Advanced Hub Main Navigation Menu
An Optoelectronic Memristor Based on Proton-Involved Photoreduction for Bimodal Sensing, Memory, and Processing
Abstract
Advanced devices and systems with integrated sensing, memory, and processing functionalities similar to those of the human nervous system are highly desirable for efficient artificial intelligence applications. In this study, an optoelectronic memristor with integrated bimodal sensing, memory, and processing based on graphite oxide (GO)/titanium dioxide (TiO2) nanocomposite is presented. The resistive switching behavior of the memristor is based on proton-involved photoreduction, and the memristor exhibits humidity-dependent optical resistive switching and synaptic behaviors similar to an artificial optoelectronic synapse. The plasticity of the artificial synapse can be further modulated by a heterosynapse with an external bias caused by electric field-driven oxygen migration. These features equip the artificial synapse with not only a combined light/humidity information sensing and memory but also contrast enhancement and attention-driven functionalities similar to those of the human visual memory system. Moreover, as a proof of concept, a sensory–motion system is constructed, which sends the synaptic feedback of the optoelectronic memristor to direct responses in a robotic arm. This work could provide a fundamental unit for the future development of perception systems in efficient humanoid robots.
1 Introduction
Biological nervous systems are highly efficient given the integration of their sensing, memory, and processing functionalities.[1-3] The development of hardware with such capability is desirable for advanced humanoid robot and artificial intelligence applications. However, artificial nervous systems based on complementary metal oxide semiconductor technology in separate units pose substantial challenges in terms of energy consumption and time latency.[4-6] In-sensor and in-memory computing within a single cell are promising technologies for developing artificial nervous systems.[7-10] However, these devices often lack nonvolatile memory or sensing capabilities, and additional separate components are still needed. The development of a device that simultaneously satisfies the requirements of sensing, memory, and processing remains a key research topic. The ability of the nervous system to fuse multiple sensing modes is essential for a human to comprehensively recognize an objective.[11-13] For example, the environmental relative humidity (RH) has been found to affect the recognition accuracy of the human visual system.[14]
Emerging memristor has been proposed as promising building blocks for artificial synapse with fused sensing, memory, and processing.[15-18] The continuous modulation of device conductance allows the integration of memory and processing functions in a single cell. Toward the development of artificial nervous systems, optoelectronic memristors with light-sensitive switching characteristics are promising because they can combine sensing functionality with the inherent memory and processing capabilities of the device.[19-24] For example, Chai et al.[20] developed an ultraviolet (UV) light-sensitive optoelectronic memristor with image sensing, memory, and contrast-enhancement functions. Han et al.[14] developed a multimodal memristor that combines optical/protonic sensing and processing (noise suppression and filtering). However, there remains an urgent need to explore memristors with multimodal sensing capabilities and higher processing capabilities. Attention-modulated visual memory is an important ability for improving efficiency and avoiding damage.[25] This enables a system to develop a weak or strong memory after viewing input images with low or high attention for the same period of time. This behavior resembles the heterosynapse modulation in biological functions. Taking all of this into account, the exploration of an optoelectronic memristor that can integrate the multimodal sensing, memory, and heterosynapse modulation functionalities is desirable.
In this work, we report an optoelectronic memristor with a graphite oxide (GO)/titanium dioxide (TiO2) nanocomposite. As an artificial optoelectronic synapse, its synaptic behaviors can be co-modulated by light, humidity, and electric fields, taking advantage of the proton-involved photoreduction and electric field-driven oxygen migration effect. This feature equips the memristor with multimodal sensing, memory, and processing capabilities. As a proof of concept, the synaptic feedback of the memristor was sent to direct a robotic arm to demonstrate a sensorimotor function. Notably, this work can implement multimodal sensing, memory, and processing in a single cell, which will help reduce the energy consumption and time latency in artificial nervous systems.
