Introduction

Ferroelectrics, characterized by spontaneous polarization that can be switched by a directional external electric field, have demonstrated promising applications in non-volatile memories, logic devices, and sensors1,2,3,4. In the past few decades, most efforts have focused primarily on traditional ferroelectrics, including perovskite oxides, hybrid perovskites and organic compounds5,6,7. However, robust ferroelectricity at room temperature in ultrathin traditional ferroelectrics is rarely sustained due to critical depolarization thickness or interfacial effects8,9. In contrast, two-dimensional (2D) van der Waals (vdW) ferroelectrics with dangling-bond-free surface, weaker interlayer vdW coupling, stable polarization and atomic-scale integration offer significant advantages for potential applications in non-volatile memory and next-generation nanodevices10.

Recently, several emerging intrinsic 2D ferroelectric materials have been experimentally verified, including CuMP2S6 (M = In, Cr)11,12, SnX (X = S, Se, Te)13,14,15, α-In2Se316, Bi2O2Se17, WTe218, d1T-MoTe219, GaSe20, and 3 R MoS221. Among these, 2D transition metal phosphorus chalcogenides (TMPCs) have garnered significant attention due to their unique electronic structures and novel physical properties22. In particular, CuMP2S6 (M = In, Cr), which exhibits strong room-temperature ferroelectricity, has been extensively studied, with the corresponding ferroelectricity attributed to the predicted Cu ion migration effects11,12,23,24,25. However, several issues remain to be addressed for the development of 2D ferroelectricity: (1) the synthesis of ultrathin 2D ferroelectrics for direct applications in nanodevices is still challenging; (2) an experimental understanding of the ferroelectric mechanism has not been achieved. CuVP2S6 (CVPS), an analog incorporating magnetic V cations into the TMPCs system, emerges as a potential candidate for realizing theoretically predicted room-temperature multiferroics and non-volatile functions, with promising applications in neuromorphic computing systems26,27,28,29. However, the direct synthesis of ultrathin ternary 2D CVPS remains challenging due to its complex crystal structure, inferior stability and intricate chemical reaction processes. Furthermore, while theoretical models have provided valuable insights into the ferroelectric mechanisms of 2D ferroelectric materials, a comprehensive experimental understanding of these ferroelectric mechanisms has not yet been achieved. Therefore, the synthesis and investigation of the ferroelectric ordering in vdW layered CVPS are crucial for understanding the ferroelectric mechanisms and developing novel electronic devices.

In this study, we report the synthesis of CVPS for the first time via the spatially confined chemical vapor deposition (SC-CVD) method. The as-synthesized CVPS flakes exhibit robust ferroelectricity with a high transition temperature of 650 K and maintain switchable out-of-plane (OOP) ferroelectric polarization down to 15 nm. Importantly, we identify a ferroelectric mechanism originating from the movement of Cu ions induced by applied in-situ external electric fields, resulting in different structures, including CuxVP2S6 (x:1 ~ 2), due to the multi-occupations of Cu ions. Furthermore, ferroelectric synaptic devices based on few-layer CVPS exhibit bias-controllable ferroelectric switching behavior and demonstrate notable electrical and optoelectronic synaptic properties. By utilizing up to 214 levels of fine conductance modulation, these devices can effectively simulate letter recognition and neural machine translation. Our study not only demonstrates the ferroelectric properties of CVPS but also elucidates the ferroelectric mechanism of TMPCs, underscoring the promising applications in non-volatile memory devices.

