Interface engineered V₂O₅-based flexible memristors towards high-performance brain-inspired neuromorphic computing

Abstract

Bio-inspired neuromorphic computing offers a revolutionary approach by replicating brain-like functionalities in next-generation electronics. This study presents two flexible resistive memory devices fabricated using DC magnetron sputtering, D₁(Nb/V₂O₅/Ni) and D₂(Nb/NbO_x/V₂O₅/Ni). Device D₁ exhibits abrupt SET and gradual RESET switching, while D₂ demonstrates fully gradual resistive switching (GRS), highly desirable for analog synaptic behavior. Mechanistically, D₁ is primarily governed by oxygen vacancies, whereas D₂ benefits from the synergistic interplay between oxygen vacancies and interfacial NbO_x/NiO layers, confirmed by XPS depth profiling. These interfacial layers significantly enhance D₂'s GRS performance and synaptic fidelity. Both devices exhibit temperature-dependent control of oxygen vacancies, which dynamically increases the memory window, lowering the ON/OFF ratio. Multilevel resistive states are generated in both devices by controlling the compliance current, with D₂ outperforming D₁ by exhibiting a higher memory window (∼552) and exceptional endurance beyond 7000 cycles. Moreover, both devices effectively replicate biological synaptic functions such as LTP and LTD. However, D₂ also mimics complex neural dynamics, including spike time-dependent and rate-dependent plasticity. Simulation of D₂'s artificial neural network demonstrates ∼86.75% excellent accuracy level, attributed to its linear, symmetric analog weight modulation and multiple conductance states. These results highlight the potential of V₂O₅-based devices for high-performance neuromorphic computing.

---

New concepts

This study demonstrates a breakthrough in flexible neuromorphic device engineering through the design of interface-engineered V₂O₅-based memristors employing a dual-interface structure with NbO_x and NiO_x layers. This configuration enables precise control of oxygen vacancy migration, resulting in fully gradual resistive switching behavior essential for analog synaptic operation unlike conventional abrupt-switching V₂O₅ or single-layer devices. The integration of both engineered and self-formed interfacial oxides modulates electric field distribution and stabilizes filamentary conduction, achieving superior switching uniformity, multilevel conductance states, and mechanical flexibility on nickel substrates. This approach provides a fundamental differentiation from existing research by simultaneously realizing flexibility, analog tunability, and endurance in a single device architecture fabricated via scalable DC magnetron sputtering. Beyond device performance, the conduction analysis reveals a transition from Ohmic to trap-controlled space-charge-limited conduction, offering new physical insight into defect-mediated analog switching in oxide-based systems. The successful emulation of synaptic behaviors such as LTP, LTD, STDP, and SRDP, together with ANN-based image recognition, establishes this concept as a significant advance in nanoscience for developing energy-efficient, flexible, and brain-inspired neuromorphic hardware.

1. Introduction

Conventional Von Neumann systems have limitations in the emerging field of artificial intelligence, as the physical separation between memory and the computing unit results in low processing speed and high power consumption.^1–7 These restrictions make it more difficult for them to satisfy the rising need for enormous processing power.^8–11 Researchers have developed an artificial electronic synapse that integrates memory, computation, and signal detection into a single unit, mimicking the parallel processing capability of the human brain. Neural impulses are transmitted between pre- and post-synaptic terminals via electrical and chemical signals in a biological synapse. Neurotransmitters (Ca²⁺, Na⁺) facilitate information transmission by controlling the synaptic weight of biological memory synapses (BMS).^8,12,13 Artificial memory devices have a comparable bionic structure, which includes top and bottom electrodes resembling pre- and post-synapses and vacancies, ions, and electrons like neurotransmitters in BMS.^14–17

Filamentary-based resistive random-access memory (RRAM), with fast switching (∼1.2 ns), low power consumption (∼1.0 fJ), high-density integration via crossbar arrays, and excellent endurance and retention, exhibits synaptic behavior. Its metal–insulator–metal (MIM) structure enables easy fabrication.^{11,12,18–21} Bipolar resistive switching (BRS) can be categorized into abrupt resistive switching (ARS) and gradual resistive switching (GRS) modes. ARS exhibits a high ON/OFF ratio and strong retention but suffers from poor endurance and limited synaptic plasticity. In contrast, GRS offers multiple conductance states with smooth and controllable modulation, making it more suitable for neuromorphic computing applications. This analog-like behavior enhances the linearity of long-term potentiation (LTP) and depression (LTD), enabling precise weight updates required for spike-timing-dependent plasticity (STDP). A. K. Jena et al. reported that gradual switching reduces circuit complexity, while M. Ismail et al. demonstrated that improved LTP/LTD linearity significantly enhances computational accuracy in neuromorphic systems.^9,22 Moreover, GRS devices generally operate at lower voltages and currents, leading to reduced power consumption and improved reliability. Their superior endurance, uniformity, and scalability make them promising candidates for large-scale neuromorphic hardware. The gradual modulation of resistance states is typically governed by the progressive formation and migration of oxygen vacancies, contributing to stable analog performance. Fabricating ARS devices is more straightforward than fabricating GRS devices.²³ To convert ARS to GRS, researchers have optimized key parameters using various techniques. Likewise, a thermally enhanced layer is used to regulate vacancy distribution, thereby modulating conductivity, while interfacial layers restrict charge carrier mobility, both of which facilitate the transition from ARS to GRS.^20,24–27

Bilayer memristor devices are promising candidates for mimicking biological synapses due to their low power consumption and ability to emulate key synaptic behaviors such as long-term potentiation (LTP), long-term depression (LTD), paired-pulse facilitation (PPF), spike-timing-dependent plasticity (STDP), and spike-rate-dependent plasticity (SRDP).^28,29 Several studies have demonstrated the effectiveness of interfacial engineering in enhancing synaptic performance. For example, Ryu et al. reported bilayer ReRAM devices using HfO₂ as interfacial layers, achieving excellent analog switching and synaptic characteristics.³⁰ Wan et al. introduced SiO₂ and graphene oxide layers to realize analog switching and neuromorphic behavior, while Ismail et al. demonstrated gradual switching in TiO₂/HfO₂ bilayers using a tailored pulse scheme.^28,31 In our work, we successfully achieved multilevel analog switching at low operating voltages and emulated essential synaptic functions in a V₂O₅-based ReRAM device by introducing a NbO_x interfacial layer, further validating the effectiveness of bilayer structures in neuromorphic computing applications. Niobium-oxide (NbO_x) is an attractive option, as the material exhibits rich and complex electrically memristive behaviors that can vary based on the film composition, structure (e.g., crystallinity), and mode of electrical operation. In addition, it creates an oxygen vacancy gradient in the oxide layer, which enhances the switching performance. Furthermore, various transition metal oxides such as TiO₂, ZnO, and NiO have demonstrated resistive switching (RS) characteristics. Although these memristors exhibit stable switching behavior and good scalability, they often suffer from limited analog tunability. In contrast, vanadium pentoxide (V₂O₅), a layered oxide with unique optical, electrical, and structural properties, offers significant potential for achieving gradual, analog-type switching behavior, making it highly promising for neuromorphic computing applications.^32–39 Zhenni Wan et al. reported resistive switching in V₂O₅ with the effect of annealing temperature and abrupt SET/RESET behavior.⁴⁰ However, researchers have not fully explored the tunability of oxygen vacancies and interfacial layers for achieving gradual SET/RESET, which holds significant potential for neuromorphic applications. Gradual resistive switching (GRS) behavior facilitates smooth and predictable variations in conductance, essential for implementing analog weight updates in neuromorphic systems. Nevertheless, achieving dependable GRS in single-layer systems sometimes proves difficult due to constraints in stability and variability. Bilayer memristor architectures have been developed to address this limitation. These architectures employ a regulated interfacial layer to govern the migration of oxygen vacancies and the development of filaments. These bilayer configurations maintain the slow switching characteristics of GRS while enhancing endurance, homogeneity, and analog tunability, all of which are crucial for precise synaptic emulation. Thus, the present study examines V₂O₅-based bilayer memristor devices, including a NbO_x interfacial layer to achieve robust and reproducible GRS behavior suitable for neuromorphic computing.^4,9,33 The diffusion coefficient and mobility of oxygen vacancies in active materials and vacancy-driven filament diameter modulation play a crucial role in synaptic measurements and facilitate analog switching behavior, ensuring stable endurance and long retention.^41–45

