Keywords

1 Introduction

Large resistance change with voltage application was widely known in Perovskite type complex oxides , binary oxides such as NiO, TiO2, HfO2 and Ta2O5, and solid electrolyte (e.g. Ge-S) with Cu or Ag. In recent years, many works have been reported to apply this phenomenon on non-volatile resistive RAM s (ReRAMs) and on neuromorphic devices [1–4]. To guarantee the operation of these devices, clarification of the switching mechanism is strongly required. Based on the electrochemical discussion of their electric properties, it is a common knowledge that the resistance switching in binary oxides and solid electrolytes is caused by formation and rupture of the conductive filament [1, 2]. However, details of its operation mechanism are still ambiguous and difficult to be clarified only using electric properties. Dynamical observations in real space using in-situ transmission electron microscopy (TEM) are a method to overcome this difficulty, where the electric measurements and TEM observations can be performed synchronously [5–10]. Several types of in-situ TEM holders became commercially available in recent years, and the number of reports in this category has increased. However, works on realistic switching operations such as the current-voltage (I-V) switching cycle, the switching repetition and the pulse endurance are rare [11, 12].

In this report, our recent works on ReRAMs using the in-situ TEM are reviewed, which are categorized as the conductive bridging RAM (CBRAM) such as GeS-Cu [13, 14], Cu/MoOx [11, 15, 16] and Cu/WOx [17, 18]. In all of these CBRAM systems, the Cu conductive filament was experimentally confirmed to appear in the Set process giving the low resistance state (LRS, on-state) and to disappear in the Reset process giving the high resistance state (HRS, off-state). No drastic change in the geometry of the conductive filament was seen at the Set/Reset switching moment, and the filament growth/erasure was seen during continuation of the current flow.

2 Experimental Procedure

Three types of ReRAM s were used for investigation, which were composite GeS-Cu(8–60 nm) deposited on a wedge-shaped PtIr, Pt(100 nm)/Cu(30 nm)/MoOx(50nm) or Pt(100nm)/Cu(30nm)/WOx(20nm) trilayer films on a TiN/Si wafer chip. For all samples, the film deposition was done using conventional RF sputtering (Ar) or reactive sputtering (Ar-20 %O2) of metal targets at room temperature (RT). In the latter two samples, the Cu and the TiN layers worked as the top electrode (TE) and the bottom electrode (BE), respectively. The GeS-Cu sample was used for in-situ TEM without any further processing, because the tip of the PtIr wedge was thin enough for TEM observations. On the other hand, the ReRAM samples of MoOx and WOx were processed using the ion-shadow method [19] which is a kind of the ion-milling method.

In-situ TEM experiments were carried out using the system shown in Fig. 1 composed of a hand-made piezo-driven TEM holder and a JEM-2010 microscope (C s = 0.5 mm). A movable Pt-Ir probe was contacted to one of the ReRAM devices formed on the TEM sample, and the I-V switching cycles were measured using a commercially available source-measure-unit (SMU, Yokogawa GS610). To prevent eternal device destruction caused by the strong overshoot current in the Set process, current limitation (compliance current; I comp) was introduced using this SMU. In the GeS-Cu sample, the wedge-shaped PtIr substrate was biased while the probe was grounded. On the other hand, in MoOx and WOx samples, the TiN/Si substrate was grounded while the probe was biased. Geometrical change of the filament was dynamically recorded using a CCD camera (30 ms/frame).

Fig. 1
figure 1

In-situ TEM system

Full size image

3 Results and Discussion

3.1 Filament Formation in Cu-GeS

The I-V curve measured in TEM is shown in Fig. 2, where the horizontal axis corresponds to the potential of the substrate relative to the probe [13, 14]. The voltage was varied in a sequence of 0 V, +7 V, −2 V and 0 V, and a clear hysteresis curve was obtained. Though the retention time (memory holding time) of the LRS was not long enough for the practical memory application, formation and erasure of the conductive filament was clearly confirmed.

