vacancy (OV) concentration within the films [17]. The TiOx films, on the other hand, are subjected to post-deposition implantation by Ar+-ions of varying fluences with the aim to introduce increased density of OV defects [11, 18]. A detailed AFM-based local probe analysis reveals the role of CZO film's resistivity in facilitating stable RS operation, whereas defect modification in TiOx films by ion implantation reduces the operating voltages, affirming their potential for energy-efficient devices. We further demonstrate forming-free resistive switching along with emulation of the fundamental bio-synaptic functionalities in both devices at nanoscale that not only marks their potential to develop artificial neural networks with ultra-high density integration for in-memory computation [5, 11], but also establishes AFM as a versatile tool for optimizing nano-sized memristors and electronic synapses.

2 Experimental Details

CZO and TiOx films were grown on ultrasonically cleaned ITO/glass (Merck) and Pt/TiO2/SiO2/Si (platinized-Si, Radiant Technology, UK) substrates, respectively, using a pulsed laser deposition (PLD) system at room temperature (RT). The PLD has several advantages including versatility, control over the growth rate, stoichiometric transfer and unlimited degrees of freedom in the ablation geometry compared to other deposition techniques [19]. In this study, a KrF excimer laser (wavelength: 248 nm, pulse duration: 20 ns, and repetition rate: 10 Hz) with a fixed laser energy density of 1.7 J cm−2 was employed to ablate a hot-pressed ceramic target of the respective material. To control the concentration of the intrinsic oxygen vacancy defects, two different methods are adopted. For CZO films, different O2 flow rates (viz. 0, 5, and 10 SCCM) were introduced during deposition. On the other hand, the TiOx films are implanted post-deposition by 5 keV Ar+-ions at an incident angle of 45° with respect to the surface normal of the sample with different ion fluences in the range of (0.5−3) × 1017 ions cm−2  at RT using an electron cyclotron resonance (ECR)-based ion source at Institute of Physics, Bhubaneswar, India.

The thicknesses of the as-grown CZO and TiOx films are 100 and 80 nm, respectively, as determined using a profilometer. The surface morphology of the films was studied by using tapping mode AFM (Asylum Research, Model: MFP-3D). Electrical measurements were performed using conductive atomic force microscopy (cAFM) with a conducting Ti/Ir-coated tip (ASYELEC.02-R2, Electric-lever, Olympus) having a radius of curvature ~ 25 nm, stiffness ~ 42 N m−1, and a resonance frequency ~ 285 kHz. During the cAFM measurements, the bias was applied to the bottom electrodes (BE), while the cAFM tip was kept virtually grounded. In order to investigate the resistive switching properties of the metal oxide films, local current–voltage (I-V) measurements are performed using cAFM technique by placing the cAFM tip on a randomly chosen spot on the film. In this process, the cAFM tip works as a top electrode and ITO (Platinized-Si) as a bottom electrode, fulfilling the basic metal/insulator/metal (MIM) configuration for a two-terminal memristive device. The I-V measurements are carried out on a large number of randomly chosen spots on the films, which exhibit similar behaviours.

In addition, Kelvin probe force microscopy (KPFM) was performed in dual-pass tapping mode with a user-defined lift height of 40 nm to estimate the local work function of the films. All the measurements were carried out in a closed enclosure having very low humidity [12].

