Traditional Von Neumann architecture is limited by the time delay and the power consumption due to the separation of processing and memory units^[1−3]. Reservoir computing (RC), as inspired by the functions of human brain, is considered as a promising architecture to reduce latency and energy consumption though temporal and sequential data processing^{[4, 5]}. Synapses and neurons are the foundations for human brain^[6]. Various emerging electronic devices have been proposed for artificial synapse applications, such as memristor, transistor, and phase change memory^[7−10]. Among them, memristor stands out with simple structure, low power consumption, and fast switching speed^[11−13]. Optoelectronic memristors with hybrid electrical- and optical-singles as stimuli provides a platform for coupling of photons and electrons for efficient computing^[14−17].

The development of optoelectronic memristor is still at initial stages. Various optoelectronic materials have been proposed to develop memristors for emulating synaptic plasticity, such as two-dimensional materials, metal oxides, perovskite and carbon materials^[18−23]. For example, Li et al. proposed CeO₂/MoS₂ optoelectronic memristor with light sensing and memory capability^[18]. Yang et al. developed α-In₂Se₃-based optoelectronic memristors with short-term plasticity to construct multi-mode RC system^[20]. Lu et al. developed carbon dot optoelectronic memristor with integrated storage and computing capabilities for neuromorphic computing^[21]. In previous reports, many amorphous materials (e.g., niobium based compounds) have been used in memristor, which has outstanding performance in ultralow operating voltage and stability^[24−26]. Amorphous carbon (a-C) is one kind of promising electronic materials with large area preparability and complementary metal−oxide−semiconductor (CMOS) compatibility^{[27, 28]}. However, a-C is not sensitive to light, which hinders its optoelectronic applications. As one of the emerging materials in electronic and optoelectronic applications, tellurium (Te) has extraordinary carrier mobility and significant optical absorption^[29−31]. More importantly, Te can be employed as a potential optoelectronic infrared material due to superior characteristics of narrow-bandgap semiconductor with air stability. This is an advantage to the development of infrared technology in critical applications such as night vision, military communication and autonomous vehicle^{[32, 33]}. The introduction of Te into some materials has proved to be an effective means to boost their optoelectronic properties^[34].

In this work, an optoelectronic memristor based on Te doped a-C (a-C:Te) thin film is developed for multi-mode RC application. The device with Au/a-C:Te/Pt structure exhibited volatile switching behaviors under electrical and optical stimulation. Basic synaptic behaviors including excitatory post-synaptic current (EPSC) and pair-pulse facilitation (PPF) were emulated by the memristor. More importantly, the device exhibited distinguishable current states by controlling the electrical/optical pulse sequences, which can be used as hybrid optical-electrical reservoir states. A multi-mode RC system was constructed to emulate human tactile-visual fusion recognition of Chinese character and an accuracy of 98.7% was achieved.

The Au/a-C:Te/Pt optoelectronic memristor were prepared as follows: Firstly, the a-C:Te thin film was deposited on Pt/Ti/SiO₂/Si substrate by radio frequency (RF) sputtering of a carbon target (power: 100 W) and a Te target (power: 40 W) at room temperature. The chamber pressure was 1 Pa with Ar atmosphere. During the 60 min carbon sputtering process, Te was sputtered for 15 s at 30 min because of its fast sputtering rate. Secondly, the sample was annealed at 150 °C for 30 min in nitrogen environment. Finally, Au electrodes with circular pad were evaporated through a shadow mask.

The switching behaviors of the Au/a-C:Te/Pt device were measured by TGA12104 arbitrary waveform generator, Tektronix DPO70404C digital oscilloscope, Keysight 2636B sourcemeter and infrared lamp.