Fig. 2
figure 2

I-V curve of Cu-GeS

Full size image

The TEM images extracted from the video are shown in Fig. 3. There are a Cu-GeS amorphous film containing Ge nanocrystals on the Pt-Ir substrate (Fig. 3a). In the first quadrant of the I-V graph, where the substrate was positively biased relative to the probe, the current increased gradually. There was drastic increase at the voltage of +2.6 V, and its resistance was converted to the LRS (Set operation). Correspondingly to this change, a filamentary dark contrast appeared (Fig. 3b). In the selected area diffraction (SAD) pattern during the Set process, Debye rings of Cu appeared. In addition, the energy dispersive X-ray spectroscopy (EDX) measurements were done during the Set process with positive bias voltage to the substrate (but on another filament than that in Fig. 3). In Fig. 4, the EDX spectra before (Fig. 4a) and during (Fig. 4b) the Set operation are compared. It was clearly identified that the deposit contained much Cu. Considering these two results, the deposit appeared in the Set process was concluded to be made of Cu nanocrystals. Afterwards, a negative voltage sweep was done, and the deposit gradually disappeared and completely erased at about −2.8 V (Fig. 3c, d). The resistance showed a sudden increase at this moment. This is the Reset process which converts the LRS to the HRS. The deposit observed here worked as a conductive filament for the CBRAM switching.

Fig. 3
figure 3

Microstructure corresponding to Fig. 2. Video a in the initial state, b and c with positive or negative voltage, and d after the I-V cycle

Full size image
Fig. 4
figure 4

EDX spectra a without and b with positive voltage application. The deposit contained much more Cu than other regions

Full size image

Repeating switching cycles, the Set voltage (V set) gradually decreased. This indicates that the Cu-GeS layer in the HRS after the Reset operation changed from the initial state as confirmed in Fig. 5. First, a filament was formed (Fig. 5a) and erased. Afterwards, the Pt-Ir probe was shifted by 100 nm (Fig. 5b left panel), and a positive voltage was applied to generate a filament again (Fig. 5b right panel). The filament appeared from the probe toward the position of the first filament. It did not reach the nearest position on the substrate. The region in which the filament was formed once must be a priority position for the formation and the growth of a filament. It is expected that there are some residuals such as Cu nanocrystals even after filament erasure, which can behave as filament nuclei. This is thought to make the filament formation easy and to induce reduction of V Set.

Fig. 5
figure 5

Filament growth in sequential switching cycles (a and b), where arrows indicate corresponding position. In each figure, the left panel is before the switching cycle, and the right panel is with positive voltage. The filament appeared almost in same place

Full size image

The results mentioned above suggest that the Cu filament is formed/erased by application of positive/negative voltage, and the CBRAM operation is achieved. Checking the TEM video, the filament grew from the probe (negative voltage side) with positive voltage, and it was ruptured and erased from the substrate (negative voltage side) with negative voltage. This dynamics fits to the electrochemical model (ECM) reported so far, where the Cu ions drifted along the electric field are reduced to be the metal by receiving electrons on the cathode surface, and the Cu filament is formed towards the anode [1]. However, the current was small (<1 μA), and the retention property of the LRS was not good enough. The probe used as the BE was mechanically contacted to the solid electrolyte switching layer (Cu-GeS), and this is different from the realistic CBRAM device. To investigate more realistic switching operation, a Cu/MoOx/TiN trilayer CBRAM will be discussed in the next subsection.

3.2 Set/Reset Switching of Cu/MoOx/TiN

The sample was Pt/Cu/MoOx/TiN/Si where Cu and TiN were the TE and BE, respectively [11, 15, 16]. The device size was 350 nm in diameter. The electric properties were measured between the Pt cap layer (biased) and the Si substrate (grounded). An example of the I-V curve is shown in Fig. 6a, which was the second switching cycle after the initialization process called as Forming. A shaper Set than in Cu-GeS is seen in the first quadrant. The corresponding TEM images extracted from a video are presented in Fig. 6b. The resistance state was the HRS between the states A and B, and there was no clear change in the image. Just after the Set switching (state C), a small dark region (nucleus of the filament) appeared near the TiN BE. Though the current was almost the compliance value (I comp), there was no dark region binding the TE and BE. The ionic current of the Cu anion is thought to contribute to electric conduction. Afterwards, the filament grew from the TiN BE towards the Cu TE. The bridging of the filament between the TE and the BE was completed within 200 ms.

