RF-wave induced CBRAM characteristic modulation specific to sheet-like conductive filaments

The shape of conductive filaments in CBRAM is important for resistance switching and conductance modulation, especially in applications like neuromorphic and reservoir computing that use conductance as weight. We report on RF-induced modulation of CBRAM using Ge2Sb3.4Te6.2 with sheet-like filaments and compared it to those with dendritic filaments. RF input below 100 MHz reduced SET and RESET voltages, similar to CBRAM with dendritic filaments, but showed significantly different resistance changes. Repeated RF on/off input gradually increased the resistance of low-resistance state, unlike the dendritic filament CBRAM, where the resistance decreased. The increased resistance suggests RF-induced denser sheet-like filaments. Furthermore, the resistance of the high-resistance state showed a peculiar RF-induced resistance change not observed in dendritic filaments. The resistance decreased during RF input and increased to nine times the initial value when RF was switched off. The results show that the conductance modulation by RF input strongly depends on the filament type.


Introduction
Conductive bridge memory (CBRAM) is a type of nonvolatile electrochemical metallization device among resistance random access memory (RRAM) devices, and has the advantages of scalability, fast operation speed and multi-value storage [1][2][3][4].CBRAM has also attracted attention in recent years as a device for neuromorphic system [5,6] and for reservoir computing system [7][8][9][10].CBRAM's rich dynamics and nonlinearity appear to make it a viable candidate for new neuromorphic or reservoir computing systems that are currently being developed to replace traditional computing architectures to cope with increased demands for data processing [11][12][13][14].CBRAM also offers efficient radio frequency switching solutions.Electrical resistance control programming of CBRAM RF switches could benefit the development of electronically controlled nonvolatile RF attenuators and other reconfigurable RF devices [15,16].
Ag-chalcogenide-based CBRAMs have been widely studied because of their low set/reset voltages, large memory window, good retention characteristics [17][18][19], and multiple resistive states suitable for use as analog synapses in neuromorphic systems.Optimizing the conductance linearity response to multiple pulse cycles of Ag-Ge-Se-based CBRAM synaptic devices has demonstrated improvement in the recognition accuracy of neural networks [6].GeTe-based CBRAMs with low forming voltage, fast switching speed and self-compliance properties have been used in both memory and neuromorphic computing applications [20][21][22].CBRAM characteristics, with these intriguing applications, rely on the formation and rupture of conductive filaments.Thus, controlling the filament shape and elucidating the relationship with CBRAM properties holds great significance for neuromorphic computing, reservoir computing, and RF device applications.Systems that imitate conductive filaments as biological synapses have shown a strong correlation between filament shape and short-and long-term plasticity [23].However, the relationship between conductance modulation by RF waves and filament shape has not been investigated at all.In this paper, we study RF-induced resistance modulation of a CBRAM with Ag-doped Ge 2 Sb 3.4 Te 6.2 with sheet-like conductive filaments.This study is the first to show that the conductance modulation by RF waves intrinsically depends on the filament shape.[26] although in other chalcogenides like GeTe with dendritic filaments precursor filament growth prevents reproducible Faraday current observation.In this study, we demonstrate significantly different RF-induced CBRAM modurations between Ag-Ge 2 Sb 3.4 Te 6.2 with sheet-like filaments and Ag-GeTe with dendritic filaments, despite similar basic CBRAM properties like the SET/RESET voltages and the resistances of low and high resistance states (LRS/HRS).We discuss the CBRAM characteristic modulation specific to sheet-like filaments in terms of RF-enhanced Ag migration.

