2D MoTe2/MoS2−xOx Van der Waals Heterostructure for Bimodal Neuromorphic Optoelectronic Computing

The von Neumann bottleneck has long been a significant obstacle to the advancement of the era of intelligent computing. Neuromorphic devices are considered a promising solution to overcome this bottleneck. These devices draw inspiration from the information processing and computing capabilities of neurons in the human brain. Nevertheless, biomimetic synaptic devices used in neural network computing encounter significant challenges, including high nonlinearity in regulation, limited abundance of state conductance, and restrictions in unidirectional plasticity. Here, a memristor synaptic device is reported that utilizes the ion migration properties of MoTe2/MoS2−xOx heterojunction interface. This device demonstrates remarkable exceptional linearity, extensive dynamic regulation, and bidirectional independently controllable synaptic plasticity when subjected to bimodal regulation using electrical and optical signals. In addition, it shows significant paired‐pulse facilitation, empirical learning, and spike‐timing‐dependent properties. Furthermore, a deep learning framework is constructed to evaluate the reliability of devices in neuromorphic computation. The electronic synapses achieve high accuracy rates of 99.3% and 96.5% in recognizing digits and floral graphics, while photonic synapses achieve 95.3% and 91.5%. These findings emphasize the superior performance of photonic synapses in synaptic computation and provide a potential methodology for integrating multimodal neuromorphic hardware with artificial intelligence computing systems.


Introduction
With the rapid development of artificial intelligence, researchers have initiated extensive studies on brain-like computing in both the software and hardware domains.The traditional computer architecture based on storage and computation separation system is facing a severe test.The information explosion has led to a substantial increase in data, while the required computing power to process it significantly lags behind.In contrast, the neuronal system of the human brain composed of a vast network of synapses enables parallel processing of computations on valid data while filtering out unnecessary information.Each neuronal synapse serves as a storage and computational unit, contributing to a total energy consumption of only 20 W. [1,2] Motivated by this, researchers have started to conduct extensive studies on brain-like computing across both the software and hardware domains.
Nevertheless, the current research indicates that hardware development is lagging behind software updates. [3,4]Recent theoretical and experimental findings have demonstrated that memristors are highly promising devices for overcoming these challenges. [5][8] Artificial synapses can be classified into two types, electronic and photonic, depending on their triggering mechanism.Through electrical impulses, electronic artificial synapses generate synaptic plasticity events, whereas photonic synapses achieve a similar function by responding to light waves. [9,10]The continuous tunability of currents expressed by artificial synapses provides a hardware architecture for simulating operations with multi-level computing, through electrical impulses, while relying on weight coefficients of connected neurons in Neural Network for training intelligent models.13][14] Currently, the photoelectric transport mechanism involving 2D transition metal sulfide is commonly employed in numerous studies for the preparation of memory resistors and the simulation of synaptic functions.This mechanism encompasses the tunneling effect, phase transition effect, charge trapping, release, among others.[17][18][19][20] Consequently, the development of optoelectronic devices that provide high linear modulation and a multi-rich state response can effectively advance neural control technology and meet the requirements of high-performance neural network computing.
In this study, we developed a 2D MoTe 2 /MoS 2−x O x heterostructure that utilizes the excellent electrical transport properties and narrow absorption band gap of MoTe 2 , as well as the domain-limiting effect of the oxygen ion defect state in MoS 2−x O x on carrier transport.Compared to monolithic materials, the heterostructure exhibits stronger electrical signal modulation, a broader range of dynamic conductance, increased optical responsiveness, and a wider spectral range of inputs.Experimental results demonstrate that the MoTe 2 /MoS 2−x O x heterostructure device can simulate continuously tunable biological synapses in both electrical pulse and visible-near-infrared (NIR) laser modes.This includes excitatory/inhibitory synaptic currents (E/IPSC), long-term inhibition (LTD), long-term potentiation (LTP), and spike timing-dependent plasticity (STDP).Unlike previous studies that reported artificial synapses with a single input function, [16,[21][22][23][24] our device can independently achieve bidirectional continuous conductance tunability of synapses under both optical and electrical stimulation.The device's bimodal operation was verified through simulated image identification, yielding excellent accuracy results in neuromorphic computing.These findings provide promising research strategies for implementing multimodal control of neuronal computing hardware.

