Photoelectric synapses based on all-two-dimensional ferroelectric semiconductor heterojunction

Photoelectric synapses are attracting intensive attention due to its low power consumption and adaptive learning. However, traditional ferroelectric field effect transistors are not conducive to the integrated application in artificial intelligence systems. Here, we design the all two-dimensional photoelectric synapse device based on WSe2/MoS2/α-In2Se3 ferroelectric van der Waals heterojunction, which has high memory capacity (memory on/off = 105) and synaptic function. In addition, we simulate an artificial neural network to modify the handwritten digit recognition of the National Institute of Standards and Technology. In particular, the recognition rates are 92.4% and 93.6% for electrical synapse and photoelectric synapse, respectively. This work provides an effective strategy for achieving stable integration of neuromorphic computing.


Introduction
With the development of the information age, the next generation of electronic devices need to process the exponential growth of information [1][2][3].The traditional computing technology based on complementary metal-oxide-semiconductor (CMOS) circuits and von Neumann architectures is facing the von Neumann bottleneck, which can not meet the demand of next generation information technology [4,5].The human brain works as a control center of the nervous system and has the capable of efficient complex computation [6][7][8].Hence, Brain-like computation that mimics artificial synapses is expected to overcome the von Neumann bottleneck [9][10][11].Recently, great efforts have been made to explore and simulate the artificial synaptic devices, such as memristor [12][13][14][15], ferroelectric field-effect transistors (Fe-FETs) [16][17][18], and floating-gate transistors (FGTs) [19][20][21][22][23][24].Among them, the Fe-FETs have good non-volatility and are considered to have great potential in future electronic device applications [25][26][27].
Although remarkable progress has been shown in simulating electronically controlled artificial synapses with Fe-FETs, the operation speed of electronic synapses is still affected by resistance delay, power loss and high energy consumption.Studies show that introducing light into synaptic devices can overcome these limitations [28], the photoelectric synapses are more applicable for smart sensors for Internet of Things and biomedical electronics [29,30].The semiconductor materials with photoresponse are used as ferroelectric regulatory layers to simulate photoelectric synaptic devices.However, the material of ferroelectric layer is bulk material, which is not conducive to the integration of Fe-FETs.Therefore, two-dimensional (2D) semiconductor ferroelectric materials are considered to be an ideal platform for small size integrated photoelectric synaptic devices.Among them, 2D α-In 2 Se 3 has stable ferroelectricity at room temperature, which is suitable for the preparation of ferroelectric photoelectric synaptic devices for easy integration.In addition, MoS 2 and WSe 2 are 2D materials with excellent photoelectric properties [31,32], and combining them with α-In 2 Se 3 is expected to build a device with good performance.
In this work, we develop the all 2D photoelectric synapse device based on the WSe 2 /MoS 2 /α-In 2 Se 3 van der Waals heterojunction (vdWH), in which the α-In 2 Se 3 and WSe 2 act as ferroelectric layer and photonic gate layer.The device behaviors excellent characteristics, such as high on/off ratio (10 5 ), paired-pulse facilitation (PPF), short-term plasticity (STP), long-term potentiation (LTP) and learning-forgetting-learning behavior.Particularly, the recognition rate of an artificial neural network is higher, the recognition rate of the electrical synapse is 92.4% and the photoelectric synapse is 93.6%.This work provides an effective strategy to realize the integration of artificial synaptic neural networks.

Device fabrication
The Si/SiO 2 substrate was cleaned with alcohol and acetone for 1 h.The MoS 2 , α-In 2 Se 3 and WSe 2 flakes were obtained via mechanical exfoliation method.Under the help of a 2D material transfer platform, the three kinds of materials were superimposed.Finally, standard electron beam lithography was used to carve the electrode, and thermal evaporation method was used to prepare the 30 nm electrodes of Au.

Electric measurement
The electric characteristics and basic memory behaviors were measured by semiconductor parameter analyzer (Keithley B1500) at room temperature.The optical microscope (Olympus BX51 M) was used to gain surface topography.The thickness of the MoS 2 , WSe 2 , and α-In 2 Se 3 thin films were performed by AFM (Veeco Multimode).The quality of the MoS 2 , WSe 2 , and α-In 2 Se 3 were defined through Raman spectroscopy (Renishaw InVia, 532 nm excitation laser).

