Introduction

To overcome the inherent power wall of von Neumann architectures, there is an increasing emphasis on neuromorphic computing techniques that emulate the energy-efficient information processing observed in biological neural networks1,2,3,4,5,6. Memristive switching devices, considered for the next generation of digital memory, offer unique properties such as high switching speeds, high scalability, and high integration density7,8. Notably, the conductance of memristive devices can be gradually modulated by external electrical pulses, akin to the updating of synaptic weights in biological synapses9,10,11,12,13. In contrast to conventional metal oxides used for memristive devices, atomically thin two-dimensional (2D) materials offer advantages such as power efficiency, reliability, tunability, heterointegration, and multifunctionality, which are essential for nonvolatile memory (NVM) devices14,15.

In the current research, 2D materials employ mechanisms such as grain boundary-induced vacancies16,17,18, phase transitions19,20,21, and Schottky barriers22,23,24 to achieve memory functions in three-terminal NVM devices. However, the characteristics of memristors based on these mechanisms primarily depend on the intrinsic quality and manufacturing processes of the active material. Two-dimensional materials that have been previously reported often exhibit uncontrollable material properties, including grain boundary density and grain size25,26. In addition to the abovementioned methods, substitutional doping is another controllable and efficient approach for modifying materials to achieve resistance-switching devices. This process enables the precise tuning of electronic properties, imparting rich physical functionality27,28. In memristive devices employing doped 2D materials, the concentration of electrons in the material is modulated through significant voltage tunability, influencing conductance via the capture and release of electrons at defect centers to achieve high performance. For example, Appenzeller et al. reported an electric-field-induced structural transition in 2H-MoTe2- and Mo1−xWxTe2-based resistive random-access memory (RRAM) devices that achieved on/off current ratios of 106 and low programming currents in a selectorless RRAM architecture via an Al2O3/MoTe2 stack29. Wang et al. created MoS2–xOx via O2 dry etching in a plasma system and designed a graphene/MoS2–xOx/graphene heterostructure that exhibited excellent switching performance with an endurance of up to 107 30. Additionally, Cheng et al. fabricated artificial synaptic transistors using heavily V-doped MoS2 as the channel material, demonstrating synaptic potentiation, depression, and repetitive learning processes through gate-tunable changes in channel conductance with abundant vanadium atoms trapping/detrapping electrons31.

The role of atomic doping in Transition Metal Dichalcogenide (TMDC) materials in the memristor mechanism of devices is unclear, with the ongoing debate being whether the trapping/detrapping mechanism of doped atoms or the impact of substrate oxide defects plays a more significant role in synaptic transistors. Clarifying the effect of dopant atoms in alternative doping methods is essential for optimizing neuromorphic devices based on TMDCs. In this work, we propose a p-type doping strategy in WS2 to increase the conductance ratio and improve the linearity of weight updates in synaptic transistors. Using liquid-phase precursor chemical vapor deposition (CVD), we achieved uniform incorporation of Nb atoms into WS2 crystals. Characterization techniques, including scanning transmission electron microscopy (STEM), Raman spectroscopy, photoluminescence (PL), and X-ray photoelectron spectroscopy (XPS), confirmed doping-induced alterations in the atomic structure. The synaptic transistor, which was fabricated with a monolayer of Nb-doped WS2 as the channel material, demonstrated an enhanced switching ratio, surpassing that of undoped WS2 by a factor of 100. When applied to the recognition of the Modified National Institute of Standards and Technology (MNIST) handwritten digit sets, the synaptic device achieved an accuracy of 92.30%.