Results

Synthesis and microstructure characterization

To synthesize high-quality quaternary CVPS, we propose an SC-CVD method with a specific precursor ratio (see schematic diagram in Fig. S1 and details in the Experimental Section). The growth kinetics of CVPS can be well-regulated by adjusting the growth parameters. Compared to traditional CVD methods, a controllable temperature gradient and reaction time are crucial for achieving equilibrium in the gas-solid reversible chemical reaction of CVPS. At higher temperatures, CVPS samples with thick flower-like morphology are easily obtained in the low-temperature region due to excessive precursors and prolonged reaction time (Fig. S2). After optimizing the amounts of the precursors, growth temperature and reaction time, the growth of bulk crystals was greatly avoided by decreasing the reaction rate and nucleation density on the substrate surface, making it possible to obtain CVPS flakes with different thicknesses in the low-temperature zone. Figure 1a shows the optical image of CVPS flakes with a transverse size of ~20 μm on a SiO2/Si substrate. Figure 1b presents the top view of the CVPS, where the sulfur and Cu atoms retain the traditional characteristics of the phosphorus-sulfur compound. The metal atoms are confined within octahedral frames formed by sulfur atoms and arranged in an orderly manner in the ab-plane within the ab-plane to form a honeycomb lattice. However, in the side view shown in Fig. 1c, a difference between V atoms (blue balls) and Cu atoms (red balls) is observed, where the V ions are nearly centered in each vdW layer and the Cu ions are positioned at the edges of the vdW layers.

Fig. 1: Crystal structure and characterizations of CuVP2S6.
figure 1

a Optical image of CuVP2S6. b, c Top and side views of the atomic structure of CuVP2S6. d Thickness-dependent Raman spectra of CuVP2S6 flakes. Four peaks with Eg and Ag modes are observed. e-h XPS spectra of Cu 2p, V 2p, P 2p, and S 2p orbitals, which are in line with the corresponding standard characteristic peak positions.

Full size image

Raman spectroscopy was performed to confirm the quality and the crystal structure of ultrathin CVPS. Figure 1d depicts the Raman spectra of CVPS samples with different thicknesses, revealing four prominent Raman peaks of P1 (~222 cm−1), P2 (~298 cm−1), P3 (~375 cm−1) and P4 (~578 cm−1) at room temperature. Fig. S3a presents the parallel (HH) and cross (VH) polarized Raman spectra of CVPS samples, illustrating that the P2 and P3 peaks correspond to the Eg mode, whereas the P1 and P4 peaks correspond to the Ag mode. The temperature-dependent Raman spectra shown in Fig. S3b indicate no phase transition within the measured temperature range. The X-ray diffraction analysis of CVPS single crystals (Fig. S4) reveals two peaks at 13.83° and 27.88°, corresponding to the (002) and (004) planes, respectively, indicating high crystallinity along the c-axis of the obtained CVPS samples. The X-ray photoelectron spectroscopy (XPS) measurements were conducted to characterize the chemical states of elements in CVPS crystals (Fig. 1e-h). The characteristic peaks at ~953.3 and 932.8 eV correspond to Cu 2p1/2 and Cu 2p3/2, respectively. The peaks corresponding to V 2p1/2 and V 2p3/2 are located at around 522.3 and 514.8 eV. The fitting peaks of P located at 134.9 and 133.8 eV correspond to the P 2p1/2 and 2p3/2, while the peaks for S 2p1/2 and S 2p3/2 core level orbitals are detected at 164.3 and 163.2 eV, respectively. The aforementioned characterizations confirm the high quality of the synthesized CVPS samples.

Second harmonic generation characterization and ferroelectric switching measurement

