In situ observation of oxygen ion dynamics in topological phase change memristors through self-assembled interface design

Abstract

Resistive switching memory represents a potentially transformative advancement in next-generation nonvolatile memory and neuromorphic technologies. Recently, vertically aligned nanocomposites (VANs) have been proposed to optimize the memristive performance of single-phase memories. However, the microscopic mechanism of dynamic resistive switching in these VAN architectures is still elusive. Here, we built up a VAN structure with brownmillerite SrCoO_2.5 (BM-SCO) and magnesium oxide (MgO), where the topological phase transformation in BM-SCO provides a well-defined facile vertical path for oxygen ion migration within the vertical interfaces between BM-SCO and MgO phases. Compared with the BM-SCO memristor, the (BM-SCO)_0.5:(MgO)_0.5 VAN memristor exhibits advantages in resistive switching and simulates various synaptic functions, achieving high accuracy in image recognition tasks. Using in situ scanning transmission electron microscopy, we revealed the microscopic mechanism of oxygen ion migration dynamics along the vertical interfaces. Our work substantially advances the understanding of resistive switching mechanism and further demonstrates the great potential of VAN architectures for practical application in high-performance resistive memory.

Visualizing the oxygen ion migration in vertically aligned nanocomposite using in situ STEM.

INTRODUCTION

Neuromorphic computing—an approach inspired by the human brain’s parallel processing capabilities—has emerged as a promising alternative to overcome the “von Neumann bottleneck” (1–8). Of diverse neuromorphic device platforms, memristors are considered as building blocks for not only mimicking the synaptic functionalities in artificial neural networks but also offering the compactness and low power required for large-scale integration (9–12). Various types of memristors have been explored for artificial synapse applications, such as phase change memory (13, 14), ferroelectric random access memory (15–17), magnetoresistive random access memory (18), and resistive random access memory (RRAM) (19). Among these, RRAM with a two-terminal metal-insulator-metal (MIM) structure has become a promising candidate in terms of low cost, low energy consumption, high scalability, and fast switching speed (20, 21). By applying voltage across the metallic electrodes in MIM nanostructures, the electrical resistance of an insulating layer can be modulated into two or more states. This feature closely mirrors the adjustable synaptic weights in biological systems, making RRAM very attractive for artificial synaptic applications. Numerous single-phase transition-metal oxides including HfO₂ (22, 23), TiO₂ (24), and Ta₂O₅ (25) have been broadly investigated for RRAM devices. Despite significant progress, memristors still suffer from major issues of long-term reliability and uniformity that hinder their practical applications. Therefore, several strategies were proposed to address these problems, such as tuning the microstructure of functional layer (26), choosing deposition techniques to optimize material properties (27), or designing well-engineered interfaces (28).

Constructing vertically aligned nanocomposites (VANs) have been regarded as an effective approach for manipulating functional materials’ magnetic and ferroelectric properties (29, 30), as well as resistive switching characteristics (31). VAN structures are typically composed of two immiscible phases that self-assemble into vertically aligned nanocolumns during growth. Because of the structural incompatibility of nanopillar and matrix, this structure offers regularly arranged and spatially confined conductive channels at the vertical interfaces between the two phases (32, 33). It mitigates the inherent randomness and uncontrollability of formation and rupture in conducting filaments, thereby improving the uniformity and stability of devices. Compared with traditional transverse multilayer frameworks, such as superlattices and multilayers, VAN architectures not only have higher integration density and higher strain adjustability but are also readily manufactured through self-assembly (34). Previous works, such as SrTiO₃ (STO):MgO (35), STO:Sm₂O₃ (34), STO:Sm-doped CeO₂ (36), BaTiO₃ (BTO):ZnO (37), and BTO:CeO₂ (38), have demonstrated that self-assembled VAN films have great potential for applications in designing high-performance neuromorphic devices. However, some open questions still remain unresolved. The inherently low oxygen vacancy content in the matrix materials of these reported VAN structures may constrain the switching speed and resistance state tunability of devices. Resistance switching behavior of BTO-based VAN films is determined by the joint action of ferroelectricity and oxygen vacancy migration along vertical interfaces. This combined mechanism makes it difficult to distinguish their respective contributions, challenging to independently study the impact of vertical interfaces on memristive behavior in VAN structures. There is no direct observation of oxygen ions preferring to move along vertical interfaces, making oxygen ion migration dynamics mechanism in VAN architecture still a conjecture. In situ transmission electron microscopy (TEM) technique offers a powerful method for directly observing ion dynamics (metal atoms, oxygen vacancies, etc.) in an atomic-scale resolution in memristive materials (25, 39–41). By contrast, in the above-reported VAN architectures, the use of in situ TEM can be challenging because of the low image contrast between oxygen-rich and oxygen-deficient regions. This limitation makes it arduous to clearly distinguish between areas of varying oxygen concentration, complicating the direct observation of oxygen ion migration in these VAN architectures. Therefore, innovations in materials and devices with novel structures are urgently desired to extract more precise information about resistive switching behaviors in memristive materials for VAN architectures.