Results and Discussion

WS2 monolayer flake growth and Nb atom substitutional doping

The growth of Nb-doped WS2 layers was conducted through atmospheric pressure CVD, as outlined in Fig. 1a. A precursor solution with a concentration of 1.8 mg/mL was created by combining niobium oxalate (Nb(HC2O4)5·xH2O) with sodium tungstate (Na2WO4·2H2O). To prepare Nb-WS2 with nominal doping concentrations of 0% (undoped), 2%, and 5%, we precisely controlled the ratio of W and Nb atoms in the precursor reagent. The prepared materials and devices were named based on the precursor concentrations, 2% Nb-WS2 and 5% Nb-WS2, but these designations do not represent their actual doping concentrations. After the SiO2/Si substrate was spin-coated with the precursor solution, the substrate was placed in a one-inch quartz tube and heated to 875 °C. Monolayer WS2 crystallites with triangular morphologies were obtained by adjusting the hold duration and Ar gas flow rate. After a 15-minute growth period and natural cooling, WS2 monolayer flakes were uniformly distributed across the entire silicon wafer, as shown in Fig. 1b. Pristine WS2 yielded equilateral triangles with side lengths exceeding 200 μm (Fig. 1c). In comparison, the maximum lateral dimensions of the thin films with 2% Nb dopant and 5% Nb dopant added reached 117 μm and 75 μm, respectively (Fig. 1d, e). We systematically examined 100 triangular flakes across three different doping concentrations and recorded their dimensions. The findings revealed that the average sizes of WS2, 2% Nb-WS2, and 5% Nb-WS2 were 125.3 μm, 62.5 μm, and 31.6 μm, respectively (Fig. S1). Interestingly, as the doping concentration increased, both the average and maximum edge lengths noticeably decreased. This trend suggests a correlation between the doping level and the geometric properties of the Nb-WS2 flakes, highlighting the material’s sensitivity to dopant-induced structural changes. The difference in lateral dimensions may be attributed to the absence of impurity atoms, resulting in minimal atomic distortion and facilitating the formation of larger flakes. The thickness of the CVD-grown WS2 was assessed via atomic force microscopy (AFM), revealing a typical monolayer flake with a height of 0.86 nm (Fig. S2).

Fig. 1: Material growth and sample photomicrograph.
figure 1

a Illustration of the CVD growth process for both doped and undoped WS2. b Low-magnification optical image revealing undoped WS2 with uniformly distributed flakes on the growth substrate. c-e Optical images of undoped WS2, 2% Nb-doped WS2, and 5% Nb-doped WS2, respectively

Full size image

Characterizing the crystal structure changes induced by doping

To evaluate the quality of the grown WS2 flakes, we employed a comprehensive characterization approach involving Raman, photoluminescence (PL), and XPS measurements. In the Raman spectra, characteristic vibration peaks of E’ and A1’ are observed at 349 cm−1 and 416 cm−1, aligning with the distinctive monolayer WS2 peaks (Fig. 2a)32,33. The normalized PL spectra of WS2 with various doping concentrations are depicted in Fig. 2b. The emission position redshifted from 626.9 nm to 637.5 nm with increasing Nb ratio up to 5%, which is indicative of p-type doping induced by Nb. The reflection contrast spectrum (Rsample-Rsub)/Rsub was used to characterize the synthesized Nb-WS2. Here, Rsample represents the reflectivity of monolayer WS2 on a SiO2/Si substrate with a thickness of 285 nm, whereas Rsub corresponds to the reflectivity of the substrate. The spectrum indicates easy excitation of photons within the material at wavelengths between 630 and 640 nm (Fig. 2c), which is in good agreement with the PL peaks. The broad peak at approximately 530 nm is attributed to the gradually changing background, which is primarily induced by the 285 nm-thick SiO2/Si substrate34. These results confirm the successful growth of monolayer Nb-doped WS2. Figure 2d–f displays Raman mapping images of WS2 with different doping concentrations. The well-defined and homogeneous emission patterns across all the pristine and doped WS2 samples are illustrated in the Raman mapping results. These graphs provide evidence of the excellent Raman properties of Nb-WS2, emphasizing its uniform spatial distribution.