Fabrication and measurement of CBRAM
Our CBRAM comprises the Ti/Au/Ge 2 Sb 3.4 Te 6.2 /Ag/Pt/Ti/Au multilayer structures [figure 2(a)].First, the bottom electrode and contact pad of Ti/Au (10 nm/50 nm) was deposited on a 300 nm thick oxidized Si wafer substrate by using electron-beam evaporation with background pressure better than 3 × 10 −5 Pa, and patterned by using photolithography and the lift-off process.The SiO 2 layer was inserted to ensure no leakage current to the Si substrate.Then, a 50 nm thick Ge 2 Sb 3.4 Te 6.2 film, a 50 nm thick Ag layer, and a 20 nm thick Pt to electrode layer were deposited by room temperature RF magnetron sputtering at the power of 50 W with background pressure better than 3 × 10 −5 Pa, and patterned to form a square area of 50 × 50 μm 2 .Finally, the Ti/Au (10 nm/50 nm) top contact pad was formed by electron beam deposition.The DC I-V characteristics during RF input were measured using two bias-Ts, a GSG probe [figure 2(b)], and a synthesizer to supply the RF input.Spurious harmonics from the synthesizer were suppressed by a low-pass filter to less than −60 dBm.The power of the harmonics generated by the CBRAM was measured using a spectrum analyzer.The S-parameters of the CBRAM were measured by a network analyzer.

Results and discussion
3.1.Endurance and retention Figure 3(a) shows the endurance of the Ag-Ge 2 Sb 3.4 Te 6.2 CBRAM.The currents of HRS and LRS read at 10 mV show no obvious resistance degradation in Ag-Ge 2 Sb 3.4 Te 6.2 for over 10 4 switching cycles.Figure 3(b) shows the retention during the RF input of 150 MHz at 5 dBm.The negative HRS current near 0 V during RF application can be explained by the higher presence of Ag ions induced by the RF wave, leading to an increased Faraday current.As in cyclic voltammetry, the Faraday current for the return of the oxidation reaction is negative because a reduction process begins before the origin.Both current at LRS and HRS were reduced by RF application compared to before RF application; the details of the effect of RF waves are described in section 3.3.Even during the RF input and at 100 °C, the currents at LRS and HRS remained the same after over 10 4 seconds.The endurance and the retention were enough to study the effect of RF input in the CBRAM.

S-parameters and frequency multiplication
The LRS resistance of our CBRAM without RF application was approximately 60 Ω, and the HRS resistance was approximately 1 kΩ, which is quite low for a CBRAM and advantageous for impedance matching with an RF source.In order to investigate the RF transmission characteristics of the CBRAM, we measured the S-parameter of the Ag-Ge 2 Sb 3.4 Te 6.2 based CBRAM using a network analyzer with a ground-signal-ground RF probe.Figure 4(a) shows the RF transmittance, reflectance, and absorption of the CBRAM obtained from the S-parameters, as The transmittance of the device without the Ag-Ge 2 Sb 3.4 Te 6.2 layer was around 0.8 in all measured frequency ranges from 300 kHz to 5.5 GHz, ensuring that the CBRAM electrode design allows for RF transmission.With Ag-Ge 2 Sb 3.4 Te 6.2 layer, the transmittance was over 0.5 (S 21 > 3 dB) in both LRS and HRS when the RF frequency was higher than 500 MHz.The data above 500 MHz is to show the transmission characteristics of this device for RF and to demonstrate that this device structure allows RF waves to be applied into the device with low reflection when RF waves are not absorbed or modulated by Ag ion dynamics.The main RF frequency that affects the CBRAM characteristics is below 200 MHz.Below 500 MHz, the transmission decreased significantly with decreasing frequency, and the reflection and absorption gradually increased.Below 500 MHz, the transmittance, reflectance, and absorption weakly depended on the state of resistance; at 100 MHz, the absorption of HRS was about 0.5 and the absorption of LRS was about 0.6.The nonlinear response of Ag-Ge 2 Sb 3.4 Te 6.2 to an incident RF wave was evaluated by measuring the frequency multiplication.
Figure 4(b) shows the measurement system of the frequency multiplication.The details of the measurement method are described in [27].Figure 4(c) displays the frequency multiplication spectrum, and figure 4(d) shows the frequency dependence of the harmonics output.Above 200 MHz, both outputs from the LRS and HRS outputs consisted mainly of fundamental harmonics.As the frequency decreased, particularly below 100 MHz, the power of harmonics of HRS and LRS increased.At 20 MHz, HRS generated fifth-order harmonics while LRS generated second-order harmonics.The frequency dependence of the harmonic power being qualitatively similar to the that of  absorption in figure 4(a) suggests that the Ag-Ge 2 Sb 3.4 Te 6.2 layer partially absorbed RF waves below 100 MHz to cause a nonlinear response to the RF waves.The increase in absorption below 100 MHz was larger, and the orders of generated harmonics were higher than those of Ag-GeTe with dendritic filaments although the frequency dependence of the S-parameters and the frequency multiplication were qualitatively similar.In Ag-GeTe, the absorption and nonlinear response of the RF wave at low frequencies were attributed to the dynamics of the Ag ions; at frequencies lower than a time constant, Ag + ions follow changes in the electric field of the RF wave, inducing absorption and nonlinear response, and resulting in migration enhancement.The resistance state-dependent S-parameters and frequency multiplication could not be explained by differences in DC resistance only, but were attributed to differences in the RF wave electric field applied to the Ag ions between the connected and ruptured filaments.Therefore, similar Ag + ion dynamics occurred more actively in Ag-Ge 2 Sb 3.4 Te 6.2 than Ag-GeTe.

