Near-Infrared Synaptic Responses of WSe₂ Artificial Synapse Based on Upconversion Luminescence from Lanthanide Doped Nanoparticles

Abstract

Near-infrared (NIR) photoelectric synaptic devices show great potential in studying NIR artificial visual systems integrating excellent optical characteristics and bionic synaptic plasticity. However, NIR synapses based on transition metal dichalcogenides (TMDCs) suffer from low stability and poor environmental performance. Thus, an environmentally friendly NIR synapse was fabricated based on lanthanide-doped upconversion nanoparticles (UCNPs) and two-dimensional (2D) WSe₂ via solution spin coating technology. Biological synaptic functions were simulated successfully through 975 nm laser regulation, including paired-pulse facilitation (PPF), spike rate-dependent plasticity, and spike timing-dependent plasticity. Handwritten digital images were also recognized by an artificial neural network based on device characteristics with a high accuracy of 97.24%. In addition, human and animal identification in foggy and low-visibility surroundings was proposed by the synaptic response of the device combined with an NIR laser and visible simulation. These findings might provide promising strategies for developing a 24/7 visual response of humanoid robots.

1. Introduction

With the development of applications in real-time pattern recognition and edge intelligence, the demand for computing efficiency is increasing rapidly. Traditional von Neumann architecture faces bottleneck problems of transmission delay and high energy consumption due to storage and computing separation [1,2,3]. In contrast, the human brain possesses an efficient and low-consuming superiority for integrating memory and computation with neurons and synapses to mitigate the von Neumann bottleneck [4,5]. It is worth noting that the visual system plays a leading role in external information perception. It undertakes more than 80% of the information acquisition tasks by transmitting signals from the retina to the brain via neurotransmitters [6,7,8]. To mimic the visual sensing function of the human brain, researchers have successfully developed photoelectric synaptic devices using various photosensitive materials such as perovskites [9], polymers [10,11], transition metal dichalcogenides (TMDCs) [12,13,14,15]. Among them, two-dimensional (2D) TMDCs attract much attention due to their perfect advantages of high carrier mobility, strong absorption of visible light, low toxicity, and high stability [16,17,18,19]. Notably, the layered structure of WSe₂ endows unique interfacial properties, demonstrating significant potential in constructing photo-responsive heterostructures. Most TMDC-based photoelectric synaptic devices reported to date operate primarily in the visible range, as their response to the near-infrared (NIR) region is limited by the intrinsic bandgap [20,21,22]. Compared with visible photoelectric synapses, NIR photoelectric synapses are not limited by day and night light conditions, allowing them to work stably at night or in dark environments, which makes it possible for the collaborative application of infrared night vision systems [23]. In addition, thanks to its long-wavelength characteristics, NIR has the advantage of deeper signal penetration in haze and smoke environments. Thus, NIR photoelectric synapses can maintain stable recognition even in low-visibility environments (e.g., sandstorms or hazy weather), and have important value in key areas of autonomous driving, security monitoring, and military applications. Therefore, there is an urgent need to develop NIR photoelectric synaptic devices.

To improve the NIR response of TMDC-based photoelectric synapses, more attention has been paid to heterojunction structures, such as CoTe₂/ZnO/WS₂, In₂Se₃/MoS₂, and Ta₂NiSe₅/SnS₂. Ray et al. expanded the response of the TMDC synapse to 800 nm by interfacial band alignment of CoTe₂, ZnO, and WS₂ [24]. Hu et al. extended the NIR response to 1060 nm by constructing an In₂Se₃/MoS₂ heterostructure [25]. And Li et al. proposed a physisorption-assistant optoelectronic synapse for Ta₂NiSe₅/SnS₂ with an extended modulation at 1310 nm [26]. Yet, the complex preparation process and sensitivity of light and water make it difficult to control interface defects and environmental sensitivity. Another method was explored with lanthanide-doped upconversion nanoparticles (UCNPs) and 2D TMDC nanosheets. UCNPs convert low-energy NIR photons into high-energy visible or ultraviolet light by an upconversion process, which could be absorbed by TMDCs. Thus, the light response range of TMDCs was expanded to NIR. Han et al. proposed a NIR neuromorphic computing based on UCNP-sensitized MoS₂ nanocomposites floating gate dielectric with pentacene as the semiconductor channel [27]. Long-term potentiation and depression (LTP/D) were achieved for the first time with an accuracy of 70% in handwritten character recognition tasks. However, the MoS₂ nanosheets reduced the transport performance and the sensitivity of pentacene to light and air, also confining the practical applications. Thus, it is urgent to improve the NIR response of TMDC-based photoelectric synapses with a simple process and good stability.

