Memristors Based on Many-Layer Non-Stoichiometric Germanosilicate Glass Films

Abstract

Nonstoichiometric GeSi_xO_y glass films and many-layer structures based on them were obtained by high-vacuum electron beam vapor deposition (EBVD). Using EBVD, the GeO₂, SiO, SiO₂, or Ge powders were co-evaporated and deposited onto a cold (100 °C) p⁺-Si(001) substrate with resistivity ρ = 0.0016 ± 0.0001 Ohm·cm. The as-deposited samples were studied by Fourier-transformed infrared spectroscopy, atomic force microscopy, X-ray photoelectron spectroscopy, and Raman spectroscopy. A transparent indium–tin–oxide (ITO) contact was deposited as the top electrode, and memristor metal–insulator–semiconductor (MIS) structures were fabricated. The current–voltage characteristics (I–V), as well as the resistive switching cycles of the MIS, have been studied. Reversible resistive switching (memristor effect) was observed for one-layer GeSi_0.9O_2.8, two-layer GeSi_0.9O_1.8/GeSi_0.9O_2.8 and GeSi_0.9O_1.8/SiO, and three-layer SiO₂/a–Ge/GeSi_0.9O_2.8 MIS structures. For a one-layer MIS structure, the number of rewriting cycles reached several thousand, while the memory window (the ratio of currents in the ON and OFF states) remained at 1–2 orders of magnitude. Intermediate resistance states were observed in many-layer structures. These states may be promising for use in multi-bit memristors and for simulating neural networks. In the three-layer MIS structure, resistive switching took place quite smoothly, and hysteresis was observed in the I–V characteristics; such a structure can be used as an “analog” memristor.

1. Introduction

For the development of digital technologies, in particular for storage technologies and for processing a huge amount of information (Big Data), more and more memory is required, while the requirements for the speed of reading and rewriting data are increasing. To store and process large amounts of information, new memory elements are needed, since the characteristics of old types of memory, including flash memory, are not enough. The memristor effect is a reversible switching process between states with a high resistance state (HRS) and a low resistance state (LRS) [1], occurring when a current of a different polarity flows through it. Thus, the resistance of the memristor depends on the current pulse that has passed through it; we can say that the memristor “remembers” the charge that has flowed through it. Memristors are promising candidates for creating a universal non-volatile memory for the future [2]. The requirements for the universal memory of the future (ideal memory) are as follows: non-volatility; scalability to reach Terabits or more per chip; high performance (reading and rewriting time up to 100 picoseconds); low power consumption (a small amount of energy needed to rewrite one bit, ideally less than one femtojoule); the number of rewriting cycles (endurance) up to 10¹²; and an information storage time (retention) of 10 years or more at a temperature of 85 °C.

In recent years, neuromorphic computing has attracted much attention; it is assumed that due to the possibility of parallel signal processing, they can be more efficient than classical digital calculations in the von Neumann architecture of computers [3]. In addition, interest in artificial neural networks is inspired by the fact that they can simulate the operation of biological neural networks. Electronic synaptic devices are new solid-state devices designed to implement neuromorphic computing in hardware systems [4,5]. There is much research aimed at obtaining the function of biological synapses in solid-state devices using new non-volatile devices. Among them, Resistive Random Access Memory (RRAM) devices are the most promising candidates for electronic synaptic devices, due to their simple structure and high performance. The RRAM device has an adaptable conductance that is similar to the biological renewal of a synapse’s weight. Conventional neuromorphic circuits require a combination of several transistors and capacitors to simulate a single synaptic function, which increases power dissipation and limits integration density. In contrast, a single RRAM-based memristor can simulate a synapse. Thus, simulating biological synaptic behavior with an RRAM-based memristor can significantly improve integration density and processing efficiency, as well as reduce the power consumption of the artificial neural network [6].

A memristor with two or more resistance states can be created based on metal–insulator–semiconductor (MIS) structures, where a two-layer structure can be used as an insulator. For example, a memristor based on titanium oxides is created from two films of titanium oxide, where the first layer is TiO₂ and the second layer is oxygen-poor TiO_2−x material [1]. This configuration allows the formation of two resistive states, that is, a classic one-bit memristor. At present, the formation and destruction of conductive filaments is considered to be the main mechanism of resistive switching, but there are also non-filament memristors based on ferroelectrics [7].

