Solution-processed SnO₂/SnS₂ bilayer-based robust memristors for reliable neuromorphic computing

Abstract

The development of scalable, low-power, and high-density resistive memory devices is crucial for next-generation computing architectures, particularly in neuromorphic applications. Here, we report solution-processed SnO₂/SnS₂ bilayer thin films as functional layers for memristors and synaptic devices. The incorporation of the SnO₂ layer enables the formation of oxygen-vacancy conductive filaments that act as virtual electrodes, which effectively guide the nucleation and rupture of sulfur-vacancy filaments in the two-dimensional (2D) SnS₂ layer. This synergistic mechanism significantly enhances resistive switching performance, yielding an ON/OFF ratio exceeding 200, stable endurance over 10⁴ cycles, and robust retention. Beyond conventional memory behavior, the bilayer devices emulate essential synaptic functions, including excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), and spike-timing dependent plasticity (STDP), and achieve ∼93% inference accuracy in artificial neural network tasks.

---

New concepts

Two-dimensional sulfides hold great promise for neuromorphic hardware, yet their application in vertical memristor architectures is fundamentally limited by sluggish ion kinetics and the difficulty of forming uniform sulfur-vacancy filaments. Here, we introduce a conceptually new bilayer strategy in which an oxide layer with easily generated oxygen-vacancy filaments acts as a virtual electrode to guide the ordered formation of sulfur-vacancy pathways in an overlying SnS₂ layer. This vertically coupled SnO₂/SnS₂ system enables synergistic, composite conductive channels that overcome the intrinsic transport bottlenecks of 2D sulfides, yielding highly stable switching with excellent endurance and a large ON/OFF ratio. Beyond electrical performance, the emergent filament coupling provides analog, brain-like plasticity that supports efficient hardware learning. Demonstrated by achieving 93% accuracy on MNIST, this work establishes a new paradigm for vacancy-engineered bilayers, offering a simple, solution-processed, and CMOS-compatible route toward scalable neuromorphic processors.

Introduction

With the advent of the big data era, brain-inspired parallel computing and complex information processing demand storage technologies that combine high density with low power consumption, enabling massive data handling at the nanoscale. Memristors, as emerging electronic devices capable of emulating biological synaptic functions while providing nonvolatile memory, have attracted considerable attention as promising candidates for next-generation memory architectures and neuromorphic systems.^1–4 Their advantages, such as ultra-high integration density, low energy consumption, fast switching speed, and strong compatibility with CMOS technology, make them particularly appealing for future computing architectures.^5–7 Structurally, a conventional memristor features a simple metal–insulator–metal (MIM) configuration, with a functional layer sandwiched between two electrodes. Despite this apparent simplicity, the resistive switching mechanism is complex and highly dependent on the properties of the functional layer, which govern switching dynamics, device reliability, and overall performance. Recent advances have significantly expanded the range of functional layer materials, including transition metal oxides, perovskites, 2D sulfides, and conductive polymers. Each class enables distinct switching mechanisms, such as ion migration, phase transitions, or charge trapping/detrapping, thereby providing versatile platforms for performance optimization and functional diversification.^8–11

Notably, recent studies have shown that rational materials engineering and device-architecture design can markedly improve memristive performance. For instance, Ga-doped Ge₂Sb₂Te₅ phase-change memristors enable controllable switching and improved neuromorphic accuracy through compositional tuning, while core–shell Ag nanowire memristors achieve quasi-2D filament confinement, leading to ultralow power consumption and enhanced uniformity.^12,13 These advances underscore the critical role of defect engineering and nanoscale architectures in regulating filament dynamics. However, many of these high-performance devices rely on complicated material synthesis and complex fabrication processes, limiting their cost-effective and scalable implementation. In this context, solution processing has emerged as a promising route for thin-film memristor fabrication, offering clear advantages in simplicity, scalability, and compatibility with large-area electronics.^14–16 More importantly, it allows flexible control over device architectures and defect states, providing a practical pathway toward bridging the gap between high-performance memristors and manufacturable neuromorphic hardware.

Among functional materials, 2D sulfides such as MoS₂, WS₂, and SnS₂ have attracted increasing attention for memristor and neuromorphic applications owing to their layered structures and favorable physicochemical properties.^17–21 Their atomically thin nature offers high surface area, mechanical flexibility, and quantum confinement effects, enabling device miniaturization and performance enhancement.^22,23 However, their strong environmental sensitivity often limits stability and reproducibility. In addition, weak interlayer van der Waals interactions hinder vertical ion transport, resulting in sluggish sulfur-ion kinetics and difficulty in forming uniform sulfur-vacancy filaments.²⁴ As a result, most 2D sulfide-based memristors adopt lateral device configurations, which partially alleviate transport limitations but suffer from low integration density and poor filament controllability.²⁵ Developing vertical architectures is therefore essential for improving integration and achieving more reliable switching behavior.²⁶ At present, 2D sulfides are mainly used as dielectric components in hybrid devices, where their properties are modulated through interactions with other functional layers, while simple and scalable vertical structures remain challenging to realize.^27,28