Fig. 2: Characterization of p-type doping and homogeneity of CVD-grown Nb-WS2.
figure 2

a Raman spectra of Nb-WS2, exhibiting peaks at 349 cm−1 and 416 cm−1. b PL spectra of Nb-WS2, revealing emission characteristics at approximately 630 nm. The emission wavelength of undoped WS2 is 626.9 nm, whereas the emission wavelengths of 2% and 5% Nb-doped WS2 redshifted by 3.7 nm and 10.6 nm, respectively. c Differential reflectivity of CVD Nb-WS2 flakes. The neutral exciton peak was observed at 630 nm. d-f Raman mapping graphs illustrating the high homogeneity of undoped WS2, 2% Nb-WS2, and 5% Nb-WS2. XPS spectra and curve fitting of g Nb, h W, and i S

Full size image

Figure 2g–i present high-resolution XPS spectra of Nb, W, and S in the as-grown flakes. As shown in Fig. 2g, no Nb peak was discernible in the XPS pattern of undoped WS2. In contrast, both the Nb 3d3/2 and Nb 3d5/2 peaks were observed for the 2% and 5% doped WS2 samples. Notably, the peaks in the plots redshifted for both W and S due to atom doping. Compared with those of undoped WS2, the W 4f5/2 and W 4f7/2 orbitals of 2% Nb-doped WS2 redshifted by 0.2 eV (Table S1). Moreover, the S 2p1/2 and S 2p3/2 orbitals exhibited a redshift of 0.25 eV. These shifts suggested that the bonding states of some W and S atoms were changed by the nearby substituted Nb atoms. Compared with those of 2% Nb-WS2, the binding energies of the Nb 3d3/2 peak and Nb 3d5/2 peak observed for 5% Nb-WS2 redshifted by 0.4 eV and 0.2 eV, respectively. Similarly, the binding energies of W and S in the 5% Nb-doped WS2 sample further redshifted from those of the 2% Nb-WS2 sample. These shifts in the lower binding energy direction indicated a shift in the Fermi level (EF) of WS2 toward the valence band upon increasing the doping concentration, thereby confirming the successful attainment of p-type doping through the introduction of Nb35. On the basis of the results of the PL spectra and XPS measurements, the impact of atom doping on the electronic structure of WS2 is illustrated in Fig. S3a. Upon p-doping, the valence band maximum shifts closer to EF36. The energy difference between the valence band maximum and the EF decreases with increasing doping concentration. This upward shift in the valence and conduction bands is further supported by density functional theory (DFT) calculations (Fig. S3b, c). In addition, the presence of defects induced by Nb doping in the material was confirmed through in situ low-temperature PL analysis. Fig. S4 presents a typical curve fit of the 5% Nb-WS2 PL spectrum at 100 K. The introduction of Nb atoms increases the number of defects in the material, leading to the detection of the characteristic XB peak of defects, located at ~1.88 eV37. The results of the above assessments confirmed the successful incorporation of Nb into the WS2 film via the addition of niobium oxalate to the precursor, effectively achieving p-type doping. As the doping concentration increased, the peak positions for Nb, W, and S all redshifted, indicating the occurrence of p-type doping subsequent to the introduction of Nb. Additionally, an increased doping concentration correlates with an increase in hole concentration.

The lattice structure of Nb-WS2 was systematically characterized via spherical aberration-corrected transmission electron microscopy (TEM). The transferred flakes exhibit excellent thickness uniformity, which is consistent with the monolayer observed in the AFM results (Fig. S5a). High-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) images showed distinct brightness contrast for W, S, and Nb atoms, with the contrast directly correlating to the atomic number of the elements. In undoped WS2, a flawless six-membered ring lattice was evident, as illustrated in the pseudocolor image (Fig. 3a). The intensity in the annular dark-field (ADF) image directly corresponded to the atomic number, enabling spatially resolved chemical identification through quantitative analysis of image intensity. Brighter atoms (W) exhibited a contrasting hue, whereas darker atoms (S) manifested as different shades in the pseudocolor representation, displayed as yellow and blue, respectively. The selected area electron diffraction (SAED) pattern exhibited a hexagonal structure with threefold symmetry, featuring two sets of spots, Ka (indicated by blue arrows) and Kb (indicated by red arrows) (Fig. S5b)38. Upon the incorporation of niobium oxalate into the precursor, the presence of darker atoms indicative of Nb was revealed, as shown in Fig. 3b, c and validated by the intensity profile analysis in Fig. 3d–f. On the basis of atomic counts, films synthesized from precursors with a nominal Nb content of 2% presented a measured doping concentration of approximately 0.2%. Conversely, films synthesized from 5% Nb precursors presented a relatively high doping concentration of approximately 0.3%. The atomic pseudocolor plots in Fig. 3g–i feature labeled crystal plane spacings of Nb-WS2. Notably, the crystal plane spacing of the (10-10) crystal plane is approximately 0.27 nm, whereas the range of crystal plane spacings for the (11–20) crystal plane is 0.16–0.18 nm, which is consistent with previously reported data39,40. The lattice parameters suggest that the introduction of Nb atoms as dopants did not induce alterations in the lattice structure of the WS2 flakes.