RF-induced modulation of CBRAM characteristics
The change in CBRAM characteristics due to RF application was measured by I-V measurements during the onoff of RF waves, as shown in figure 5(a).The frequency of the RF wave was varied in the order of 500 MHz to The SET/RESET voltage under 100 MHz application is not much different from that before RF application, and the magnitude of the SET/RESET voltage under 20 MHz RF application is found to be reduced compared to 100 MHz. Figure 5(e) shows the RF frequency dependence of the SET and RESET voltages.The RF wave application started at 500 MHz from the high frequency side.After one RF wave on/off measurement, it was followed by RF application at progressively lower frequencies, down to 200 kHz.The results for repeated RF wave on/off at a single frequency are described in the next section.As the frequency decreased, the SET and RESET voltage magnitude reduced more largely, with the most significant reduction observed around 5 MHz for the SET voltage and around 20 MHz for the RESET voltage.The RF input at 5 MHz at 3 dBm decreased the SET voltage magnitude from 0.08 V to 0.01 V, and the RF input at 20 MHz at 3 dBm decreased the RESET voltage magnitude from 0.13 V to 0.07 V.When the RF input was turned off, the SET voltage returned to its initial value; the RESET voltage returned to almost its initial value after three I-V voltage cycles between +0.5 V and −0.5 V.These SET and RESET voltage variations for RF input were similar to those of Ag-GeTe with dendritic filaments, and occurred over a narrower frequency range than in Ag-GeTe.On the other hand, the RF-wave-induced changes in the DC resistance of the HRS and LRS were very different from those of Ag-GeTe.
Figures 6(a) and (b) show the frequency dependence of the HRS and LRS resistances at 0 V.The LRS resistance increased from its initial value with RF input across all measured frequencies, with a more pronounced increase at lower frequencies.Turning off the RF wave further increased the resistance.For example, at 200 kHz, the LRS resistance increased to approximately 162 Ω, which was about 2.6 times larger than the initial resistance of 63 Ω.Without the RF input, the LRS resistance was 275 Ω, which was about 4.4 times larger than the initial resistance.We note that the change in LRS resistance due to RF input below 100 MHz includes the change resulting from RF on/off repetition, as described in section 3.4.The RF-induced increase of the LRS resistance contrasts to the case of Ag-GeTe with dendritic filaments, where the RF input decreased the resistance of both LRS and HRS.
The HRS resistance, with an initial value of approximately 1 kΩ, changed greatly for RF input from 0.2 MHz to 30 MHz.The change was greatest at 20 MHz.The HRS resistance decreased to approximately 800 Ω with the RF input, and increased to about 9 kΩ when the RF was turned off.The decrease in HRS resistance with relatively low-frequency RF input was similar to that of Ag-GeTe.However, the resistance change after RF was turned off was very different from that of Ag-GeTe.In Ag-GeTe, the reduced HRS resistance only recovered to about 80% of the initial value after RF was turned off, while in Ag-Ge 2 Sb 3.4 Te 6.2 , it increased to nine times the initial value.Unlike the broad RF frequency dependence in Ag-GeTe, the resistance increase in Ag-Ge 2 Sb 3.4 Te 6.2 was resonant-like, with a significant rise occurring only around 20 MHz.