The approach to form a multi-bit memristor is as follows. Let us assume that filaments can be formed and destroyed in each of the two layers. Then, several intermediate states of conduction are possible: the lowest resistance, when filaments are formed in both layers (two “short-circuited” series resistances); and the highest resistance, when filaments are absent in both layers. The intermediate state, is possible when filaments are present in only one layer. Therefore, the presence of HRS and LRS in each layer should lead to the sequential switching of resistive states in the entire structure consisting of several layers, that is, it should lead to multi-bitness. A multi-bit memristor, with many intermediate resistive states that form a quasi-continuous resistance spectrum, is sometimes referred to as an analog memristor [8].

As already mentioned, the first memristor based on the two-film TiO₂/TiO_2–x system was created in 2008 [1]. Since silicon oxides have been used in microelectronics for more than sixty years, and the technology of their manufacture is well developed, almost immediately after the discovery of the first memristor, researchers started working on the study of memristor effects in SiO_x films [9]. Since then, significant progress has been made in the use of SiO_x-based memristors [10]. Recently, the interest of memristors based on materials with low bond enthalpy has increased. It is assumed that in such materials, due to weaker bonds, filaments are more easily formed and destroyed, and the energy required to rewrite one bit should is less. Germanium [11] and tin [12] oxides have a lower bond enthalpy than silicon oxide [11]. However, memristor effects in germanium oxides have been poorly studied; there are only a few works on this subject [13,14,15].

The advantage of germanosilicate glasses (SiGe_xO_y) in multi-bit memristors is that their deposition technology is simple, inexpensive, and fully compatible with silicon technology. The SiGe_xO_y has the ability to form two different nanoscale potential fluctuations. The gaps in the density of states in SiO₂ (8–9 eV) and GeO₂ (4–5 eV) differ significantly. This makes it possible to modulate the parameters of traps of the following type: the inclusion of germanium oxides in silicon oxide. Another possibility is the controlled formation of precipitates with an excess of germanium atoms. The germanium nanoclusters are deep traps for electrons, and are holes both in the GeO_x matrix and in the SiO_x matrix. The memristor effect in germanosilicate films with different stoichiometries was first discovered in 2019 [15]. It is known that SiGe_xO_y, with an excess of germanium and a deficiency of oxygen, contains nanoclusters of amorphous germanium [16]. It is assumed that conducting filaments can be formed between germanium clusters, and the clusters themselves can be the centers of the formation of such filaments [17]. Thus, by forming a multi-layer structure based on SiGe_xO_y films with different stoichiometries, it is possible to achieve multi-bitness in memristors based on these structures.

This work aims to search for the possibility of forming intermediate resistive states in MIS structures, based on several layers of nonstoichiometric SiGe_xO_y glasses.

2. Materials and Methods (Experimental Methods)

Films of non-stoichiometric GeSi_xO_y germanosilicate glasses, and multi-layer structures based on them, were obtained by the high-vacuum electron beam vapor deposition (EBVD) technique. It is possible to control the stoichiometry of the films using the co-evaporation of various targets, in our case, these were GeO₂, SiO₂, SiO, and Ge powders. In addition, the power of the electron beam evaporating each target can be controlled. The vacuum in the growth chamber was maintained at 10^–8 Torr. During the evaporation of the targets, the pressure increased by one or two orders of magnitude. The temperature of the substrate (silicon of p⁺–type with specific conductivity ρ = 0.0016 ± 0.0001 Ohm·cm and (100) orientation) was kept low and amounted to 100 °C. The deposition rate was controlled by microbalance and was approximately 0.1 nm per second. It is known that oxygen and germanium monoxide are more volatile compounds than germanium dioxide. Thus, the evaporation products of the GeO₂ target are mainly oxygen and germanium monoxide. Not all the evaporated oxygen interacts with the germanium monoxide deposited on the cold substrate. Therefore, the stoichiometry of the resulting film is close to that of GeO, but contains slightly more oxygen. According to infrared (IR) absorption data and Energy Dispersive X-ray Spectrometry (EDXS) data, these are GeO_z films with z from 1.1 to 1.2 [18]. During the evaporation of the SiO₂ and SiO targets, films of the stoichiometric composition SiO₂ and SiO, respectively, were deposited. Thus, during the co-evaporation of GeO₂ and SiO₂ sources, in the case of the same evaporation rate of both sources, films of nonstoichiometric germanosilicate glass with a composition of approximately GeSiO_2+z are deposited. During the co-evaporation of GeO₂ and SiO, in the case of the same evaporation rate of both sources, films of nonstoichiometric germanosilicate glass with a composition of approximately GeSiO_1+z are deposited. The growth conditions and stoichiometry of the deposited films are discussed in more detail in [17,19,20].