Fig. 3: TEM image of a monolayer WS2 lattice substitutionally doped with Nb.
figure 3

a Aberration-corrected Z-contrast STEM image of monolayer undoped WS2. b-c STEM images of monolayer 2% and 5% Nb-WS2, respectively. The intensity profile of the brightness of atoms, marked by yellow rectangular regions in Panels a, b, and c, is displayed in Panels d, e, and f, respectively. g–i Magnified STEM images of typical regions. The yellow sites represent tungsten atoms, whereas the green sites represent sulfur atoms

Full size image

Performance of Nb-WS2 synaptic transistors and their suitability for neuromorphic computing

We subsequently employed p-type doped materials in the fabrication of artificial synaptic transistors. As depicted in Fig. 4a, monolayer Nb-WS2 served as a channel material and was transferred onto predeposited metal electrodes through a wet etch transfer approach (refer to the Experimental section and Fig. S6 for details and Fig. S7 for typical optical images of the devices). The devices were constructed on a 285 nm SiO2/Si substrate, which functions as the back gate. In Fig. 4b, the transfer curves (Ids-Vgs) of devices based on undoped WS2, 2% Nb-WS2 and 5% Nb-WS2 exhibit a continuous shift in the threshold voltage toward a positive voltage with increasing doping concentration. This observation reaffirms the enhancement of p-type transport behavior, validating the efficacy of the doping process. An on/off current ratio of 105 is achieved for the WS2 field effect transistor (FET) within a sweeping range of 60 V. The switching ratios of the 2% Nb-WS2 and 5% Nb-WS2 FETs are only 104 and 103, respectively. Notably, varying the channel length due to the triangular shape of the Nb-WS2 flakes may influence the on-state and off-state currents (Fig. S8). However, the geometric parameters have little effect on the on/off ratio. We also investigated the electrical properties of Nb-WS2 films with a precursor Nb doping concentration of 7%. Our findings revealed a rapid decline in electrical and synaptic performance as the doping concentration increased (Fig. S9). Figure 4c illustrates the potential reason for the decreased on-current caused by defects attributed to Nb atom doping41,42,43. Hot electrons from the metal electrode surmount the effective Schottky barrier in the WS2 FET and transfer to the mobile state of the conduction band in WS2. Compared with that in the WS2 FET, the downward shift of the channel layer’s EF in the Nb-WS2 FET leads to a flatter band at the Schottky barrier, increasing the effective width of the Schottky barrier and reducing the injection of hot electrons from the electrode. When Vgs > 0 V, the introduction of defects caused by doping increases carrier scattering, thereby reducing electron mobility. Importantly, all the transfer characteristic curves demonstrate substantial hysteresis windows. While pristine WS2 has a hysteresis window of only 6.59 V, 2% and 5% Nb-WS2 have hysteresis windows of 20.16 V and 27.63 V, respectively. Considering the uniformity of the substrates, transfer processes, and contact metals across all the devices, the increased hysteresis window in the Nb-doped devices can be attributed to the trapping and detrapping of electrons at the doping site. At a negative gate voltage (Vgs = −60 V), carriers in the channel are entirely depleted, and trapped electrons lack the necessary thermal energy for capture. Conversely, with a positive gate voltage (Vgs= 60 V), an abundance of free carriers quickly fills traps, resulting in a rightward shift in the threshold voltage and the establishment of clockwise hysteresis. Owing to the sensitivity of trap states to surface adsorbates in an air environment, we investigated the operational conditions of the transistors and the influence of environmental factors (Fig. S10). In Fig. 4d, pulse voltages of −30 V were applied to the gate to augment the current, whereas +30 V pulse voltages were employed to suppress the current. The current values were read at Vgs = 0 V after the pulses. For undoped WS2, the synaptic transistor switching ratio is merely 10 on the basis of the starting current. Remarkably, with minor Nb doping, the switching ratio increases by a factor of 100. P-type doping modulation enhances the switching ratio of the synaptic transistor via two main factors. First, doping introduces defects into the material, inducing a hysteresis curve in the FET by trapping and detrapping electrons. Second, p-type doping positively shifts the threshold voltage of the transistor, significantly enhancing the current ratio between high-resistance and low-resistance states at Vgs = 0 V. In summary, engineering hysteresis is achievable by increasing the Nb doping concentration, with 5% Nb-WS2 exhibiting the largest hysteresis loop, making it a suitable candidate for synaptic transistor construction in the context of TMDC materials.

