【AIGC One Sentence Reading】:Laser processing enhances graphene oxide memristors for flexible reservoir computing, achieving accurate predictions with optimized laser power.
【AIGC Short Abstract】:Laser processing of chaotic graphene oxide films induces nonvolatile memory, creating rCGO memristors suitable for reservoir computing. Optimized laser power reduces oxygen content, enhancing memristor performance. The device exhibits short-term memory and accurate prediction capabilities, paving the way for flexible RC networks.
Note: This section is automatically generated by AI . The website and platform operators shall not be liable for any commercial or legal consequences arising from your use of AI generated content on this website. Please be aware of this.
Abstract
Graphene oxide, as a 2D material with nanometer thickness, offers ultra-high mobility, chaotic properties, and low cost. These make graphene oxide memristors beneficial for reservoir computing (RC) networks. In this study, continuous-wave (CW) laser processing is used to reduce chaotic graphene oxide (CGO) films, resulting in the non-volatile storage capability based on the reduced chaotic graphene oxide (rCGO) films. Laser power significantly impacts the characteristics of the rCGO memristor. Material characterization indicates that laser radiation can effectively reduce the oxygen content in CGO films. With optimized laser power, the rCGO memristor achieves a large ratio at 18 mW laser power. Benefiting from the short-term memory characteristics, distinct conductive states are achieved, which are further utilized to construct RC networks. With a third control probe, the rCGO memristor can express rich reservoir states, demonstrating accuracy in predicting the Hénon map with an NRMSE below 0.3. These findings provide the potential for developing flexible RC networks based on graphene oxide memristors via laser processing.
1. Introduction
To break the bottleneck in the traditional von Neumann architecture where storage and processing are performed separately, brain-inspired neuromorphic computing is proposed as a promising solution. Reservoir computing (RC) originates from recurrent neural network (RNN) theory, only the reading weights need to be trained to read the state of the fixed reservoir[1], which significantly reduces the computing power and cost and provides an efficient and energy-saving calculation scheme[2]. Memristors due to their advantages of ultra-high speed, high density, low power consumption, and good non-volatility[3]. Traditional memristors still have limitations in some aspects, such as high manufacturing costs, poor stability, and slow write speeds. To overcome these problems, graphene memristors have been proposed as an emerging memory device and have attracted extensive attention.
As a two-dimensional (2D) material, graphene oxide (GO) offers outstanding scalability, nanometer thickness, and the potential for large-scale fabrication[4, 5]. GO films are heavily oxygenated, bearing hydroxyl and epoxide functional groups on their surface[6]. These functional groups make GO films electrically insulating, posing a challenge for their application in memristor devices. The band gap of GO can be modulated by controlling the density of oxygen-related functional groups between the interlayer and interface, which provides the tunable ability on the electrical properties[7]. Various strategies have been reported in the literature, including chemical[8] and electrochemical processes[9], thermal annealing[10], microwave[11], and laser processing[12]. Except for laser processing, these reduction strategies face challenges in controlling the reduction degree and achieving uniform distribution. Laser processing stands out as a superior option, offering precise control over the location and intensity of the reaction area with the merits of rapid processing, and no pollution[13]. Laser processing technology effectively removes oxygen-containing functional groups from the surface of GO, converting it into reduced graphene oxide (rGO). However, the investigation of rGO memristor to be implemented in the RC network is rarely reported.
In this work, Ag/rCGO/Ag memristors based on reduced chaotic graphene oxide (rCGO) are fabricated through laser processing. Different laser powers result in varying degrees of reduction in the chaotic graphene oxide (CGO) films, significantly impacting the devices' resistive switching characteristics. Scanning electron microscopy (SEM) and energy dispersive spectroscopy (EDS) results indicated that the oxygen element density decreased as laser power increased. At a laser power of 18 mW, the rCGO memristor exhibited the ON/OFF resistance ratio of 1.73 × 103 (Vread = 1 V). The reservoir computing (RC) system can be constructed by using rCGO memristors. The chaotic and short-term memory characteristics of the rCGO memristor facilitate the achievement of diverse reservoir states. Additionally, a third control probe is used to enhance the richness of the reservoir states further. These lead to a low normalized root means square error (NRMSE) below 0.3 in Hénon map prediction. These findings may open new avenues for developing tunable RC networks by using rCGO memristors.