Results and discussion
In order to describe the device structure, figure 1(a) presents the structure diagram of the device, and the channel of the device is MoS 2 .In addition, the WSe 2 and α-In 2 Se 3 are located in the upper and lower parts of the channel, respectively.The substrate of the device is SiO 2 /Si substrate with 300 nm thick SiO 2 , and the gate connects to the Si layer.
The optical image of the device is shown in the figure 1(b), the green orange and violet line outlines the material WSe 2 , MoS 2 and α-In 2 Se 3 , respectively, and the interface of the device is clean.The device is built in the super clean room, so the interface defects and the trapped charge at the junction of the device are less, which can not affect the performance of the device.To confirm the presence of these components and the junction mass of the material, the Raman spectroscopic measurements were performed.The Raman spectra of WSe 2 , MoS 2 ,   [35,36], the proposed energy band diagrams of the heterojunction can be shown in figures 1(e)−(f), in which the electron affinity (χ) and workfunction (j) are also given.The electron affinity (χ) and workfunction (j) are approximately 4.15 and 1.82 eV, respectively, for the MoS 2 and 3.55 and 2.61 eV, respectively, for WSe 2 (figure 1(e)).The WSe 2 is p-type material and MoS 2 is n-type material, the energy band shows the p-n junction type (figure 1(f)).However, α-In 2 Se 3 is the ferroelectric material, which is to generate a ferroelectric field.Therefore, the schematic diagram of the section structure can more clearly show the movement routes of the electrons and holes, the working mechanism is explained by the section structure in the figure 2.
In figure 2, the memory characteristics of the device are investigated.Figure 2(a) shows the dual-sweeping transfer curves of the device under dark.The figure shows that the device has a memory window where the switching ratio reaches 10 5 , which shows that the device has the property of memory.Also, figure S2 presents the I-V curve the dual-sweeping transfer curves of the device under the dark, showing that the related device performance is slightly better under vacuum conditions.Therefore, in order to further explore the memory ratio of the device, figure 2(b) tests the relationship between the device current in the erase state and the program state with time.The current in both states can be maintained for a long time, and the memory ratio of the device is the current of erase/programming, which indicates that the device has a memory ratio of 10 5 .Here the program voltage is high due to the utilization of 300 nm SiO2/Si substrate as dielectric layer.Nevertheless, the operating voltage can be decreased by introducing high-k materials.In addition, figure 2(c) also tests the current change of the device after a single laser pulse, the current is small when the device is programmed, and the current suddenly rises after light irradiation and can maintain in a high current state for a long time.The related optical synapses can be simulated by using the optical memory properties.The working mechanism of the device is explained as follows, the upper and lower parts of the figure 2(d) show the carrier distribution of the device before and after the gate voltage.In the original state, the ferroelectric domain direction of α-In 2 Se 3 is disorganized and does not affect the number of free holes of channel MoS 2 .However, when the device is applied a negative gate voltage, the ferroelectric domain direction of α-In 2 Se 3 is downward.The side that contacts MoS 2 gathers a large number of negative polarization charges, which adsorbs the free holes in the MoS 2 .Hence, the concentration of free electrons in the channel increases, the current of the device rises and the device is in the erase state.In contrast, in the figure 2(e), the ferroelectric domain direction of α-In 2 Se 3 is upward when a positive gate voltage is applied to the device.The side of MoS 2 gathers a large amount of positively polarized charge, which adsorbs the free electrons in the MoS 2 .Therefore, the concentration of free electrons in the channel is reduced, the current of the device is reduced and the device is in a programming state.In addition, when the device is illuminated in the programming state, MoS 2 and WSe 2 generate photogenerated electron-hole pairs.In the MoS 2 , the electrons adsorbed by the upward polarized ferroelectric layer are neutralized by photogenerated holes, the photogenerated electrons remain in the MoS 2 .In the WSe 2 , the photogenerated electrons enter into the MoS 2 , and the photogenerated holes remain in the WSe 2 , which act as positive gate pressure.Therefore, remove the light, the electron concentration of MoS 2 is rises, the free current of the device becomes larger, and the device is in the optically programmed state.