2 Results and Discussion
In biological visual systems, light information is sensed by the retina and transmitted via the afferent visual nerve pathway to the visual cortex for memory and processing (Figure S1, Supporting Information).[26, 27] In this study, an optoelectronic memristor with Au/GO-TiO2/indium tin oxide (ITO) sandwich structure (Figure 1a) is proposed to emulate the in-situ sensing, memory, and processing capabilities of biological visual systems. The device fabrication is presented in detail in the Section 4, and the transmission electron microscopy image and light-absorption property of the GO-TiO2 nanocomposite are shown in Figure S2, Supporting Information. UV light at a wavelength of 380 nm was selected as the optical stimulus. First, the light and humidity co-modulated electrical characteristics of the memristor at a reading voltage of 0.05 V were studied. Figure 1b shows the response current of the memristor under 80 consecutive UV light pulses (width: 2 s) with different intensities (1.5, 3.0, and 4.5 mW cm−2) at 25% RH. The retention characteristics of the memristor within 100 s after removing the UV signal were also examined. As shown in Figure 1c, it was observed that a higher UV intensity produces a larger response current and longer relaxation time, confirming the ability of the memristor to simulate the transition of synaptic characteristics from short-term plasticity (STP) to long-term plasticity (LTP). In addition, the current response and relaxation time of the device were found to be highly dependent on the environment RH (Figure 1e,f). As the environment RH increases from 45 to 85%, the response current is increased and the decay rate is slowed. The STP behaviors in response to a single UV pulse stimulation at different intensities and widths as well as different RH are presented in Figure S3 and S4, Supporting Information. The response current of the memristor was modulated from STP to LTP as the frequency and number of light pulses increased (Figure 1i and Figure S5, Supporting Information). The relative current variation (ΔI/I0, where ΔI is the response current change and I0 is the initial current) of the memristor as a function of the UV intensities and RH levels is summarized in Figure 1d. The ΔI/I0 increases steadily with respect to light intensity and RH. Light-induced synaptic plasticity was studied at a light intensity of 4.5 mW cm−2 and RH of 85%, which can induce large ΔI/I0. Paired-pulse facilitation (PPF) is an important feature of synaptic plasticity. Given a paired stimulus, the amplitude of the response to the second stimulus should be larger than that of the first one. The response current of the memristor to paired light pulses with an interval (ΔT) of 2 s is shown in Figure 1g. The second current spike (A2) is higher than the first current spike (A1), which is similar to the PPF behavior of a biological synapse. The PPF index is defined as A2/A1 × 100% and is usually related to the ΔT between the two light pulses. As shown in Figure 1h, the PPF index gradually increases as ΔT decreases. These results demonstrate the potential light/RH information bimodal sensing and memory capability of the memristor.
The working mechanism of the device was studied using Raman spectra and X-ray photoelectron spectroscopy (XPS) measurement. The results of Raman mapping using the intensity ratio of the D and G bands (ID/IG) of the composite under different light intensities and RH conditions are compared in Figure 2a–c. When the memristor was in its initial state (S-I: RH: 25%, intensity: 0 mW cm−2), the entire area was dominated by low ID/IG (blue tones) with scattered regions of higher ID/IG (light red). A large number of locally high ID/IG regions (red spots) formed after the device was exposed to UV light with the RH of 25% (S-II: RH: 25%, intensity: 4.5 mW cm−2), as shown in Figure 2b. The high ID/IG (red) area was further enlarged when the RH increased to 85% (S-III: RH: 85%, intensity: 4.5 mW cm−2; Figure 2c). The D peak is usually associated with defects, mainly because of the various oxygen functional groups dangling on the basal graphene plane. It has been reported that the detachment of oxygen groups can create more defect regions (reduced graphene oxide (RGO) domains) and result in an increase in ID/IG.[28-30] The results of the ID/IG mapping images represent the generation of small-sized RGO domains, which can be attributed to the oxygen functional group dissociation of GO films. The XPS measurements were used to reveal the percentage of different carbon bonds of the composite. As shown in Figure 2d, the C1s XPS spectra of the composite can be divided into four peaks at 284.6, 286.7, 287.5, and 288.6 eV, corresponding to CC, CO, CO, and COO, respectively.[31] The atomic percentage of oxygen functional groups gradually decreases with increases in light intensity and RH (Figure 2e), which is consistent with the Raman mapping results. The origin of the resistive switching mechanism could be proton-involved photoreduction (Figure 2f). According to previous reports,[30, 32, 33] the photo-assisted reduction of GO is believed to be associated with photocatalytic reactions at the surface of TiO2 nanoparticles because pure GO film is not sensitive to short-time UV irradiation. Under UV irradiation, TiO2 nanoparticles are excited and generate electron–hole pairs. The holes react with surface-adsorbed water to generate oxygen and protons, whereas the electrons can be efficiently transferred to the GO. The process may initiate the reactions that dissociate oxygenated functional groups together with protons. As a result, the sp2 regions will gradually expand and merge to form a conductive channel throughout the film. The experimental results supporting this working mechanism are shown in Figure S6, Supporting Information.
