Impact of oxide and electrode materials on the switching characteristics of oxide ReRAM devices

Resistive switching random-access memory (ReRAM) is one of the most promising technologies for non-volatile memories. Thanks to the low power and high speed operation, the high density CMOS-compatible integration, and the high cycling endurance, the ReRAM technology is becoming a strong candidate for high-density storage arrays and novel in-memory computing systems. However, ReRAM su ﬀ ers from cycle-to-cycle switching variability and noise-induced resistance ﬂ uctuations, leading to insu ﬃ cient read margin between the programmed resistive states. To overcome the existing challenges, a deep understanding of the roles of the ReRAM materials in the device characteristics is essential. To better understand the role of the switching layer material in controlling ReRAM performance and reliability, this work compares SiO x - and HfO 2 -based ReRAM at ﬁ xed geometry and electrode materials. Ti/HfO 2 /C and Ti/SiO x /C devices are compared from the point of view of the forming process, switching characteristics, resistance variability, and temperature stability of the programmed states. The results show clear similarities for the two di ﬀ erent oxides, including a similar resistance window and stability at high temperatures, thus suggesting a common nature of the switching mechanism, highlighting the importance of the electrodes. On the other hand, the oxide materials play a clear role in the forming, breakdown, and variability characteristics. The discrimination between the role of the oxide and the electrode materials in the ReRAM allows ReRAM optimization via materials engineering to be better explored for future memory and computing applications. In this work, we compare SiO x -based ReRAM and HfO 2 -based ReRAM on the same test vehicle, which is a one-transistor/one-resistor (1T1R) structure which enables a tight control of the programming current, namely the compliance current I C . A comprehensive comparison is conducted, covering the forming operation, the switching characteristics, and the retention behavior at elevated temperature. We also investigate cycle-to-cycle variability by means of the relative standard deviation, namely the ratio between the standard deviation s R and the resistance R , which is the standard metric for stochastic variation of programmed states in ReRAM. Our results indicate that the SiO x and HfO 2 ReRAM show remarkably similar device properties when compared on the same test vehicle. Similar performance includes the large resistance window and the good retention at elevated temperature. This performance similarity suggests that the ReRAM behavior is mostly dictated by the Ti TE and its migration controlling the  lamentary switching in the two devices. On the other hand, the forming, breakdown and variability characteristics show remarkable dependence on the oxide material, which enables discrimination between the role of the oxide and of the electrode materials in the ReRAM switching mechanism.


Introduction
In recent years, resistive switching memory (ReRAM) devices have emerged as strong candidates for next generation high-density storage technologies. [1][2][3] The resistance switching mechanism enables not only high-density non-volatile memory with high speed, high endurance, and low power, but can also be exploited for implementing novel computing paradigms such as in-memory computing and neuromorphic computing. [4][5][6] Currently, ReRAM technologies are facing signicant challenges due to switching variability and resistance uctuations, which require major breakthroughs at the levels of materials engineering and device physics. [7][8][9] In previous works, we proposed a silicon oxide-based ReRAM, in which an ultra-thin layer of SiO x is sandwiched between an inert carbon bottom electrode (BE) and an active titanium top electrode (TE). 10 Electrical switching between a high-resistance state (HRS) and a low-resistance state (LRS) was attributed to the oxidation of titanium in the TE and migration of Ti ions into the solid electrolyte, where they react to form a Ti-rich conductive lament (CF). This type of switching is equivalent to the well-known conductive bridge RAM (CBRAM), or electrical metallization cell (EMC). The adoption of a TE made of Cu and Ag was shown to lead to volatile switching, which might be applicable for selector devices in high density crosspoint arrays 11 and synaptic devices. 12 In the static regime, SiO x -based ReRAM shows an outstanding resistance window exceeding 10 4 . On the other hand, a ratio of about 10 is typically found in HfO 2 -based ReRAM, where switching is generally assumed to be due to the oxidation/reduction of metallic elements, known as the valence change memory (VCM) mechanism. 13 The larger resistance window of the SiO x ReRAM is in agreement with previous studies which attribute a larger window to CBRAM-type devices. 14 A high resistance window is essential in ReRAM to allow for sufficient read margin between the HRS and the LRS, ensuring a sufficient immunity from post-programming uctuations of resistance. 9 SiO x ReRAM has also shown a relatively low variability, an excellent cycling endurance, and an outstanding stability at elevated temperature, 10 compared with HfO 2 ReRAM. 15,16 To enable a high device reliability and a deeper understanding of the materials parameters controlling the resistance window, ReRAM materials should be compared at xed geometry and top/bottom electrode types.