To further enhance resistive switching performance, various strategies have been explored, among which the use of virtual electrodes has proven particularly effective. By redistributing the electric field and ion-migration pathways, virtual electrodes enable more uniform filament formation and controllable rupture, thereby improving device-to-device consistency and cycling stability.^29,30 Metal oxides such as TaO_x, HfO₂, and TiO₂ are widely employed owing to their abundant oxygen vacancies and high oxygen-ion mobility, with resistive switching governed mainly by oxygen-ion migration and vacancy evolution.³¹ Typically, filament formation requires lower energy than rupture, as reflected by V_set < V_reset.^32,33 Under appropriate operating conditions, oxygen-vacancy filaments can therefore act as effective virtual electrodes to guide sulfur-vacancy filament formation in oxide/2D sulfide bilayer systems. Moreover, suitable deposition methods allow controlled modulation of vacancy states, offering a practical route to further improving device performance.^34,35

Guided by these considerations, we fabricated bilayer SnO₂/SnS₂ films via a spin-coating process and applied them to memristor and synaptic emulation devices. In this structure, oxygen-vacancy filaments in the SnO₂ layer act as “seeds” to induce the ordered growth of sulfur-vacancy filaments in the SnS₂ layer, thereby forming composite conductive channels with synergistic effects. The bilayer devices exhibit excellent resistive switching characteristics, with an ON/OFF ratio of up to 10² and endurance exceeding 10⁴ cycles, demonstrating both stability and reliability. Furthermore, they successfully emulate synaptic plasticity behaviours such as long-term potentiation (LTP) and paired-pulse facilitation (PPF). When evaluated on the MNIST dataset for supervised learning in handwriting recognition, the device-based neural network achieved an accuracy of 93%, confirming the strong potential of this system for neuromorphic computing and pattern recognition.

Results and discussion

Structure and component of memristor devices

Three types of memristor devices were fabricated using a liquid-phase spin-coating method: (i) ITO/SnO₂/ITO (oxide-based memristor, OM), (ii) ITO/SnS₂/ITO (sulfide-based memristor, SM), and (iii) ITO/SnO₂/SnS₂/ITO (oxide/sulfide bilayer memristor, OSM), as illustrated in Scheme S1 of the SI. Surface morphology characterization by SEM (Fig. S1, SI) confirmed that all devices formed uniform, dense, and continuous films on the ITO substrate, with consistent microstructures providing a solid foundation for subsequent electrical studies. Furthermore, AFM (Fig. S2a–d, SI) revealed that the root-mean-square (RMS) roughness of the SnS₂ surface decreased markedly from 5.87 nm to 4.54 nm following the deposition of the SnO₂ functional layer. This reduction improved the electrode-resistive layer interfacial contact, thereby lowering contact resistance and enhancing device stability.

As shown in Fig. 1b, XRD analysis clearly identifies the crystalline features of the three materials. For OMs, the sharp diffraction peak at 2θ = 33.7° corresponds to the (101) plane of SnO₂,³⁶ confirming its tetragonal rutile structure. In contrast, SMs display a characteristic peak at 2θ = 14.9°, which matches well with the (001) plane of SnS₂,³⁷ indicative of its layered disulfide structure. OSMs exhibit both characteristic peaks at 14.9° (SnS₂) and 33.7° (SnO₂) with sharp, unshifted profiles. This observation suggests that the heterojunction interface does not undergo lattice distortion, while the interlayer van der Waals interactions are preserved, thereby confirming the structural compatibility of the two components in the composite system. Raman spectroscopy further supports the XRD results (Fig. 1c). The characteristic band at 630 cm⁻¹ in OMs corresponds to Sn–O vibrations of the rutile SnO₂ phase,³⁸ while the 314 cm⁻¹ peak in SMs arises from the in-plane Sn–S vibrational mode of the SnS₂ layers.³⁹ In OSMs, both peaks at 314 cm⁻¹ and 630 cm⁻¹ appear simultaneously, further confirming the coexistence and structural integrity of SnO₂ and SnS₂ within the bilayer film.

Fig. 1 (a) Device structure, (b) XRD patterns, and (c) Raman shift spectra of OMs, SMs and OSMs; XPS spectra of (d) oxygen (O 1s, after ∼15 nm etching), (e) sulfur (S 2p, after ∼30 nm etching), and (f) tin (Sn 3d, after (I) ∼15 nm and (II) 30 nm etching).

Depth profiling of the OSM functional layer using high-resolution XPS provides detailed insight into the chemical states and elemental distribution of the SnO₂/SnS₂ bilayer film. At a depth of 15 nm (Fig. 1d), oxygen and tin were detected with an atomic ratio of 64.3 : 35.7. The O 1s spectrum, after peak fitting, exhibited two peaks at 532.01 eV and 530.6 eV, corresponding to surface-adsorbed oxygen and lattice oxygen, respectively. At 30 nm (Fig. 1e), sulfur and tin were observed with an atomic ratio of 67.2 : 32.8, and the S 2p spectrum showed peaks at 162.66 eV (S 2p1/2) and 161.45 eV (S 2p3/2), assigned to S²⁻ ions in SnS₂. The high-resolution Sn 3d spectrum (Fig. 1f) displays binding energy peaks at 495.15 eV (Sn 3d3/2) and 486.75 eV (Sn 3d5/2). At 15 nm, these peaks correspond to Sn⁴⁺–O bonds in SnO₂, whereas at 30 nm they correspond to Sn⁴⁺–S bonds in SnS₂. Collectively, these results confirm that Sn maintains a stable +4 oxidation state throughout the SnO₂/SnS₂ bilayer structure.

