Introduction

In the existing Von Neumann structure, information is exchanged through a data bus in the form of separate central processing units and memory. However, as the amount of data increases exponentially, memory bottlenecks become inevitable, preventing memory from maintaining the computational speed of the central processing unit [1, 2]. Therefore, the limitations of the von Neumann structure are revealed in algorithms that require the processing of vast amounts of data, such as artificial intelligence and deep learning [3]. Neuromorphic computing is a system that has been proposed to solve this problem. Neuromorphic computing mimics the biological functions of the human brain, and the implementation of synaptic devices and neuron circuits is essential [4,5,6,7,8,9,10,11]. Among these, synaptic devices have the most important analog characteristics that can express synaptic strength in various ways because of the number of signals or temporal correlation. The concept of a memristor was introduced to express these analog characteristics [12]. A memristor refers to an element whose resistance value changes according to an applied signal pulse and may also serve as a memory for storing the same. Memristor elements include ReRAM, magnetic random access memory (MRAM), and phase-change random access memory (PRAM), among which ReRAM has been actively studied as an artificial synapse device owing to its simple structure and high compatibility with complementary metal-oxide semiconductor (CMOS) [13,14,15,16,17]. HfO2 is a material with excellent reproducibility and repeatability and has long been studied in non-volatile storage memory applications as well as the high-k dielectric of CMOS [18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33]. An ideal synaptic memristor for neuromorphic computing should have high speed, endurance, and uniformity.

In this study, we propose a Ru/HfOx/TiN single layer ReRAM. Our proposed device showed nanosecond speed (20 ns, 50 ns) and stable retention characteristics (85 ℃, 2 h). We also showed ultra-low nonlinearity at the speed of nanoseconds and stable endurance that remains even after 50 k number of pulses. Based on these advantages, the recognition rate of 94.7% was confirmed in the MNIST-based object recognition simulation.

Experimental

The fabricated resistive switching (RS) devices exhibited a crossbar-like structure with a 16 um2 cell size. To fabricate the RS device, a 150 nm-thick TiN layer was sputtered on a SiO2/Si substrate and patterned to a shape of crossbar-like structure type bottom electrode (BE). We patterned TiN BE using photolithography and then deposited the HfOx layer using the physical vapor deposition (PVD) method. The 10 nm-thick HfOx layer was formed using the radio-frequency (RF) reactive sputtering method. The base and working pressures for sputtering were maintained at 5e−7 and 1e−3 torr, respectively. The oxygen flow rate for HfOx deposition was 6 sccm. After the HfOx layer formation, the Ru top electrodes (TE) were patterned using the same process as that used for the BE, and the TE and BE were crossed perpendicularly to make the crossbar-type device. All electrical properties of our RS devices were tested using a semiconductor parameter analyzer (SPA, Keithley 4200 SCS) and an arbitrary function generator (AFG, Agilent 81150A). The RF circuit-switching module alternately accesses the two-terminal electric circuits between the SPA and AFG. All measurements were performed at room temperature under an ambient atmosphere. A bias was applied to the top Ru electrode, whereas the bottom TiN electrode was grounded. The Ru/HfOx/TiN structure is shown in Fig. 1a.

Fig. 1
figure 1

a Structure of the Ru/HfOx/TiN device. b I-V characteristics of the device by a negative set and positive reset. c Retention characteristics of the HRS and LRS

Full size image

Results and discussion

DC bias sweep of 100 cycles were applied to the device to characterize the typical I-V curves for electrical analysis, as shown in Fig. 1b. It shows a negative set and positive reset during − 0.5 V ~ 1.5 V. DC bias sweeps were applied to the device after the forming process. As shown in Fig. 1c, the retention characteristics of the HRS and LRS states were evaluated to confirm the storage maintenance characteristics of the device at 85℃. Several tests, such as long-term potentiation (LTP) and long-term depression (LTD), have been conducted on Ru/HfOx/TiN devices to mimic the synaptic functions essential for artificial electronic devices. The microscopic structure of the Ru/HfOx/TiN film was characterized using high-resolution transmission electron microscopy (HR-TEM), as shown in Fig. 2a. Figure 2b shows the distributions of the components (Ru, Hf, O, Ti and N) obtained through energy-dispersive X-ray spectroscopy (EDS) mapping.