Today's world is driven by an unending need for wearable, flexible, implanted sensors, rollable screens, wearable health diagnostic systems, and portable electronics. Thus, researchers focus on developing low-cost, multifunctional, and mechanically stable flexible electronics. Stainless steel, metal foils, and polyimide are examples of flexible substrates that are being thoroughly studied for the development of RRAM devices.^36,46 Flexible metal foils garner significant interest due to their high-temperature applications. Flexible Ni (nickel) foil has shown great promise due to its additional features, including low cost, magnetostrictive property, broad manufacturing area capability, and high fracture strength. However, attributed to the temperature imbalance that arises from interfacial chemical effects, the direct deposition of an active V₂O₅ layer on Ni foil produces an interfacial layer.^8,13 Nb was selected as the top electrode due to its high conductivity and reactivity. This layer enhances the switching performance of the device, making it a promising candidate for RRAM memory applications.

This work reports the development of two flexible RRAM devices, D₁ and D₂, fabricated on nickel substrates via DC magnetron sputtering, designed for neuromorphic computing inspired by biological synaptic behavior. Device D₁ exhibits abrupt SET and gradual RESET transitions, enabling multilevel current–voltage (I–V) characteristics under compliance control. In contrast, device D₂ demonstrates fully gradual resistive switching (GRS) behavior in both SET and RESET states, attributed to introducing an engineered NbO_x interfacial layer. This layer improves switching uniformity and enhances analog modulation, an essential feature for high-fidelity synaptic emulation. It emulates key synaptic behaviours such as LTP, LTD, STDP, and SRDP. These attributes collectively position the V₂O₅-based ReRAM device D₂, exhibiting superior endurance and mechanical flexibility, as a promising candidate for next-generation wearable neuromorphic electronic devices.

2. Experimental section

2.1. Materials and fabrication

The magnetron sputtering technique was used to fabricate the MIM stack of Nb/V₂O₅ and Nb/NbO_x/V₂O₅ over a flexible Ni substrate using a commercial Vanadium, Niobium, sputtering target (5 mm thick and 2-inch diameter).^46–49 The Ni substrate was ultrasonically cleaned thoroughly before thin film deposition. Initially, the V₂O₅ layer was deposited directly onto the flexible Ni substrate under optimized sputtering conditions, followed by the deposition of an ultrathin NbO_x interfacial layer to improve RS performance. The deposition parameters were carefully tuned and summarized in Table 1. After that, the top electrode of Nb was deposited using a shadow mask of diameter 100 µm, and the resulting devices with oxygen vacancies are named D₁ (Nb/V₂O₅/Ni) and D₂ (Nb/NbO_x/V₂O₅/Ni). In addition, the V₂O₅/Ni heterostructure was annealed at 450 °C in a pure argon atmosphere for 35 minutes to create oxygen vacancies. Nb top-electrode was deposited thereafter, and the resulting devices had oxygen vacancies in devices D₁ and D₂.

Table 1 Optimized sputtering parameters of the as-fabricated V₂O₅-based RRAM devices

Sputtering parameters	V2O5 (Switching layer)	Nb	NbOx (Top - electrode)
Substrate temperature	450 °C	RT	550 °C
Target-substrate distance	5 cm	5 cm	5 cm
Pressure (Torr)	2.06 × 10^−6	3.1 × 10^−6	2.2 × 10^−6
Deposition-pressure	7 mTorr	6 mTorr	6 mTorr
Sputtering gas	Ar + O2	Ar	Ar + O2
Sputtering power	130 W	50 W	130 W
Deposition time	3.5 min	40 s	30 s

---

2.2. Formation mechanism of nano-rod like – V₂O₅ structures

During sputtering, Ar⁺ ions transport V and O atoms to the substrate, forming initial nucleation sites for V₂O₅ thin films. In the early stages, the films grow layer-by-layer until a critical thickness is reached. Beyond this point, the V₂O₅ structure transitions into a nanoworm morphology, driven by the intrinsic nature of its 2D layered crystal structure, which is prone to curling and edge dislocations. Localized heating during deposition further contributes by introducing slight curvatures and dislocations that serve as energetically favorable sites for vertical growth. As deposition continues, these curved regions act as nucleation sites, promoting anisotropic growth and resulting in the formation of nano rod-like structures. The final entangled morphology arises from variations in local growth conditions, filament curvature, and strain, leading to the development of the observed nanorod features.

2.3. Material characterizations

The crystalline orientation and phase composition of the V₂O₅/Ni structure were investigated using X-ray diffraction (XRD) with a Cu K_α radiation source (λ = 1.54 Å) at a grazing incidence angle of 2°. Surface morphology and cross-sectional film thickness were characterized using field emission scanning electron microscopy (FESEM), providing high-resolution imaging of the heterostructure architecture. Elemental composition and atomic distribution were determined via energy-dispersive X-ray spectroscopy (EDX), integrated with FESEM, to confirm the presence and stoichiometry of constituent elements. X-ray photoelectron spectroscopy (XPS, PerkinElmer Model 1257) was employed to investigate the chemical states and bonding environments of elements in the active layers. The electrical and synaptic properties of the fabricated D₁(Nb/V₂O₅/Ni) and D₂(Nb/NbO_x/V₂O₅/Ni) structure were measured using a semiconductor characterization system (SCS 4200, Keithley) and Keithley source meter 2400, 2450. The magnetic hysteresis loops were recorded using a physical property measuring system (PPMS, DYNACOOL).