The SEM cross-sectional image in Fig. 2a shows that the thickness of the OSM film is approximately 50 nm. To further verify both the thickness and elemental distribution of the SnO₂/SnS₂ bilayer, HRTEM and EDS mapping analyses were conducted (Fig. 2b and c). The total functional layer thickness was confirmed to be ∼50 nm, consisting of a ∼20 nm SnO₂ layer and a ∼30 nm SnS₂ layer (Fig. S3a, SI). Meanwhile, cross-sectional scanning electron microscopy (SEM) images (Fig. S3b and c, SI) reveal that the thicknesses of SM and OM are 30 nm and 21 nm, respectively, which are comparable to the monolayer thickness of OSM devices, with a well-defined heterojunction interface being clearly observed. In particular, the HRTEM image in Fig. 2c reveals a lattice fringe spacing of ∼0.58 nm for the 2D SnS₂ layers, which is consistent with previously reported interlayer distances,⁴⁰ thereby confirming the atomic-level structural fidelity of the material. The bilayer configuration was further corroborated by EDS mapping (Fig. 2d), which distinctly resolves the SnO₂ and SnS₂ layers. In addition, Uv-vis-NIR spectroscopy (Fig. S4, SI) demonstrates that OM, SM, and OSM all exhibit excellent optical transparency in the visible range (400–760 nm). Notably, the OSM achieves >90% transmittance across the visible spectrum, attributed to the unique bilayer architecture. This level of transparency significantly surpasses that of comparable composite materials, highlighting the potential of the OSM for application in transparent electronic devices such as memristor arrays and window layers for optoelectronic systems.

Fig. 2 (a) SEM cross sectional and (b) HRTEM images of the OSM device, (c) high-resolution image of the SnO₂/SnS₂ bilayer, and (d) the EDS mapping images of the SnO₂/SnS₂ bilayer.

Memristive performance

To elucidate the performance advantages of the SnO₂/SnS₂ bilayer architecture in resistive random-access memory (RRAM), systematic electrical measurements were carried out. As shown in Fig. 3a–c, the I–V characteristics of SMs, OMs, and OSMs were evaluated over 50 consecutive voltage sweep cycles. Memristors based on 2D sulfides (SMs) exhibited a narrow switching window (<10) with large variability in the I–V curves (Fig. 3a), indicating poor stability. In contrast, oxide-based devices (OMs) displayed a wider switching window (∼100), though their I–V curves showed considerable non-overlapping (Fig. 3b), reflecting limited uniformity. Moreover, the set voltage of OMs was markedly lower than the reset voltage, consistent with the low energy barrier for oxygen-vacancy filament formation compared to the higher barrier required for rupture, highlighting the self-sustaining nature of oxygen-vacancy conductive filaments. Building upon this, SnO₂/SnS₂ bilayer-based devices (OSMs) exhibited superior nonvolatile memory performance (Fig. 3c). It should be noted that the SnO₂ layer functions as a regulation layer, where oxygen vacancies form virtual electrodes that direct the nucleation and growth of sulfur-vacancy filaments in the SnS₂ layer. This cooperative process enables more efficient and controllable resistive switching. As a result, OSMs showed highly reproducible I–V curves with extended switching windows of nearly two orders of magnitude (∼10²), demonstrating enhanced contrast and durability compared with SMs and OMs.

Fig. 3 The 100 I–V sweeps from the single-layer and multi-layer devices: (a) SM, (b) OM, and (c) OSM; (d) resistance distribution, (e) set/reset voltage distribution, (f) endurance, and (g) retention test of the memristor devices.

Statistical analysis of switching parameters across 50 cycles (Fig. 3d and e) further confirmed that OSMs exhibit substantially reduced standard deviations in threshold voltages and resistance states, indicating excellent uniformity and stability. This robustness reflects strong resistance to fluctuations from both external stimuli and internal defects. Such results not only validate the oxide virtual electrode-induced switching mechanism but also highlight the potential of OSMs for advanced applications, including multilevel data storage and neuromorphic synaptic emulation. In addition, endurance and retention tests (Fig. 3f and g) revealed that OSMs achieved markedly improved reliability, sustaining up to 10⁴ cycles compared with 800 for SMs and 200 for OMs, while maintaining stable data retention for over 10⁵ seconds. OSM devices exhibit excellent performance at 80 °C and −30 °C, with their endurance and retention maintained at room-temperature levels and minimal fluctuation (see Fig. S5a–f, SI), confirming their applicability under extreme operating conditions. These improvements stem from the bilayer design: the oxide layer induced the sulfur-vacancy filament growth of the 2D sulfide layer and stabilized the otherwise less reliable oxide-only channels. Consequently, the OSMs significantly enhance both cycling endurance and data retention, underscoring their potential for high-density, low-power nonvolatile memory applications.

Device conduction mechanism

For elucidating the carrier transport mechanism of OSMs, I–V curve fitting was performed (Fig. S6a, SI). At forward bias below 0.16 V (region 1), the I–V response is linear, indicating ohmic conduction dominated by thermally excited carriers. Under these conditions, neither the SnO₂/SnS₂ interface nor the material surface imposes a significant barrier, and carrier migration remains unrestricted. When the applied voltage exceeds 0.16 V, the device exhibits a stable dielectric constant of 2.04 (region 2), as derived from the space charge-limited current (SCLC) fitting of the I–V characteristics (Fig. S6). This response is attributed to the filling of deep-level traps, such as sulfur vacancies in the SnS₂ layer, leading to a reduced effective dielectric constant and a quadratic (V₂) current dependence. With further voltage increase, the slope parameter (λ) rises sharply to 6.56 (region 3), reflecting strong space charge accumulation at the electrode–film interface. This localized high-field environment promotes the formation of oxygen-vacancy filaments in SnO₂, which act as virtual electrodes to induce the growth of sulfur-vacancy filaments in SnS₂, thereby driving the resistive transition. During reverse bias scans, the I–V curve maintains ohmic conduction below 0.12 V (region 4), attributed to residual conductive paths from disrupted filaments. At higher reverse bias, λ increases to 2.73, consistent with SCLC behaviour during partial trap refilling and filament reconstruction, confirming that trap states are more readily occupied after filament rupture. Temperature-dependent measurements (310–350 K, Fig. S6b and c in the SI) further support a hopping conduction mechanism, as indicated by near-unity slopes in log I–log V plots and a slight thermally activated current increase. Collectively, these findings reinforce the virtual electrode-induced filamentary switching mechanism in OSMs.