To study the intrinsic ferroelectricity of CVPS, polarization-electric field (P-E) hysteresis tests were conducted at room temperature. Measurements under a triangular waveform at 1 kHz revealed a well-defined P-E hysteresis loop (Fig. S6b), confirming the intrinsic polarization of CuVPS. The residual polarization of the CVPS flake is ~0.96 μC/cm². Detailed experimental details are shown in Fig. S6. The piezoresponse force microscopy (PFM) amplitude and phase signals reflect the local electromechanical responses and distinguish different out-of-plane polarization orientations of ferroelectric domains, thereby confirming the existence of ferroelectricity on the nanoscale. Spontaneous ferroelectric domains were observed in the CVPS samples, revealing distinct regions with different polarization orientations, shown in Fig. S7. To comprehensively detect the OOP ferroelectric responses, CVPS nanoflakes of various thicknesses were transferred onto gold-coated SiO2/Si substrate, and the substrate was grounded during the testing process. We performed Dart-PFM hysteresis measurements on the samples at room temperature. The off-field phase loop of the 25 nm CVPS flake exhibits a typical signature of polarization switching (red line in Fig. 2g). The corresponding off-field piezoelectric loop displays a characteristic butterfly-shaped response (blue line in Fig. 2g). The PFM phase hysteresis and butterfly-shaped amplitude-voltage loops demonstrate the OOP ferroelectric properties of CVPS. To further confirm the ferroelectricity of CVPS, the polarization reversal in ferroelectric domains is manipulated through PFM electric “read and write” operations. Figure 2h and i show the amplitude and phase images of OOP PFM image after writing box-like patterns with opposite DC bias applied through the tip (large box at −7 V, small box at +7 V). The distinct phase and amplitude contrast in the writing regions further demonstrate the ferroelectricity in CVPS. The PFM phase hysteresis and butterfly-shaped amplitude hysteresis of CVPS with different thicknesses (15 nm and 69 nm), along with the box-in-box patterns, are shown in Fig. S8. The AFM image of the corresponding sample after the domain writing test is shown in Fig. S9. It can be seen that the surface morphology of the sample does not change significantly after DC voltage domain writing. The PFM results demonstrate that the ferroelectricity of CVPS can be sustained down to a few-layer thickness.

Fig. 2: SHG and ferroelectric characterizations of CuVP2S6.
figure 2

a SHG spectra of CuVP2S6 as a function of excitation power, with the incident laser wavelength of 1064 nm. b Power-law relationship between SHG intensity and excitation power, yielding a fitting coefficient of 1.95. c Thickness-dependent SHG intensity of CuVP2S6 nanoflakes. d, e Angle-resolved polarized SHG spectra in parallel (d) and perpendicular (e) configurations, featuring experimental data (dots) and corresponding fitted curves. f Temperature-dependent SHG intensity of CuVP2S6 nanoflakes, indicating the Curie temperature of 650 K. g The PFM amplitude (blue) and phase (red) hysteresis loops by applied a reverse DC voltage of ±7 V in CuVP2S6 with thickness of 45 nm. h, i The PFM OOP phase (h) and OOP amplitude (i), images of CuVP2S6 nanoflake after writing a box-in-box pattern with opposite DC bias of ±7 V, respectively.

Full size image

Scanning transmission electron microscope characterizations

The high-angle annular dark-field scanning transmission electron microscope (HAADF-STEM) technique is employed to investigate the atomic structure and ferroelectric mechanism of CVPS. Figure 3a presents a low-magnification image of the CVPS sample. The HAADF-STEM atomic structures along the [001] and [010] directions, along with the corresponding atomic models, are shown in Fig. 3b, c, respectively. Notably, in the CuVP2S6 atomic model, Cu alternates between the VS6 octahedral frameworks. However, the HAADF images obtained at high voltage show Cu ions fully occupying the spaces between the VS6 octahedra, forming a Cu2VP2S6 structure (Fig. S10). This discrepancy suggests that the energy of the electron beam plays a significant role in the movement of the highly mobile Cu ions within the CVPS layers. To better illustrate the behavior of Cu ions migration between the VS6 octahedral frameworks, the corresponding atomic model is shown in Fig. 3d.

Fig. 3: Atomic structure of CuVP2S6 and the in-situ process.
figure 3

a Low-magnification image of the CuVP2S6 sample, showing regular dissociation surfaces. b Atomic structure of CuVP2S6 in [001] direction with relative atomic model. The red, blue, mauve, and yellow dots represent Cu, V, P, and S atoms, respectively. c Atomic structure of CuVP2S6 in [010] direction with relative atomic model. d Atomic model illustrates the Cu ion migration process. e Schematic of low-dose e-beam irradiation on the CuVP2S6. f Atomic structures of the irradiation process at 2 s, 4 s, and 6 s. The yellow rectangle and array show the Cu ion migration direction from right to left, and the blue rectangle and array show the Cu ion migration direction from left to right. g Schematic diagram of the in-situ STEM system equipped with electrical biasing capabilities h HAADF-STEM image of a CuVP2S6 sample under +9 V bias, showing leftward polarization reconfiguration. Yellow arrows and dashed boxes indicate the migration trajectories of Cu ions toward the left. i HAADF-STEM image of the pristine CuVP2S6 sample (zero bias), displaying the equilibrium distribution of Cu ions within the van der Waals layered structure. j HAADF-STEM image of the CuVP2S6 sample under −9 V bias, demonstrating rightward polarization reversal. Blue arrows and boxes highlight the coordinated displacement of Cu ions toward the right, consistent with field-driven ferroelectric switching dynamics.