In this work, we constructed a novel UCNPs-WSe₂ optoelectronic synapse device based on upconversion luminescence (UCL) from lanthanide-doped nanoparticles with spin-coating UCNPs on the surface of WSe₂, which extends the photoresponse to 975 nm. Direct spin-coating of UCNPs onto large-area WSe₂ ensures continuous WSe₂ channels to avoid discontinuous electrical transportation. Meanwhile, direct contact between WSe₂ and UCNPs reduces the sensitivity of organic linkers. The prepared synaptic device successfully simulated core biological synaptic functions, including paired-pulse facilitation (PPF), spike rate-dependent plasticity, and spike timing-dependent plasticity. Then, the handwritten digit recognition was achieved with 97.24% accuracy on the basis of two types of LTP in UCNPs-WSe₂ optoelectronic synapse. Moreover, a dual-mode scheme was proposed to identify the target in foggy environments. These findings provide an innovative solution for autonomous driving, industrial inspection, and disaster relief.

The prepared synaptic device successfully simulated core biological synaptic functions, including paired-pulse facilitation (PPF), spike rate-dependent plasticity, and spike timing-dependent plasticity. Then, the handwritten digit recognition was achieved on the basis of two types of LTP in UCNPs-WSe₂ optoelectronic synapse. Moreover, a dual-mode scheme was proposed to identify the target in foggy environments. These findings provide an innovative solution for autonomous driving, industrial inspection, and disaster relief.

2. Results and Discussion

2.1. NIR-Responsive Artificial Synapse

In biological neural networks, synapses, as the functional basis unit, can enhance synaptic connections by conducting neurotransmitters released by presynaptic neurons and generate excitatory postsynaptic current (EPSC). Taking the human retina as an example, this biological structure can transmit external light signals to the brain through neurotransmitters to store and process (Figure 1a(i)) [28]. However, the biological retina cannot directly perceive NIR stimulation due to the light response range of opsin. To enhance NIR response, an artificial UCNPs-WSe₂ optoelectronic synaptic device (Figure 1b) was constructed. Those NIR responsive synapses will help to recognize the kitten silhouettes if a robot is equipped with this device (Figure 1a(ii)). Under the condition of light response characteristics of the NIR band regulated by UCNPs and the visible region responded to by WSe₂, the simulated robot system can adapt to the all-weather environment of super bionic vision.

Figure 1b shows a schematic diagram of the UCNPs-WSe₂ photoelectric synaptic device. When the device is irradiated with NIR, the visible light emitted from UCNPs excited by 975 nm is reabsorbed by WSe₂ to achieve precise modulation of NIR synaptic performance. By characterizing the WSe₂ material prepared by physical vapor deposition (PVD), it can be seen that the synthetic WSe₂ is a homogeneous and high-quality crystalline material with a bilayer structure, as shown in Figure S1. To study the synaptic characteristics of the device, we first conducted an output performance test. Figure 1c shows the output hysteresis curve of the UCNPs-WSe₂ device under the dark and 975 nm states, which indicates the synaptic potential. Compared with dark conditions, the 975 nm irradiation hysteresis window was significantly larger. This improved performance originated from the UCL effect of UCNPs, which activates WSe₂ photoresponse characteristics. Higher carrier concentrations enhanced the probability of capturing electrons, leading to longer electron release during backward scanning and a larger hysteresis window. And the defect-state in WSe₂ enhanced the carrier capture/release effect. This phenomenon suggests that devices play a greater role in regulating synaptic plasticity under NIR irradiation.