To further probe the role of oxygen-vacancy concentration in filament regulation, OSM devices were fabricated without annealing and after argon annealing for 30, 60, and 120 min. As shown in Fig. S7a–d, the resistance of the high-resistance state (HRS) increases progressively with annealing time, a trend corroborated by XPS analysis (Fig. S7e–h) that reveals a systematic reduction in oxygen-vacancy density. Unannealed devices exhibit unstable switching, with incomplete filament rupture during reset, due to the excessive oxygen-vacancy content that strengthens the virtual electrode effect and prevents full filament disconnection in SnS₂. In contrast, devices annealed for 30 min display the most optimized switching behaviour, suggesting that this condition yields an ideal oxygen-vacancy concentration for virtual electrode formation. Longer annealing (60–120 min) weakens this effect, reducing the ability to activate sulfur vacancies across the SnS₂ layers and degrading resistive switching performance. These results confirm the pivotal role of oxygen vacancies in the oxide layer in modulating sulfide filament behaviour.

Unexpectedly, we also observed that precise regulation of compliance current (I_cc) and set voltage (V_set) enables reversible transitions between nonvolatile and volatile switching (Fig. S8 and S9, SI). Specifically, I_cc constrains ion migration flux by limiting carrier transport,^41,42 while V_set dictates the threshold for electric-field-driven ion-electron transport. Both parameters govern the formation efficiency and field strength of the virtual electrode. Lowering I_cc or V_set reduces oxygen-vacancy migration and injection efficiency in SnO₂, shrinking the active region of the virtual electrode and weakening the local field. Consequently, filament nucleation and growth are suppressed: (i) ion migration flux becomes insufficient to form continuous conductive pathways, and (ii) metastable filaments rapidly fracture once the field is removed. This dual inhibition causes the low-resistance state to relax quickly back to the high-resistance state, resulting in volatile switching. Such tunability demonstrates that OSM devices can flexibly switch between nonvolatile and volatile states, offering a pathway toward multifunctional neuromorphic devices and polymorphic memory systems.

Scheme 1 illustrates the resistive switching mechanism of OSMs. Owing to the device's negative set polarity, applying a negative bias induces rapid evolution of oxygen vacancies within the upper SnO₂ layer. The hexagonal crystal structure of SnO₂ provides efficient channels for ionic migration, lowering the energy barrier for oxygen-vacancy formation and thereby accelerating the nucleation and growth of oxygen-vacancy filaments. In contrast, the formation of sulfur-vacancy filaments in the lower SnS₂ layer proceeds more sluggishly (Scheme 1b), as the layered structure of SnS₂ imposes a strong spatial barrier that hinders sulfur-vacancy migration. The rapid formation of oxygen-vacancy filaments in SnO₂ establishes a virtual electrode within the device, which optimizes the internal electric-field distribution and lowers the activation energy for sulfur-vacancy filament formation in SnS₂. Under this synergistic effect, sulfur vacancies in SnS₂ gradually nucleate, aggregate, and grow into root-like conductive filaments with robust electrical conductivity, completing the set process (Scheme 1c). Concurrently, oxygen atoms migrate downward into the SnS₂ lattice, partially substituting sulfur atoms, with enrichment localized near the upper SnS₂ region but not penetrating deeper layers (Scheme 1(c1)). This process is corroborated by the EDS elemental distribution shown in Fig. 2d. When a positive bias is applied, the oxygen-vacancy filaments in SnO₂ begin to dissolve, while the sulfur-vacancy filaments in SnS₂ rapidly disconnect (Scheme 1d). Simultaneously, some oxygen atoms that had migrated into the SnS₂ lattice move upward, enabling sulfur atoms to reoccupy their lattice positions (Scheme 1(c1)). With increasing forward voltage, sulfur-vacancy filaments in SnS₂ break first, followed by complete rupture of oxygen-vacancy filaments in SnO₂, thereby completing the reset process.

Scheme 1 Schematic illustration of the internal structure during electro set/reset: (a) initial state, (b) set-1 (oxygen-vacancy conductive filaments in SnO₂ are rapidly formed to constitute a virtual electrode), (c) set-2 (formation of S-vacancy filaments in the lower SnS₂, and (c1) oxygen atoms replacing sulfur atoms), and (d) reset-1 (the S-vacancy conductive filament in the lower SnS₂ is rapidly broken, (d1) sulfur atoms returning to their original positions).