An exemplary material is strontium cobaltite SrCoO_x (SCO), which can be transformed between insulating brownmillerite SrCoO_2.5 (BM-SCO) and conducting perovskite SrCoO₃ (PV-SCO) through losing or gaining oxygen, accompanied by a prominent change in electrical conductivity (42, 43). This model offers an exceptional platform for in-depth investigation of the topotactic phase transition–related switching mechanism, enabling a comprehensive correlation between electrical and structural characterizations. Benefiting from these characteristics, SCO films have received huge interest for resistive switching memory devices in recent years (44–46). BM-SCO has an intrinsic one-dimensional ordered oxygen vacancy channels along the [010] crystallographic direction. Thus, SCO-based VAN structures are beneficial for confining uncontrollable oxygen ion migration along the vertical interfaces. This will also help elucidate the role of vertical interfaces through phase transition. In this work, we designed a VAN structure composed of brownmillerite BM-SCO and MgO. Compared with the BM-SCO memristor, the (BM-SCO)_0.5:(MgO)_0.5 (S₅₀M₅₀) VAN memristor exhibits improved resistance switching performances, such as electroforming-free, higher on/off ratio, enhanced cycling endurance of a single device, and uniformity of multiple devices, and mimics biological synapses functions, such as paired-pulse facilitation (PPF), paired-pulse depression (PPD), long-term potentiation (LTP), and long-term depression (LTD). A simulated neural network based on S₅₀M₅₀ devices achieves a high accuracy of 97.9% for traffic sign recognition, and it maintains 96.3% even with data augmentation. Using atomic-level in situ scanning TEM (STEM), we observed a distinct phase transition from BM-SCO to PV-SCO at the vertical interfaces between the BM-SCO and MgO phases, together with distinct electric resistance change, providing direct evidence for the preferential vertical interfacial oxygen ion migration for resistive switching mechanism. Our findings contribute significantly to the understanding of the resistive switching mechanism in metal oxide–based nanocomposite films, highlighting the great potential of VAN-based structures for constructing high-performance resistive switching memory.