3. Results and discussion

Fig. 1a schematically illustrates the DC magnetron sputtering process for fabricating V₂O₅-based ReRAM devices (D₁ and D₂) on a flexible Nickel (Ni) substrate. The X-ray diffraction (XRD) pattern of Nb₂O₅/V₂O₅/Ni in Fig. 1b reveals diffraction peaks at 2θ = 15.44°, 20.04°, 21.7°, 24.92°, 26.18°, 31.06°, 32.5°, 34.42°, 38.14°, 41.26°, 45.46°, 47.38°, 49.6°, 53.66°, 55.68°, 61.16°, and 62.16° angles. These peaks are attributed to (200), (001), (101), (201), (110), (301), (011), (310), (401), (002), (411), (600), (112), (220), (021), (321), and (710) diffraction planes, confirming the successful synthesis of V₂O₅ with orthorhombic crystal structure (JCPDS Card No: 89-2482).^34,50,51 Additional peaks at 2θ = 22.8°, 28.8°, and 29.3° correspond to the (001), (180), and (200) planes of orthorhombic Nb₂O₅ (JCPDS No. 30-0873). Furthermore, the diffraction peaks observed at 44.18° and 51.3°, corresponding to the (111) and (200) crystallographic planes, respectively, are in good agreement with the standard diffraction pattern of Ni substrate (JCPDS Card No: 01-087-0712).⁵² The alignment of diffraction peaks with established data underscores the structural precision and reliability of the fabrication process.

Fig. 1 (a) Schematic diagram for DC magnetron sputtering and as-fabricated ReRAM devices, namely Ni/V₂O₅/Nb (D₁), and Ni/V₂O₅/NbO_x/Nb (D₂), along with their digital photographs, surface FE-SEM images. (b) XRD pattern of Nb₂O₅/V₂O₅/Ni. (c) XPS full survey spectrum for V₂O₅/Ni. (d and e) XPS high-resolution spectra of O 1s and V 2p regions. (f) XPS depth profile. (g) Raman spectra.

X-ray photoelectron spectroscopy (XPS) was employed to analyze the chemical composition and electronic bonding states of the V₂O₅ thin film. The survey spectra confirmed the presence of vanadium (V), oxygen (O), and carbon (C), as shown in Fig. 1c.⁵³ The C 1s peak at 284.7 eV, attributed to adventitious carbon contamination, was utilized as an internal reference for binding energy calibration. High-resolution XPS spectra of the V 2p region revealed two distinct peaks at binding energies of 524.1 eV and 517.3 eV, corresponding to the V 2p_1/2 and V 2p_3/2 spin–orbit components, respectively (Fig. 1d). Quantitative elemental analysis was performed by integrating the areas under the deconvoluted peaks. Further insight into the oxidation states of vanadium was obtained through deconvolution of the V 2p_3/2 peak, which revealed components at 516.4 eV and 517.4 eV, corresponding to V⁴⁺ and V⁵⁺ oxidation states, respectively. The estimated V⁴⁺/V⁵⁺ ratio of ∼10.2% confirms the existence of oxygen vacancies within the V₂O₅ active layer, which plays a critical role in modulating the resistive switching behavior of the device.⁵⁴ The O 1s core-level spectrum in Fig. 1e exhibited a primary peak at 530.2 eV, indicative of stoichiometric V₂O₅, and an additional peak at 531.6 eV, which is associated with V–O bonding involving non-lattice (dangling) oxygen atoms, signifying oxygen vacancy-related states. Fig. 1f presents the X-ray photoelectron spectroscopy (XPS) depth profile of the Nb/NbO_x/V₂O₅ multilayer structure deposited on a Ni substrate. The depth profile of elemental concentration confirms the existence of a NbO_x layer at the Nb–V₂O₅ interface and formation of an additional NiO_x interfacial layer at the V₂O₅–Ni interface. The comprehensive depth-resolved elemental analysis reveals a well-defined multilayer structure comprising Nb/NbO_x/V₂O₅/NiO sequentially stacked on the flexible Ni substrate. Raman spectra in the 100–1200 cm⁻¹ wave number range were used to analyze the fabricated vanadium pentoxide (V₂O₅), as shown in Fig. 1g. The crystalline and layered structure of V₂O₅ is confirmed by wave numbers 145, 197, 284, 304, 407, 490, 690, and 994 cm⁻¹.⁵⁵ The Raman peaks at 994 and 490 cm⁻¹ represent the doubly coordinated oxygen bond between vanadium and oxygen (V=O) and the stretching mode of V–O₃–V, respectively. The vibration at 145 is due to the VO₅–VO₅ mode. The 284 and 407 cm⁻¹ peaks confirm the bending vibration mode of V=O.^41,42

High-resolution transmission electron microscopy (HRTEM) was systematically performed to elucidate the nanostructural features of the V₂O₅ active layer integrated within the RRAM architecture. HRTEM images (Fig. 2a–c) reveal a nanosheet-like morphology that facilitates multiple ionic conduction paths, promoting robust filament formation. A zoomed-in region of Fig. 2b and c at 20 and 10 nm scale was selected and analyzed using fast Fourier transformation (FFT) to assess the lattice orientation. The resulting FFT pattern (right side of inset Fig. 2e and f) displays well-defined diffraction spots corresponding to the (110) and (010) planes of orthorhombic V₂O₅. Inverse FFT (IFFT) images of the selected region are further generated (left side of inset Fig. 2e and f), highlighting the coherence and periodicity of the atomic planes. HRTEM images in Fig. 2e and f confirm the layered crystalline structure, with interplanar spacings of d₁₁₀ = 0.341 nm and d₂₀₀ = 0.575 nm identified via FFT and inverse FFT analysis. Additionally, the selected area electron diffraction (SAED) pattern (Fig. 2d) shows polycrystalline diffraction rings corresponding to the (200), (110), (301), (310), (411), (112), (021), (412), (710), and (712) lattice planes, further confirming the phase purity and crystallinity of V₂O₅. Energy-dispersive X-ray spectroscopy (EDX) elemental mapping (Fig. 2g–i) reveals a homogeneous spatial distribution of vanadium (V) and oxygen (O) elements across the examined region.

Fig. 2 (a–c) Surface HRTEM image showing porous V₂O₅ nanosheets recorded at different magnifications. (d) SAED pattern. (e and f) HRTEM image showing a zoomed-in region of V₂O₅ nanosheets with resolved planes (Inset shows the FFT and IFFT image). (g–i) EDX elemental mapping of V₂O₅.