To further validate the proposed switching mechanism, systematic experimental characterization studies were performed. Electron paramagnetic resonance (EPR) spectroscopy was carried out on devices in the initial, set, and reset states after 50 I–V cycles (Fig. 4a I–III). Because semiconductor defects such as oxygen vacancies are paramagnetic, the EPR spectra provide direct insight into their concentration. All samples exhibited a characteristic oxygen-vacancy signal in the SnO₂ layer at g = 2.003.⁴³ The set state device displayed a markedly stronger EPR signal than the pristine device, indicating that repeated cycling leads to an increased density of oxygen vacancies, consistent with the stabilization of conductive filaments. The enhanced vacancy density promotes higher ion mobility, increases active sites, and accelerates interfacial reaction kinetics. After 50 cycles, the reset state still showed a higher signal than the initial state, suggesting that oxygen-vacancy filaments in SnO₂ are “easy to form but difficult to rupture,” thereby underpinning the virtual electrode effect. Complementary XPS analyses (Fig. 4b I–III) further corroborated these findings. Characteristic test sites within a ∼100 µm area were selected on three devices, corresponding to the initial state, the set state after 50 I–V sweeps, and the reset state after 50 I–V sweeps, respectively. No significant binding-energy shifts were observed for Sn, S or O core levels, excluding changes in the overall chemical composition and confirming the structural stability of the heterojunction. Importantly, the intensity of the oxygen-vacancy-related peak in the set state increased compared to the pristine state, providing strong evidence for the formation of the virtual electrode. In contrast, only a slight reduction in the oxygen-vacancy content was observed in the reset state, reinforcing the notion that the conductive filament network is robust and not easily disrupted. Taken together, the EPR and XPS results collectively confirm the critical role of the SnO₂ layer in establishing a stable oxygen-vacancy-based virtual electrode, which serves as the physical foundation for the proposed switching mechanism. This virtual electrode not only facilitates the nucleation and growth of sulfur-vacancy filaments in the SnS₂ layer but also overcomes the intrinsic challenge of vertical ion migration, thereby providing direct experimental evidence for the proposed switching mechanism and offering key insights for optimizing device performance.

Fig. 4 (a) EPR spectra of OSM: (I) as-prepared, (II) set state after 50 I–V cycles, and (III) reset state after 50 I–V cycles; (b) O 1s XPS spectra of OSM: (I) as-prepared, (II) set state after 50 I–V cycles, and (III) reset state after 50 I–V cycles.

Synaptic plasticity

The synaptic plasticity of OSM devices is illustrated in Fig. 5a, encompassing short-term plasticity (STP), long-term plasticity (LTP), PPF, paired-pulse depression (PPD), and STDP. Such a modulation of synaptic behaviours provides important insights into learning and memory processes in biological neural networks.^44–46 To investigate EPSC behaviour, devices were stimulated with a fixed pulse width of 30 ms. Similar to biological synapses, STP was characterized by the relaxation time required for the current to return to the baseline, which increased as the pulse amplitude decreased (−0.3, −0.5, and −0.7 V). This phenomenon arises from the role of the virtual electrode formed in the SnO₂ layer, which induces the generation of sulfur-vacancy filaments in the SnS₂ layer. Variations in pulse amplitude modulate the kinetics of filament formation and rupture, thereby tuning the relaxation time. Notably, the device transitions from STP to LTP at a pulse amplitude of −0.5 V (Fig. 5b), attributed to the stabilization of sulfur-vacancy filaments at this voltage, leading to persistent conductance modulation.

Fig. 5 (a) Schematic illustration of synapse emulation; (b) dependence of potentiating and depressing EPSCs on pulse amplitude under a continuous low read voltage of 100 mV (the stimuli consist of single positive pulses with a constant width of 30 ms and varying amplitudes of −0.3, −0.5, and −0.7 V), (c) PPF, and (d) STDP of OSM.

In addition, both short-term depression (STD) and long-term depression (LTD) behaviours were observed (Fig. S10). STD was induced using positive pulses (0.3 V, 30 ms), and the relaxation time was prolonged with increasing amplitude (Fig. S10b). At a pulse amplitude of 0.5 V, the device switched from STD to LTD (Fig. S10c), highlighting its ability to undergo controlled transitions between short- and long-term synaptic states. These results underscore the device's capability to emulate key neuromorphic behaviours with advanced cognitive relevance. For spiking neural network (SNN) applications, PPF and STDP were further investigated. As shown in Fig. 5c, when two consecutive presynaptic pulses (0.1 V read voltage, −1 V pulse voltage, separated by 1 ms to 1 s) were applied, the second pulse consistently elicited a higher current, demonstrating PPF. The PPF index exhibited strong dependence on the inter-pulse interval and could be well fitted by a biexponential function. For STDP, symmetric pulse pairs were employed (−0.7 V/10 ms for potentiation and +0.7 V/10 ms for depression) at varying intervals (Fig. 5d). The results revealed both LTP and LTD, depending on pulse timing and order, consistent with the asymmetric Hebbian learning rule. Importantly, the use of symmetric pulse pairs simplifies the peripheral circuit design for STDP implementation, thereby enhancing the device's applicability in neuromorphic systems.