The processing capability of biological visual systems is critical. Preprocessing is a key function of the retina of the human eye, and is required to improve the efficiency and accuracy of subsequent tasks.[34] The light intensity- and RH-dependent switching characteristics (Figure 1d,f) reveal that the ΔI/I0 and retention time increase significantly with increase in RH. As for input images composed of different gray levels (light intensity normalization), it is reasonable to infer that RH can enhance the contrast of the input images (represented by ΔI) after repeated light stimulation. Hence, a 3 × 3 pixel memristor array was developed to demonstrate this contrast-enhancement function. The stimuli for the image consisted of 80 light pulses with a width of 2 s and a frequency of 0.25 Hz. As shown in Figure 3a, during the light training process, the array was divided into four regions with different light intensities: 0 mW cm−2 (region IV), 1.5 mW cm−2 (region III), 3.0 mW cm−2 (region II), and 4.5 mW cm−2 (region I). The light intensity was normalized to a range of 0–1 (Figure 3a(i)). The normalized output current images under different RHs (25, 45, and 85%) after UV light training was removed for 100 s are shown in Figure 3a(ii). The differences in the four regions are amplified at 85% RH, suggesting the contrast enhancement capability of the memristor. More importantly, it was found that the response current of the memristor can be modulated by applying an external bias (this mechanism is discussed later). The bias voltage was applied to the ITO electrode while keeping the Au electrode grounded during measurement. Figure 3c shows the device response current and relaxation characteristics under light stimulation (intensity: 4.5 mW cm−2, RH: 85%) with different biases (−0.50→0.05→0.50 V) applied on the ITO electrode. It was observed that the response current can be enhanced or suppressed using a positive or negative bias, respectively (Figure 3d). Heterosynaptic modulation plays an important role in achieving efficient and high-level biological neural processing. Kandel and Tauc[35] proposed that the regulation of synaptic weight between pre-synapse and post-synapse may be the result of the activation of a third synapse (modulatory synapse). The regulation can lead to an increase (modulatory facilitation) or decrease (modulatory inhibition) in synaptic strength. In neuroscience, heterosynaptic plasticity can be divided into non-associative and associative forms. The non-associative form is purely heterosynaptic, while the associative form is the results from the co-activity of modulatory and presynaptic input.[36, 37] In this work, light serves as the input of the artificial optoelectronic synapse. Bias was introduced to co-regulate the synaptic weight (response current). The regulation is similar to that of associative form of heterosynaptic plasticity (Figure 3b). Based on the bias-regulated characteristic, attention-driven function of human visual memory was mimicked. As well-known, the activity of visually tuned neurons can be modulated to improve memory retention when one pay attention to the target.[38] In this study, attention to cherry, banana, and apple fruits was assumed to increase in that order. The fruit patterns were recorded using a 14 × 14 optoelectronic memristor array stimulated by light pulses at 85% RH with biases of −0.50, 0.05, and 0.50 V applied to the ITO electrode. The output currents were recorded after the light stimulation was removed for 80 s. As expected, the recorded images exhibit enhanced intensity according to the increased attention (Figure 4d). The multimodal sensing, processing, and storage of the reported memristors are in Table S1, Supporting Information. Compared with other reported memristors,[7, 14, 16, 24, 39, 40] the advantage of our device is its multimodal sensing ability without the integration of additional sensors. More importantly, the device exhibits long-term memory, suggesting the potential multimodal sensing, memory, and processing capabilities of the device. As for the bias-modulation mechanism, the electric field-induced oxygen migration between the GO-TiO2 layer and the ITO layer is taken into account. It has been reported that ITO can behave as an oxygen reservoir for supplying or receiving oxygen ions.[41, 42] By applying a positive bias on the ITO electrode, some oxygen ions will drift out of the GO-TiO2 nanocomposite into the ITO. As a result, the RGO percolating conduction channel is strengthened. Conversely, the oxygen ions in the ITO can be injected into the GO-TiO2 nanocomposite, which weakens the RGO channel.