The upper electrode on the resulting MIS structures was obtained by depositing a layer of indium tin oxide (ITO) using magnetron sputtering through a mask. The ITO layer thickness was 200 nm. The sheet resistance of the ITO layer was 40 Ohm/□, and the size of the contacts was 0.7 mm × 0.7 mm, so the contact area was approximately 0.5 mm².

The stoichiometric composition of the GeSi_xO_y films was determined via the analysis of X-ray photoelectron spectroscopy (XPS) data. The XPS data were obtained using a SPECS spectrometer equipped with an Al/Ag double anode X-ray source, a FOCUS-500, an ellipsoidal X-ray crystal monochromator, a PHOIBOS 150 semispherical electron analyzer, and an ion source. The spectra were obtained using monochromatic Al Kα radiation (hν = 1486.74 eV) at the analyzer pass energy, equal to 20 eV. The binding energy of the experimental peaks was calibrated using the C1s peak at 284.8 eV, associated with hydrocarbons on the surface of the samples. The technique is described in more detail in [21].

The surface morphology was studied by atomic force microscopy (AFM) on Dimension Icon (Bruker, Santa Barbara, CA, USA) in PeakForce tapping mode with the help of ScanAsyst–Air probes (Bruker, Santa Barbara, CA, USA); this had a tip radius of approximately 2 nm and a spring constant of approximately 0.4 N/m. The measurement was performed with a 0.1 Hz small velocity and a small force of tip-sample interaction, in order to protect the tip apex from damage. The imaging range was 10 µm × 10 µm, with an image resolution of 512 × 512 pixels.

The structural properties of GeSi_xO_y films and GeSi_xO_y-based multi-layer structures were also studied using vibrational spectroscopy methods, such as Raman scattering and Fourier Transformed InfraRed (FTIR). The Raman spectra were registered using a T64000 spectrometer (Horiba Jobin Yvon, France) at room temperature in backscattering geometry. For excitation, a fiber laser GFL-515-0200-FS (Inversion–Fiber, Novosibirsk, Russia) was used and the wavelength of the laser radiation was 514.5 nm. The spectral resolution was no worse than 2 cm^–1. The power of the laser beam reaching the sample was 1.3 mW, and the spot diameter was 10 μm. Since the absorption of light, with a wavelength of 514.5 nm in GeSi_xO_y films, is low, the measurement mode used did not lead to the local heating of the samples during the measurement. The Fourier spectrometer FT–801 (SIMEKS, Novosibirsk, Russia) was used to register IR absorption spectra. The spectral range of the device is from 650 to 4000 cm^–1; the spectral resolution is 4 cm^–1. A silicon substrate without films was used as a reference signal.

To measure the current-voltage (I–V) characteristics of the studied MIS structures, we used a setup based on an Agilent B2902A multimeter. A two-contact measurement scheme was used. The back side of the Si substrate covered with indium–gallium paste was used as the lower electrode. For better contact, the back side of the substrate with the indium–gallium paste was pressed against the metal sheet, to which the probe of the multimeter was in contact. The ITO contacts were used as the top electrode, to which the other probe of the multimeter was in contact. The experimental setup allows automatic measurements of currents in the ON and OFF states of the memristors (LRS and HRS correspondingly) during cycling.

3. Results and Discussion

Figure 1 shows a scheme of the four studied MIS structures. Further, these samples will be referred as A, B, C, and D.

