Atomic Nb-doping of WS₂ for high-performance synaptic transistors in neuromorphic computing

The study addresses the challenge of energy-efficient neuromorphic computing by leveraging atomically thin TMDCs for synaptic transistors. Conventional von Neumann architectures face power limitations, and memristive devices with gradual conductance modulation offer a path to emulate synaptic weight updates. Prior 2D material-based neuromorphic devices often rely on mechanisms (grain boundary-induced vacancies, phase transitions, Schottky barriers) that are sensitive to material quality and fabrication variability. Substitutional doping presents a controllable route to tune electronic properties and engineer defects for reliable synaptic behavior. The research question focuses on whether p-type substitutional Nb doping in WS₂ can enhance synaptic transistor performance (e.g., switching ratio, linearity of weight updates) and clarify the role of dopant-induced traps versus substrate oxide defects. The study proposes a p-type doping strategy via CVD-grown Nb-doped WS₂ and evaluates its structural, electronic, and device-level impacts, including neuromorphic task performance.

Previous works demonstrated memory functions in 2D materials using mechanisms such as grain boundary vacancies, phase transitions, and Schottky barriers, but device characteristics depend strongly on intrinsic material quality. Appenzeller et al. reported electric-field-induced structural transitions in 2H-MoTe₂ and Mo₁₋ₓWₓTe₂ RRAM with on/off ratios of 10⁶ and low programming currents using Al₂O₃/MoTe₂ stacks. Wang et al. formed MoS₂₋ₓOₓ via O₂ plasma etching and realized graphene/MoS₂₋ₓOₓ/graphene heterostructures with endurance up to 10⁷ cycles. Cheng et al. built artificial synaptic transistors with heavily V-doped MoS₂ showing potentiation, depression, and repetitive learning via gate-tunable trapping/detrapping. However, the role of atomic dopants versus substrate oxide defects in TMDC-based synaptic transistors remains debated. This work aims to clarify dopant contributions using Nb-doped WS₂.

Material synthesis: Nb-doped WS₂ monolayers were grown via atmospheric pressure CVD using a liquid-phase precursor. A 1.8 mg/mL solution was prepared from niobium oxalate (Nb(HC₂O₄)₅·xH₂O) and sodium tungstate (Na₂WO₄·2H₂O). SiO₂/Si substrates (285 nm SiO₂) were spin-coated with the precursor, placed in a one-inch quartz tube, and heated to 875 °C under Ar flow. Growth duration was 15 minutes followed by natural cooling. Nominal Nb precursor ratios targeted 0% (undoped), 2%, and 5% (also explored 7%). Triangular monolayer flakes formed with sizes dependent on nominal doping. AFM confirmed monolayer thickness (~0.86 nm). Characterization: Raman spectroscopy identified E′ and A₁′ modes at ~349 and ~416 cm⁻¹. PL spectroscopy showed emission redshift with increased Nb content (from 626.9 nm undoped to 630.6 nm at 2% and 637.5 nm at 5%). Differential reflectance indicated a neutral exciton around 630 nm. XPS detected Nb 3d peaks in doped samples; W 4f and S 2p binding energies redshifted with doping (e.g., W 4f by ~0.2 eV; S 2p by ~0.25 eV; Nb 3d₃/₂ and 3d₅/₂ further redshift at higher doping), indicating p-type doping and Fermi level movement toward the valence band. Low-temperature PL (100 K) revealed a defect-related X peak at ~1.88 eV. HAADF-STEM and SAED confirmed monolayer lattice quality and substitutional Nb at W sites with distinct Z-contrast. Quantitative intensity analysis yielded actual Nb concentrations of ~0.2% (from 2% nominal) and ~0.3% (from 5% nominal). Measured lattice spacings: (10-10) ~0.27 nm; (11-20) ~0.16–0.18 nm. DFT calculations supported the upward shift of valence and conduction bands upon p-doping. Device fabrication: Artificial synaptic transistors used monolayer Nb-WS₂ as the channel, transferred via wet-etch to prepatterned metal electrodes on heavily doped Si/285 nm SiO₂ substrates serving as back gates. Devices exhibited transfer characteristics (I_ds–V_gs) measured over a 60 V sweep. Due to triangular flakes, channel lengths varied, though on/off ratios were reported to be weakly affected by geometry. Neuromorphic evaluation: Synaptic behavior (potentiation, inhibition, repetitive learning) was assessed via gate-modulated conductance changes attributed to dopant-induced electron trapping/detrapping. Recognition performance was evaluated on the MNIST handwritten digit dataset; training involved 125 iterations leading to the reported accuracy.