Fig. 4: FET performance when undoped WS2 and Nb-WS2 are used as channel materials.
figure 4

a Schematic of the Nb-WS2-based three-terminal device built on a 285 nm SiO2/Si substrate: Nb-WS2 as a semiconducting channel, a 285 nm SiO2 layer as a dielectric layer and p++ Si as a back gate. Bias voltages were applied on the drain electrode and back gate, with the source electrode being grounded. b Transfer characteristics of the Nb-WS2-based device at Vds = 0.5 V, showing different shifts and hysteresis in the threshold voltage, indicating that hysteresis can be engineered by changing the doping concentration of Nb atoms. c Schematic diagram of the causes of hysteresis. d Modulation of conductance under the continuous application of potentiation pulses (25 pulses of −30 V) and depression pulses (25 pulses of +30 V). The inset is a diagram of the applied pulse

Full size image

The synaptic behavior of the 5% Nb-WS2 transistor was further investigated. Figure 5a shows the long-term potentiation (LTP) and long-term depression (LTD) characteristics of the device during 25 potentiation (−30 V)–25 depression (+30 V) gate pulse trains. Following 25 cycles of negative gate voltage pulses, the transistor current consistently increased nearly 103 times, increasing from 1.7 × 10−13 A to 1.5 × 10−10 A (Fig. 5a). With another 25 cycles of positive gate voltage pulses (+30 V), the current gradually receded to approximately 1.8 × 10−13 A. Thus, the Nb-WS2 transistor emulates both synaptic potentiation and depression responses. Even after 11 cycles of stimulation, the ratio between the high- and low-conductance states remains above 103, and the consistent curve shape demonstrates the commendable reliability of this synaptic transistor (Fig. 5b). The Nb-WS2 synaptic transistor exhibited minimal baseline drift and low cycle-to-cycle variation, confirming its outstanding electrical durability. As depicted in Fig. 5c, throughout 50 repetitive transfer curve measurements, the hysteresis remains distinctly discernible. Furthermore, we demonstrated that dynamically controlling the gate bias during pulse training can significantly improve the linearity of the synaptic transistor. Figure 5d shows the current update with the stimulus bias gradually increasing from −1 to −25 V (in steps of −1 V), and the depression pulses are illustrated with a gradually decreasing bias from 1 to 25 V (in steps of 1 V).