As shown in Fig. 6a, the OSM devices exhibit both potentiation and depression behaviours under different pulse conditions, which are essential for artificial neural network (ANN) applications. Specifically, the application of 100 consecutive negative pulses leads to a gradual increase in device current, representing potentiation, whereas consecutive positive pulses induce a steady current decrease, corresponding to depression. Nonlinear conductivity modulation interferes with precise weight adjustment during brain-inspired system training, reducing learning efficiency. To achieve controllable and gradual conductivity tuning, the potentiation (P) and depression (D) behaviors of OM, SM, and OSM devices under the same pulse pattern were investigated, as shown in Fig. S11(a)–(c) and Table S2, SI. Compared with previously reported memristors, the OSM device exhibits a significantly lower nonlinearity factor (NLF), demonstrating superior linear conductivity modulation capability that is more suitable for synaptic weight simulation.⁴⁷ Furthermore, the OSM device can achieve long-term plasticity of LTP and LTD stimulated by pulse train with varying pulse amplitudes for 5 consecutive cycles (Fig. S11d), indicating excellent synaptic behavior stability and cycling endurance, which guarantees its long-term reliable operation in brain-inspired systems. The multilayer perceptron (MLP) architecture based on these characteristics is illustrated in Fig. 6b, consisting of an input layer with 28 × 28 neurons for MNIST or fashion-MNIST images, followed by 256 hidden neurons and 10 output neurons corresponding to digit or clothing categories.⁴⁸ The simulations were performed using the standard MNIST dataset from the torchvision library, which includes 60 000 images for training and 10 000 images for testing. Fig. 6c demonstrates the recognition performance for the digit “7” under varying noise levels, where the visual clarity decreases progressively as the noise level increases.⁴⁹ After 50 training epochs, the device achieved an accuracy of 93.18% in the absence of noise, and 92.53%, 90.68%, and 86.84% at 20%, 40%, and 60% noise levels, respectively (Fig. 6d and e). Similarly, for fashion-MNIST recognition of the “shoe” class, accuracies of 86.06%, 83.52%, 79.52%, and 75.44% were obtained at 0%, 20%, 40%, and 60% noise, respectively (Fig. S12). As training progressed, the loss consistently decreased and eventually converged, confirming effective learning. The incremental learning process of the ANN is further reflected by the diagonal colour evolution of the confusion matrix (Fig. 6f–i). After 50 training cycles, the predicted outputs for each category aligned precisely with the actual labels, achieving robust recognition of all digits from 0 to 9. Notably, comparative studies of SM-, OM-, and OSM-based networks (Fig. S13 and S14) demonstrated that OSM devices yielded the highest recognition accuracy and the lowest loss. These results underscore the strong potential of OSM-based neuromorphic systems for noise-tolerant preprocessing and high-precision pattern recognition in next-generation intelligent computing applications.

Fig. 6 (a) Potentiation/depression of OSMs under 100 consecutive voltage pulse conditions, (b) neural network diagram for digit recognition, (c) recognition output images under different noise levels, (d) recognition accuracy of the number “7” after 50 training sessions at 0%, 20%, 40%, and 60% noise levels; (e) the loss of number “7” after 50 training sessions at 0%, 20%, 40%, and 60% noise levels; confusion matrix for recognition of the number “7” under different noise conditions: (f) 0%, (g) 20%, (h) 40%, and (i) 60%.

Conclusions

In conclusion, SnO₂/SnS₂ bilayer thin films composed of a 2D sulfide and a hexagonal oxide stack were successfully fabricated via a solution-based spin-coating process and applied as functional layers in memristive and synaptic devices. Benefiting from the introduction of the SnO₂ layer, the bilayer devices exhibit significantly enhanced resistive switching performance, achieving an ON/OFF ratio exceeding 200 and stable operation over 10 000 DC cycles, outperforming previously reported stacked-layer structures (Table S1). Distinct from conventional oxide-assisted bilayer memristors, in which filament regulation is typically limited to a single defect system or simple dual-ion effects, the present SnO₂/SnS₂ architecture introduces a cross-defect-type coupling mechanism. In this design, oxygen-vacancy filaments in the oxide layer act as virtual electrodes that directionally regulate the formation and rupture of sulfur-vacancy filaments in the sulfide layer, enabling both improved switching stability and reversible modulation between nonvolatile and volatile memory states. This represents a clear conceptual advance toward multifunctional, interface-engineered neuromorphic devices. The devices further demonstrate reliable neuromorphic functionality, achieving ∼93% inference accuracy in artificial neural network simulations and emulating essential synaptic behaviours such as short- and long-term plasticity. Correlative EPR and XPS analyses reveal that the robust performance originates from the facile formation and high stability of oxygen-vacancy filaments in SnO₂, which guide sulfur-vacancy filament evolution in SnS₂. By tuning the compliance current and set voltage, the devices can be reversibly switched between nonvolatile and volatile operation modes, enabling on-demand control of storage characteristics. Overall, this work provides new insights into the design of oxide–sulfide bilayer memristors, demonstrates the feasibility of scalable solution processing for transparent and environmentally benign synaptic devices, and highlights the potential of OSMs as a versatile platform for high-density memory arrays and neuromorphic electronic systems.

Author contributions

Xiuyang Tang fabricated the devices and performed the I–V measurements and material characterization. Xinming Ma, Sizhu Ha, Weifang Sun, and Niwei He helped with I–V measurements and data analysis. Song Xue helped build the theoretical models and analyse the results of Raman measurement. Gangri Cai and Jin Shi Zhao provided guidance in experimental designs and understanding the phenomena and supervised the entire research program. All the authors participated in writing the manuscript.

Conflicts of interest

The authors declare no conflicts of interest.

Data availability

All data supporting the findings of this study are available within the article and its supplementary information (SI). Supplementary information: the description of the experimental setup, SEM and AFM images of the deposited films, UV-vis spectra of optical transparency tests, and I–V electrical measurements for memristive performance and neuromorphic computing analyses. See DOI: https://doi.org/10.1039/d5nh00779h.

Additional raw data are available from the corresponding authors upon reasonable request.