2.2. Mechanism Analysis of UCNPs-WSe₂ Synaptic Device

To gain an in-depth understanding of the working mechanism of UCNPs-WSe₂ synaptic devices, we considered the schematic diagram of the energy levels (Figure 2). Figure 2a shows the energy band diagram of Ag and WSe₂. Before contacting (Figure 2a), the work function (W_m) of Ag is 4.2 eV, and the electron affinity (χ_s) of WSe₂ is 4.0 eV [29,30,31]. The bandgap (E_g) of WSe₂ is 1.59 eV [32]. Because of the p-type semiconductor characteristics of WSe₂, according to the transfer characteristic curve in Figure S2, the Fermi level is located near the valence band. When Ag is in contact with WSe₂, electrons from Ag are injected into WSe₂, resulting in bending of the WSe₂ surface energy band, as shown in Figure 2b. When a forward bias is applied, the potential gradient between the source (left) and the drain (right) causes the semiconductor energy band to gradually tilt downward from the source to the drain [33,34,35]. Under this bias voltage, the holes in the valence band and electrons in the conduction band migrate to the source and drain electrodes along the inclined energy band, respectively. For deep energy-level traps formed by Se vacancy defects in WSe₂, lots of electrons are captured. And the captured electrons will be rereleased when thermal excitation or an electric field drives them. The difference in relaxation times during the capture–release process results in a hysteresis of current response [12], consistent with the test results in the dark conditions of Figure 1c. As shown in Figure 2c, under light conditions, photons will generate electron–hole pairs, leading to a higher current. Meanwhile, part of the photons generated electrons in the conduction band are captured by the Se vacancy defects, and their dynamic capture–release process results in the redistribution of the interface carriers, ultimately causing the device to present a larger hysteresis window (red curve in Figure 1c).

To explain the energy transfer mechanism between WSe₂, UCNPs-WSe₂ and UCNP absorption and photoluminescence (PL) spectra of WSe₂, UCNPs, and UCNPs-WSe₂ composites were characterized. Two absorption peaks at 595 nm and 762 nm were detected, corresponding to A and B exciton resonance characteristics (Figure 2d) [36,37], and indicating the strong absorption capacity of the material in the visible light region. An absorbed band around 980 nm was presented from NaYF₄: 20%Yb, 0.5%Tm UCNPs, originating from the ²F_7/2 → ²F_5/2 transition of Yb³⁺, and a weak absorption peak at 980 nm was observed in UCNPs-WSe₂ composites, suggesting the potential of an NIR response of the UCNPs-WSe₂ device. Further, PL spectra were conducted to reveal the effects of UCNPs on 2D WSe₂ (Figure S3), showing no significant effect of UCNPs on the luminescent properties of WSe₂. Notably, UCNPs exhibit emission peaks of Tm³⁺ at 452 nm, 476 nm, 648 nm, 695 nm, and 792 nm, upon excitation by 980 nm, corresponding to the transitions of ¹D₂ → ³F₄, ¹G₄ → ³H₆, ¹G₄ → ³F₄, ³F₃ → ³H₆, and ³H₄ → ³H₆, respectively (Figure 2e) [38]. Notably, the luminescence intensity of these emission peaks was significantly reduced or even completely quenched when UCNPs were composited with 2D WSe₂, indicating an efficient energy transfer from UCL of UCNPs to WSe₂ [39]. Figure 2f shows the schematic diagram of UCL from Yb³⁺ to Tm³⁺ and energy transfer to WSe₂. When 975 nm NIR is applied, Yb³⁺ ions effectively absorb 975 nm photons from the ground state ²F_7/2 to the excited state ²F_5/2, then transfer energy to the excited state ³H₅ of Tm³⁺, followed by relaxation to ³F₄ state by non-radiative relaxation. And the ³F₄ state of Tm³⁺ will be pumped to the ³F_2,3 excited states after transferring the second photon energy from Yb³⁺. Thereby, ¹G₄ and ¹D₂ states will be generated after absorbing the third and fourth photon energy from Yb³⁺. And radiative transitions from the ¹G₄ and ¹D₂ states provided visible luminescence around 452 nm, 476 nm, 648 nm with ¹D₂ → ³F₄, ¹G₄ → ³H₆, ¹G₄ → ³F₄ transitions, respectively. Meanwhile, the visible luminescence was reabsorbed by WSe₂ by radiative reabsorption, enabling the NIR response of WSe₂. This multi-level energy regulation lays the physical foundation for the bionic modulation function of photo-controlled synaptic devices.