Fig. 5: Performance of transistors simulating artificial synapses with 5% Nb-WS2 as the channel material.
figure 5

a Synaptic potentiation occurs during the first 25 pulses, whereas synaptic depression takes place during the subsequent 26th to 50th pulses, both of which are induced by voltage pulses of −30 V and 30 V. The current was assessed following the gate pulses (Vgs = 0 V). The inset is a diagram of the applied pulse. b Cyclic repetition of the potentiation and depression processes by alternating negative and positive pulses. c Device durability was assessed by conducting a 50-cycle transfer characteristic curve test. d Potentiation and depression behaviors are demonstrated by adjusting the gate bias. e Network structure of the five-level CNN for simulating pattern recognition. The evaluation is based on a Fashion-MNIST dataset. f Fashion-MNIST pattern recognition accuracy when potentiation/depression data are used. g Schematic diagrams depicting a three-layer ANN and simulated synaptic weight array for classifying MNIST handwritten digits. h The simulated accuracy of pattern recognition is depicted for synaptic transistors via 20 × 20-pixel handwritten digit image datasets

Full size image

To evaluate the neuromorphic computing performance of the Nb-WS2 transistor, we utilized performance data from single devices to simulate the behavior of the Nb-WS2 transistor arrays. We evaluated its convolution operation capabilities by constructing a five-layer convolutional neural network (CNN) for image recognition via the Fashion-MNIST dataset. As illustrated in Fig. 5e, the CNN structure comprises convolutional and pooling layers for feature extraction, along with fully connected layers for classification. By utilizing 28 × 28-pixel grayscale images as inputs, the CNN employs 3 × 3 kernels for convolutional operations. Subsequent to convolution, a pooling function facilitates downsampling, generating feature maps connected to the fully connected layer for image classification. In simulation experiments with the Nb-WS2 transistor array, the CNN achieved a final accuracy of 83.4% after 125 training epochs (see Fig. 5f). Moreover, within the framework of a three-layer artificial neural network (ANN) constructed with the Nb-WS2 transistor array, superior performance was observed in recognizing the MNIST handwritten digit dataset. Figure 5g illustrates the use of a multilayer perceptron neural network for performance benchmarking tests. A backpropagation algorithm was employed to assess the theoretical learning and recognition performance traits of artificial synaptic transistors fabricated from Nb-WS2. The ANN architecture comprises 400 input neurons (corresponding to 20 × 20 image pixels), 100 hidden layer neurons, and 10 output neurons (representing the 10 categories of Arabic numerals from 0 to 9). The diverse parameters for the Nb-WS2 artificial synaptic transistors were subsequently determined via the following equations:

$${{G}_{LTP}={\rm{B}}(1-{{\rm{e}}}^{(-\frac{P}{A})})+{G}_{{min }}}$$
(1)
$${G}_{{LTD}}=-{\rm{B}}(1-{e}^{\left(\frac{P-{P}_{max }}{A}\right)})+{G}_{max }$$
(2)
$${\rm{B}}=({G}_{{max }}-{G}_{{min }})/(1-{e}^{(-\frac{{P}_{{max }}}{A})})$$
(3)

where GLTP and GLTD represent the conductances for LTP and LTD, respectively. The parameters Gmax, Gmin, and Pmax are derived from experimental data, indicating the maximum conductance, minimum conductance, and maximum number of pulses required for the device to switch between its minimum and maximum conductance states. Parameters A and B are the nonlinear fitting coefficients of LTP and LTD. In the context of handwritten digit dataset recognition, the device achieves a recognition accuracy of 92.3% after 125 training iterations, which is closely comparable to the 95.3% accuracy observed for the ideal synaptic device (Fig. 5h). These findings collectively underscore the robust endurance and stability of the synaptic transistor. This effectively mimics the repetitive learning progression exhibited by synapses, confirming that Nb-WS2 is a fitting material for the construction of artificial synaptic devices.

Conclusion

In summary, we devised an effective methodology for achieving controlled Nb doping in monolayer 2D WS2. The doping concentration of Nb atoms can be tailored by manipulating the concentration of the liquid precursor. The uniformly distributed Nb atoms within WS2 establish a stable doping effect and electron characteristics, acting as trap centers to modulate the conductivity of WS2. By employing Nb-WS2 as the channel material, we fabricated gate-tunable synaptic transistors that effectively simulate synaptic potentiation, depression, and repetitive learning processes. The doping of Nb leads to a substantial increase in the switch ratio of the synaptic transistor, increasing it from 10 to 103. In the recognition of the MNIST handwritten digit dataset, the Nb-WS2 synaptic transistor attains 92.30% accuracy after 125 training iterations. This work demonstrates controlled doping in 2D materials, offering versatile tuning of doping concentrations and providing a suitable platform for designing functional electronic devices, such as artificial synaptic transistors.