Results and Discussion
Figure 1a shows a MoTe 2 / MoS 2−x O x heterostructure memristor with dual-mode signal operation, the right panel corresponds to a local enlargement of the molecular structure in the vertical stacking region.The Raman spectra (laser excitation wavelength 532 nm) of MoTe 2 and MoS 2 and their stacked heterostructures are shown in Figure 1b.The two peaks of MoS 2 (dark green) occur at 405.39 cm −1 (interlayer A 1g vibration mode) and 380.83 cm −1 (in-plane E 1 2g vibration mode), in MoTe 2 (olive green), two distinct peaks located at 232.96 and 173.93 cm −1 can be assigned to two vibrational modes, intraplane E 1 2g and interlayer A 1g Raman activity of the vibrational group located at 290.47 cm −1 is weak and therefore an insignificant weak peak appears.All peaks are represented in the MoTe 2 /MoS 2−x O x (red line) heterostructure region, demonstrating the formation of a good interface of van der Waals heterostructures, and significant Raman redshifts were observed at 409.15 and 383.08 cm −1 , which indicated that the embedding of oxygen ions had some disruption of the lattice symmetry of MoS 2 .
To further determine the chemical state of MoTe 2 /MoS 2−x O x heterostructures, we conducted XPS spectroscopy.Figure 1c shows the O 1s elemental spectrum in MoS 2−x O x , where the four peaks that split into 531.1 eV (C═O), 532.0 eV (V o ), 532.4 eV (Mo─O), and 533.1 eV (─OH).Among the Mo 3d peaks in MoS 2−x O x , 229.5 eV (Mo─S) and 232.7 eV (Mo─S) can be due to Mo 3d 3/2 and Mo 3d 5/2 , while Mo─O bonds produced by UVozone preparation appear near 235.9 eV (Figure 1d).Within Figure 1e, the two peaks at 162.3 and 163.5 eV denote S 2P 3/2 and S 2P 1/2 , respectively, at 165.7 and 168.3 eV can be classified as S─O bonds, where the sulfur element is in the intermediate valence level.More oxygen-binding peaks appear in these UV-ozone treated MoS 2 samples and a full spectrum comparison shows an increase in the XPS peak states of the associated elements, in addition, we also include an optical microscope top view of the device as well as the thickness for testing tests (Figure S1, Supporting Information).An analysis of the elements of MoS 2 is given in Figure S2 (Supporting Information), which can be interpreted as a structural evolution toward MoS 2−x O x with lattice defects arising in MoS 2 .Figure 1f,g represents the Mo 3d and Te 3d elemental spectra in MoTe 2 , respectively, and the two peaks in Mo 3d can be attributed to Mo 3d 5/2 near 228.3 eV and Mo 3d 3/2 near 231.5, 572.4,and 576.7 eV in Te 3d, which indicate Te 3d 5/3 , 583.0 and 587.1 eV for Te 3d 3/2 . [25,26]][29][30] Carriers within the material undergo three sequential steps of transportation.First, in the initial state (as shown in Figure 2ai), where MoS 2−x O x easily participates in redox reactions due to some S elements in the intermediate valence state and the influx of oxygen ions from ozone disposal (see XPS spectroscopy for details).Applying a positive external bias voltage cause free sulfur and oxygen ions to move positively toward the electrode, at which time the device is in the initial make relatively high resistance (HRS) state.Second, Figure 2aii indicates that the sulfur vacancies and oxygen vacancies inside the MoS 2−x O x structure accumulate close to the MoTe 2 heterostructure interface after the forward bias voltage scan, forming a conical CF with abundant percolation channels, the device has switched from the initial HRS to the LRS.As shown in Figure 2aiii, applying a negative bias at both ends of the device fills the vacancy defects within the material with surrounding oxygen and sulfur ions.This closure of the conducting path between the positive and negative electrodes transforms the device from a low resistance state (LRS) to a high resistance state (HRS), creating a reset state.It is worth noting that the p-type semiconductor nature of 2H-MoTe 2 exerts a certain binding effect on the migrating oxygen and sulfur ions at the heterostructure boundary, thereby causing a delayed relaxation time of the free electrons at the heterostructure interface.Additionally, ozonation of the surface and interior of MoS 2−x O x is ozonated to produce defects, trapping electrons and holes on the defect energy level for recombination.[33][34] Figure 2b illustrates the MoTe 2 /MoS 2−x O x memristor synaptic device structure.To achieve greater light absorption, we utilized MoTe 2 with a smaller band gap and superior electrical transport properties.This was achieved by placing it on top of MoS 2 through mechanical exfoliation and target transfer (see Methods and Materials section for details).The MoTe 2 /MoS 2−x O x heterostructured device without the UV-ozone process shows rectification curve characteristics, which is a benefit caused by the dominance of the p-n junction at the heterostructure stack interface.The inset result shown on a logarithmic scale yields a rectification ratio close to one order of magnitude, consistent with previous reports in the literature (see Figure 2c).Figure 2d shows the memristor I-V characteristics of the device after the UV-ozone process.By applying a cyclic direct current sweep voltage (starting at 0 V with 0.1 V steps and 25 ms intervals), a classical analogue memristor characteristic diagram can be obtained.Stages 1-2 represent the transition from HRS to LRS, known as set, while stages 3-4 demonstrate the reset process from LRS to HRS.We also performed a test on the device's switching characteristics for 2000 cycles.The test yielded a resistive storage characteristic with a switching ratio of approximately one order of mag-nitude and a relatively stable resistance distribution (see Figure S3, Supporting Information).
Positive feedback produces a current benefit when the interval between successive applications of positive voltage is shorter than the relaxation time of the carriers, resulting in excitatory postsynaptic current (EPSC).Conversely, applying a continuous negative voltage leads to the generation of negative feedback inhibitory postsynaptic current (IPSC).37][38] When applied, paired pulsed voltage stimuli showed that the later pulse evoked a significantly higher current response compared to the previous pulse.This disparity decreased as the time interval increased, replicating the characteristics of paired-pulse facilitation (PPF) observed in biological synaptic weights.shows the application of five consecutive pairs of pulses, with pulse intervals ranging from 20 to 100 ms.The first and second response currents generated by each pair of pulses are denoted as A x and A y , respectively.The data results indicate that all pairedpulse stimuli resulted in A y > A x , indicating an enhancement of the weighted association between synapses.The increase in synaptic weight is dependent on the time interval between paired pulses.Longer intervals between paired pulses result in smaller increases in synaptic weight.This can be attributed to the diminishing learning and memory properties of the organism as the time interval increases.The PPF index can be defined by the following formula: Application of the pairwise pulse interval increased from 20 to 100 ms and the PPF index decayed from 47.59% to 9.04% (Figure 3c), a correspondence of this phenomenon can be made with the information transfer decay relationship of the organism, which can be visually reflected by fitting the following double exponential weight decay formula to the PPF growth index change: The PPF behavior of memristor synaptic devices can be split into two phases: fast and slow.The characteristic relaxation times for fast decay and slow decay are represented respectively by  1 and  2 , and the equation yields a  1 / 2 value of 27.34/33.42ms.This is combined with the relaxation scales observed in biological synapses.[41][42] Short-range synaptic plasticity is subdivided into short-range enhancement and depression, which can be modeled by applying a delta wave voltage to the device with a pulse width of 0.5 s and amplitudes of 5 and −5 V for this characteristic.Figure 3d shows that the current response of the heterostructured device increases with successive stimulation of the forward triangle wave, while successive stimulation of the reverse triangle wave leads to a continuous decay in device current (Figure 3e).Additionally, we conducted successive unidirectional scans at 5 and −5 V, respectively, and observed a continuous increase (decrease) in positive (negative) conductance (see Figure S4, Supporting Information).Based on the experimental results described above, it has been confirmed that MoTe 2 /MoS 2−x O x structure exhibits short-term synaptic plasticity in both pulsed and DC scanning voltage dual modes.