2.3. Properties of UCNPs-WSe₂ Synaptic Device

Based on the fact that WSe₂ can effectively absorb the visible light emitted by UCNPs, the synaptic function of the UCNPs-WSe₂ device was simulated with the excitation at 975 nm. The detailed mechanism of optical energy transfer to synaptic electroplasticity is shown in Figure S4. When the NIR pulse was applied, the device produced a typical EPSC response. Single-pulse-triggered EPSC under different power and pulse width conditions was increased significantly with higher optical power density and longer pulse width (Figure 3a,b and Figure S5a,b). In addition, compared with the WSe₂ device without spin-coating, the photoresponse was significantly enhanced after spin-coating UCNPs (Figure S6), which demonstrated that the bandgap mismatch between WSe₂ and NIR was effectively overcome.

PPF is another typical short-term potentiation (STP) behavior that is conducive to the signal transmission efficiency between neurons. When two consecutive pulses were applied, the second pulse caused a greater change in synaptic weight than the first one. This was attributed to the increased synaptic weight due to the failed recovery of carriers to their initial states within the time interval. Here, PPF behavior was demonstrated successfully under two 975 nm pulses with a width of 3 s and an interval of 1 s. The PPF index was defined and calculated by the following formula [40,41,42,43]:

$$\mathrm{P}\mathrm{P}\mathrm{F} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}={\displaystyle \frac{{\mathrm{A}}_2}{{\mathrm{A}}_1}}\times 100\mathrm{\%}$$

(1)

where A₁ and A₂ represent the ΔEPSC amplitudes triggered by the first and second pulses, respectively. PPF reached a maximum of 160.15% with a time of 23 s. The correlation between the PPF index and time interval can be fitted using the bi-exponential equation (Figure 3c) [40]:

$$\mathrm{P}\mathrm{P}\mathrm{F} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}=1+{\mathrm{B}}_1\times \mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{\mathsf{\Delta }\mathrm{t}}{{\mathsf{\tau }}_1}}\right)+{\mathrm{B}}_2\times \mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \frac{\mathsf{\Delta }\mathrm{t}}{{\mathsf{\tau }}_2}}\right)$$

(2)

where B₁ and B₂ are the initial excitation amplitudes, ${\mathsf{\tau }}_1$ and ${\mathsf{\tau }}_2$ are the fast and slow characteristic relaxation times during decay, respectively. The calculated ${\mathsf{\tau }}_1$ and ${\mathsf{\tau }}_2$ values are 4.38 s and 84.24 s, respectively, which are consistent with the characteristics of biological synapses [44].

The transformation from STP to LTP is of great significance to the development of neuromorphic computing systems with adaptive learning capabilities. This is because STP mainly implements dynamic signal coding and temporary information cache, while LTP is a persistent change, providing the basis for long-term memory storage and neural network function reshaping. Therefore, synaptic plasticity was investigated with different power densities, numbers, and frequencies of light pulses (Figure 3d–f and Figure S7). The transition of EPSC amplitude from STP to LTP was realized by increasing incoming light intensity from 4.95 to 31.8 W/cm², the number of pulses from 600 to 3000, and pulse frequency from 20 to 200 Hz. The fundamental reason for this is that larger light intensity/pulse numbers/pulse frequency induces a larger concentration of carriers in 2D WSe₂, resulting in capturing more electrons in defective states and a longer recovery [12,45,46].

2.4. Learning Experience Behavior and Handwritten Digital Image Recognition

A UCNPs-WSe₂ synapse can simulate learners’ characteristics triggered by 975 nm pulses (Figure 4a), where the learning–forget–release process can be achieved. After applying 4000 pulses, ΔEPSC increased from 0 to 2.23 nA, corresponding to the first learning process (i). When the pulses were removed, ΔEPSC was attenuated to 0.75 nA within 7 s, representing the forgetting process (ii). In the re-learning phase (iii), only 1600 pulses were required to achieve the same ΔEPSC triggered by the 4000 pulses in the first learning phase.