Fig. 6
figure 6

a I-V characteristics of Cu/MoOx/TiN and b TEM images during in-situ observation

Full size image

The Reset operation was started at 5 min. after the Set process was completed. The LRS was kept after this interval, because I comp was large (400 μA). Two small current jumps between −1 and −2 V in Fig. 6a correspond to the Reset switching. The Reset curve was uneven as frequently observed in conventional CBRAM devices. This may suggest that Reset is more complicated process than Set. In the corresponding TEM image (Fig. 6b), there is no clear structure change just after Reset switching (state F). This switching is thought to occur very locally, and it was not identified in the TEM image. Increasing negative current to about −300 μA (at −2.6 V), the filament was ruptured near the interface between MoOx and TiN and shrank toward the Cu TE (states G). This process with additional negative current is called the over-Reset process in this report. The filament erasure here cannot be explained by the ECM model where the filament shrinks from the Cu TE to the BE. In the present work, the surface of the TiN BE was oxidized. The resistance at this oxidized region is high, and the local temperature must increase by Joule heating. Thus, it is thought that the electrochemical reaction of Cu in this region preferentially occurred, and the Cu filament shrank toward the TE. The current of 100 μA (or more) was used for switching operation in many CBRAM devices reported earlier. While the details are still obscure, it must be sure that the Joule heat influences the Reset operation.

In this paragraph, the filament position and size will be discussed. Images after Set of five switching cycles are compared in Fig. 7, where I comp was increased stepwise. Here, the filament was almost completely erased by performing over-Reset. The initial state was the HRS, and the contrast in Fig. 7a is due to the matters such as the outer shape of the specimen. In the first switch with I comp = 200 μA, a filament having a width of 7 nm was formed (Fig. 7b). By increasing I comp, the filament width increased to 35, 45 and 100 nm in Figs. 7c–e respectively. This increase can be used for the multi-bit memory operation. However, the filament position changed as seen in these figures. This position instability is thought to be caused by almost complete filament erasure during over-Reset.

Fig. 7
figure 7

TEM images of Cu/MoOx/TiN. a Initial state and b–e LRS with increasing I comp

Full size image

3.3 Device Degradation of Cu/WOx/TiN

The investigated sample was Pt/Cu/WOx/TiN/Si [17, 18]. Examples of I-V graphs measured during in-situ TEM are shown in Fig. 8a–c where the horizontal axis denotes the potential of the Cu electrode relative to TiN/Si. Figure 8a is the Forming curve. There was a current jump at 2.4 V. This resistance change to the LRS was much quicker than in MoOx. Correspondingly, a filament was formed between the Cu and TiN electrodes within one video frame (30 ms). The LRS was kept just before the Reset operation which was performed at 9.5 min after Set, and the retention time was long enough for in-situ TEM experiments. By carefully checking the video during the Reset operations, clear filament erasure was not identified though sharp switching to the HRS was realized. No strong over-Reset was adopted in this switching series as seen in Fig. 8b, c, and thus the filament was not erased (but with switching to the HRS). This result suggests that ReRAM switching occurs in local area of the filament. The resistance change in switching repetition (endurance graph) is shown in Fig. 8d. The resistance ratio of HRS/LRS was kept about 100, and a reasonable CBRAM endurance property is seen. However, as seen in this graph, the resistance in the HRS gradually decreased with the switching cycle. This degradation is known also in realistic ReRAM devices, and it has been reported to induce the endurance failure where the ReRAM switching cannot be achieved [20, 21]. This is believed to be caused by imbalance between Set and Reset. In Fig. 9, TEM images after Set/Reset switching cycles are compared. During switching repetition with increasing I comp, Cu moved into the WOx layer in a wide area, and the effective thickness of the WOx layer became thin. This is thought to be the origin of the endurance failure.

Fig. 8
figure 8

a–c I-V curves during in-situ TEM of Cu/WOx/TiN and d the cyclic endurance where the resistance was estimated at ±0.3 V

Full size image
Fig. 9
figure 9

TEM images of Cu/WOx/TiN. a Initial state and b–e LRS after the continuous Set operations with increasing compliance current

Full size image

4 Conclusion

In this work, in-situ TEM was applied on three CBRAMs to investigate the microstructure during the switching operation. It was experimentally confirmed that the Cu filament is formed in Set, ruptured in Reset and erased in over-Reset. The filament becomes thick with increasing the operation current, but its position becomes instable. It was also confirmed that Cu moved into the solid electrolyte and thinned the switching layer. This reduces the resistance of the HRS and finally induces the endurance failure. The retention property must improve by increasing the Set current, and large resistance ratio of HRS/LRS is expected by over-Reset. However, these treatments induce instability of the filament position and movement of excess Cu in wide area. These can be origins of device failure. Selection of materials and the device structure for easy Set/Reset balancing must be important for further development of ReRAM devices.