In the human visual system, the light signal reflected on the object is transmitted to the retina, and the retina transmits the light signal to the brain through the neurons, then the object can be observed by humans.Thus, the behavior of visual neurons and synapses is stimulated by light signals, as shown in figure 3(a).In our device, light signals can be used as stimulus signals applied to devices to simulate the behavior and function of synapses.The most typical behavior in synaptic plasticity is Paired-pulse Facilitation (PPF), when two electrical pulses are continuously applied, the time interval between the pulses is 1.5 s, as shown in the illustration of figure 3(b).It is found that the current after the first pulse stimulation increases to 16.84 pA, and the current after the second pulse stimulation is 28.7 pA, which is greater than the effect of the first pulse stimulation.In order to study the temporal correlation between two pulses, the current amplitude of the first pulse is A 1 , and the current amplitude of the second pulse is A 2 .The PPF can be calculated using the following formula: the figure 3(b), making the amplitude and pulse width unchanged and changing the time interval of the two pulses to 1 s, 1.5 s, 3 s, 6 s and 10 s respectively.It is found that the larger the time interval, the smaller the PPF, and this reduction is non-linear, first fast and then slow.The law of attenuation can be described by a function [37], The τ 1 and τ 2 are the time constants about fast and slow decaying terms, which are calculated to be 0. In addition, the human process is simulated by applying a series of light pulse (figure 3(e)).At the first learning phase, the 20 consecutive light pulses are used to stimulate synaptic weights.After 40 s, the same sequence of light pulses is applied to the device again.When the light pulses numbers is 6, the same current level as the first learning phase can be obtained.After the same time interval of 40 s, the same current level require the minimum number of light pulses (n = 4).The past experiences influence subsequent learning, which is similar to the human brain learns and remembers.
Finally, the convolutional neural network (CNN) is built for supervised learning in order to simulate artificial neural network.As shown in the figure 4(a), the CNN structure mainly consists of three parts: convolution layer, pooling layer and fully connected layer [38,39].The supervised learning of CNN is a modified National Institute of Standards and Technology (MNIST) handwritten data set, which contains many handwritten digital images.MNIST is often used by various image recognition systems to train and evaluate machine learning performance.The MNIST contains 60000 images for training the network and 10000 images for evaluating the recognition accuracy of the trained network.The left part of figure 4(a) presents the example of an image from the MNIST dataset, which is typically normalized to a black and white image containing 28 28 ´pixels.Therefore, for simulation, a two-layer multilayer perceptron (MLP) neural network with 784 input neurons, 100 hidden neurons, and 10 output neurons is utilized.Through further experimental tests, the conductivity changes of the device can be well controlled by the number of electrical and light pulses.As shown in figure 4(b), when the V g is −5 V, the conductance (G) is continuously increased, which is long-term enhancement (LTP).In addition, when there is a positive peak (V g = 0.01 V), the G will gradually decline, which is the long-term decline (LTD).According to the optical response characteristics of the device in figure 3, the light-LTP and electricity-LTD of the device are measured in figure 4(c).The LTP and LTD are important parts of synaptic plasticity, meaning that the weight of the synapse is adjustable, the device can be used in artificial neural networks.For the ideal device, the LTP and LTD conductance curves should be linear.The nonlinearities (NL) of the LTP and LTD curves are extracted from the following equations [40]: where the G n+1 and G n represent the synaptic conductance of the device in the present and updated states, and the parameters α and β denotes the changing step sizes of the conductance and nonlinearity, respectively.G max and G min are the measured maximum and minimum values of G, respectively.To evaluate the synaptic performance of our device, we calculate the NL of photoelectric synapses and electrical synapses by equations (3) and (4), where the NL are shown in the figures 4(b) and (c).
In the simulation, the accuracy of the electrical synaptic simulation reach up to 92.4% [green circle in figure 4(d)] under the training of the device-based neural network.In addition, the highest value of photoelectric synapse accuracy is 93.6% [red circle in figure 4(d)].In figures 4(b) and (c), the nonlinearity of photoelectric synapses is lower than that of electrical synapses.Therefore, the recognition performance of photoelectric synapses is better than that of electrical synapses.This demonstrates both the high performance and versatility of our synaptic device and the high efficiency of our designed neural network.