Field-emission scanning electron microscopy (FE-SEM) images of the V₂O₅ layer at different magnifications (25 kx, 50 kx, and 100 kx), as shown in Fig. 3a and b, reveal a uniform nanorod-like morphology of the active switching layer. This well-aligned architecture suggests a high degree of surface regularity, which is advantageous for consistent filamentary switching behavior in resistive memory applications. Fig. 3c presents the FESEM cross-sectional image revealing the structural configuration of the device. The energy dispersive X-ray spectroscopy (EDX) spectrum in Fig. 3c confirms the presence of V and O with atomic percentages of 30.82% and 69.18%, respectively, values in close agreement with the ideal stoichiometry of V₂O₅. The uniform elemental distribution of V and O across the active layer is further verified through elemental electron imaging, as shown in Fig. 3d–g. The combined XRD, Raman, XPS, FE-SEM, and HRTEM results confirm the successful synthesis of V₂O₅.

Fig. 3 (a and b) FE-SEM images of V₂O₅ layer illustrating the nanorod-like surface morphology. (c) Shows the FESEM cross-section image of NbO_x/V₂O₅/Ni. (d) EDX analysis and elemental composition of V₂O₅ thin film. (e–g) EDX elemental mapping of vanadium (V) and oxygen (O) elements.

To investigate the role of interfacial engineering in resistive switching characteristics, two MIM-structured RRAM devices (D₁ and D₂) were designed and fabricated using vanadium pentoxide (V₂O₅) as the active functional layer.¹² Both devices employ a nickel (Ni) substrate as the bottom electrode (BE) and niobium (Nb) as the top electrode (TE); however, device D₂ is uniquely engineered with an ultrathin NbO_x interlayer between the V₂O₅ and Nb layers (Fig. 4a). The resulting configurations are denoted as D₁(Nb/V₂O₅/Ni) and D₂(Nb/NbO_x/V₂O₅/Ni). The electrical characterization (Fig. 4b and c) demonstrates a significant disparity in switching behavior between the two devices. Device D₁ exhibits a highly abrupt SET transition followed by a relatively gradual RESET process, demonstrating asymmetric BRS behavior governed by stochastic filament dynamics and electric field. This abrupt switching is often associated with uncontrolled filament formation and rupture, leading to variability in device performance. However, the D₂ demonstrates a continuous, gradual conductance modulation during both SET and RESET characteristics. This gradual switching behavior suggests enhanced control over the migration of oxygen ions and vacancy-mediated filament. The improved switching uniformity and reversibility in D₂ are attributed to the functional role of the inserted NbO_x interfacial layer, which acts as an energy barrier. The NbO_x layer redistributes the local electric field at the V₂O₅ interface, smooths potential gradients, and suppresses abrupt filament growth, facilitating controlled resistive switching.

Fig. 4 (a) Schematic presentation of electrical measurement. (b and c) I–V characteristics of Ni/V₂O₅/Nb (D₁) and Ni/V₂O₅/NbO_x/Nb (D₂) (inset shows an analytical model for SET and RESET of D₁ and D₂).

Structurally, the NbO_x interlayer stabilizes the growth of the conductive filament and promotes improved charge injection symmetry by introducing a customized energy environment through interfacial engineering. This design successfully lowers switching variability, increases device durability, and improves analog tuning capabilities, all of which are critical features for multilevel memory systems and neuromorphic computing applications. Moreover, in both D₁ and D₂ configurations, the Ni bottom electrode chemically interacts with the V₂O₅ active layer during the film deposition, leading to interfacial modifications. XPS depth profiling reveals the emergence of a thin interfacial nickel oxide (NiO_x) layer at the V₂O₅/Ni interface. This spontaneous reaction is driven by interfacial oxygen exchange, resulting in a modified electrode interface that can influence barrier height and interfacial defect chemistry. The NiO_x layer may act as an additional resistive element, further modulating the bottom interface's electrochemical properties and impacting the overall switching kinetics. Dual-interface engineering via intentional NbO_x insertion at the top and spontaneous NiO_x formation at the bottom offers a synergistic pathway for fine-tuning resistive switching behavior in V₂O₅-based RRAM devices. These findings underscore the critical importance of atomic-scale interface control in governing switching reliability, endurance, and performance consistency in emerging memory technologies.

When a positive voltage is applied to the Nb top electrode (TE) while the Ni bottom electrode (BE) is grounded in devices (D₁), Oxygen ions (O²⁻) migrate toward the TE. This ion migration creates oxygen vacancies in the V₂O₅ and NiO_x layers, both acting as active switching layers.⁵³ The oxygen vacancies tend to align and drift downward, forming a conductive filament. This filament formation results in an abrupt transition from a high-resistance state (HRS) to a low-resistance state (LRS), representing the SET process as shown in Fig. 4b and supported by an analytical model that explains the SET process (inset of Fig. 4b). Conversely, applying a negative voltage to the TE causes oxygen vacancies in D₁ to migrate away from the filament (Fig. 4b). As a result, the device gradually transferred from LRS to HRS, corresponding to the RESET process, as illustrated in the inset of Fig. 4b. In contrast, Device D₂ exhibits gradual SET and RESET switching when positive and negative voltage is applied to TE. This behavior is described as the introduction of an additional interfacial NbO_x layer. The observed difference arises from variations in Gibbs’ free energy, activation energy, and diffusion coefficients of NbO_x, NiO, and V₂O₅, which govern the mobility of the oxygen vacancies, leading to non-uniform filament formation.^28,56 Upon the application of the positive bias, oxygen vacancies accumulate asymmetrically near the conductive filament, resulting in an increase in filament diameter, as depicted in the inset of Fig. 4b. Instead of undergoing abrupt rupture, the filament experiences partial and gradual dissolution, initiated by the outward migration of oxygen vacancies. This controlled degradation reduces filament conductivity progressively, thereby facilitating a smooth transition from a low-resistance state (LRS) to a high-resistance state (HRS), as illustrated in the inset of Fig. 4c. Furthermore, device D₂ exhibits lower conductivity and a higher ON/OFF ratio due to an additional interfacial layer NbO_x, which influences charge transport and resistive switching behavior.