LTP and LTD are extensively researched forms of activitydependent persistent synaptic plasticity.These forms of plasticity are induced by different patterns of synaptic activity, primarily through the conversion of short-term plasticity into long-term plasticity by repeated high-frequency stimulation.Figure 4a,b demonstrates the devices' outstanding conductance (current) dynamic gain (G max /G min ≈ 55) and the presence of ≈150 effective synaptic plastic states when continuously applying pulses with amplitudes of 5 and −5 V and widths of 50 ms are applied continuously.These performance indicators surpass those of previous studies in this field.Importantly, the heterostructure devices attain stable and highly linear LTP/LTD synaptic properties without requiring the design and optimization of intricate programming voltages, which are influenced by the time interval between test pulses and the carriers' relaxation time.The high linearity is advantageous for minimizing the training loss of the matrix model, thereby enhancing accuracy, a critical aspect in neuromorphic computations.Experiential learning is a significant neural activity in organisms, characterized by the continuous enhancement of neuronal connections through learning and relearning.In this study, we simulate this behavior by externally applied pulses, with pulse intervals of 0.125 s and durations gradually increasing from 0.125 to 0.750 s with an amplitude of 5 V.The current of the device gradually increases with the duration of the stimulus, as shown in Figure 4c.Additionally, the next pulse maintains the memory current state of the previous pulse, influenced by the pulse interval time.This indicates the strong learning and memory capability of the device.STDP is widely recognized as fundamental to biological neuronal information processing, data storage, and is considered one of the manifestations of LTP/D, which is caused by tight temporal correlations between pre-synaptic and post-synaptic neuronal spikes. [43,44]Figure 4d presents the results of the test involving programming voltage and feedback.It can be observed that when the presynaptic membrane action potential responds to precedes the postsynaptic membrane action potential, repeated stimulation can induce LTP.Conversely, LTD is induced when the postsynaptic membrane action potential precedes the presynaptic membrane action potential.This property depends on the time sequence and interval of the anterior and posterior membranes.Experimental data demonstrate that the device fits well with a temporally asymmetric Hebbian learning form of the STDP phenomenon, which is fitted with the following equation: In MoTe 2 /MoS 2−x O x heterostructures, the photo-generated electrons are mainly localized in the MoTe 2 layer when the 1064 nm laser light is shone in the region near the positive electrode.This results in these electrons not easily recombining due to the built-in potential difference in the heterostructure region, which continuously separates the holes.[47][48][49] Conversely, when the laser stimulus is irradiated in the negative electrode region, the photogenerated electrons from the MoS 2−x O x interface layer become trapped in the MoTe 2 and MoS 2−x O x stacking region.This causes a decrease in the number of transferred free electrons and a continuous decrease in the negative photocurrent.To test the short-term synaptic plasticity of the devices, we performed experiments using 532 and 1064 nm NIR laser sources without an applied bias electric field.By varying the position of the spot irradiation on the positive/negative electrodes, it was found that successive current signals collected at low optical frequencies can produce a short-term enhancement/inhibition effect (Figure 5a,b).This addresses the limitation of most photosynaptic devices that can only be modulated in one direction.Unlike electronic synapses, photosynapses have the ability to undergo various cross-modal modulations such as wavelength, polarization, and spatial modes In fact, we have also tested the optical responsiveness of the devices (Figure S5, Supporting Information).In contrast to the optical response of MoS 2 semiconductor material, MoTe 2 /MoS 2−x O x heterostructure devices not only have improved optical responsiveness (G MoTe2/MoS2−xOx /G MoS2 ≈ 150) but also demonstrate a broad-spectrum response in the visible and near infrared range.Obviously, this is related to the smaller absorption bandgap of the MoTe 2 material.By simultaneously adjusting the position of the laser light source and increasing the optical frequency, we were able to obtain current signals that corresponded to the continuously adjustable LTP/D characteristics of the photonic synapses (Figure 5c,d), which are related to the separation of photogenerated electron-hole pairs and vacancy defects within the material.Optical modes exhibit a 14-fold maximum enhancement and 2.73-fold maximum attenuation over a tunable range of dynamic conductance (current), and the photovoltaic effect produced by the device is also observed in tests of I-V curves for 405, 532, 808, and 1064 nm lasers (see Figure S6, Supporting Information, for details).All these results demonstrate the excellent optical performance of MoTe 2 /MoS 2−x O x devices.
To verify the reliability of our memristor synaptic devices in neuromorphic computing, we built a Convolutional Neural Network (CNNs) learning framework for image recognition using the Python language package and the PyTorch tool.Front and back input signals in the handwritten character recognition project shown in Figure 6a corresponds to the connections of front and back neurons, where the hidden neurons represent the convolution and pooling operations processing of the signals, which will be done vicariously by the studied synaptic plasticity data.As shown in Figure 6b, a deeper neural network is also established to adjust the appropriate convolution kernel size, channel number, and full-connection parameters for feature extraction of image information to test its usability in complex image classification, which puts higher requirements on malleable parameters.Handwritten digit training sets from the Modified National Institute of Standards and Technology (MNIST) and floral recognition training sets from Data Flicr, Google images, and Yandex images were used here.Detailed logic diagrams of the synaptic dot product arrays and simulated hardware architectures are depicted in Figure S7 (Supporting Information).
In the handwritten digit recognition project, accuracy values of 99.3% and 96.5% were achieved by electronic and photonic synapses, respectively (Figure 6c).The electronic synapse showed a faster learning rate than the photonic synapse due to its near ideal linearity of synaptic tunability under electrical pulse conditioning.To train the model, the measured synaptic plasticity data was substituted into the algorithm weight layer.The trained model was then used on the test set to produce accurate results.In the floral recognition project, both synapses achieved accuracy values of 95.3% and 91.5%, respectively (Figure 6d), with a difference of only 3.8% in accuracy between the two.These impressive results were made possible by the rich conductance state and dynamic modulation range of the photoelectric synapses.[56][57] Several key properties of this work compared to similar neuromorphic work are shown in Table 1, including the triggering mechanism, operating wavelength, synaptic function, and recognition accuracy in the neural network.