Figure 2 shows an overview XPS spectrum of sample C, the upper layer of which was obtained by the co-evaporation of GeO₂ and SiO₂ powders.

The survey spectrum was obtained at a fixed analyzer transmission energy of 50 eV. As can be seen, intense photoelectron peaks are observed in the spectrum. These peaks are specific for germanium, silicon, oxygen, and carbon. Since the samples are stored in a normal atmosphere, the surface always contains organic materials as an impurity. The organic materials contain carbon and oxygen. Specifics for other element peaks were not found. It should be noted that the XPS method is surface-sensitive; in our case, we obtained information on the composition of the upper part of the layer, since the depth of analysis was several nanometers. Thus, the composition of only the upper layer of sample C was determined. The relative content of elements near the surface of the samples and the ratio of atomic concentrations were determined using the integral intensities of photoelectron peaks, corrected for the corresponding atomic sensitivity factors based on calculations of the Scofield cross sections [22]: 1.42 for Ge3d, 0.817 for Si2p, 1.0 for C1s and 2.93 for O1s. Thus, for the as-deposited GeO[SiO₂] layer, the O/Si ratio is 3.32, and the Ge/Si ratio is 1.16. The surface composition of the sample is presented in

Table 1. Elemental composition of the layer obtained by co-evaporation of GeO₂ and SiO₂ powders.

Ge	C	O	Si
18.2	13.6	52.4	15.8

. It should be noted that the oxygen content in the near-surface region is usually higher, which is associated with the oxidation of the near-surface region of the samples during their storage in the air. If one identifies the oxygen that is bound to carbon (using the approach described in [21]), then the composition of the upper layer in samples A, C, and D can be estimated as GeSi_0.9O_2.8. Unfortunately, there was no experimental possibility to determine the composition of the lower layers, formed by the co-evaporation of GeO₂ and SiO powders. However, assuming that the evaporation rates of SiO and SiO₂ powders are equal, we can estimate the composition of this layer as GeSi_0.9O_1.9.

Figure 3 shows the results of the AFM studying of the relief of the upper layers of samples B and C as two-dimensional scans of morphology. AFM allowed the measurement of the numerical values of the root-mean square (RMS) roughness of the samples, which appeared in the same order as reported in [14]. The RMS roughness for the sample with GeO₂ and SiO₂ (Figure 3a) was 0.47 nm, while for the layer deposited by evaporation of SiOx powder (Figure 3b), it was 0.57 nm. At the same time, the RMS for the Si substrate was 0.32 nm. Thus, the surfaces became slightly more developed after the deposition of layers, although retained the necessary smoothness of a semiconductor thin film.

Figure 4 shows the Raman spectra of all studied samples. For comparison, the peak from the silicon substrate is also shown.

As can be seen, in almost all spectra, one can observe features associated with two-phonon inelastic scattering from the silicon substrate. This is a consequence of the fact that the studied layers are thin and semitransparent. The most intense feature of the Raman spectrum from the substrate is the peak at ~305 cm^–1, due to two-phonon scattering by transverse acoustic phonons (2TA) [23]. The Raman spectral region for long-wavelength phonons of the Si substrate (520.6 cm^–1) is not shown in the Figure 4. One can see a broad peak with a maximum at about 275 cm^–1 only in the spectrum of sample D (containing a layer of amorphous germanium). This peak can certainly be attributed to amorphous Ge [24].

Previously, we found that amorphous germanium clusters are formed in the layers obtained by the co-evaporation of GeO₂ and SiO powders. This occurs due to the redox reaction GeO+SiO → Ge+SiO₂ [15,16,17]. However, in the case of samples B and D, containing such layers, their Raman spectra do not contain a ~275 cm^–1 peak from amorphous germanium. This is because in this case, the thickness of the layers is an order of magnitude smaller than in the articles [17]. Obviously, in our case, there are so few amorphous germanium clusters and they are so small that the Raman signal from them is lower than the detection limit.

Figure 5 shows the IR absorption spectra for all samples. Absorption (or optical density) is: A = −ln(T/100), where T is the transmission percentage.