As aforementioned, the most important features of our memristor are its multimodal sensing and synaptic processing. The device exhibits two behaviors, humidity-dependent and electric field-modulated optical switching/synaptic behavior, which enable the memristor to emulate neuromorphic humidity-modulated and attention-driven visual behavior, respectively. As well known, personal attention affect the ocular system and environmental humidity indirectly affect visual performance by influencing eye comfort,[38, 43-45] which in turn affects subsequent judgments and actions. Therefore, as a proof-of-concept, a sensorimotor system was constructed by transmitting the synaptic feedback of the optoelectronic memristor to a robotic arm, which mimics neuromorphic attention-driven visual motion function at high and low humidity, respectively (Figure 4a). The unique features of this system are its simple circuit and flexible modulated response. In traditional multimode sensorimotor systems, different types of sensors, memory, and processing units are required. By contrast, the proposed device includes multimodal sensing, memory, and processing capabilities, and this feature reduces the circuit complexity of the system. Devices with mechanical apertures are essential for visual modulation in traditional sensorimotor systems. For example, the opening of the pupil must be controlled to stabilize the received light intensity in a biological visual system. However, mechanical aperture devices are complicated in structure and require high operating voltages. Because the sensorimotor system has a bias-modulated current response, its response can be flexibly modulated by applying a small external voltage. In our sensorimotor system, the response ΔI of the optoelectronic memristor can be modulated by UV light pulses at different RHs and biases. Subsequently, the response ΔI was converted into action instructions through the Python program, and transmitted to the robot arm to control its spatial position. The corresponding relationship (Equation (1)–(3)) between ΔI and the spatial position of the robot arm is shown in Figure 4b. It should be noted that a coordinate (X, Y, Z) of the robot arm is its spatial position after stimulation. Here, the initial position of the robot arm is set to (0, 0, 0), and the target object is placed at (0, Dy, Dz) in front of the robot arm, where Dy and Dz were 5 and 20 cm, respectively. (In−I0)a% represents the change of the response current from the initial to the terminal state after the optoelectronic memristor received n light pulses at a% RH. In addition, ΔIt is a preset threshold (0.2 nA) that represents the minimum change in response current required to successfully grasp the target. We first demonstrated the system grabbing red objects (high attention, 0.50 V bias) and green objects (low attention, −0.50 V bias) in a high-humidity environment (85% RH; Figure 4c,d, respectively). The response ΔI of the memristor is shown in Figure S7, Supporting Information. Four states (the initial state, state 1, state 2, and the terminal state) were recorded. In state 1, state 2, and the terminal state, the memristor was stimulated by one, two, or three light pulses (intensity: 4.5 mW cm−2; width: 2 s; frequency: 0.25 Hz). As shown in Figure 4c,d, the robot arm moved toward the target object during light stimulation. The red object with high attention was finally grabbed and the green object with low attention was not grabbed. The robotic arm's response to the high-attention red object was also tested in a low-humidity environment (25% RH, bias: 0.50 V). The movement of the robot arm was significantly decreased, and the object was not grabbed (Figure 4e). The entire movement process of the robot arm is shown in Movie S1. The sensorimotor system not only possesses the capability to integrate multiple environmental information but also simulates the influence of attention on visual functions, demonstrating promising application potential for future humanoid robotics.