Full size image

To further investigate the migration of Cu ions, we first examined the effect of electron beam dose on Cu ions migration, as shown in Fig. S11. The Cu ions concentration at the observation zone decreases with the reduction of electron beam dose, confirming the sensitivity of CVPS material to electron beam dose. To balance both resolution and the observation of Cu ions migration, we selected a 6 pA electron beam dose for the subsequent series of experiments, as depicted in Fig. 3e. The atomic evolution images over time in Fig. 3f illustrate the movement of Cu ions. It is evident that, with increasing irradiation time, Cu ions gradually migrate into the electron beam-irradiated region. These Cu ions randomly occupy the positions on either the left or right side of the vdW layers, referred to as Cu-left and Cu-right, respectively, forming Cu atomic columns as indicated by the bright spots in Fig. 3f. Due to the sensitivity of CVPS to electron beam dose, we can only capture images within a short possible time to ensure optimal imaging results.

Since the ferroelectricity of CuMP2S6 (M = In, Cr, V) strongly depends on the ordered occupation of Cu ions, direct observation of Cu ions migration at the atomic scale is crucial for understanding the ferroelectric mechanism in two-dimensional materials. Thus, we studied the atomic structure and ferroelectric mechanisms using an in-situ applied external electric field, as shown in Fig. 3g-j. The low-magnification STEM image of the sample with in-situ applied voltage is shown in Fig. S12. The schematic of the in-situ voltage-applied STEM setup is shown in Fig. 3g. The HAADF-STEM images in Fig. 3h and j reveal two opposite ferroelectric polarizations of the CVPS, where the yellow and blue boxes indicate the positions of Cu atoms on the left or right side of each vdW layer, resulting in non-zero electric dipole moments directed to the left or right, respectively. Traces of Cu ions movement forced by e-beam irradiation and voltage show solid evidence of the Cu movement between the VS6 octahedral skeletons, which vividly exhibits the influence of atomic evolution on the ferroelectric properties. The findings of Cu ions migration dynamics provide key insights into the ferroelectric performance at the atomic scale, offering valuable guidance for the future design and performance tuning of CuMP2S6 (M = In, Cr, V) devices.

Electronic and optoelectronic synapse based on ferroelectric CuVP2S6

To further characterize the ferroelectricity of CVPS, Kelvin probe force microscopy measurements (KPFM) were conducted to elucidate the contact potential of the device, with the optical image shown in Fig. S13a. As shown in Fig. S13b, the contact potential is ~90 mV higher than that of gold. Additionally, we performed Conductive Atomic Force Microscopy (C-AFM) measurements to investigate the OOP ferroelectric polarization switching capability of 9 nm thick CVPS nanoflakes (Fig. S13d, with the corresponding AFM image shown in Fig. S13c). The I-V hysteresis obtained corresponding to the voltage sweep from −5 to +5 V is shown in Fig. S13d and is attributed to the modulation of Schottky barrier height between the AFM tip and CVPS due to the ferroelectric polarization switching.