Fig. 1(a) shows the schematic diagram of Au/a-C:Te/Pt memristor. The cross-sectional scanning electron microscope (SEM) image of the device is shown in Fig. 1(b). It can be observed that the thickness of deposited a-C:Te layer between two electrodes is ~60 nm. Figs. 1(c) and 1(d) show X-ray photoelectron spectroscopy (XPS) spectra of the a-C:Te thin film. The Te doping concentration of about 0.44% atomic ratio was acquired based on the XPS data. The Te 3d_3/2 and Te 3d_5/2 peaks centered at 586.4 and 576.0 eV can be assigned as Te−C, and that centered at 582.8 and 572.4 eV can be assigned as Te⁰. The result indicates Te⁰ is the dominant component. They may be able to form Te nanoparticles in the thin film. During thermal treatment, Te⁰ atoms would migrate randomly and are thermally unstable because of their high surface energy. They tend to nucleate and stick together to form nanoparticles to minimize the surface energy. C sp²-satelite, C sp³, and C sp² were observed, which is in agreement with the a-C^{[35, 36]}. The absorption spectras of the pure a-C film and a-C:Te film are shown in Figs. 1(e) and 1(f). Compared with the pure a-C film, the a-C:Te film shows absorption capability in the near infrared region, demonstrating the infrared photosensitive property of the a-C:Te film. Light beam with a wavelength of 800 nm was applied as optical stimulus.

The synaptic plasticity of the memristor was studied. A series of electrical pulses were applied on the Au electrode as shown in Fig. 2(a). Fig. 2(b) shows the EPSC behaviors of Au/a-C:Te/Pt memristor under varying stimulation amplitudes. Response current increases with higher voltage amplitude, indicating EPSC behavior dependent on voltage amplitude. The device can emulate PPF function with volatile characteristic by paired pulses (V = 4 V, W = 20 μs) with an interval time (Δt) of 40 μs (Fig. 2(c)). It can be observed that the response current increases under the electrical pulse stimulation and exhibits relaxation process as the stimulation removal. Meanwhile, the maximum response current of the second electrical stimulation (A2) was larger than that of the first one (A1). Here, PPF index was defined as (A2 − A1)/A1 × 100%. Fig. 2(d) shows the PPF index as a function of Δt. When the time interval decreases from 160 to 10 μs, PPF index increases from 20.0% to 46.7%. The Au/a-C:Te/Pt memristor also shows volatile behavior by applying different number of electrical pulses (#4, #6, #8, and #12), as shown in Fig. 2(e). The peak value of the response current increases with the increase of pulse numbers at a fixed amplitude (V = 4 V, W = 10 μs, Δt = 10 μs) and eventually returns to the initial state after the pulse stimulus is removed. Fig. 2(f) shows the response current of the 4-bit input electrical pulses 1000, 1110, and 1111, where the presence and absence of an electrical signal were denoted as "1" and "0", respectively. The results of different current states indicate that the final state of the memristor depends on both of the last stimulation and the historical stimulation. The energy consumption of one reservoir state can be as low as ~60 pJ. Furthermore, the Au/a-C:Te/Pt device exhibits distinguishable current responses under 4-bit electrical pulse sequences, which can be used as reservoir states of RC to perform image recognition task (Fig. 2(g)). As a proof of concept, the input image ("0"−"9") is divided into 5 lines and each line is converted into a 4-timeframe input signal (Fig. 2(h)). In each timeframe input signal, the white pixel "0" and the red pixel "1" represent zero bias and 4.0 V/20 μs voltage pulse, respectively. For example, the conversion array of image "1" includes five timeframe input signals: "0010". Under the stimulation of different timeframe input signals, the five final response current combinations representing each digit are significantly different, verifying the capability of the reservoir to recognize these 10 digits.

In addition to electrical stimulation, the optical response of the Au/a-C:Te/Pt memristor was studied. Fig. 3(a) shows the schematic diagram of the Au/a-C:Te/Pt device with optical stimulation (wavelength: 800 nm), and the response current of the device was monitored by a 0.1 V reading voltage. Fig. 3(b) shows the EPSC response of the memristor under different intensities (25, 35, 60, 100, and 150 mW/cm²) of optical pulse (W = 5 s). It is obvious that the response current is enhanced with the increase of light intensity. In Fig. 3(c), PPF function was achieved by paired optical pulses with Δt of 5 s. The relationship between the optical PPF index and Δt is similar to that of electrical stimulation. Response current of the device stimulated by 4-bit optical pulse was also studied. As shown in Fig. 3(d), the distinguishable states indicate that the Au/a-C:Te/Pt memristor has the potential to implement optical RC.