An artificial neural network (ANN) for handwritten digit recognition was constructed based on the synaptic plasticity of the device in the Vis-NIR range. According to the strong absorption characteristics of WSe₂ in the visible range [47], the synaptic performance of the UCNPs-WSe₂ device was modulated by 405 nm visible light (Figures S8 and S9). The synaptic weights were regulated by timing-controlled 405 nm and 975 nm NIR pulses. First, 10 pulses at 405 nm (26.68 mW/cm², width 0.5 s) were applied, followed by 10 975 nm pulses (27.84 W/cm², width 1 s). Conductance values were extracted by simulating the continuous variation of LTP with the timing of the two wavelength pulses (Figure S10). The successful simulation of two types of LTP by optical pulses highlights its unique advantages in the field of photo-controlled synaptic bionics. Then, a three-layer ANN was constructed containing 784 input neurons, 300 hidden neurons, and 10 output neurons, as shown in Figure 4c. Training and recognition of 28 × 28 pixel handwritten digits from the Modified National Institute of Standards and Technology (MNIST) database were performed using the CrossSim simulator. During the training process, 784 input neurons corresponded to a handwritten digital image, while 10 output neurons matched 10 numbers (0–9). The single-layer structure of the network performs vector matrix multiplication and sum operations by mapping the cross-switch array hardware (Figure 4d) to achieve efficient computing, thereby enabling large-scale parallel signal processing. As shown in Figure 4e, the recognition accuracy corresponding to the regulated synaptic weights in Figure 4b increases gradually with the increase in training rounds. The accuracy rate finally reached 97.24% after 20 training rounds, giving great potential in the field of neuromorphic vision.

2.5. Possible Application Directions of UCNPs-WSe₂ Device

Inspired by the sensing and information processing mechanism of the viper’s vision system, the identification of birds in a day–night environment was proposed according to visible and NIR response characteristics of the UCNPs-WSe₂ photoelectric synaptic device by setting the recognition EPSC threshold. In viper’s vision system, NIR and visible signals are integrated through neural integration to generate dynamic thermal imaging maps, allowing them to achieve millimeter-level spatial positioning in day–night environments, which is particularly prominent when tracking mobile targets [48,49,50]. The threshold for “recognized” and “unrecognized” was set based on test results. For the identification of birds in daytime mode, when 405 nm pulsed light (26.68 mW/cm², width 3 s) was applied, the EPSC of UCNPs-WSe₂ synapse exceeded the preset threshold, indicating that the capability of identifying the bird (Figure 5a(i),b). When the environment switches to night mode, the system automatically enables the NIR sensing mechanism. The induced EPSC of UCNPs-WSe₂ synapse also reached the recognition threshold triggered by NIR 975 nm laser (31.8 W/cm², width: 3 s), which shows no response to a 36 °C living body.

Then, the robot’s recognition of riders on a foggy day was proposed, triggered by visible and NIR pulses. The threshold for “recognized” and “unrecognized” was set based on test results. The EPSC intensity in visible mode is lower than the preset recognition threshold, causing failure to identify the rider, as shown in Figure 5c(i),d(left). By combining NIR excitation, the EPSC intensity exceeded the threshold, enabling clear contour recognition of the rider (Figure 5c(ii),d(right)). This dual-mode optical fusion strategy allows robot vision systems to maintain robust recognition capabilities under harsh conditions.

3. Materials and Methods

3.1. WSe₂ Synthesis

The 2D WSe₂ was grown by PVD [51]. High-purity WSe₂ powder (99.9%, Alfa Aesar, Shanghai, China) was placed in a quartz boat as raw material and placed in the center of a 1-inch quartz tube. The 285 nm SiO₂/Si substrate (Lijing Electronics Co., Ltd., Shenzhen, China) was placed 20 cm downstream from the raw material. At a continuous rate of 20 sccm Ar (99.999%, Wangzhou Gas Co., Ltd, Nanning, China), the tube furnace was heated to 1185 °C for 5 min. Then, the furnace temperature was naturally cooled to room temperature.