Fig. 2
figure 2

a Cross-sectional high-resolution transmission electron microscopy (HR-TEM) image of Ru/HfOx/TiN. b EDS elemental mapping

Full size image

The ReRAM is largely divided into filamentary and interface types. The filamentary type has advantages, such as low energy and fast switching on the nanosecond scale during programming [34,35,36,37,38,39,40,41,42]. In general, it rapidly changes to a low-resistance state when the filament is formed (set) and gradually changes to a high-resistance state when the filament is broken (reset). Analog switching characteristics must be used in a device; therefore, research is actively underway to use only the reset process, which causes gradual resistance changes, or to gradually change the rapid set process. The filamentary-type ReRAM is divided into oxide-based resistive memory (OxRAM), which is switched owing to defects in the oxide film, and conductive bridge random access memory (CBRAM), which is switched by metal ions. Unlike the filamentary-type ReRAM, the resistance of the interface-type ReRAM changes owing to the movement of oxygen ions at the interface between the metal and the insulator layer rather than the formation of a local channel in principle [43,44,45,46,47,48]. Therefore, because uniform switching characteristics are observed and the conductivity changes according to the amount of ion movement, it may have multiple resistance states. It also has the advantage of being able to secure the characteristics of gradually changing conductivity according to pulses. Our proposed device is an interface-type ReRAM caused by the movement of oxygen ions at the interface between the metal and insulator and shows gradual behavior in both set and reset operations. In addition, it differs from other interface types ReRAM in that it shows nanosecond-scale switching, which is an advantage of the filamentary type.

When potentiation and depression operations are performed, the linearity characteristics deteriorate if the conductivity of the device changes rapidly. In this case, the uniformity of the conductivity decreases and a negative effect on the accuracy of the pattern recognition rate becomes inevitable. The device may break down when abrupt resistance changes occur during the reset operations. To prevent this, it is common to set current compliance (C.C). However, in a pulse operation, it is impossible to set the current compliance itself; therefore, other methods, such as attaching an external load resistor, are required. In contrast, our device has the advantage of self-compliance without current compliance or an external load resistor. This self-compliance was due to the formation of oxygen reservoir at the interface between HfOx and TiN [49, 50]. TiON produced by oxidation at the interface serves as an oxygen reservoir. In addition, owing to the gradual set and reset operations, a relatively constant change in conductance can be expected during the LTP and LTD operations.

X-ray photoelectron spectrometer depth profile (XPS) analysis of identical material stacks with HfOx/TiN devices was performed to determine the microscopic role of the TiON layer in the device. Figure 3a shows a schematic of the sample structure used for XPS analysis. Figure 3b and c shows the variations in the in-depth XPS spectra of Ti 2p and O 1s, respectively. The direction of sputtering is denoted by an arrow on the right side of Fig. 3. As a result, the binding energy of Ti 2p decreased from 461.9 to 461.4 eV, and the binding energy of O 1s decreased from 530.6 to 530.1 eV. This analysis indicates that TiON was formed between the HfOx and TiN layers. The set behavior occurring in the negative region was also formed by this TiON layer, which can perform a gradual set operation by acting as an oxygen reservoir [51, 52].