- Successful substitutional Nb doping at W sites in WS₂ with uniform distribution across monolayer flakes, confirmed by HAADF-STEM Z-contrast and XPS. - Optical/electronic signatures of p-type doping: PL redshift from 626.9 nm (undoped) to 630.6 nm (2% nominal) and 637.5 nm (5% nominal); XPS redshifts in W 4f (~0.2 eV) and S 2p (~0.25 eV) peaks with increasing Nb; Nb 3d peaks present only in doped samples and further redshift at higher doping. DFT supports valence band maximum shift toward Fermi level. - Low-temperature PL reveals defect-related emission at ~1.88 eV in doped samples, indicating increased defect/trap density due to Nb incorporation. - Morphology trends: Average lateral flake size decreases with higher nominal doping (undoped: 125.3 µm; 2%: 62.5 µm; 5%: 31.6 µm). AFM thickness ~0.86 nm confirms monolayers. - Actual dopant concentrations measured from STEM intensity profiles are lower than nominal (~0.2% for 2% nominal; ~0.3% for 5% nominal). - Electrical behavior: Back-gated FET transfer curves show threshold voltage shifts to more positive V_g with increased doping, consistent with enhanced p-type transport. FET on/off ratios: undoped WS₂ ~10⁵; 2% Nb-WS₂ ~10²; 5% Nb-WS₂ ~10³ (over 60 V sweep). - Synaptic switching: Nb-WS₂ synaptic transistors exhibit an improved switch ratio of ~10³, approximately 100× larger than undoped WS₂ synaptic devices, attributed to dopant-induced trapping/detrapping. - Neuromorphic task: MNIST recognition accuracy of 92.30% after 125 training iterations using Nb-WS₂ synaptic transistors. - Higher nominal doping (7%) degrades electrical and synaptic performance.

The results demonstrate that substitutional Nb dopants in WS₂ introduce controllable trap states that effectively mediate electron trapping and detrapping, enabling stable and tunable conductance modulation under gate control. This addresses the prior ambiguity regarding whether synaptic behavior in TMDC-based devices primarily arises from substrate oxide defects or intrinsic dopant-related mechanisms; the observed p-type shifts in PL/XPS and consistent STEM evidence support the central role of dopants. Although the FET on/off ratio decreases with doping, the synaptic switching ratio and weight update linearity improve significantly due to optimized trap density and energy levels. The uniform large-area CVD growth and reproducible doping provide a practical route for scalable neuromorphic hardware. The achieved 92.30% MNIST accuracy after modest training demonstrates functional relevance for neuromorphic computing, and the device exhibits key synaptic behaviors such as potentiation, depression, and repetitive learning.

This work introduces a practical p-type substitutional doping strategy for WS₂ using Nb, achieving uniform monolayer films with controlled electronic structure and defect profiles suitable for synaptic transistors. Structural (STEM), spectroscopic (Raman, PL, XPS), and electronic measurements confirm successful p-type doping and substitution at W sites. Nb-WS₂ synaptic transistors show a markedly enhanced switching ratio (~10³), about 100× improvement over undoped synaptic devices, and deliver 92.30% accuracy on MNIST after 125 training iterations. The methodology provides a scalable pathway for engineering channel materials in neuromorphic devices and clarifies the role of dopant-induced traps in synaptic functionality. Future directions include optimizing dopant concentration (avoiding degradation observed at higher levels), exploring alternative dopants and co-doping schemes, engineering device geometries and gate dielectrics for improved linearity and energy efficiency, and systematically evaluating endurance, retention, variability, and large-scale array integration.

- Actual Nb incorporation levels were substantially lower than nominal precursor ratios (~0.2–0.3%), indicating limited dopant uptake control. - Increased doping reduced lateral flake size, potentially affecting device scalability and uniformity. - FET on/off ratios decreased with doping, reflecting a trade-off between transistor metrics and synaptic performance. - Very high nominal doping (e.g., 7%) led to rapid performance degradation. - Channel lengths varied due to triangular flake geometry, which may influence current levels, though reported to have minimal impact on on/off ratios. - The study does not report detailed endurance, retention, noise/variability statistics, or energy per operation, which are important for practical neuromorphic systems.