Repeated RF wave on/off at single frequency
We conducted repetitive measurements of resistive switching during the RF input of 20 MHz and 0 dBm on the same device as the one shown in figures 5 and 6, where an RF wave of 200 kHz-500 MHz and 3 dBm had been applied.The repetitive measurements were performed about one month after the measurements in figures 5 and 6.During 20 MHz RF input, I-V measurements with ±0.5 V were conducted as shown in figure 5. Subsequently, I-V measurements between −0.5 V and 0.5 V were performed after turning off the RF, and this process was repeated 11 times, as shown in figure 7(a).Figure 7(b) shows the changes in the SET/RESET voltages with the 11 cycles of RF ON/OFF.The SET voltage did not change much with repeated RF wave on/off cycles.The RESET voltage magnitude after RF input gradually increased with the number of cycles, stabilizing at around −0.17 V from the fifth cycle onwards.
Figure 7(c) shows that the LRS resistance increased monotonically with increasing RF on/off cycles.This is in contrast to Ag-GeTe, where there was no significant change with the repetition rate.Note that during the monotonically increasing LRS resistance, the SET voltage did not change significantly.Figure 7(d) shows that during the first four cycles the HRS resistance alternated between decreasing with RF ON and increasing greatly with RF OFF, as observed in the frequency-dependent measurement at around 20 MHz in figure 6(b).After the fourth cycle, there was no significant increase after RF OFF although the resistance during RF ON showed some variations.The resistance difference from RF on/off almost disappeared, and the resistances stabilized around 5 kΩ.