To further investigate the synaptic properties of CVPS, we fabricated vertical devices with Graphene electrodes. The schematic diagram of the CVPS-based device is shown in Fig. 4a. The I-V curves of the device display distinct hysteresis windows, which progressively increase with scan voltages increasing (Fig. 4b). We exclude the effect of trap states through the invariant hysteresis loop window at different scan rates (Fig. S14). Figure 4c illustrates the electrical pulse characteristics of the device under pulse voltages ranging from −1 V to −2 V with a pulse width of 10 ms, measured at 0.1 V. Following the negative pulses, the postsynaptic current (PSC) noticeably decreases and then recovers to its original state, demonstrating a short-term depression (STD) effect. In addition to the synaptic properties induced by single-pulse stimulation, the synaptic characteristics observed under paired-pulse stimulation are also interesting. The impact of paired-pulse facilitation (PPF) and paired-pulse depression (PPD) indicate alterations in neurotransmitter release, which are critical for the function of biological synapses30. Figure 4d and Fig. S15 illustrate the PPD ratio decreasing with the intervals between paired pulses increasing. The PPD ratio associated with A1 and A2 can be fitted using a double-exponential decay function: C0  + C1exp(-Δt/τ1) + C2exp(-Δt/τ2), where C0, C1 and C2 represent the initial facilitation magnitudes of facilitation, τ1 and τ2 denote characteristic relaxation timescales. Figure 4e further shows the inhibitory PSC response to a series of continuous electrical pulses of equal length under different frequencies. The cumulative inhibition effect increases as the pulse frequency increases from 10 Hz to 50 Hz, leading to a reduced inhibitory PSC. Figure 4f illustrates the conductance modulation of CVPS, including long-term potentiation (LTP) and long-term depression (LTD). The improved LTP/LTD curves were obtained when the incremental voltage pulses (excitation: 0.1 V to 2.02 V, 20 mV steps, 10 ms pulse width; inhibition: −0.02 V to −2.38 V, 20 mV steps, 10 ms pulse width) were applied as the synaptic device’s input, with modulated conductance recorded at the output. Long-term conductance modulation of CVPS exhibits excellent linearity and a broader range of conductance states compared to In2Se3 synapses31, thereby enhancing the precision of CVPS-based synaptic devices in recognition tasks.

Fig. 4: Electronic characteristics and optoelectronic properties of the ferroelectric CuVP2S6-based synapse.
figure 4

a Schematic illustration of the ferroelectric graphene/CuVP2S6/graphene device. b Short-term potentiation effects triggered by electrical pulses of varying amplitudes (−1, −1.3, −1.5, −1.7 and −2 V, 10 ms) with 0.1 V bias for state reading. c Short-term potentiation effects stimulated by pulses of diverse widths (10, 20, 30, 40 and 50 V, 1.5 V), device states read at 0.1 V voltage. d PPD index plotted against pulse interval time, defined as (A2-A1)/A1 × 100%, with the red fitted curve via a double exponential function. Inset: PPD effect induced by voltage pulses with definitions of A1, A2, and time intervals Δt. e Excitatory PSC induced by a series of electrical pulses with frequencies of 10, 20, 25, and 50 Hz and device states read at 0.1 V bias, indicating more pronounced accumulation effects induced by higher pulse frequencies. f For LTP, 96 pulses are applied with a voltage range of 0.1 to 2.02 V for 10 ms, whereas for LTD, 118 pulses are used with a voltage range of −0.02 to −2.38 V for 10 ms, with state readout at 0.1 V bias. g PSC variations with different light pulse illumination times (0.1, 0.5, 1, 1.5, 2 and 2.5 s) are tested at 117 W cm⁻², read at 0.1 V. h PPF index plotted against pulse interval time, given by (A2 - A1)/A1 × 100%. The red curve fitted by a double exponential function. Inset: PPF effect induced by light pulses with definitions of A1, A2, and time intervals Δt. i Excitatory PSC versus time stimulated by N presynaptic spikes (N = 10, 20, and 30), with a pulse laser power density of 117 W cm−2, pulse width of 500 ms, read at 0.05 V.