It is known that analysis of the absorption of IR radiation on the vibrations of the Si–O and Ge–O polar bonds can provide information on the stoichiometry of the suboxides SiO_x, GeO_x, and GeSi_xO_y. However, the p⁺–type silicon substrate with specific conductivity ρ = 0.0016 ± 0.0001 Ohm·cm is opaque for IR absorption, due to strong absorption by free holes. To analyze the IR absorption spectra, satellite samples were used, that is, the same structures obtained together with the “side-by-side” structures under study in the same growth process, but on an n-type silicon substrate with a specific electric resistance of 5.5 ± 0.1 Ohm·cm. The same substrate was used to record the reference spectrum (reference sample) when measuring the IR absorption spectra. The spectra contain absorption peaks on vibrations of oxygen bound both to silicon and germanium. The spectrum of sample B contains a broad absorption line from 1000 to 1050 cm^–1, which was attributed to the Si–O–Si valence mode in SiO films [25]. Thus, the stoichiometry of the upper layer in this sample is close to that of SiO.

In the spectra of all samples, which include layers obtained by the joint evaporation of GeO₂ and SiO powders and GeO₂ and SiO₂ powders, there are absorption peaks from vibrations of the valence modes of the Ge–O–Si and Ge–O–Ge bonds. The frequency of vibrations of the valence modes of the Ge–O–Si bonds is about 990 cm⁻¹ [26,27], and the frequency of vibrations of the valence modes of the Ge–O–Ge bonds depends on the stoichiometric composition of the films [28,29]. Therefore, according to FTIR spectroscopy, it can be approved that our layers contain Si–O–Si, Ge–O–Si, and Ge–O–Ge bonds.

Figure 6 shows the first three I–V cycles for sample A. The applied voltage range was from −2 to 2 V. The voltage sweep started at −2 V. The current limit was 10⁻² A. It is worth noting that the I–V characteristics of this structure are similar to the I–V characteristics of an equivalent structure with a larger film thickness [17]. One can clearly see the presence of two resistive states (ON and OFF) or the memristor effect. The ratio between the current in the OFF and ON states (memory window) is about 10³. It can also be noted that switching ON occurs abruptly, and intermediate states in resistance practically do not appear. When turned OFF, one intermediate state is clearly observed. The characteristic spread of switching voltages is approximately 0.5 V.

For sample A, resistive switching cycles (endurance characteristics) were studied. The results of the study and the measurement scheme for each cycle are shown in Figure 7. The turn-on voltage V_set was 3 V, the turn-off voltage V_reset was −3 V, and the readout voltage V_read was −0.5 V. The turn-on and turn-off pulse times were 10 ms. The length of each cycle was 100 ms.

It is worth noting that Figure 7 does not show the first switching cycles. After several switchings, the current in the OFF state increased. After that, the memory window was maintained for the next 5000 cycles. However, the ratio of currents in the ON and OFF resistive states decreased from 3 to 1.5 orders of magnitudes compared to the first cycles shown in Figure 6. It can be seen that the current during multiple rewriting cycles increases both in the OFF and ON states. In this case, the memory window even after 5000 cycles remains more than an order of magnitude in size.

Figure 8 shows the first 3 cycles of the I–V curve characteristics of sample B. The applied voltage range was from −4 to 2 V. The current limit was set to 10^–2 A.

An analysis of the I–V characteristics showed that switching is unstable. In this case, the conductivity in HRS increased with each cycle until the switching disappeared. After several cycles, the switching OFF became impossible. Therefore, during the first I–V curve cycle, a certain memory window (less than one order of magnitude) was observed. However, after a certain number of switching cycles, the switching ceases to be reproduced and the current remains at the ON-state level; the memristor does not turn OFF. Apparently, the very bad endurance is associated with a thick layer of SiO, and more precisely, with its weak memristor properties [30]. It is probable. that there is too much excess silicon in this layer, and the conductive filaments in it are too stable; they do not break down. Presumably, a structure with a thick SiO_x (where x > 1.5) will be more stable than a structure with SiO [31].

The I–V curve characteristics of sample C are shown in Figure 9. The applied voltage range was from −3 to 1.5 V.