Furthermore, Fig. 5a and b illustrates the manifestation of the multilevel switching (MLS) behavior as a direct function of varying compliance currents during the RS process. This behavior stems from the dynamic modulation of the filamentary conduction pathway within the switching dielectric layer, governed by the magnitude of the compliance current. A higher compliance current induces a more pronounced soft breakdown in the dielectric medium, promoting enhanced ion migration and accumulation, leading to a more robust and wider conductive filament. This increase in filament diameter reduces the effective resistance in the low-resistance state (LRS) due to the enlarged conductive cross-section. Moreover, the rise in filament density reduces charge carrier hopping distances, facilitating more efficient charge transport across the filament. Consequently, the compliance current modulates the LRS and high-resistance state (HRS) levels, reflecting the filament's morphological evolution and its impact on the overall resistance states. The endurance and retention characteristics of devices D₁ and D₂ under varying compliance currents are presented in the inset of Fig. 5a and b, respectively. These plots highlight the multilevel endurance behavior, demonstrating the stability and reproducibility of discrete resistance states over repeated switching cycles. Notably, the device D₁ exhibits endurance of 7000 cycles and retention of 6000 s, whereas D₂ demonstrates endurance of 5000 cycles and retains data for up to 4500 s. Fig. 5g illustrates the re-plot of flexible RRAM devices D₂ having I–V characteristics on a double logarithmic (log I–log V) scale to provide a deeper understanding of the underlying conduction mechanism. The analysis reveals that the conduction process is primarily governed by the space charge-limited conduction (SCLC) mechanism, which implies that electron trapping and detrapping in oxygen-related traps/defects are intimately connected to the conduction process. In LRS, the slope of the log I–log V curve is approximately 1.02 across the entire voltage range, indicating the Ohmic behavior (I ∝ V), a characteristic of thermally generated free carrier conduction.¹² In the high-resistance state (HRS), the device exhibits three distinct conduction regions. In the low-voltage region (0 < V < 0.15 V), the slope of 0.902 indicates Ohmic conduction, suggesting that the current is limited by thermally generated free carriers. As the voltage increases, the conduction mechanism transitions to Schottky emission in the intermediate region (0.15 < V < 0.35 V), which is confirmed by the linear behavior observed in the ln(I) versus √V graph plotted in Fig. 5h (The detailed analysis is provided in the SI).^23,57 At higher voltages, the slope increases to approximately 4.55, indicating a transition from the trap-filled limit (TFL) to the Child's law region of space-charge-limited conduction (SCLC), where the current follows the relation I ∝ Vⁿ (n > 2). These progressively increasing slope values reflect a gradual trap-filling process, in which charge transport becomes increasingly constrained by the availability of free carriers and trap states. Therefore, the dominant conduction mechanism in device D₂, particularly in the HRS regime, is attributed to trap-controlled SCLC, governed by carrier injection and transport through defect states within the switching layer. Furthermore, the combined effects of temperature and compliance current (CC) on the I–V characteristics of devices D₁ and D₂ were investigated, as shown in Fig. 5c and d. Fig. 5e and f illustrate the temperature-dependent behavior alone, where an increase in conductivity with rising temperature was observed. This phenomenon can be attributed to the enhanced generation and mobility of oxygen vacancies, which facilitate charge transport within the switching layer.

Fig. 5 (a and b) Multilevel I–V behavior in devices D₁ and D₂ under compliance current (Inset shows the endurance and retention characteristics of D₁ and D₂). (c and d) combined effect of Temperature-dependent and CC I–V response in D₁ and D₂ (inset shows device-to-device measurement of D₁ and D₂). (e and f) Temperature dependent I–V response with constant CC (g) log[I] versus log[V] curve for D₂. (h) (b) Schottky conduction of the HRS (high-voltage region of negative bias data) by fitting. (i) Memory window versus temperature response of D₁ and D₂.

The device-to-device variability measurements for devices D₁ and D₂ are presented in the inset of Fig. 5c and d, respectively. The device D₂ demonstrated superior stability, attributed to the valence change mechanism (VCM) and an additional interfacial NbO_x layer. This interfacial layer plays a crucial role in enhancing the RS mechanism by stabilizing the diameter of conductive filaments, thereby improving the overall reliability and performance of the device. In devices D₁ and D₂, the memory window initially increases with temperature, reaching a maximum at 335 K. This enhancement is attributed to the thermally activated generation of oxygen vacancies, which facilitates the formation of stable conductive filaments, thereby improving the ON/OFF ratio. Within the 300–335 K range, the gradual increase in oxygen vacancy concentration leads to a wider memory window. However, beyond 335 K, particularly between 335–350 K, the memory window decreases due to excessive vacancy accumulation. This surplus disrupts the filament formation dynamics, leading to instability in resistive switching and a decline in performance (Fig. 5i). Thus, while oxygen vacancies initially enhance switching behavior, their overaccumulation at higher temperatures negatively impacts device reliability.

Fig. 6a illustrates the key components of a biological synapse in the human brain, including dendrites, axons, pre- and post-synaptic neurons, the synaptic cleft, and neurotransmitters. In the artificial device, the top and bottom electrodes correspond to pre- and post-synaptic neurons, while electrons, oxygen ions, and vacancies function analogously to neurotransmitters. The device's electrical conductivity variations, representing weak and strong conductive states, correlate with the modulation of synaptic weight (strengthening and weakening) in biological memory systems (BMS).^9,13 Fig. 6b, the sequential application of five positive voltage pulses to the top electrode of device D₂ leads to a gradual increase in current. In comparison, five negative voltage pulses decrease the previously incremented current. This current and voltage versus time response is related to synaptic functions such as potentiation and depression of synaptic weight.

Fig. 6 (a) Visual representation of synaptic activity between the pre-synaptic and post-synaptic neurons and ionic memristor interface on flexible Ni substrate. (b) I–V characteristics of multiple consecutive voltage sweeps of device D₂. (c and d) LTP and LTD cycles of D₂ and D₁. (e) Comparison of LTP and LTD for D₁ and D₂. (f and g) LTP and LTD at different amplitudes and pulse widths in D₁ and D₂.

To investigate electronic synaptic behavior, 40 identical positive (1.1 V, 1 ms) and negative (−1.1 V, 1 ms) voltage pulses were applied to devices exhibiting gradual (D₁ and D₂) switching. The LTP and LTD cycles of devices D1 and D2 are shown in Fig. 6c and d. D2 achieves higher LTP/LTD cycle counts than D1, highlighting the enhanced filament control and repeatability enabled by the interfacial engineering. Fig. 6e shows conductance modulation in D₁ and D₂, where gradual and linear LTP/LTD occurs due to defects, vacancies, ion migration, and the interfacial layer, which is beneficial for neuromorphic applications. The conductivity of D₂ is lower than that of D₁ due to the interfacial layer NbO_x.⁵⁸ Fig. 6f and g shows the LTP and LTD measurements for device D₂ under different pulse width trains and amplitudes, leading to more efficient synaptic weight control and indicating synaptic amplitude-dependent plasticity.

After successfully emulating a biological synapse, we investigated a fundamental synaptic learning mechanism, paired-pulse facilitation (PPF). This process involves the modulation of synaptic conductance by the cumulative effect of consecutive input pulses and stimuli, thereby enabling the memristive device to mimic short-term neuromorphic functionalities such as synaptic adaptation and temporal filtering. For the PPF measurements, two consecutive voltage pulses of 3V amplitude and 10 ms duration were applied, with the inter-pulse interval (Δt) varied from 0.1 to 15 ms (Fig. 7a). A significant increase in output current was observed in response to the second stimulus relative to the first, indicating facilitation. This enhancement occurs because the synaptic response from the first pulse does not fully decay before the arrival of the second pulse. When the second stimulus arrives within this short recovery window, it results in an elevated conductance response analogous to biological synaptic behavior. The observed increase in current reflects a temporary modulation of the device's conductance induced by the applied pulse sequence. This behavior is attributed to the migration of oxygen vacancies (V_o) towards the conductive filament region, even after removing the potentiating pulse.^9,58 When the interval between two successive pulses is shorter than the relaxation time of the oxygen vacancies, the residual ionic redistribution from the first pulse enhances the effect of the second, leading to a higher conductance state. Conversely, when the inter-pulse interval exceeds the relaxation time, the vacancies return to their equilibrium positions, and the facilitation effect saturates or diminishes. The PPF can be calculated as

where A_f and A_i are the post-synaptic current readouts after the first and second pulses, respectively.