3 Results and Discussion

3.1 Memristive Performance of CZO Films

Figure 1a depicts the I-V characteristics of the CZO-based memristors grown under the oxygen flow rate of 10 SCCM. The I-V curves are acquired on a large number of randomly chosen point locations on the sample surface. From the figure, the SET process, i.e. a transition from the initial high resistance state (HRS) to a low resistance state (LRS) and subsequent RESET process (transition from LRS to HRS) are distinguishable. However, the RS loop for the film grown in the absence of any oxygen partial pressure (i.e. pO2 = 0 mbar) disappears within 3 cycles. Although not stable, the other film grown under an O2 flow rate of 5 SCCM reflects improvement in the RS stability and a loop opening up to 20–30 cycles (data not shown). Interestingly, the device grown under 10 SCCM O2 flow shows an impressive RS behaviour for more than 100 cycles. The inset in Fig. 1a shows the variations in the SET and RESET voltages for 100 RS cycles, ensuring the excellent stability of this device. The introduction of O2 during the growth of CZO films compensates for the excessive defects present in the form of OV, making the film more resistive that helps in improving the stability of the RS process [20]. In this respect, KPFM can be a useful tool to estimate the local work function and therefore, the resistivity of the films, which is an important metric for the stability of a device and charge transport across the hetero-junctions involved [21]. From KPFM analysis, the work function is observed to rise from 4.25 to 4.43 to 4.53 eV for the films grown under 0, 5, and 10 SCCM of O2 flow rates, respectively. The systematic rise in work function, therefore, indicates an increase in the film resistivity (confirmed by four-probe measurements) when grown under increasing O2 flow rates because of the compensation of intrinsic OV defects that would otherwise increase the electron density in the conduction band of the film [20]. Here, it is important to mention that although the CZO films grown under higher O2 flow rates exhibit stable resistive switching behaviour, the respective operating voltages increase with increasing O2 flow rate (5.3–8–9.9 V for 10, 15 and 20 SCCM), which is not desirable for their practical applications. Therefore, the CZO films grown under 10 SCCM O2 flow can be considered as the optimized ones in terms of switching stability and operating voltages [Fig. 1a].

Fig. 1
figure 1

a Depicts I-V curves for 100 sweep cycles at a single point location on the CZO film grown under an oxygen flow rate of 10 SCCM. The inset shows variations in SET/RESET voltages for 100 cycles. b Surface topography, and c local current map obtained by applying a write voltage (−5 V) and then a read voltage of + 1 V. d VCPD map acquired after applying + 5 and −5 V successively to the AFM tip in contact mode, during KPFM measurements

Full size image

Henceforth, we make use of this film for further investigations. The surface morphology image acquired using tapping mode AFM of this film shows a granular structure with a decent root mean square roughness of 2.5 nm [Fig. 1b]. For a clear visualization of the read/write operation on the optimized CZO film, a cAFM current map is acquired by scanning a 2 µm × 2 µm region using −5 V. This is a write process that decreases the resistance of the scanned region up to LRS. Then, a bigger area (4 µm × 4 µm) was scanned with a read voltage of +1 V. The higher current values observed in the region of writing clearly represent a successful write/read operation into the memristor [Fig. 1c]. It is worth mentioning here that each nano-sized conducting grain detected within the LRS region in Fig. 1c resembles one end of the channel/filament that spans from below the AFM tip location through the CZO active layer up to the bottom electrode (BE) [22, 23]. The presence of such conducting grains in the cAFM current map points towards a filamentary RS mechanism within the device [24]. In another useful application, KPFM can be exploited to get insights into the nature of charge storage capability (unipolar to bipolar) of any memristor. Here, we have employed field-induced KPFM measurements to record the contact potential difference (VCPD) map of the film’s surface. As shown in Fig. 1d, the biasing with different polarity of voltages (here ± 5 V) results in a noticeable increase (or decrease) in the VCPD value in the respective regions, which confirms the bipolar charge storage capability of the CZO film under consideration [21, 22].

Next, we demonstrate the emulation of synaptic functionalities in the optimized CZO-based memristors by showing a possible analogy between a single electronic nano-synapse probed by cAFM and a biological synapse [Fig. 2a]. Here, the bottom ITO electrode works as a pre-synaptic neuron and the cAFM tip behaves as a post-synaptic neuron. Interestingly, the lateral scanning of the conducting tip inherently generates artificial nano-synapse (Ti/Ir tip/CZO/ITO) at the point of measurement on the film’s surface, the conductance of which can be modulated under successive input electrical stimulations of different amplitudes, widths, and intervals. Figure 2b depicts the current response under ten identical negative (−4 V) and positive (+4 V) pulses having width and interval of 200 ms and 10 ms, respectively. The observed modulation in current (∆I = IN–I1; N = 1 to 10 denotes the pulse number) under the negative (positive) pulses shows an excellent analogy with the potentiation (depression) of a bio-synapse where the synaptic current level enhances (reduces) with pulse number. The observed potentiation phenomena in the present CZO memristors can be explained by means of current augmentation in subsequent pulses of the same polarity. In contrast, if consecutive depressing spikes are applied, dissolution of the localized filament-like conducting chains reduces the current with successive pulse numbers.