Materials and methods

CVD growth of WS2 and Nb-doped WS2

Si substrates measuring 15 mm × 15 mm with a 285 nm oxide layer underwent ultrasonic cleaning with acetone and isopropanol (IPA). The precursor solutions, which were composed of Nb(HC2O4)5·xH2O (99.99%, Aladdin) and Na2WO4·2H2O (99.99%, Aladdin) at atomic molar ratios, were spin-coated onto oxygen plasma treated substrates at 3000 rpm. Undoped WS2 and WS2 doped with nominal concentrations of 2% and 5% were grown while maintaining a constant concentration of 1.8 mg/ml to regulate the ___domain density. The Si substrates were positioned within the center of a quartz tube. Following a 20-minute Ar purge at a rate of 300 sccm, the furnace was heated to 875 °C within 30 min under the protection of 80 sccm Ar. A sufficient S source (200 mg, 99.95%, Aladdin) was placed upstream at the edge of the temperate zone. Upon reaching 780 °C in the main temperature zone, the S precursor was heated to 150 °C synchronously with the main temperature zone by a heating belt. After the growth process continued for 15 min, the furnace tube was naturally cooled to room temperature.

Transfer of the sample onto a TEM Grid

The transmission samples were subjected to a polymer film-free method to minimize residues on the WS2 flake. Under microscopic guidance, the carbon film side of the copper mesh was gently positioned on the target sample. A drop of IPA was applied to the flake, which was allowed to evaporate naturally. The assembly was subsequently dried on a hot plate at 80 °C for 10 min to enhance the binding of WS2 to the carbon film. The silicon substrate, carrying the copper mesh, was immersed in a 1.5 mol/L KOH solution and etched at 60 °C for several hours until the substrate separated from the copper mesh. The copper mesh, retrieved with self-locking tweezers, was briefly soaked in deionized water to eliminate residual KOH and then subjected to a vacuum pump for 10 min to completely evaporate the adsorbed water.

Materials characterization

Raman and PL measurements were conducted via a Raman spectrometer (Renishaw, U.K.) equipped with a 532 nm excitation laser. Prior to characterization, the system underwent calibration using standard Si with a peak at 520.5 ± 0.2 cm−1. The laser powers were maintained below 1 mW to prevent sample damage. The surface chemical state of the WS2 flakes was analyzed via Axis Ultra XPS (Kratos, UK) with an Al Kα X-ray source. The thickness of each WS2 flake was examined via AFM (Bruker, Dimension ICON). The atomic structure was investigated via ADF-STEM (FEI Titan ChemiSTEM) with an accelerating voltage of 200 kV. The SAED patterns were recorded with an FEI Tecnai F20 system.

Device fabrication and measurement

The source and drain electrodes were defined through standard electron beam lithography. Electrodes were deposited via e-beam evaporation of 5 nm Cr followed by 10 nm Au. By employing a polymethyl methacrylate (PMMA)-assisted transfer method, WS2 sheets were transferred onto the prepared electrodes. Initially, PMMA (MicroChem, 950 A4) was coated on the substrate and spin-coated at 3000 rpm for 60 seconds. The samples were subsequently dried at 80 °C for 20 min and allowed to rest for 24 h to ensure the robustness of the polymer film. The samples were subsequently etched in a 1.5 mol/L KOH solution, after which the solution was heated to 60 °C to expedite the etching process. Upon separation of the PMMA film from the substrate, the film was picked up via tweezers and submerged in deionized water for 10 min to remove the KOH residue. This process was repeated three times. The meticulous transfer of WS2 nanosheets onto a substrate with predeposited electrodes was then executed under microscopic observation (see the details and schematic diagram in Fig. S6). The device measurements were conducted with a semiconductor characterization system (4200-SCS, Keithley, USA) using a vacuum probe station (10−5 Pa).