Fig. 7 (a) Schematic illustration of the PPF rule of bio-synapse between pre- and post-synaptic neurons and diameter modulation of the artificial device. (b) PPF ratio measurement for device D₂ using a pulse of 3 V and 10 ms. (c and d) STDP and SRDP for device D₂.

Fig. 7b illustrates that the PPF value progressively decreases as the interval between consecutive pulses increases. At shorter intervals, the diameter of the conductive filament formed by the first pulse partially persists, resulting in an enhanced response to the second pulse and a larger change in current. In contrast, the filament partially dissolves at longer intervals before the second pulse arrives, leading to a minor increase in filament diameter and a lower current response, thereby reducing the PPF ratio value. The extracted fitting parameters are c₁ = 26, c₂ = 103.5, τ1 = 1.519 ms, and τ2 = 468.69 ms, here, τ1 corresponds to the fast relaxation process, typically associated with short-term synaptic response dynamics, while τ2 represents the slower relaxation process, linked to longer-lasting synaptic modulation. The obtained values are consistent with the characteristic time scales reported in memristive synapses, further validating the reliability of our measurements. The increasing and decreasing characteristics of PPF with time intervals closely mimic biological synaptic behavior.

To study the SRDP and STDP, which are fundamental associative learning rules governed by Hebbian plasticity, it is crucial to analyse the role of frequency-dependent and temporally structured input stimuli in modulating synaptic weight. In the case of SRDP, synaptic weight adaptation can be achieved by varying the average firing rate of input pulses. Conversely, the relative timing between pre- and post-synaptic spikes determines STDP-induced synaptic weight modulation. In this experiment, SRDP has been simulated by varying the time interval between voltage write pulses (+1.5 V) and erase pulses (−2.0 V), both applied with a pulse width (PW) of 1 ms. The read pulses were delayed by 1 ms relative to the write/erase pulse train. Fig. 7c displays the changes in synaptic weight (Δω) as a function of the time interval (Δt), calculated using the formula.

where K_f = current value at final V_read, K_i = Current value at initial V_read.

The weight changes for both LTP and LTD were well-fitted with the equation:

where, τ is the time constant for LTP and LTD in SRDP.

For STDP, similar experiments have been measured utilizing an adjustment in pulse mode as shown in Fig. 7d, which is explained by the relationship between time interval (Δt) and synaptic weight (Δω). The gap in time between them is described as

Δt = t_pre − t_post (4)

where t_pre and t_post were the pulse times that came to the post-synaptic and pre-synaptic electrodes. A pair of pulses was applied to the pre-synaptic electrode (1.1 V, 1 ms) and post-synaptic electrode (−1.1 V, 1 ms), and initial and final V_read (0.5 V, PW = 1 ms). The synaptic weight connection increases, indicating the relationship between the two synapse connections is potentiated. However, synaptic weight decreases, indicating that the connection between two synapses is weakened.

In addition, the Modified National Institute of Standards and Technology (MNIST) handwritten digits dataset is used for the simulation. The simulation is performed using MATLAB and a three-layer network (one hidden layer) with 400 input neurons, corresponding to the input black and white images with 20 × 20 pixels encoded from the MNIST dataset, and 100 hidden neurons. The pattern recognition results (0–9) are presented through the 10 output neurons (Fig. 8a). The inner product between the input neuron signals and the corresponding synaptic weights obtained from experimental measurements is computed by the first synapse array and transmitted to the hidden neuron layer. A similar computation is performed between the hidden and output layers via the second synapse array, completing one training epoch. During each epoch, the artificial neural network (ANN) is trained using 60 000 randomly selected images from the MNIST training dataset, and its recognition accuracy is evaluated on a separate test subset (detailed analysis in SI).⁹ Fig. 8b shows the device structure of the device D₂. Our synaptic device achieves a validation accuracy of 95.5%, closely approaching the ideal software-based implementation, which attains a validation accuracy of 93.5%, in the pattern recognition task (Fig. 8c and d).

Fig. 8 (a) Illustration of a hardware-based neural network configuration employing memristive synapses for real-time recognition of handwritten digits, mimicking brain-inspired computation for efficient pattern classification. (b) Schematic of the D₂(Nb/NbO_x/V₂O₅/Ni). (c) Accuracy versus epoch plot showing the learning progression during image recognition. (d) Comparison of the accuracy of the ideal and D₂ device.

The schematic illustration of the flexible V₂O₅-based ReRAM devices architecture highlights its mechanical adaptability (Fig. 9a). The device's mechanical flexibility has been tested in three bending modes: tensile, compressive, and no bending, which shows no change in I–V characteristics, as well as LTP and LTD (Fig. 9b and c). As presented in Fig. 9d, the device exhibited one bending mode that maintains stable performance around 1300 cycles between LRS and HRS, thus confirming the futuristic memory application of flexible RRAM. To further investigate angular stability, the device was bent at multiple angles (30° → 45° → 60° → 90° → 120° → 150°), with no observable degradation in either the LRS or HRS, as presented in Fig. 9e. The measurement setup used for this angular flexibility assessment is schematically depicted in Fig. 9f. Furthermore, cyclic bending tests were performed to probe under mechanical stress. The performance metrics summarized in Table 2 confirm that the Nb/NbO_x/V₂O₅/Ni-based device achieves reliable resistive switching along with outstanding mechanical resilience and data stability, making it a strong candidate for flexible neuromorphic applications.

Fig. 9 (a) Schematics of the present flexible RRAM device. (b) Gradual characteristics of heterostructure (D₂). (c) LTP and LTD of RRAM devices (D₂). (d) one bending cycle. (e) Endurance characteristics at various bending angles (D₂). (f) Digital photographs showing the device at different bending angles.