2. Experiment
2.1. Device fabrication
The fabrication process and structure of the Ag/rCGO/Ag memristor are shown in Fig. 1. The rCGO memristor devices are fabricated on a p-Si wafer with a native oxide layer and the main process flow is shown in Fig. 1(a). First, the 100 nm Ag bottom electrode (BE) is deposited on clean Si substrates by vacuum evaporation. Then, the CGO films are uniformly coated on the bottom electrode with 2 mg/mL graphene oxide dispersion, and the thickness is approximately 900 nm, which can avoid the laser treatment on the Ag BE. As shown in Fig. 1(b), the rCGO film is treated with laser power of 16, 18, 20, 22, 24 and 26 mW by laser processing. Finally, the top electrode of 100 nm Ag is deposited by vacuum evaporation through a shallow mask, which is square with a length of 100 nm.
Figure 1.(Color online) (a) The fabrication process for the rCGO memristor. (b) The schematic images of the rCGO memristor structure.
The structural material characterization of the rCGO memristor is carried out by scanning electron microscopy-energy dispersive spectrum (SEM-EDS, VEGA LMS-Explore15) analysis to assess the elemental distribution in the rCGO memristor samples. Electrical measurements are performed by using a probe station (CasCade Summit 12000) and semiconductor device parameter analyzer (Keysight B1500A) with two waveform generator fast measurement units (WGFMU).
3. Results and discussions
3.1. Structural characteristics
The reduction of GO to different degrees can be achieved by adjusting the laser power. The SEM images of the rCGO memristor are shown in Fig. 2. The surface morphology of the CGO film is influenced by laser processing at 22 mW. Before irradiation, the surface of the CGO film was relatively flat. After laser processing, the oxygen-containing functional groups are removed by laser processing, resulting in a rough and porous surface[14].
Figure 2.(Color online) SEM images of the rCGO memristor. (a) Non-laser-processed region. (b) Comparison of non-laser-treated and laser-treated regions. (c) Laser-processed region.
Elemental analysis shows that after laser processing, the percentage of carbon atoms increases, with carbon content rising and oxygen content decreasing (Table 1). This demonstrates that laser processing effectively reduces CGO by eliminating oxygen-related functional groups. The density of these functional groups directly affects the resistance characteristics of the rCGO memristor.
laser power
C (Wt%)
O (Wt%)
Si (Wt%)
S (Wt%)
Cl (Wt%)
Ag (Wt%)
Total (Wt%)
18 mW
wo
57.97
31.83
1.17
1.61
0.62
6.81
100.00
w
59.17
29.66
1.26
1.69
0.49
7.73
100.00
22 mW
wo
55.28
24.85
7.21
0.67
1.12
10.86
100.00
w
58.55
18.71
9.45
0.45
0.59
12.26
100.00
Table 1. The elemental analysis of laser-processed region and non-laser-processed region.
To characterize the electrical properties of the rCGO memristor, positive and negative direct current (DC) sweep are applied. The DC voltage scanning characteristic curves for devices treated with different laser powers (0, 18, and 22 mW) are shown in Fig. 3. For the untreated device (0 mW laser power) in Fig. 3(a), the current remains low, and the device is in a high resistance state (HRS), exhibiting no significant resistive switching behavior. When the laser power is increased to 18 mW (Fig. 3(b)), the device displays typical bipolar resistive switching behavior, and the OFF/ON resistance ratio is 1.73 × 103 (Vread = 1 V). During the forward sweep, the current suddenly increases at a set voltage (Vset) of 1.9 V, transitioning the device from HRS to a low resistance state (LRS). Conversely, during the reverse sweep, the current suddenly decreases at a reset voltage (Vreset) of −2 V, switching the device back from LRS to HRS. The LRS and HRS remain unchanged till 100 s or 10 times I−V tests (not shown), which deserves further improvement in the next study. At a laser power of 22 mW (Fig. 3(c)), the device shows less pronounced resistance changes in the LRS, indicating the memory window (MW) degrades. According to the I−V characteristic curves of the device, the resistance values at different laser powers are shown in Fig. 3(d). These results indicate that the electrical switching properties of the rCGO memristor are significantly influenced by the laser power. Lower laser power results in a device that remains mostly in HRS, while high laser power leads to constant LRS with a significantly reduced MW. The laser processing induced memory capability can be understood by the multiphoton trap-assisted tunneling (MTAT) model. After the laser treatment, the resistive behavior of GO films is related to the conversion between sp2 and sp3 hybrid orbitals, which corresponds to LRS and HRS. Note that setting a reasonable current limit is crucial to protect the device and achieve reliable switching properties. The laser processing needs to be carefully optimized in further study.