Conclusion
In summary, we fabricated the Fe-FETs based on the all 2D WSe 2 /MoS 2 /α-In 2 Se 3 vdWH, in which the α-In 2 Se 3 is used as ferroelectric material.Based on α-In 2 Se 3 and WSe 2 provide ferroelectric field and grating action, the devices exhibit good memory performance and photoelectric synaptic plasticity, such as write/erase rates above 10 5 , STP and LTP, and conversion between the STP and LTP.These operating modes can be selected by varying the intensity, pulse width, and number of input light pulses.In addition, the device also has the ability to simulate learning-forget-relearning.At the same time, the device can simulate artificial neural network and has high image recognition accuracy.This research has the potential to solve the integration problem of FE-FETs in ferroelectric synaptic applications, which helps to achieve better performance of neural network integrated systems.

Figure 1 .
Figure 1.(a) Schematic diagram of the device structure.(b) Optical microscopy image.(c) Raman spectrum.(d) AFM topography image.(e) The band distribution before the material contact.(f) The band distribution after the material contact.

α-In 2
Se 3 and the overlap are shown in figure 1(c).Obviously, the Raman characteristic peaks of WSe 2 , MoS 2 and α-In 2 Se 3 are clearly observed in the WSe 2 /MoS 2 /α-In 2 Se 3 heterojunction device, which corresponds well with the characteristic peaks of single material, indicating that the high quality of the junctions has been formed.For the two-dimensional (2D) layered materials with dangling-bond-free surface, different two-dimensional layered materials can be combined to create the van der Waals heterostructures (vdWHs) without the constraints of conventional lattice matching and processing compatibility [33, 34].Moreover, figure S1 presents the TEM image and high-resolution cross-sectional image of α-In 2 Se 3 /MoS 2 /WSe 2 heterostructures, which exhibit the sharp interface between interface and less lattice mismatch appears.In addition, figure 1(d) shows the AFM diagram of the device, where the thickness of MoS 2 is 10 nm.Based on the reported literatures

Figure 2 .
Figure 2. (a) Dual-sweeping transfer curves of the device under the dark.(b) Retention capability of the device in erase state (V g = 80 V for 0.2 s) and program state (V g = −80 V for 0.2 s).(c) Optical response memory characteristics of a device under a single laser pulse.(d) Mechanism diagram of device memory and optical erase:(d) electrical erasing state, (e) electrical programming state, (f) lighterasing state.

Figure 3 .
Figure 3. (a) Schematic diagram of the human visual nervous system.(b) PPF index as a function of pulse interval (Δt).The curve is fitted with a double exponential decay function.(c) The current and relaxation process triggered by light pulses of different intensity (0.48 ∼ 2.59 mW).(d) The current evoked by ten light pulses with different pulse frequency (1 ∼ 20 HZ).(e) Simulation of multiple learning.

(
1 and 4.56 s by fitting experimental data in figure 3(b).This attests to our device has the short term synaptic plasticity property.As shown in figure 3(c), keeping the wavelength and the pulse width of the light pulse are 405 um and 1 s, the current of the device increases with the increase of the intensity of the pulse.And the intensities of the light pulse are to 0.48 mW, 0.812 mW, 1.198 mW, 1.886 mW and 2.59 mW.In addition, figure 3(d) shows the current changes under different frequency light pulses, a single frequency pulse sequence consists of 10 pulses.It can be found from the figure that the current becomes larger with the increase of pulse frequency.Therefore, changing the intensity and frequency of the light pulse can regulate the weight of synaptic connection.

Figure 4 .
Figure 4. (a) Schematic diagram of the convolutional neural network (CNN).(b) Long-term potentiation (LTP) and long-term depression (LTD).(c) The conductance change of the device under 15 light pulses and positive grid voltage pulses.(d) The accuracy based on CNN simulation.