Acknowledgements

This work was supported by the Natural Science Foundation of Tianjin (Grant No. 23JCYBJC00340 and 23JCQNJC00590) and Beijing-Tianjin-Hebei Natural Science Foundation Cooperation Project (No. 25JJJJC0006 and 25JJJC0011).

Notes and references

1. W. Wan, R. Kubendran, C. Schaefer, S. B. Eryilmaz, W. Zhang, D. Wu, S. Deiss, P. Raina, H. Qian, B. Gao, S. Joshi, H. Wu, H.-S. P. Wong and G. Cauwenberghs, Nature, 2022, 608, 504–512.
2. L. Zou, Z. Peng, H. Sun, Y. Yi, C. Zhu, J. Xu, J. Zhang, X. Hu, Y. Dang, L. Ye and X. Miao, Adv. Funct. Mater., 2025, 35, 2416325.
3. M. Jain, M. J. Patel, L. Liu, J. Gosai, M. Khemnani, H. J. Gogoi, M. Y. Chee, A. Guerrero, W. S. Lew and A. Solanki, Nanoscale Horiz., 2024, 9, 438–448.
4. A. Serb, J. Bill, A. Khiat, R. Berdan, R. Legenstein and T. Prodromakis, Nat. Commun., 2016, 7, 12611.
5. D. Cui, M. Pei, Z. Lin, H. Zhang, M. Kang, Y. Wang, X. Gao, J. Su, J. Miao, Y. Li, J. Zhang, Y. Hao and J. Chang, Light: Sci. Appl., 2025, 14, 161.
6. H. Hu, A. Scholz, C. Dolle, A. Zintler, A. Quintilla, Y. Liu, Y. Tang, B. Breitung, G. C. Marques, Y. M. Eggeler and J. Aghassi-Hagmann, Adv. Funct. Mater., 2024, 34, 2302290.
7. L. Hu, S. Fu, Y. Chen, H. Cao, L. Liang, H. Zhang, J. Gao, J. Wang and F. Zhuge, Adv. Mater., 2017, 29, 1606927.
8. W. Deng, X. Yan, L. Wang, N. Yu, W. Luo and L. Mai, Nano Energy, 2024, 128, 109861.
9. P.-L. Lin, Z.-S. Liao, S.-M. Chen and J.-S. Chen, Nanoscale Horiz., 2025, 10, 379–387.
10. S. K. Vishwanath, B. Febriansyah, S. E. Ng, T. Das, J. Acharya, R. A. John, D. Sharma, P. A. Dananjaya, M. Jagadeeswararao, N. Tiwari, M. R. C. Kulkarni, W. S. Lew, S. Chakraborty, A. Basu and N. Mathews, Mater. Horiz., 2024, 11, 2643–2656.
11. J. Kim, C. Im, C. Lee, J. Hwang, H. Jang, J. H. Lee, M. Jin, H. Lee, J. Kim, J. Sung, Y. S. Kim and E. Lee, Nanoscale Horiz., 2023, 8, 1417–1427.
12. Q. Wang, R. Luo, Y. Wang, W. Fang, L. Jiang, Y. Liu, R. Wang, L. Dai, J. Zhao, J. Bi, Z. Liu, L. Zhao, Z. Jiang, Z. Song, J. Schwarzkopf, T. Schroeder, S. Wu, Z. Ye, W. Ren, S. Song and G. Niu, Adv. Funct. Mater., 2023, 33, 2213296.
13. E. Wu, Y. Wang, S. Huo, J. Xu, M. Sheng, H. Liu, L. Zhong, J. Gao, Y. Xie and C. Pan, Adv. Funct. Mater., 2025, e18764.
14. Z. Zhang, X. Tang, Y. Gao, X. Ma, K. Chang, L. Chen, S. Xue, G. Cai and J. S. Zhao, Chem. Eng. J., 2025, 523, 168557.
15. J. Li, X. Tang, Q. Liu, Y. Jiang and Z. Tang, J. Adv. Ceram., 2021, 10, 1001–1010.
16. J. Shi, S. Kang, J. Feng, J. Fan, S. Xue, G. Cai and J. S. Zhao, Nanoscale Horiz., 2023, 8, 509–515.
17. P. Saha, M. Sahad E, S. Sathyanarayana and B. C. Das, ACS Nano, 2024, 18, 1137–1148.
18. Z. Wang, B. Peng, X. Huang, Z. Yang, Y. Liu, K. Chen, B. Zhang, S. Zhang, Z. Ao, X. Zhou, Z. Zhang, X. Tang and J. Yue, Adv. Funct. Mater., 2025, 35, 2501372.
19. S. Rehman, M.-S. Kim, M. F. Khan and S. Kim, Nano Energy, 2024, 127, 109764.
20. X. Liu, C. Zhang, E. Li, C. Gao, R. Wang, Y. Liu, F. Liu, W. Shi, Y. Yuan, J. Sun, Y. Lin, J. Chu and W. Li, Adv. Funct. Mater., 2024, 34, 2309642.
21. X. Chen, D. Yang, G. Hwang, Y. Dong, B. Cui, D. Wang, H. Chen, N. Lin, W. Zhang, H. Li, R. Shao, P. Lin, H. Hong, Y. Yao, L. Sun, Z. Wang and H. Yang, ACS Nano, 2024, 18, 10758–10767.
22. W. Ahn, H. B. Jeong, J. Oh, W. Hong, J. Cha, H. Y. Jeong and S. Choi, Small, 2023, 19, 2300223.