Full size image

The inherent optoelectronic properties of CVPS are analogous to those of electrical pulse stimulation, indicating potential applications in optoelectronic synapses. Figure 4g illustrates that applied to the device, the excitatory PSC characteristics resemble electronic synapses under 520 nm laser pulse illumination with an equivalent power density of 117 W/cm² and a small bias voltage of 0.1 V. For shorter illumination durations (e.g., 0.1 s), the device exhibits short-term plasticity (STP) and excitatory PSC characteristics. As the illumination time increases, the optoelectronic synapse exhibit LTP characteristics, potentially due to increased trap filling, leading to higher current levels within the equivalent decay time. For a constant pulse width of 500 milliseconds and varying intensities (7.8, 39, 78, 117, 156, and 195 W/cm2), the peak photocurrent increases with light intensity increasing, as shown in Fig. S16. Similar to electronic synapses, the optoelectronic synapse achieves a PPF effect by applying a pair of light pulses, where the second pulse is applied before the PSC of the first pulse has fully relaxed, resulting in increased photocurrent, as illustrated in Fig. 4h and S17. The PPF index in Fig. 4h decreases with longer interval times, fitted by a double-exponential decay function. As shown in Fig. 4i, by emitting multiple presynaptic spikes consecutively in a short period, the PSC level rises dramatically and decays within a few seconds, similar to a single spike. This demonstrates that our graphene/CVPS/graphene artificial synapse can achieve long-term potentiation. Time dynamics of both electrical and optical stimulation were separately achieved in the same synaptic device, attributed to the ferroelectric and optoelectronic properties of CVPS, highlighting its potential for in-sensor computation.

CuVP2S6 synapses for optical character recognition and neural machine translation

The functional performance of artificial neural networks is fundamentally governed by the diversity and precision of synaptic weight distributions, analogous to the role of synaptic strength variability in biological systems32. Here, we utilized the exceptional properties of CVPS materials, which possess up to 214 levels of adjustable conductance (Fig. 4f), to introduce a brain-inspired neural network chip structure that integrates 2D synaptic devices. While the modulation range of conductance is modest, it is sufficient to demonstrate the chip’s ability to perform tasks such as Optical Character Recognition (OCR) and Neural Machine Translation (NMT) using Convolutional Neural Networks (CNN) and Transformer architectures33, as shown in Fig. 5a, b. The details of the neural networks are discussed below.

Fig. 5: Handwritten letter recognition and neural machine translation.
figure 5

a The convolutional neural network architecture of the OCR module achieves image-to-text conversion. b The Transformer network architecture of the NMT module accomplishes English text to Chinese text translation. c The accuracy of OCR module in recognizing the 26 letters of “a-z” normalized confusion matrices. d, e Simulation results of the NMT module for ideal and device, showing BERT-scores for translations of 100 English sentences across various training epochs. f Comparison of BERT scores for the translations of sentences numbered 90 –100, assessing the final simulations of the NMT modules in both ideal and device-based scenarios.

Full size image

We constructed distinct convolutional neural networks34 for OCR and employed a Transformer-based seq2seq model for NMT33. The weights of these ideal models (using floating-point precision) were derived by training the EMNIST35 dataset for the OCR module and the AI Challenger dataset for the NMT module36 on the Graphics Processing Unit (GPU). In the simulated hardware platform, CVPS neural network weight units replace the weights of the convolutional kernels and fully connected neurons in the OCR module. Similarly, they also replace the weights of the multi-head attention and fully connected layers in the Transformer architecture of the NMT module. Comprehensive details regarding the neural network simulations, including dataset specifics, model parameters, training procedures, and simulation methodologies are detailed in Supplementary Note 1.

The simulation results of each module in the model are illustrated in Fig. 5c-f. We evaluated the performance of the OCR module using 20,800 test images from the EMNIST dataset, achieving a high recognition accuracy of 92.42%. The evaluation results are presented through a normalized confusion matrix, in Fig. 5c. The OCR module benefits from 500 adjustable weights in CVPS synaptic devices, ensuring high-precision recognition, with the variation of recognition accuracy shown in Fig. S18 across training epochs. The recognition outputs from the OCR module are utilized as inputs to the NMT module. For an illustration of the manner in which an electrical signal representing the result of an operation is transmitted between two modules, please refer to SI Note 1.4 Synapse Device Model Simulation. Bert-score37 was employed to measure the translation quality of the NMT module in English-to-Chinese translation tasks, where a score closer to 1 indicates higher translation quality. We selected 100 specific English sentences and translated them using both ideal models trained at different epochs and device-simulated models, comparing them against reference translations with Bert-score results presented in Fig. 5d and e. Figure 5f shows the translation performance of ideal models and device-simulated models on samples 90–100 after 11 training epochs. Overall, observations from Fig. 5d-f demonstrate that the multi-weight-modulated NMT module based on CVPS devices exhibits performance closely approaching that of the ideal computer model in NMT tasks. For some sentence samples, the CVPS synaptic device produces even higher Bert scores than the ideal model, which is attributed to the superior translation efficiency given by its multi-level weight adjustment. Our 2D synaptic devices simulate the complex processes of optical character recognition and real-time translation, laying the foundation for energy-efficient neuromorphic systems capable of multimodal sensory processing and real-time cognitive tasks.