Mechanism of RF-induced CBRAM characteristic modification
First, we discuss the variation of SET and RESET voltages caused by RF input on the sheet-like filaments in Ge 2 Sb 3.4 Te 6.2 .As shown in figures 5 and 7, the SET and RESET voltage magnitudes decreased with the input of RF waves, and the SET voltage returned to nearly its pre-RF applied value when the RF waves were turned off.The RESET voltage returned to its initial value for up to four cycles of application, but with multiple applications, its absolute value increased.The RF-induced variations in SET/RESET voltages, and resistances are attributed to the potential modulation, rather than a temperature rise.This is because the Ag-Ge 2 Sb 3.4 Te 6.2 showed minimal changes in SET/RESET voltages and resistances from room temperature to 100 °C.
We assume that Ag migration is caused by hopping through the migration potential barrier as shown schematically in figure 8(a).The modulation of electrode potential can produce a local electric field through redistribution of electrons and ions.The local electric field reduces the potential barrier for ion migration [29].For RF input frequencies above 100 MHz, the movement of Ag + ions cannot follow the modulation of the migration potential caused by the RF input, and it has little impact on CBRAM characteristics.For RF input  onset, not geometric factors like spacing between ruptured filaments, which can result in hysteresis and irreversibility.Turning off RF restores the height of the migration barrier, accounting for the reversible change in SET voltage determined by the migration onset.The SET voltage being determined by the migration start voltage also accounts for the SET voltage's stability against RF on/off repetition.On the other hand, the RESET voltage is the voltage needed to rupture the filaments from their connected state, and therefore depends on the filament strength against rupture.We suggest that RF-induced Ag migration promotes Ag aggregation and increases filament strength.Ag ion migration prompted by RF primarily aligns with the DC electric field, with some random movement as well.This migration promotes the aggregation of Ag clusters, moving to energetically favorable positions.This increases the density of the filaments, stabilizes the filament, and increases its strength against rupture.During this aggregation, if the total Ag ion amount forming the filaments remains unchanged, the cross-sectional area of the filaments decreases [figure 8(b)].This is consistent with the increase in the LRS resistance for RF on/off repetition.This filament strengthening explains the increase of the RESET voltage magnitude for multiple RF on/off cycles.
Next, we discuss the effect of RF input on the LRS and HRS resistances.Both the LRS and HRS resistances decreased at relatively low RF frequencies below 100 MHz.The main currents in HRS arise from the electronic current in amorphous Ge 2 Sb 3.4 Te 6.2 including electron hopping [30] and the Faradaic current of Ag ions [26].The RF-induced decrease of HRS resistance is attributed to an increase in the Faradaic current due to RFenhanced Ag ion migration.The main current in the LRS arises from the tunneling current across a small gap that exists in the 'connected' state of a filament [13,31,32].The RF-induced decrease of LRS resistance is attributed to enhanced Ag migration reducing the tunnel gap or to RF-assisted tunneling [33][34][35].
We discuss the resistance changes when RF is turned off.The monotonic increase in LRS resistance observed during the repetition of RF on/off is attributed to the stable formation of filaments with smaller cross-sectional areas due to the migration enhancement caused by RF input, as described above.The dramatic increase in HRS resistance observed when RF is turned off can be explained by the enlargement of a ruptured filament gap due to the aggregation of Ag clusters near electrodes [figure 8  reaching maximum HRS resistance in the RF-off state at 20 MHz, subsequent RF-on/off measurement results in decreasing HRS resistance with lower frequencies.This means that, after the filament gap reaches its maximum, I-V measurements including resistance switching under lower RF frequencies reduce the filament gap, depending on the applied RF frequency.In other words, the Ag aggregated state or the ruptured filament shape and size during RF input depends on the RF frequency.
Finally, we discuss the change in HRS resistance during repeated measurements of RF on/off.As shown in the increase in LRS resistance and the increase of the RESET voltage magnitude for repeated RF on/off cycles, the growth of the sheet-like filaments exhibits a tendency to increase the density, resulting in aggregation inside the filaments.The reason why the HRS resistance after RF off decreased from 10 kΩ to 5 kΩ and stabilized at 5 kΩ is the competition between the tendency to aggregate inside the filaments and the tendency to aggregate near the electrodes.This competition results in the filament gap stabilizing between the initial and maximum values.The reason why the difference between RF on/off was small after the fourth cycle is that once the filaments are stabilized beyond a certain level, Ag ions cannot move with a DC bias near 0 V even during RF application, which suppresses the RF-induced Faradaic current.

Conclusion
We fabricated a CBRAM capable of RF input using Ag-Ge 2 Sb 3.4 Te 6.2 with sheet-like filaments and studied RF transmission and RF-induced modulation of the CBRAM characteristics.The LRS resistance increased during RF input unlike in Ag-GeTe.The HRS resistance decreased during RF input at 20 MHz and then increased to nine times its initial value when RF was turned off.The increase of the RF-off HRS resistance, which occurred only around 20 MHz was resonant-like, in contrast to the broad RF frequency dependence in Ag-GeTe.These resistance changes characteristic of Ag-Ge 2 Sb 3.4 Te 6.2 were explained by the Ag clusters aggregating due to RFinduced Ag migration enhancement, which increased the density of the sheet-like filaments and increased the filament gap from its initial value.We suggest that the high mobility of Ag ions in Ge 2 Sb 3.4 Te 6.2 is the cause of the sheet-like filaments and the resonant-like RF-induced changes in CBRAM characteristics.