The sample C has two layers. Unlike sample A, it has an additional GeO[SiO] layer with a thickness of 5 nm. The introduction of such a layer led to two effects. The switching voltage decreased from 1.5–2V to 1–1.5V, and secondly, several resistive states were observed in this bilayer structure. The first effect is probably related to the fact that the deposited GeO[SiO] film contains amorphous Ge clusters [16,17], which act as concentrators of the electric field. The appearance of intermediate states is related to the bi-layer structure. However, the I–V curves in this sample are not stable and it withstands a small number of switching cycles; in this connection, the cycling study of this sample was not carried out.

The I–V curve characteristics of sample D are shown in Figure 10. The applied voltage range is from −3 to 2 V. The current limit was 10⁻² A. Sample D contains several layers. An additional Ge layer 2 nm thick is separated from the substrate by a layer of tunneling thin SiO₂. It can be seen that, in sample D, as in sample A, the memristor effect is observed. Note that the switching voltages for samples A and D are comparable. However, unlike sample A (Figure 6), resistive switching in sample D (Figure 10) does not occur abruptly, but rather smoothly; the I–V curve simply has hysteresis. In addition, in the I–V curve characteristics of sample D, sections with a negative differential resistance are observed, in contrast to sample A. It should be noted that the germanium layer in this structure is a deep trap for holes, since the energy of holes in silicon is higher than in germanium, and the electron energy in germanium is higher than in silicon, the Si/Ge heterostructure is a type II heterostructure [32].

The current at a positive voltage in sample D is higher compared to the current in sample A. Perhaps this is because, when a positive potential is applied to the ITO contact, most of the applied voltage bias is to the SiO₂ layer, and the electron energy in the germanium layer can be below the bottom of the conduction bands of the silicon substrate. The band diagram for sample D is shown in Figure 11. With a positive external bias, the minority charge carriers of the substrate (electrons) can tunnel from the p–Si substrate through the thin SiO₂ layer into the germanium layer. With their negative potential, electrons reduce the voltage bias to the thin SiO₂, layer, but increase the voltage bias to the thick GeO[SiO₂] layer. In the OFF state, electrons accumulate in the germanium layer, and the voltage bias to the GeO[SiO₂] layer becomes large, which leads to an increase in the current through this layer; therefore, the current is large even in the OFF state. In the ON state, the current is supposedly limited by the tunneling current through the thin SiO₂ layer, so the I_ON/I_OFF current difference in the positive bias is small. With a negative external voltage bias (Figure 11b), the current in the structure is supposed to be determined by the conductivity of the thick GeO[SiO₂] layer, which is very different in the ON and OFF states. Therefore, the memory window at a negative voltage is larger.

With a negative external bias, electrons that reach the germanium layer can either tunnel into the silicon substrate or recombine with holes, which can also tunnel from the silicon substrate into the germanium layer. Note that the charging of the germanium layer by holes, when a negative voltage is applied to the upper ITO contact, should also affect the voltage distribution in the whole structure and, hence, the flowing current. The negative differential resistance is possibly related to this effect.

For sample D, resistive switching cycles (endurance characteristics) were studied. The results of the study and the measurement scheme for each cycle are shown in Figure 12. The turn-on voltage V_set was 3 V, the turn-off voltage V_reset was −3 V, and the read voltage V_read was −0.5 V. The turn-on and turn-off pulse widths were 10 ms. The cycle duration was 100 ms.

With repeated cycling, a decrease in the memory window is observed during the first 500 rewriting cycles. Current in the ON-state tends to decrease. The OFF-state current remains quite stable. However, compared to sample A, the endurance for sample D is lower.

4. Conclusions

In many-layer MIS structures based on nonstoichiometric germanosilicate films, the presence of intermediate resistive states was demonstrated.

In the MIS structure with a germanium layer between two barriers of a SiO₂ tunnel layer and a GeO[SiO₂] layer, resistive switching proceeded smoothly, hysteresis was observed in the I–V curve characteristics, and such a structure can be used as an “analog” memristor. It is assumed that the recharging of the germanium layer affects the distribution of the electric voltage in the structure the current in it.

Thus, it can be assumed that multilayer structures based on non-stoichiometric germanosilicate glasses are promising for the creation of multibit memristors or “analog” memristors that are applicable for the creation of artificial neuromorphic networks.