Discussion

In summary, we have successfully synthesized the ferroelectric CVPS and revealed its OOP ferroelectric polarization in an ultrathin flake with a thickness of 15 nm. The in-situ STEM results uncovered the ferroelectric mechanism of Cu migration in CVPS for the first time, which enabled the utilization of CVPS ferroelectric properties to realize a single-material ferroelectric non-volatile memory device. Importantly, we developed a single-material ferroelectric synapse device for optoelectronic synapses, achieving up to 214 levels of fine weight modulation. This device accomplished letter recognition and neural machine translation in a single-material device, offering new avenues and practical insights for advancing the innovation and development of integrated “sense, memory, compute” artificial intelligence chip technologies.

Methods

Synthesis of CuVP2S6 flakes

The schematic diagram of the SC-CVD strategy to grow CVPS flakes is shown in Fig. S1. Stoichiometric amounts (Molar ratio - Cu: V: P: S = 1: 1: 2: 6) of Cu powders (Adamas, 99.9%), V powders (Adamas, 99.9%), P powders (Macklin, 99.9%) and S powders (ZhongNuo Advanced Material Technology Co., Ltd, 99.99%) are sealed at the side of a horizontally-placed quartz tube with pressure <10 mbar. Meanwhile, SiO2/Si substrates are loaded into the other side of the vacuum quartz tube. The vacuum quartz tube was horizontally placed in a tube furnace with source zone at high temperature zone and growth zone at low temperature zone. The furnace was heated to 500 °C within 2 h. After holding for 5 h, the furnace was naturally cooled to room temperature. Finally, the target CVPS flakes were successfully obtained on the SiO2/Si substrates. The CuxVP2S6 (x:1 ~ 2) appears due to the movement of Cu ions in CuVP2S6 induced by applied in-situ external electric fields via HAADF-STEM technique.

Material characterization

The measurements utilizing AFM, PFM, and KPFM were conducted using Bruker Dimension Icon system. Dart-PFM measurement was performed using a commercial atomic force microscope (Asylum Research MFP-3D) with Pt/Ir-coated Si cantilever tips with spring constants of 0.5–9.5 N/m. Domain writing operations employed sequential DC voltage pulses of opposing polarity, while domain reading leveraged AC excitation (VAC = 3 V). Ferroelectric hysteresis loops (P-E) were characterized with a commercial test system (PREMIER II 10 V/DELTA 9023). Elemental composition profiling was conducted via X-ray photoelectron spectroscopy (Thermo Scientific K-Alpha). Microstructural analysis incorporated energy-dispersive X-ray spectroscopy and HAADF-STEM imaging using a Talos F200X G2 microscope. Raman and SHG spectra were measured by the WITec alpha 300 R Ramasn system with laser at 532 nm and 1064 nm (Laser spot diameter of 2 microns), respectively. The 532 nm and 1064 nm lasers are continuous lasers and the spectra are tested at room temperature.

Device fabrication and photoelectric measurements

The CVPS device was fabricated directly on SiO2/Si (285 nm) using the standard electron-beam lithography and lift-off techniques, with contact electrodes (Cr/Au 5/50 nm) deposited via an E-beam evaporation system (DZS500). All measurements were performed in a vacuum environment (10−6 torr) using a probe station (CRX-6.5 K, Lake Shore). Test data were recorded using a Keysight 1500 A semiconductor analyzer. Optical response tests for the devices were carried out under excitation from a 520 nm laser. The device data shown in the paper come from devices corresponding to the optical pictures. We have repeated the experiment on both perpendicular and parallel devices more than three times, all of which found the same synaptic behavior.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.