In this work, we compare SiO x -based ReRAM and HfO 2 -based ReRAM on the same test vehicle, which is a one-transistor/one-resistor (1T1R) structure which enables a tight control of the programming current, namely the compliance current I C . A comprehensive comparison is conducted, covering the forming operation, the switching characteristics, and the retention behavior at elevated temperature. We also investigate cycle-to-cycle variability by means of the relative standard deviation, namely the ratio between the standard deviation s R and the resistance R, which is the standard metric for stochastic variation of programmed states in ReRAM. Our results indicate that the SiO x and HfO 2 ReRAM show remarkably similar device properties when compared on the same test vehicle. Similar performance includes the large resistance window and the good retention at elevated temperature. This performance similarity suggests that the ReRAM behavior is mostly dictated by the Ti TE and its migration controlling the lamentary switching in the two devices. On the other hand, the forming, breakdown and variability characteristics show remarkable dependence on the oxide material, which enables discrimination between the role of the oxide and of the electrode materials in the ReRAM switching mechanism.
2 Device structures Fig. 1 illustrates the two ReRAM structures studied in this work, namely the SiO x -based ReRAM (a) and the HfO 2 -based ReRAM (b). Both devices were deposited on pre-patterned substrates with metal-oxide semiconductor (MOS) eld-effect transistors (FETs) to enable the fabrication of 1T1R structures. Aer a 300 C vacuum pre-processing of the substrates, a thin dielectric lm was deposited by electron-beam evaporation from silicon monoxide (SiO x , x z 1) or hafnium dioxide (HfO 2 ) solid targets on top of a graphitic carbon bottom electrode (BE) of area 70 Â 70 nm 2 . 10 The Ti TE was deposited on top of the oxide layer without breaking the vacuum from the previous oxide deposition. The Ti TE had a thickness of 50 nm. A tungsten plug connects the BE to the drain of the FET to enable an integrated 1T1R structure. The HfO 2 lm was deposited with a thickness of t ox ¼ 5 nm or t ox ¼ 10 nm, whereas the SiO x lm thickness was t ox ¼ 5 nm.
Aer the fabrication process, the devices were electrically characterized in a probe station by means of a Keysight B1500A semiconductor parameter analyzer.
3 Forming and switching characteristics 3.1 Forming Fig. 2(a) shows the I-V curves measured on the pristine cell for 30 devices for t ox ¼ 5 nm. The voltage was applied to the TE and increased at a rate of about 1 V s À1 . The gate of the FET in the 1T1R structure was biased with a voltage V G in the range between 1 V and 1.6 V, corresponding to a compliance current I C between 1 and 70 mA, while the source was grounded.