3.2. Preparation of UCNPs (NaYF₄: 20%Yb³⁺, 0.5%Tm³⁺)

UCNPs were synthesized by the hydrothermal method [52]. For this, 0.3 g sodium hydroxide (99.99%, Sigma-Aldrich, St. Louis, MO, USA) was first weighed and dissolved in 1.5 mL of deionized water. Subsequently, 5 mL of oleic acid (90%, Sigma-Aldrich, St. Louis, MO, USA) and anhydrous ethanol were added as additives, and the mixture was thoroughly mixed using a magnetic stirrer. Then, 2 mmol of ammonium fluoride solution (99.99%, Sigma-Aldrich, St. Louis, MO, USA) was added to the mixture and stirred to form a turbid suspension. Note: Ammonium fluoride had been weighed in advance and dissolved in 1 mL of deionized water for later use. Next, 0.4 mmol, 2 mL of rare earth nitrate solutions including Y(NO₃)₃ 6H₂O, Yb(NO₃)₃ 6H₂O, and Tm(NO₃)₃ 6H₂O (99.999%, Sigma-Aldrich, St. Louis, MO, USA), respectively, were added to the mixed solution, and the mixture was stirred for 20 min. Finally, the mixed solution was transferred into a 50 mL polytetrafluoroethylene-lined bottle, placed in a stainless steel reactor, and maintained in a constant-temperature drying oven (DHG-9076A, Shanghai Jinghong Experimental Equipment Co., Ltd, Shanghai, China) at 220 °C for 12 h. The synthesized rare earth ion-doped nanocrystals were washed three times with alcohol and cyclohexane, centrifuged (TG16-WS, Changsha Xiangrui Centrifuge Co., Ltd., Changsha, China), and dried in a constant-temperature oven at 60 °C for 6 h.

3.3. UCNPs-WSe₂ Composite Material

The prepared UCNPs were placed in a beaker, and ultra-pure water was added. The mixture was then subjected to ultrasonic treatment for 5 min to uniformly disperse the nanocrystals. Subsequently, the UCNP dispersion was spin-coated onto the 2D WSe₂ surface.

3.4. Fabrication of Transistor

The device was fabricated by applying silver paste on both sides of the material by a probe, followed by high-temperature curing to form electrodes.

3.5. Characterization

The morphology of the synthesized 2D WSe₂ was characterized using field-emission scanning electron microscopy (FE-SEM, Zeiss, Sigma500, Oberkochen, Germany), optical microscopy (OM, Olympus, BX43F, Tokyo, Japan), and atomic force microscopy (AFM, Bruker Dimension Icon, Shah Alam, Malaysia). Raman spectra were collected with a confocal Raman spectrometer (Horiba, iHR550, Loos, France) equipped with a 532 nm laser excitation source. Photoluminescence spectra were recorded using a steady-state/transient fluorescence spectrometer (FLS-1000, Edinburgh Instruments, Livingston, UK). Absorption spectra were measured by a fluorescence spectrophotometer (Agilent, Cary Eclipse, Santa Clara, CA, USA). A 975 nm laser source (LE-LS-975-5000TFCA, Shenzhen, China) was employed for optical excitation. Device performance tests were conducted on a probe station (Lake Shore TTPX, Westerville, OH, USA) integrated with a semiconductor parameter analyzer (Keithley, 4200A-SCS, Solon, OH, USA). All characterizations were performed at room temperature.

4. Conclusions

This paper achieved a NIR responsive synapse made of 2D WSe₂ and NaYF₄: 20%Yb, 0.5%Tm nanocomposites. The NIR response of the synaptic device originated from the radiation reabsorption from UCL of NaYF₄: 20%Yb, 0.5%Tm nanoparticles to 2D WSe₂, which was confirmed by absorption and luminescence spectra. The synaptic plasticity was regulated by power density, pulse width, and pulse frequency of NIR and visible bands. Handwritten digit image recognition was realized in artificial neural networks with an accuracy rate of 97.24% based on two types of LTP synapses by constructing a coordinated regulation strategy of 405 nm and 975 nm NIR. In addition, robotic visual perception in day–night and foggy environments was proposed according to the strong absorption in the visible region and the penetration ability of NIR. This work might provide an innovative technological path for the development of all-weather intelligent sensing systems with environmental adaptability.