23. X. Yan, C. Qin, C. Lu, J. Zhao, R. Zhao, D. Ren, Z. Zhou, J. Wang, L. Zhang, X. Li, Y. Pei, G. Wang, Q. Zhao, Z. Xiao and H. Li, ACS Appl. Mater. Interfaces, 2019, 11, 48029–48038.
24. M. H. Pervez, E. Elahi, M. A. Khan, M. Nasim, M. Asim, A. Rehmat, M. A. Rehman, M. A. Assiri, S. Rehman, J. Eom and M. F. Khan, Small Struct., 2025, 6, 2400386.
25. S. J. Kim, H.-J. Lee, C.-H. Lee and H. W. Jang, npj 2D Mater. Appl., 2024, 8, 70.
26. P. Zhuang, H. Yan, B. Li, C. Dou, T. Ye, C. Zhou, H. Zhu, B. Tian, Y. Chen, D. Lin, W. Cai, D. Yu, J. Liu and W. Lin, Mater. Today Nano, 2023, 23, 100353.
27. X. Tian, L. Wang, J. Wei, S. Yang, W. Wang, Z. Xu and X. Bai, Nano Res., 2014, 7, 1065–1072.
28. B. K. You, J. M. Kim, D. J. Joe, K. Yang, Y. Shin, Y. S. Jung and K. J. Lee, ACS Nano, 2016, 10, 9478–9488.
29. U. Gawai, D. Kumar, A. Singh, C.-H. Wu and K.-M. Chang, ACS Appl. Electron. Mater., 2022, 4, 4265–4272.
30. R. Zhang, H. Huang, Q. Xia, C. Ye, X. Wei, J. Wang, L. Zhang and L. Q. Zhu, Adv. Electron. Mater., 2019, 5, 1800833.
31. Y. Zhang, G.-Q. Mao, X. Zhao, Y. Li, M. Zhang, Z. Wu, W. Wu, H. Sun, Y. Guo, L. Wang, X. Zhang, Q. Liu, H. Lv, K.-H. Xue, G. Xu, X. Miao, S. Long and M. Liu, Nat. Commun., 2021, 12, 7232.
32. Y. Li, J. Chu, W. Duan, G. Cai, X. Fan, X. Wang, G. Wang and Y. Pei, ACS Appl. Mater. Interfaces, 2018, 10, 24598–24606.
33. M. M. Ganaie, A. Kumar, A. K. Shringi, S. Sahu, M. Saliba and M. Kumar, Adv. Funct. Mater., 2024, 34, 2405080.
34. X. Mao, Z. Guo, S. Yang, Y. Shen, L. Wei, C. Li, H. Jiang, H. Li, H. Li and C. Li, ACS Nano, 2025, 19, 21669–21678.
35. H. Yu, H. Liang, J. Bai and C. Li, Int. J. Hydrogen Energy, 2023, 48, 15908–15920.
36. D. Manikandan and R. Murugan, Prog. Mater. Sci., 2022, 130, 100970.
37. D. B. Mitzi, L. L. Kosbar, C. E. Murray, M. Copel and A. Afzali, Nature, 2024, 428, 299–303.
38. M. Krause, C. Romero-Muñiz, O. Selyshchev, D. R. T. Zahn and R. Escobar-Galindo, J. Mater. Chem. A, 2025, 13, 15128–15151.
39. Y. Ding, W. Zheng, X. Lu, Y. Liang, Y. Zhu, M. Jin and F. Huang, J. Phys. Chem. Lett., 2020, 11, 10094–10099.
40. J. Feng, J. Fan, Z. Zhang, Y. Gao, S. Xue, G. Cai and J. S. Zhao, Adv. Funct. Mater., 2024, 34, 2401228.
41. Z. Huang, W. Huang, C. Xu, Y. Xu, T. Du, W. Liu, J. Li, S. Xie, G. Chen and H. Chen, Adv. Funct. Mater., 2025, 2502211.
42. D. Ju, S. Kim, K. Park, J. Lee, M. Koo and S. Kim, ACS Appl. Mater. Interfaces, 2024, 16, 24929–24942.
43. M. Imran, Z. Dai, F. Hussain, R. Chen, F. Wu and L. Li, ACS Nano, 2025, 19, 29372–29382.
44. Y. Xia, C. Zhang, Z. Xu, S. Lu, X. Cheng, S. Wei, J. Yuan, Y. Sun and Y. Li, Nanoscale, 2024, 16, 1471–1489.
45. C. Weilenmann, A. N. Ziogas, T. Zellweger, K. Portner, M. Mladenović, M. Kaniselvan, T. Moraitis, M. Luisier and A. Emboras, Nat. Commun., 2024, 15, 6898.
46. D. Kumar, H. Li, D. D. Kumbhar, M. K. Rajbhar, U. K. Das, A. M. Syed, G. Melinte and N. El-Atab, Nano-Micro Lett., 2024, 16, 238.
47. M. Jain, B. Fasnick Ck, M. Khemnani, L. Kortstee, B. Andola, M. J. Patel, A. Guerrero, Y. K. Srivastava, I. E. Castelli and A. Solanki, Nano Energy, 2025, 138, 110871.
48. L. Gao, P.-Y. Chen and S. Yu, IEEE Electron Device Lett., 2016, 37, 870–873.
49. Y. Chen, Y. Li, S. Niu, X. Yang, W. Dou, C. Shan and G. Shen, Adv. Funct. Mater., 2024, 34, 2315383.

---