Initially, the pristine device shows very low leakage current, limited by the instrument resolution for a TE voltage below 1.5 V. The current increases for increasing positive voltage, as a result of Poole-Frenkel conduction at localized states. Forming appears as a sudden step in the I-V curves, at a characteristic voltage V FORM , triggering a so breakdown which brings the device in the LRS. The forming operation marks the initial formation of a CF, the size of which is controlled by the maximum current I C through the device, which also prevents the irreversible breakdown of the ReRAM device upon forming. From the I-V curves in Fig. 2(a), one can notice a large difference in V FORM , which is about z4 V for HfO 2 and z7.5 V for SiO x . Fig. 2(b) shows the cumulative distributions of V FORM for the ReRAM devices, which also reveal a higher distribution spread for HfO 2 compared to SiO x . On the other hand, the leakage current I LEAK before forming is comparable in the two dielectric materials, as conrmed by the probability distributions in Fig. 2 The leakage current shows a larger distribution spread and higher noise in HfO 2 , possibly related to the higher instability of defects such as oxygen vacancies compared to SiO x . The lower V FORM in HfO 2 results in a lower forming current, which is about 1 nA in HfO 2 compared to 10 to 100 nA in SiO x . Such a result may be due to a smaller energy barrier for defect migration in HfO 2 , compared to SiO x , which is also supported by the higher current noise and the lower breakdown voltage (see Sec. 4.1). During the reset operation, the CF is disrupted thus resulting in a transition to the high resistance state (HRS) in Fig. 3(b). During the reset transition, initiated by the application of a negative voltage to the TE, positively-ionized defects (Ti impurities and/or oxygen vacancies) migrate towards the TE in response to the applied eld, thus leaving behind a highly-resistive depleted gap of length D. 17 The application of a positive voltage to the TE induces the migration of defects from the TE into the depleted gap, leading to the reconnection of the CF. Fig. 4 shows the measured I-V curves for 50 cycles for HfO 2 with t ox ¼ 10 nm (a) and SiO x with t ox ¼ 5 nm (b). Aer the forming process the ReRAM device can be   positive feedback results in a self-accelerated dynamics of the CF growth at the origin of the abrupt set transition. 13,18 During the set process the gate of the transistor is biased to V G ¼ 1.6 V to limit the maximum current to I C ¼ 80 mA. The current limitation serves as an external negative feedback mechanism to prevent irreversible breakdown of the device. Moreover, the maximum current through the device controls the size of the CF, 19 yielding an inverse proportionality between compliance current and LRS resistance R f 1/I C . The stop voltage was the same in the two sets of experiments, namely V stop ¼ À3.5 V. HfO 2 and SiO x ReRAM characteristics show comparable switching behavior, indicating an average resistance window of 5 Â 10 3 in SiO x and 3 Â 10 3 in HfO 2 . A signicant difference lies in the HRS variability, which is signicantly higher in HfO 2 . This is further conrmed by the cumulative probability distributions of the resistance values measured at À0.5 V in Fig. 4(c) and (d) for HfO 2 and SiO x , respectively. The resulting relative standard deviation of the HRS is s R /R z 1 in the SiO x stack and s R /R z 5 in HfO 2 . The larger HRS variability can be interpreted by a higher distribution of energy barriers for defect migration controlling the variation in the reset characteristics. 8 The nature of resistive switching in SiO x is attributed to the migration of Ti cations from the TE into the switching layer, resulting in the formation of a Ti-rich CF, 10 and/or to the migration of oxygen vacancies leading to silicon nanoinclusions constituting the CF. 20 Resistive switching in HfO 2 is typically attributed to oxygen vacancies, although migration of metallic cations from the active TE into the metal oxide has been postulated 19 and experimentally demonstrated. 21 However, the similarity of the switching characteristics in Fig. 4 in terms of shape of the I-V curve and resistance window suggests that the SiO x and HfO 2 share a similar microscopic switching mechanism, which might be the migration of Ti from the TE electrode controlling the CF formation and disruption.

Impact of V stop
During the reset process, the resistance of the ReRAM gradually increases with the applied negative voltage, as a result of the inherent negative feedback mechanism regulating the growth of the depleted gap. 13,18 As a consequence, the maximum applied negative voltage V stop controls the length of the depleted gap and thus plays a key role to control the HRS and thus the resistance window.