Fig. 3
figure 3

X-ray Photoelectron Spectrometer depth profile analysis at the interface of HfOx and TiN. a Ti 2p and b O 1s peak

Full size image

Linear weight update for synaptic devices is important for achieving high learning accuracy in hardware neuromorphic systems. To evaluate the conductance updates in the LTP and LTD curves, nonlinearity (NL) was defined using the following equation:

$$\mathrm{NL}=\mathrm{average}\left(\left|\frac{\mathrm{G}-{G}_{\mathrm{Linear}}}{G}\right|\times 100 \%\right)$$

where G is the measured conductance value and GLinear is the conductance with an ideal weight update. The NL value of LTP was 2.32% and the LTD was 1.78%. We used the incremental pulse programming (ISPP) method to assess synaptic plasticity. ISPP refers to the process of increasing or decreasing the amplitude at a constant level at the same pulse width. As shown in Fig. 4a, 30 pulses were applied during the LTP operation. The pulse width was 20 ns, and the amplitude range was approximately − 1.45 to − 1.73 V. For the LTD operation, 28 pulses were applied, as shown in Fig. 4b. The pulse width was 50 ns, and the amplitude range was approximately 1.5–2.06 V.

Fig. 4
figure 4

Conductance level of the Ru/HfOx/TiN device to extract nonlinearity a LTP and b LTD

Full size image

The synapse, acting as a neuronal junction or bridge, controls the electrical or chemical signals between two neurons. Synapses regulate the connection strength and are involved in learning and memory in the human brain. Similarly, a metal–insulator–metal (MIM) structure can facilitate weight (conductance) adjustment through external pulse inputs. LTP and LTD are key features in adjusting neuronal synaptic plasticity. Figure 5a shows the 50 k pulse number endurance, which indicates stable operation. The inset image shows the 3-cycle LTP and LTD curves driven by each pulse. The read voltage was measured at 0.1 V after applying the set pulse (− 1.45 to − 1.75 V, 20 ns). For LTD, a reset pulse (1.5–2.06 V, 50 ns) was applied, and a read voltage of 0.1 V was used. The conductance values were tuned in 12 µS of 30 µS. As shown in Fig. 5b, it can be observed that there is little difference in amplitude value even in several LTP and LTD operations, which shows that it always has a similar conductance value and that data can be safely stored.

Fig. 5
figure 5

a The repetitive synaptic conductance variation characteristic of Ru/HfOx/TiN, inset image is LTP/LTD operation of the Ru/HfOx/TiN device. b Distribution of LTP/LTD amplitude authorized to ISPP

Full size image

The artificial neural network (ANN) perceptron structure using HfOx as synapse, which consists of 784 input neurons, 200 hidden neurons and 10 output neurons. The ANN structure of a typical perceptron consists of three layers (input, hidden, and output layers), as shown in Fig. 6a. The input layer receives a value and does not specifically calculate. Therefore, if the model is simply composed of an input layer and an output layer, its accuracy will not be highly evaluated. The accuracy was increased by adding a hidden layer to the artificial neural network. In this ANN architecture, the modified national institute of standards and technology (MNIST) dataset was used to learn handwritten digit classification. In the MNIST dataset, handwritten digit images are represented by grayscale 28 × 28 pixel arrays. In the ANN learning process, the LTP and LTD characteristics of the device were used as synapses for handwriting recognition. The classification accuracy was then evaluated using a test dataset at each epoch. We divided 8000 pieces of data into 100 pieces each, 80 iteration is required for 1 epoch, and 10 parameter updates occur. The resulting accuracy was 94.7%, as shown in Fig. 6b, indicating that the device successfully operated as a synaptic device, representing the synaptic weights required for neural networks.

Fig. 6
figure 6

a Schematic of the structural concept of the inference simulation. b The inference simulation result of the Ru/HfOx/TiN device

Full size image

Conclusion

In this study, we successfully fabricated an HfOx film-based ReRAM device using RF magnetron sputtering with an HfO2 target. Long-term plasticity was simulated through electrical synapses, including LTP and LTD, by applying continuous pulses to the top electrodes. The synaptic operation of 50 k was stably performed at a high speed of 20 ns and 50 ns during the long-term potentiation and depression operations, respectively. Based on this, a high recognition rate of 94.7% was achieved through the LTP or LTD characteristics with ultralow nonlinearity. Therefore, we present the possibility of using HfOx-based resistive memory devices in neuromorphic computing, particularly for tasks that require real-time processing and low power consumption.