Figure 3.(Color online) The electrical performance of rCGO memristor. (a)−(c) Resistive switching behaviors under laser processing with 0, 18, and 22 mW power. (d) The summary of resistance values with different laser power.
Fig. 4(a) shows the schematic diagram of the reservoir computing (RC) network composed of the input layer, reservoir layer, and readout layer[15]. The reservoir layer enables input signals to be mapped to a high-dimensional space, simplifying data processing in the readout layer. Only the readout layer needs to be trained, which greatly reduces the training cost and time[16]. Combined with the nonlinearity and short-term memory, the RC system is established based on the rCGO memristor. The reservoir virtual nodes are derived from the device's short-term memory (fading memory), which is related to the temporary migration or release of oxygen vacancies or interface traps under the electric field. As shown in Fig. 4(b), the corresponding current response is obtained through 16 pulse programming tests input into the memristor. A pulse "1" represents a single write pulse of 1.5 V for 5 µs, while "0" represents a single write pulse of 0 V for 5 µs. The initial voltage of −5 V for 500 µs was applied before each input. The current was read after the end of each pulse sequence with a small read voltage (amplitude: 1.5 V; width: 5 µs). The final conductivity states of different pulse sequences vary, indicating that the memristor can distinguish these sequences[17].
Figure 4.(Color online) (a) Schematic image of RC network based on the rCGO memristor. (b) The read current and the pulse train for reservoir computing of rCGO memristor.
To further improve the performance of the rCGO memristor-based RC network, it is essential for the reservoir layer to produce a large number of distinct reservoir states, which are crucial for effective classification[17]. The diversity of these states enhances the ability to process and distinguish between different input patterns, thereby improving its classification accuracy and overall computational capability. A control terminal was introduced and positioned between the two electrodes (Fig. 5(a)). By altering the position of the control terminal (location 1, 2, 3) under a constant voltage of 3 V, the device currents are shown in Fig. 5(b). When the control terminal is fixed at location 1, and applying different electric field directions (3 and −3 V), the device currents are displayed in Fig. 5(c). These configurations exhibit excellent short-term memory properties with different conductance states, demonstrating that the rCGO memristor can distinguish between these different input conditions and produce varying outputs. Furthermore, the normalized root means square error (NRMSE) results of the Hénon map are shown in Fig. 5(d).
Figure 5.(Color online) (a) The diagram of the rCGO memristor with a tunable control terminal. (b) The corresponding current of the rCGO memristor with the control terminal at locations 1, 2, and 3 (applying voltages of 3 V). (c) The corresponding current of the rCGO memristor with different electric field directions when the control probe is at location 1. (d) The NRMSE results obtained by Hénon map.
The resistive switching properties of rCGO memristors are closely related to their oxygen-containing functional groups. The graphene coating process can significantly impact the type and coverage of these functional groups on GO[18, 19]. This variation leads to device non-uniformity, which is reflected in different current−voltage characteristic curves. Additionally, the coating preparation methods can affect the interlayer spacing of GO[20], influencing charge transport and carrier movement, thus affecting the resistive switching properties of the devices. Thus, rCGO memristors can store a greater number of conductive states compared to conventional memristors. The reservoir layer implemented by rCGO memristors can achieve more distinct reservoir states when different pulse sequences are input, resulting in higher state richness. This diversity enhances the nonlinear response, improving the computational capability and flexibility of the network.
4. Conclusion
In this work, rCGO memristors are fabricated by using laser processing with different laser powers. SEM and EDS characterization showed that with the increase of laser power, the oxygen density decreased, and the reduction degree of graphene oxide increased. When the laser power is 18 mW, the resistance characteristics of the device are Vset = 1.9 V, Vreset = −2 V, and the on-off resistance ratio is about 1.73 × 103. The rCGO memristor is further implemented in the RC network with a third probe leading to richer reservoir states. A high recognition accuracy is achieved by using the Hénon map with NRMSE less than 0.3. This work provides great potential for developing a highly tunable and flexible RC network based on the rCGO memristor.