The dependence of the HRS on V stop was studied for t ox ¼ 5 nm and t ox ¼ 10 nm HfO 2 and for the reference SiO x stack, and the results are reported in Fig. 5. Fig. 5 shows the measured I-V curves (a) and the resistance values for the LRS and HRS as a function of V stop (b) for a HfO 2 ReRAM with t ox ¼ 5 nm. V stop was changed every ve set-reset cycles from V stop ¼ À2.2 V to À3.4 V, while the compliance current was kept constant at I C ¼ 50 mA. The results show that V stop controls the HRS, hence the resistance window, while the LRS remains almost unaffected by V stop . The HRS increases exponentially at increasing V stop with a slope of 1.2 dec V À1 . For V stop ¼ À3.4 V, the device undergoes breakdown under negative voltage, which was already shown to control endurance lifetime under pulsed set/reset cycling experiments. 15 In the breakdown phenomenon, the current suddenly increases, usually resulting in a LRS with an extremely low resistance due to the I C not being directly limited by the series FET. More results about the negative-voltage breakdown will be given in Sec. 4.1. Fig. 5 also shows the I-V curves (c) and the resistance values (d) for the HfO 2 ReRAM with t ox ¼ 10 nm. No breakdown phenomenon is seen at negative voltage, even up to V stop ¼ À4 V, possibly thanks to the thicker oxide layer limiting the electric eld within the oxide layer. The resistance window could be effectively increased to more than 10 4 at the highest V stop .
Data for SiO x (t ox ¼ 5 nm) are also shown for comparison, including I-V curves (e) and resistance values (f). The resistance values are similar to the HfO 2 device with t ox ¼ 10 nm, except for the tighter control of the HRS which is promising for multilevel operation of the memory. Also notice that V set increases in general with the V stop adopted in the previous reset cycle, paralleling the increase of the HRS resistance. Excessive values of V set might be detrimental for ReRAM operation, which relies on low voltage operation, which suggests that a tight control of the HRS and the associated V set is essential in ReRAM devices.

Switching variability
Statistical variability of ReRAM parameters, including the HRS/LRS resistance, V set and V reset , is crucial for the memory performance and the read window margin. To study the switching variability in the two materials, we collected cycle-to-cycle distributions of the HRS and LRS and evaluated the corresponding standard deviation s R and the average resistance R for each distribution at various conditions of I C and V stop . Fig. 6 shows the relative standard deviation s R /R as a function of R for the HfO 2 ReRAM (t ox ¼ 10 nm) and the SiO x ReRAM (t ox ¼ 5 nm). 22 In the gure, LRS data were obtained at variable I C while HRS data were obtained at variable V stop . The LRS variation shows similar behavior in the HfO 2 and SiO x ReRAMs, indicating a linear increase of s R /R with R. The latter is consistent with a variability model based on the stochastic variation of the CF shape from cycle to cycle. 23 A lower LRS variability was found in HfO 2 for relatively large resistance values (R z 10 5 U), namely for low I C . HRS variation shows a slightly different behavior in the two materials, with HfO 2 indicating an increase of s R /R with R approximately given by R 0.5 , in agreement with previous variability models based on the Poissonian statistics controlling the number of defects in the depleted gap. 7,8 On the other hand, SiO x ReRAM shows an almost constant s R /R z 1 even for the highest R values around 1 GU. These data thus support the superior performance of SiO x ReRAM in terms of cycle-to-cycle variability, which plays a key role in the ReRAM memory operation and yield. Fig. 7(a) shows I-V characteristics of HfO 2 ReRAM devices with t ox ¼ 5 nm for negative voltage, indicating the typical reset process of decreasing current, and the anomalous breakdown process where the current suddenly increases at a stochastic breakdown voltage V BD . Breakdown occurs during the negative voltage sweep and leads to an LRS with extremely low resistance. Similar breakdown phenomena were observed in 1T1R structures during pulsed cycling and Fig. 6 Relative resistance spread s R /R as a function of the resistance R for 10 nm thick HfO 2 compared to SiO x data in ref. 22. The data show comparable LRS variability, while the HRS is much more stable in SiO x , showing a larger window for the same relative spread, and a weaker dependence of s R /R on R.