Table 2 Comparison of the present active switching layer with the previously reported data for RRAM applications

Active layer	OFF/ON Ratio	Voltage (V)	Flexibility	Endurance	Synaptic measurement	SET/RESET	Ref.
VO2	60	−6 to 6	No	100	No	Both abrupt	59
VOx	>100	−0.4 to 0.4	No	10	No	Both abrupt	60
SrVOx	>100	−1 to 2	No	—	No	Abrupt/Gradual	61
V2O5	350	−1 to 1	Yes	500	No	Both abrupt	55
V2O5/WO3	—	−8 to 8	No	—	Yes	Both Gradual	62
NbOx/V2O5	5 × 10^2	−1.5 to 1.5	Yes	1300	Yes	Both Gradual	This work

---

4. Conclusions

In conclusion, this research demonstrates the successful realization of bipolar resistive switching and synaptic functionalities in V₂O₅-based flexible Re-RAM devices, namely, D₁(Nb/V₂O₅/Ni) and D₂(Nb/NbO_x/V₂O₅/Ni) structures, fabricated via DC magnetron sputtering. Device D₁ exhibits a sharp SET and a gradual RESET behavior. In contrast, device D₂ demonstrates fully gradual SET and RESET transitions, enabling multilevel RS with an enhanced memory window. The superior performance of D₂ is attributed to incorporating an NbO_x interfacial layer, which modulates oxygen vacancy dynamics and stabilizes filamentary conduction. Moreover, D₂ effectively mimics essential synaptic functionality, including LTP, LTD, SRDP, and STDP, making it highly promising for neuromorphic computing. The conduction mechanism follows a combination of Ohmic and SCLC processes, consistent with the behavior of oxygen vacancy-mediated switching. Mechanical flexibility is validated through stable performance over 1300 bending cycles, confirming excellent structural integrity and endurance. These results highlight the potential of V₂O₅-based flexible memory devices as scalable, wearable, and energy-efficient electronics for high-performance neuromorphic computing applications.

Conflicts of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.

The data supporting this article have been included as part of the supplementary information (SI). Supplementary information: provides additional details on the Artificial Neural Network (ANN) simulation, the corresponding confusion matrix, and discussions related to neuromorphic computing. It also includes an analysis of device-to-device stability and the fitting of the I–V characteristics based on the Richardson–Schottky equation. See DOI: https://doi.org/10.1039/d5nh00562k.

Acknowledgements

We appreciate the financial assistance given by the Defense Research and Development Organization (DRDO) under EP & IPR 2025 with reference no. ERIP/ER/202311006/M/01/1852. We also appreciate the financial assistance given by the Science and Engineering Research Board (SERB), Department of Science and Technology (DST), India, with reference No. CRG/2020/005265. Davinder Kaur sincerely acknowledges the Shastri Indo-Canadian Institute, India, for awarding the SMP fellowship. Kumar Kaushlendra sincerely acknowledges MHRD for the Senior Research Fellowship (SRF) award.