Conclusion
In summary, the MoTe 2 /MoS 2−x O x heterostructured memristor synaptic devices demonstrate bimodal modulation of synaptic plasticity through both light and electricity.Under electrical signal stimulation, the device exhibits a remarkable 55fold dynamic gain in conductance and a high linearity of bidirectional tunability, which is equally excellent in the optical domain.Complete long/short term synaptic plasticity can be achieved using dual visible-NIR laser modulation, thanks to the device's broad spectrum and high responsiveness.Compared to monolithic devices, these multimodal tunable synaptic devices enhance the potential application scenarios for future neuromorphic hardware, increasing their richness.Moreover, pattern recognition simulations showcase the exceptional performance of heterostructure devices in vector matrix operations.These findings open up opportunities for cross-disciplinary development between the bio-numerical fields and intelligent computing.

Experimental Section
The Si substrates containing SiO 2 layer were put into acetone, anhydrous ethanol, and deionized water in turn and sonicated for 20 min and then dried with flowing nitrogen gas.First, Strip MoS 2 from the bulk crystal using blue tape and transfer it to the substrate using a 3D micromechanical platform.Second, the stacked devices were treated with UV-ozone for 30 min.Finally, Stack MoTe 2 on top of MoS 2 in the same way.In which, the electrode patterning was processed by UV lithography.Au/Ti (50/10 nm) electrodes were deposited by a thermal evaporation process and the residual adhesive was stripped by N-methyl-pyrrolidone (NMP).The interlayer and in-plane vibration modes of chemical bonds within heterostructures were tested using microconfocal Raman spectroscopy (LabRAM HR Evolution France).The chemical states of the devices were analyzed using X-ray photoelectron spectroscopy (Escalab 250Xi UK) using a monochromatic Al K light source, the percentage of atomic content, and the C1s peak at 284.8 eV was used for charge calibration of all binding energies.Optoelectronic properties of MoTe 2 /MoS 2−x O x heterostructure devices, including DC I-V curves, pulse measurements, and optical response characteristics, were experimentally determined using a Keithley 2612B digital source meter at ambient temperature and atmospheric pressure conditions.