DC breakdown
attributed to rupture of the BE interface and a consequent defect injection from the BE. 15 Negative-set cycles are generally destructive events, causing the failure of the memory device. In fact, during the application of negative voltages the gate of the integrated series transistor is generally biased to a relatively large voltage to minimize the voltage drop across the series transistor. This results in an overgrowth of the CF during negative-set, leading to a stuck-set state. To prevent irreversible breakdown in Fig. 7(a), the current meter was set to force a compliance current of 1 mA, thus much larger than the typical operating currents I reset z I C of about 80 mA.
While HfO 2 ReRAM with t ox ¼ 5 nm systematically shows breakdown, the same is not observed in the SiO x ReRAM with the same oxide thickness, even when relatively large negative voltages are applied, e.g., V stop ¼ À10 V in Fig. 7(b). This might be attributed to the higher energy barrier for defect migration in the SiO x layer, compared to HfO 2 , which is also consistent with the higher V FORM of SiO x in Fig. 2.

Retention at elevated temperature
A key requirement for non-volatile memories is data retention at high temperatures, which might be expected in several embedded memory circuits for automotive and industrial applications. For instance, the reliability specications of embedded ReRAM require that data remain stored aer a high temperature bake annealing at 260 C for a few minutes. 24 Due to their lamentary storage, ReRAM devices can be affected by temperature-induced resistance changes, as a result of defect diffusion causing either disruption of the CF or closure of the depleted gap. 16 A careful study of retention at elevated temperature is therefore crucial for validating and comparing ReRAM materials.
We studied data retention by cumulative 1 hour-annealing processes at increasing temperature T A . Multiple HfO 2 and SiO x devices with t ox ¼ 5 nm were initially prepared in various LRS and HRS states by changing I C and V stop respectively. Aer the initial resistance measurement, each device was annealed for 1 hour at T A ¼ 120 C. The process was repeated at increasing temperature steps of 20 C up to a maximum annealing temperature T A ¼ 260 C. Aer every annealing step, resistance measurements were carried out at room temperature T 0 to avoid T-induced conductivity variations. This article is licensed under a Creative Commons Attribution 3.0 Unported Licence. Fig. 8 shows the measured resistance as a function of the annealing temperature T A for the HfO 2 material in (a), and for the SiO x in (b). The data show relatively good temperature stability for both materials in both the LRS and HRS. The latter shows larger resistance uctuations, possibly explained by the redistribution of defects within the depleted gap. However, no data loss is observed, which supports the strong reliability of SiO x ReRAM devices.
The data for HfO 2 ReRAM also show a relatively small change upon temperature stress, although a smaller resistance range was explored in this case to prevent negative breakdown (see also Fig. 5(a)). Overall, the strong stability at elevated temperature observed in Fig. 8 further supports the interpretation of a common switching mechanism in HfO 2 and SiO x ReRAM, where Ti-migration is responsible for the growth and disruption of the CF. Our results thus support a strong role of the TE and BE materials in dictating the switching and reliability of ReRAM devices.

Conclusions
This paper shows a comprehensive comparison of SiO x -and HfO 2 -based ReRAM devices, aiming at discriminating between the impact of the oxide and electrode materials in the switching and reliability performance of the device. The study is carried out for the same device geometry, electrode materials, and fabrication process, by just changing the switching layer material. The forming characteristics are considerably different in the two dielectrics, possibly highlighting a difference in energy barriers for defect migration. On the other hand, clear similarities are observed for the static I-V curves; in particular, the similar resistance window suggests a central role of the TE during resistance switching. The annealing experiment evidences good stability at high temperature for both HfO 2 and SiO x . Differently, switching variability characteristics are radically different in the two oxides, showing better performance for the SiO x device.

Conflicts of interest
There are no conicts to declare.