References

1. M. Awais, N. Othman, M. D. Shafiq, F. Zhao and K. Y. Cheong, Small, 2025, 2504260.
2. R. Sun, Z. Hou, Q. Chen, B. Zhu, C. Zhu, P. Huang, Z. Hu, L. Zhen, F. Zhou, C. Xu and J. Qin, Adv. Mater., 2025, 37, 2409017.
3. H. Na and S. Kim, Adv. Funct. Mater., 2025, 2500956.
4. H. Sun, H. Tian, Y. Hu, Y. Cui, X. Chen, M. Xu, X. Wang and T. Zhou, Adv. Sci., 2024, 11, 2406242.
5. G. Kim, D. Yoo, H. So, S. Park, S. Kim, M. J. Choi and S. Kim, Mater. Horiz., 2024, 12, 915–925.
6. H. Paik, J. Lim, H. Seo, T. K. Kim, J. Shin, H. Song, D. G. Kim, W. Lee, D. S. Kwon and C. S. Hwang, Mater. Horiz., 2025, 12, 7305–7317.
7. S.-M. Chen, L.-C. Shih, J.-C. Gao, S.-X. You, K.-T. Chen, P.-L. Lin, K.-S. Hsu, C.-C. Chen, W.-L. Chen and J.-S. Chen, Nanoscale Horiz., 2025, 10, 2068–2079.
8. S. Shankhdhar, D. Arora, H. Rahman, R. Kumar, B. Ranjan, K. Kaushlendra and D. Kaur, Nano Energy, 2025, 133, 110441.
9. A. K. Jena, M. C. Sahu, K. U. Mohanan, S. K. Mallik, S. Sahoo, G. K. Pradhan and S. Sahoo, ACS Appl. Mater. Interfaces, 2023, 15, 3574–3585.
10. R. Khan, N. U. Rehman, R. Thangappan, A. Saritha and S. Sangaraju, Nanoscale, 2025, 17, 11152–11190.
11. S. M. Chen, L. C. Shih, J. C. Gao, S. X. You, K. T. Chen, P. L. Lin, K. S. Hsu, C. C. Chen, W. L. Chen and J. S. Chen, Nanoscale Horiz., 2025, 10, 2068–2079.
12. A. Kumar, S. Pawar, S. Sharma and D. Kaur, Appl. Phys. Lett., 2018, 112, 262106.
13. K. Kaushlendra and D. Kaur, ACS Appl. Electron. Mater., 2023, 5, 3362–3372.
14. G. Hu, Z. Yu, H. Qu, Y. Yuan, D. Li, M. Zhu, J. Guo, C. Xia, X. Wang, B. Wang, G. Ma, H. Wang and W. Dong, Appl. Phys. Lett., 2024, 124, 142109.
15. X. Chen, X. Wang, Y. Pang, G. Bao, J. Jiang, P. Yang, Y. Chen, T. Rao and W. Liao, Small Methods, 2023, 7, 2201156.
16. Y. F. Jiang, H. Y. Peng, Y. Cai, Y. T. Xu, M. Y. Fu, M. Feng, B. W. Wang, Y. Q. Wang, Z. Guan, B. Bin Chen, N. Zhong, C. G. Duan and P. H. Xiang, Mater. Horiz., 2025, 12, 6928–6937.
17. Y. Wang, H. Liu, W. Wu, G. Yin, M. Zhang and J. Su, J. Mater. Chem. C, 2025, 13235–13245.
18. X. Liu, J. Chen, Y. Wang, S. Han and Y. Zhou, Adv. Funct. Mater., 2021, 31, 2004733.
19. P. P. Patil, S. S. Kundale, S. V. Patil, S. S. Sutar, J. Bae, S. J. Kadam, K. V. More, P. B. Patil, R. K. Kamat, S. Lee and T. D. Dongale, Small, 2023, 19, 2303862.
20. Z. Cao, B. Sun, G. Zhou, S. Mao, S. Zhu, J. Zhang, C. Ke, Y. Zhao and J. Shao, Nanoscale Horiz., 2023, 8, 716–745.
21. T. Guo, K. Pan, Y. Jiao, B. Sun, C. Du, J. P. Mills, Z. Chen, X. Zhao, L. Wei, Y. N. Zhou and Y. A. Wu, Nanoscale Horiz., 2022, 7, 299–310.
22. M. Ismail, U. Chand, C. Mahata, J. Nebhen and S. Kim, J. Mater. Sci. Technol., 2022, 96, 94–102.
23. R. Deb, M. G. Nair, U. Das and S. R. Mohapatra, J. Mater. Chem. C, 2023, 11, 7629–7640.
24. R. Islam, H. Li, P.-Y. Chen, W. Wan, H.-Y. Chen, B. Gao, H. Wu, S. Yu, K. Saraswat and H.-S. Philip Wong, J. Phys. D: Appl. Phys., 2019, 52, 113001.
25. K. Chatterjee, J. Stanley, V. P. Dravid, R. dos Reis and S. E. Skrabalak, Appl. Phys. Lett., 2023, 123, 154101.
26. Y. Zhu, M. Chen, H. Lu, P. Mi, D. Luo, Y. Wang, Y. Liu, R. Xiong and H. Wang, Appl. Phys. Lett., 2024, 124, 063504.
27. Z. Wang, M. Yin, T. Zhang, Y. Cai, Y. Wang, Y. Yang and R. Huang, Nanoscale, 2016, 8, 14015–14022.
28. M. Ismail, U. Chand, C. Mahata, J. Nebhen and S. Kim, J. Mater. Sci. Technol., 2022, 96, 94–102.
29. Z. Wang, S. Joshi, S. E. Savel’ev, H. Jiang, R. Midya, P. Lin, M. Hu, N. Ge, J. P. Strachan, Z. Li, Q. Wu, M. Barnell, G. L. Li, H. L. Xin, R. S. Williams, Q. Xia and J. J. Yang, Nat. Mater., 2017, 16, 101–108.
30. J. H. Ryu and S. Kim, Chaos, Solitons Fractals, 2020, 140, 110236.
31. T. Wan, B. Qu, H. Du, X. Lin, Q. Lin, D. W. Wang, C. Cazorla, S. Li, S. Liu and D. Chu, J. Colloid Interface Sci., 2018, 512, 767–774.
32. K. Gao, B. Sun, B. Yang, Z. Cao and Y. Cui, Appl. Mater. Today, 2025, 43, 102628.
33. Y. Qiao, F. Wang, W. Guo, Z. He, L. Yao, J. Li, N. Sun, Y. Wang and F. Wang, J. Alloys Compd., 2023, 967, 171712.
34. B. Ranjan and D. Kaur, ACS Appl. Mater. Interfaces, 2024, 16, 14890–14901.
35. S. Yin, C. Song, Y. Sun, L. Qiao, B. Wang, Y. Sun, K. Liu, F. Pan and X. Zhang, ACS Appl. Mater. Interfaces, 2019, 11, 43344–43350.
36. S. Sharma, K. Kumar, K. Kaushlendra and D. Kaur, ACS Appl. Electron. Mater., 2024, 6, 875–886.
37. V. R. Nallagatla, T. Heisig, C. Baeumer, V. Feyer, M. Jugovac, G. Zamborlini, C. M. Schneider, R. Waser, M. Kim, C. U. Jung and R. Dittmann, Adv. Mater., 2019, 31, 1903391.
38. K. Kaushlendra and D. Kaur, ACS Appl. Electron. Mater., 2024, 6, 4548–4559.
39. Q. Tian, X. Shan, J. Bian, Y. Cheng and J. Zheng, J. Mater. Chem. C, 2025, 12046–12065.
40. Z. Wan, H. Mohammad, Y. Zhao, C. Yu, R. B. Darling and M. P. Anantram, J. Appl. Phys., 2018, 124, 105301.
41. L. Gao, Q. Liao, X. Zhang, X. Liu, L. Gu, B. Liu, J. Du, Y. Ou, J. Xiao, Z. Kang, Z. Zhang and Y. Zhang, Adv. Mater., 2020, 32, 1906646.
42. X. Zhang, Q. Liao, S. Liu, Z. Kang, Z. Zhang, J. Du, F. Li, S. Zhang, J. Xiao, B. Liu, Y. Ou, X. Liu, L. Gu and Y. Zhang, Nat. Commun., 2017, 8, 1–8.
43. R. Yang, Chin. Phys. B, 2020, 29, 097305.
44. R. Ge, X. Wu, L. Liang, S. M. Hus, Y. Gu, E. Okogbue, H. Chou, J. Shi, Y. Zhang, S. K. Banerjee, Y. Jung, J. C. Lee and D. Akinwande, Adv. Mater., 2021, 33, 2007792.
45. G. Zhang, Y. Xu, S. Yang, S. Ren, Y. Jiao, Y. Wang, X. Ma, H. Li, W. Hao, C. He, X. Liu and J. Zhao, Nano Energy, 2023, 106, 108074.
46. B. Ranjan and D. Kaur, Small, 2024, 20, 2307723.
47. R. Kumar, B. Ranjan and D. Kaur, J. Energy Storage, 2024, 84, 110866.
48. H. Rahman, G. K. Sharma, P. Singh and D. Kaur, ACS Appl. Electron. Mater., 2024, 6, 2400–2412.
49. R. Kumar, B. Ranjan, K. Kumar, S. Shankhdhar and D. Kaur, ACS Appl. Nano Mater., 2024, 7, 7663–7673.
50. K. Kaushlendra, P. Kumar, D. Arora, B. Ranjan and D. Kaur, Ceram. Int., 2023, 49, 9543–9550.
51. H. Raha, N. Dey and P. Kumar, J. Energy Storage, 2025, 128, 117177.
52. B. Ranjan and D. Kaur, J. Energy Storage, 2023, 71, 108122.
53. B. Sharmila, P. Divyashree and P. Dwivedi, Sens. Actuators, A, 2024, 367, 114979.
54. S. G. Choi, H. J. Seok, S. Rhee, D. Hahm, W. K. Bae and H. K. Kim, J. Alloys Compd., 2021, 878, 160303.
55. B. P. Yalagala, P. Sahatiya, C. S. R. Kolli, S. Khandelwal, V. Mattela and S. Badhulika, ACS Appl. Nano Mater., 2019, 2, 937–947.
56. K.-C. Chuang, C.-Y. Chu, H.-X. Zhang, J.-D. Luo, W.-S. Li, Y.-S. Li and H.-C. Cheng, IEEE J. Electron Devices Soc., 2019, 7, 589–595.
57. S. Zhu, B. Sun, G. Zhou, T. Guo, C. Ke, Y. Chen, F. Yang, Y. Zhang, J. Shao and Y. Zhao, ACS Appl. Mater. Interfaces, 2023, 15, 5420–5431.
58. S. K. Mallik, R. Padhan, M. C. Sahu, S. Roy, G. K. Pradhan, P. K. Sahoo, S. P. Dash and S. Sahoo, ACS Appl. Mater. Interfaces, 2023, 15, 36527–36538.
59. J. Li, X. Zhou, L. Xu, J. Wang, B. Wu and C. Wang, Mater. Res. Express, 2022, 9, 035003.
60. J. Ryu, K. Park, D. P. Sahu and T. S. Yoon, ACS Appl. Mater. Interfaces, 2024, 16, 26450–26459.
61. T.-J. Lee, S.-K. Kim and T.-Y. Seong, Sci. Rep., 2020, 10, 5761.
62. B. Sharmila and P. Dwivedi, ACS Appl. Electron. Mater., 2025, 7, 3191–3200.

---