2. 1 .
Filament structure and experiment Figures 1(a) and (b) present the electron micrographs after switching voltage sweeps in lateral Ag-Pt electrode pairs on GeTe and Ge 2 Sb 3.4 Te 6.2 , fabricated for easy visualization of the filament shape [24, 25].The filament shape of the Ag-Ge 2 Sb 3.4 Te 6.2 is mainly sheet-like although that of most chalcogenide-based CBRAMs including Ag-GeTe is dendritic.The presence of sheet-like filaments in Ge 2 Sb 3.4 Te 6.2 is consistent with the results of cyclic voltammetry.In repetitive cyclic voltammetry measurements Ge 2 Sb 3.4 Te 6.2 exhibits reproducible Faradaic currents during repeated voltage sweeps

Figure 1 .
Figure 1.Optical images of (a) dendritic filaments of Ag-GeTe and (b) sheet-like filaments Ag-Ge 2 Sb 3.4 Te 6.2 with lateral electrode pairs under an applied bias of 10 V between the probes.SEM images of (c) dendritic filaments and (d) sheet-like filaments after applying several voltage cycles of ±10 V.

Figure 2 .
Figure 2. (a) Schematic of cross-section of an Ag-Ge 2 Sb 3.4 Te 6.2 CBRAM.(b) Optical image of the Ag-Ge 2 Sb 3.4 Te 6.2 CBRAM with a pair of microwave ground-signal-ground (GSG) probes for RF measurements.

Figure 3 .
Figure 3. (a) Endurance of Ag-Ge 2 Sb 3.4 Te 6.2 CBRAM at room temperature without RF input.The SET/RESET voltages for our CBRAM are around ±0.1 V. We applied +0.2 V for SET and −0.2 V for RESET.The LRS and HRS currents were read at 10 mV.(b) Retention at 100 °C with an RF input of 150 MHz at 5 dBm.The LRS and HRS currents at 10 mV were read at 10 sec intervals.

Figure 4 .
Figure 4. (a) Transmittance T, reflectance R and absorption A of the CBRAM in the LRS and HRS, the T and A without Ag-Ge 2 Sb 3.4 Te 6.2 layer are indicated by a black line and a black broken line respectively.(b) Schematic of experimental setup for frequency multiplication measurement.(c) Transmission power spectra of LRS and HRS generated by the RF input of 20 MHz at 3 dBm.(d) Power of harmonic output through LRS and HRS, plotted as a function of input frequency.

Figure 5 .
Figure 5. (a) Schematic of RF on/off cycles.(b) Bipolar resistive switching before RF input.Zoom-in views of bipolar resistive switching during the RF input 'on' and after the RF input 'off ' at 3 dBm with frequencies of (c) 20 MHz and (d) 100 MHz.Arrows indicate the SET/RESET voltages.(e) SET/RESET voltages obtained by I-V measurements of ±0.5 V during the RF input 'on' and after the RF input 'off ', plotted as a function of frequency.The RF was input in descending order from 500 MHz to 0.2 MHz.To plot the stable 'off ' values, two I-V cycles were performed after the RF input was turned off.The initial values before the RF input are indicated by dashed lines.

Figure 6 .
Figure 6.(a) LRS and (b) HRS resistances, during the RF input 'on' and after the RF input 'off ', plotted as a function of frequency.The values before the RF input are indicated by dashed lines.

Figure 7 .
Figure 7. (a) Schematic of RF on/off cycles.(b) SET/RESET voltages, (c) LRS and (d) HRS resistances under the RF input 'on' and after the RF input 'off ', plotted as a function of measurement cycles.The RF power was 0 dBm and the frequency was 20 MHz.The values before the RF input are indicated by dashed lines.
(c)].During RF input, increased Faradaic current due to enhanced Ag migration reduces the HRS resistance.When RF is turned off, the absence of RF-induced current leads to higher HRS resistance due to the enlarged filament gap.The frequency dependence of the HRS resistance for RF off suggests that the filament gap depends on the frequency.As shown in figure 6(b), after

Figure 8 .
Figure 8. Schematic of (a)RF-induced hopping through the migration potential barriers.(b) Schematic of RF-induced narrowing or thinning of a sheet-like filament explaining the LRS resistance increase.(c) Schematic of RF-induced enlargement of a ruptured filament gap due to the aggregation of Ag clusters near electrodes, explaining the HRS resistance increase.