3 Conclusion
In this study, a GO-TiO2 composite optoelectronic memristor with integrated light/humidity bimodal sensing, memory, and processing functions was proposed. The synaptic plasticity can be used to sense light/humidity information, implement memory, and emulate contrast enhancement and attention-driven functions similar to those of the human visual system. A proof-of-concept sensorimotor system with attention-driven and RH-modulated behaviors was presented that sends the synaptic feedback of the memristor to a robot arm. Notably, this work implemented multimodal sensing, memory, and processing in a single cell, which will help to reduce the energy consumption and time latency in artificial neural systems. The optoelectronic memristor has potential in the development of efficient perception systems for humanoid robots.
4 Experimental Section
Preparation of the GO/TiO2 Solution
Graphene oxide (GO) powder and titanium dioxide (TiO2) powder were obtained from Kerui Nano (Guangdong) and Ningbo Jiweina New Material Technology, respectively. The sheet diameter of the GO was 5–11 μm and the particle size of the TiO2 was 7–10 nm. The GO-TiO2 solution was prepared by dissolving 72 mg of GO powder and 8 mg of TiO2 powder in 16 mL deionized water with ultrasonic and stir treatment for 30 min each.
Device Fabrication
The GO-TiO2 nanocomposite was spin-coated on the bottom ITO electrodes with a thickness of ≈200 nm (spin speed: 3000 rpm). Gold (Au) was deposited by thermal evaporation on the GO-TiO2 nanocomposite layer, which was covered with a circular mask, 150 μm in diameter, to form the Au/GO-TiO2/ITO structure.
Characterization of the Device
The surface morphology of the GO-TiO2 nanocomposite film was examined using a scanning transmission electron microscope (Tecnai G2, Thermo Fisher) at an accelerating voltage of 200 kV. The optical signals during the measurement were generated by a xenon lamp (LA-410UV, Hayashi). UV–visible absorption spectra were recorded by a spectrophotometer (843-R, Newport). The Raman spectra and XPS analysis were performed using inVia Reflex, with a laser excitation wavelength of 532 nm, and a photoelectron spectrometer (ESCALAB 250Xi, Thermo Fisher), respectively. The device was placed in a controlled humidity chamber during the tests. The dry nitrogen gas was divided into two parts, one of which was moistened through a water saturated chamber. Then, the required RH for measurement was obtained by mixing the appropriate ratio of the dry and wet flows. The relative fluctuation in humidity did not exceed ±5%. All electrical measurements were performed using a semiconductor parameter analyzer (2636 A, Keysight) equipped with a humidity test chamber.
Acknowledgements
This work was supported in part by the in part by the Science and Technology Development Plan Project of Jilin Province, China, under grant no. 20240101018JJ; in part by National Natural Science Foundation of China for Distinguished Young Scholars under grant no. 52025022; in part by the National Natural Science Funds of China under grant nos. 52072065, 52272140, and U23A20568; in part by the Fund from Jilin Province under grant nos. 20240210001GX and 20240210002GX.
Conflict of Interest
The authors declare no conflict of interest.
Author Contributions
Qiaoling Tian: conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal); writing—original draft (equal). Xinyu Sui: conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal). Xiaoning Zhao: conceptualization (equal); formal analysis (equal); funding acquisition (equal); project administration (equal); validation (equal); writing—review and editing (equal). Ya Lin: supervision (equal); writing—review and editing (equal). Zhongqiang Wang: formal analysis (equal); funding acquisition (equal); writing—review and editing (equal). Ye Tao: validation (equal); writing—review and editing (equal). Haiyang Xu: formal analysis (equal); funding acquisition (equal); project administration (equal). Yichun Liu: supervision (equal); validation (equal). Qiaoling Tian and Xinyu Sui contributed equally to this work.
Open Research
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