Fig. 2
figure 2

a Schematic illustration of a CZO-based nanoscale electronic synapse in analogy to a biological synapse between two neurons. b The current response under 10 consecutive negative voltage pulses (−4 V, 200 ms) followed by 10 positive pulses (+4 V, 200 ms) with an interval of 10 ms, corresponding to potentiation and depression phenomena in a bio-synapse, respectively

Full size image

3.2 Memristive Performance of TiOx Films

Figure 3a, b depict the cAFM current maps (acquired at an applied bias of + 1 V) corresponding to the as-grown and Ar+-ion implanted TiOx specimen (to the highest fluence), respectively, where the current conduction sites are clearly observed, confirming the filamentary RS mechanism similar to the CZO films. Further, the current map reveals an overall improvement in conductivity of the films with increasing ion fluence, which can be attributed to the fluence-dependent increase in OV within the same [11].

Fig. 3
figure 3

a–b Show the current maps of as-grown and Ar+-ion implanted TiOx samples (to the highest fluence). c Variation in I-V characteristics and d operating voltages before and after implantation using different fluences

Full size image

To investigate the possible effects of ion implantation on RS performance, point I-V characteristics of as-grown and ion-implanted films are recorded with a cyclic voltage sweep from 0 →  + 5 → 0 →  − 5 → 0 V (ramp rate of 0.2 V s−1) using cAFM technique [Fig. 3c]. Interestingly, it is observed that with increasing ion fluence, the RS loops shift towards the lower operating voltages (cycle-to-cycle variation in operating voltages is indicated by the error bars) [Fig. 3d]. This implies that the ion-implanted films, particularly the ones implanted to the highest fluence, consume less power to execute RS operation. Similar to the CZO, these TiOx memristors do not require electroforming.

Next, to examine the feasibility of the developed TiOx memristors as electronic synapses, cAFM-based investigations are carried out following the same techniques as that of the CZO films discussed in Sect. 3.1. The variation in current under the application of a voltage pulse train (+ 1.0 V; width, Δd: 1 ms, and interval, Δt: 1 ms) for as-grown and ion implanted films are shown in Fig. 4a. It is observed that the potentiation behaviour of the memristor improves with increasing ion fluence, which is an important parameter to determine the rate of learning of a synaptic device [11]. The generated nano-synapse corresponding to the highest fluence exhibited superior potentiation/depression behaviour compared to the other ones. Further, the paired-pulse facilitation (PPF) behaviour of as-grown and the one implanted to the highest ion fluence (i.e. at 3 × 10–17 ions cm−2) are examined under pulse trains of varying intervals (amplitude: + 1.0 V, width: 1 ms, intervals: 1, 5, 10, 50, and 100 ms). The PPF is calculated as (I2—I1), where I1 and I2 are the current response of the first and the second pulse, respectively. As shown in Fig. 4b, the PPF decreases exponentially when the pulse interval increases, similar to the bio-synapses. The exponential fitting of the experimental PPF data yields the value of decay constants (τ, values are mentioned in the respective figure). Here, it is interesting to note that the rate of decay (reciprocal of the decay constant) is slow for the implanted film than that of the as-grown one. A slower decay indicates the ability of a device to achieve faster learning with fewer training cycles.

Fig. 4
figure 4

a Potentiation and depression against pulse number under identical positive voltage pulse train (+ 1 V, width and duration: 1 ms) followed by a negative one (−1 V, width and duration: 1 ms) for as-grown and ion implanted films. b PPF behaviour of as-grown and ion-implanted (ion-fluence: 3 × 1017 ions cm−2) films

Full size image

4 Conclusion

We demonstrate memristive synapses at nanoscale in CZO and TiOx− films exploiting different modes of AFM technique. A clear signature of defect modification either by changing in situ O2 flow (in case of CZO) or by ion implantation using different fluences (in case of TiOx) is inevitably present in all the acquired AFM data. The conductive mode AFM revealed forming-free and stable filamentary RS properties along with successful emulation of bio-synaptic functions in both types of films. On the other hand, KPFM analysis points towards reduced oxygen vacancy concentration within the CZO films that improves RS stability. The cAFM current maps suggest an increase in oxygen vacancy concentration after ion implantation, reducing the operating voltages, thereby making the RS operation of TiOx more power efficient. Therefore, our approach to demonstrate forming-free memristors with synaptic functionalities at nanoscale proves to be highly suitable for developing next-generation high-density, power-efficient neuromorphic devices on a chip.