Figure 1 .
Figure 1.a) Illustration of the structure of a MoTe 2 /MoS 2−x O x memristor operating in dual-mode and the molecular structure of a vertical heterostructure stacked region amplification.b) Raman spectra of MoTe 2 , MoS 2 and MoTe 2 /MoS 2−x O x heterostructures.c) XPS spectra of O 1s in MoS 2−x O x .d) XPS spectrum of Mo 3d in MoS 2−x O x .e) XPS spectra of S 2p in MoS 2−x O x .f,g) XPS spectra of Mo 3d and Te 3d in MoTe 2 .

Figure 2 .
Figure 2. a) RS mechanism of MoTe 2 /MoS 2−x O x memristors.i) Distribution of the atoms of the elements within the heterostructured device in the initial state and the migration trends of oxygen and sulfur ions under positive bias.ii) Ion aggregation in the heterojunction stacking region under forward bias and the appearance of oxygen vacancies and sulfur vacancies in the MoS 2−x O x region.iii) Ion migration direction and vacancy filling under negative bias.b) Schematic diagram of MoTe 2 /MoS 2−x O x heterojunction device structure.c,d) I-V rectification and memory resistance characteristics of the device before and after ozone treatment.

Figure 3 .
Figure 3. a) Schematic diagram of biological synaptic information transfer.b) Representation of the continuous PPF properties exhibited by the device under successive application of five pairs of pulses of different time intervals.c) Curvilinear relationship of PPF effects dependent on time.d,e) Representation of the short-term synaptic enhancement/inhibition properties of electronic synapses excited by the application of 5 and −5 V delta waves.

Figure 4 .
Figure 4. a) Positive artificial synaptic long-time enhancement.b) Negative artificial synaptic long-time inhibition.c) Forgetting and learning tests at successive time intervals, simulating empirical learning.d) STDP rule for amnestic blocking synaptic devices.

Figure 5 .
Figure 5. Photonic synaptic plasticity is excited by 1064 and 532 nm laser sources.(a,b) indicates the short-term plasticity properties of photons collected at relatively low frequencies.(c,d) shows the long-term plasticity properties of photons observed at successively relatively high frequencies.

Figure 6 .
Figure 6.a) Diagram of the neural network architecture of the digital recognition project.b) Diagram of the neural network architecture of the flower classification.c) Training results of electrical and optical synapses on handwritten digit recognition items.d) Training results of electrical and optical synapses on flower recognition items.

Table 1 .
Comparison of similar neuromorphic device research.