The possible working mechanism of the present Au/a-C:Te/Pt memristor under electrical and optical pulses is schematically illustrated in Fig. 3(e). When the light is turned off, detrapping of charge carriers may be the key to perform volatile behavior. When electrical stimulation is applied, electrons are injected from the electrode and get trapped in the Te trap sites in the a-C:Te film. The injected electrons then gradually fill the trapping sites, which will produce hopping conductive channel^[37]. Therefore, the response current of the device increases. The trapped electrons will thermally release after the removal of the electrical stimulation, resulting in the resistance of the device to return to initial state. On the other hand, light-modulated response current can be considered as the photogenerated electrons excited by Te being captured in trap sites and hopping conduction^[38]. When the light is turned off, recombination of photogenerated electrons and holes may be the key to return to the pristine state.

Human interact with external environment through multi-sensory systems. The information obtained through visual and tactile process accounts for 85% of all information. They cooperate with each other to transmit the correct information that allows the brain to process and make decision^{[39, 40]}. In traditional RC systems, multiple reservoirs are required for feature extraction due to each reservoir can only respond to one type of input signal. It is worth noting that multi-mode RC computing can be realized if the reservoir can respond to different types of input signals. Feature extraction and multi-sensory fusion can be achieved in a signal reservoir to improve efficiency. The Au/a-C:Te/Pt device exhibits volatile behavior and nonlinear dynamic properties under electrical/optical stimuli, which can be used for multi-modal reservoir computing. Therefore, a hybrid optoelectronic muti-mode RC system was constructed. The Chinese character recognition task of RC system is implemented based on MATLAB simulation^{[41, 42]}, which is divided into the following three processes. (1) The Chinese character image was converted into a 5 × 8 pixels. As shown in Fig. 4(a), the half feature of two Chinese characters prevents us from getting the correct information. The left half can be sensed by vision (optical signals), while the right half requires the help of a pressure sensor that converts the sense of touch into an electrical signal (The sensing and processing steps of the pressure sensor are omitted). (2) Each pixel were binarized into optical/electrical signals input to the Au/a-C:Te/Pt memristor, where gray pixels represented the presence of input signals and white pixels represented no signal. The image was converted into reconstructed 10 × 4 pixels though four columns from left to right in each row. The response signals of Au/a-C:Te/Pt memristor were input into the reservoir to distinguish characteristics by reading the response currents of 10 memristors. (3) Six selected Chinese characters were input into the neural network for repeated training. During the process of high-dimensional mapping, the reservoir divided the input signals into different classes. Finally, the tested Chinese characters were input into the trained neural network for recognition, and the correctness of the output Chinese characters was judged. Fig. 4(b) shows the recognition accuracy of Chinese character in visual, tactile, and tactile-visual fusion modes. Compared with single mode RC, the recognition accuracy of multi-mode RC is improved to 98.7%. Fig. 4(c) shows the confusion matrix in classifying mixed Chinese characters. The prominent diagonal represented by the red element indicates successful recognition of all Chinese characters. The results suggest the superiority of muti-mode RC system.

In summary, we reported an optoelectronic memristor with a-C:Te thin film for muti-mode RC. The device exhibited volatile switching behaviors under electrical/optical pulses stimulation, and synaptic functions including EPSC and PPF were achieved. The 4-bit electrical/optical pulse sequences can be used to modulate the response current with distinguishable states. Taking advantage of hybrid optical-electrical stimulation, a muti-mode RC system was constructed to achieve Chinese character recognition with a high accuracy of 98.7%. The optoelectronic memristors provide possible potential for highly efficient multi-mode RC application.