RESULTS

Structure characteristics

Bulk BM-SCO has an orthorhombic structure with the lattice constants of a = 5.574 Å, b = 5.470 Å, and c = 15.745 Å, which can be converted to a pseudotetragonal phase with a_t = 3.905 Å and c_t = 3.936 Å (44). Therefore, BM-SCO films can be grown on perfectly lattice-matched STO substrates (a = 3.905 Å). On the basis of previous works (47, 48), La_0.7Sr_0.3MnO₃ (LSMO) was chosen as the bottom electrode in this study. Figure 1A shows x-ray diffraction (XRD) θ-2θ scans of BM-SCO/LSMO/STO (001) and S₅₀M₅₀/LSMO/STO (001) films, in which the distinct (00l) diffraction peaks collected from BM-SCO, MgO, LSMO, and STO can be observed. Apart from the (004) and (008) diffraction peaks, the superlattice peaks of (002) and (006) were also detected for BM-SCO, originating from alternate stacking of oxygen-deficient tetrahedral CoO₄ and oxygen-rich octahedral CoO₆ layers along the out-of-plane (OOP) direction. The OOP lattice constants of bulk BM-SCO and MgO are c_t = 3.936 Å and c_MgO = 4.211 Å (49), so OOP lattice mismatch between the two phases, defined as u = (c_MgO − c_t)/c_t (50), is calculated to be 7.0%. The OOP lattice constants of BM-SCO are further calculated to be 3.930 and 3.970 Å for BM-SCO and S₅₀M₅₀ films, respectively. The enhanced OOP lattice constant of the BM-SCO phase in the S₅₀M₅₀ film is correlated with the large vertical tensile strain caused by the lattice mismatch between BM-SCO and MgO. In turn, oxygen vacancies can easily generate and accumulate at the vertical heterointerfaces between BM-SCO and MgO in VAN films due to their interfacial structural discontinuity (34, 51) and the large misfit strain (35). This is confirmed by the core-level photoemission spectra for O 1s (fig. S1A). The proportion of defect oxygen in the O 1s signal for the S₅₀M₅₀ film is significantly higher than that in the BM-SCO film, indicating that the S₅₀M₅₀ film has more oxygen vacancies. To investigate strain state, we carried out reciprocal space mappings around the asymmetric (103) diffraction of STO substrate (fig. S2). The reflection spots of both BM-SCO (1112) and LSMO (103) have the same q_x position as STO (103) before and after MgO incorporation, indicating that BM-SCO phases are coherently grown on STO substrates without in-plane (IP) strain (underlined to show which plane in BM-SCO is used). This is in accordance with previous reports on BM-SCO films grown on STO substrates (45, 47). Ex situ atomic force microscopy images display that the BM-SCO film has a long strip–like structure, while the S₅₀M₅₀ film shows MgO nanopillars embedded in BM-SCO matrix (fig. S3). The compositional ratio between BM-SCO and MgO was also varied, with the resulting structural differences shown in fig. S4 for (BM-SCO)_1−x:(MgO)_x (BM-SCO:MgO) nanocomposites at x = 0.65 and 0.77, abbreviated as S₆₅M₃₅ and S₇₇M₂₃, respectively.

Fig. 1. Structural characterizations of the films.

(A) XRD θ-2θ patterns of BM-SCO and S₅₀M₅₀ films grown on LSMO-buffered STO (001) substrates. a.u., arbitrary units. (B) Cross-sectional STEM image of the S₅₀M₅₀ VAN film projected along the [010] zone axis. (C) EDS signal intensity profiles across S₅₀M₅₀ interfaces extracted along the red line in (B). Cross-sectional views of HAADF-STEM (D) and corresponding ABF-STEM (E) images of a single MgO nanopillar in the S₅₀M₅₀ VAN film. IP oriented oxygen vacancy channels in the oxygen tetrahedral layers (D) are labeled with green arrows. The inset in (E) is the FFT pattern of BM-SCO in the VAN, in which the superlattice signals are denoted with yellow circles. (F) Crystallographic model of the BM-SCO:MgO film at the vertical interface.

The microstructures of the S₅₀M₅₀ film were further investigated using high-angle annular dark-field (HAADF)–STEM (Fig. 1B and fig. S5A). The vertical columnar structures of the S₅₀M₅₀ film are evident from the images with spontaneous phase ordering. The bright and dark areas are corresponded to BM-SCO and MgO, respectively, due to the strong atomic number–dependent HAADF images. This is further confirmed by energy-dispersive spectroscopy (EDS) (Fig. 1C). We found that MgO nanopillars with an average diameter of 4 to 7 nm are uniformly distributed and embedded in BM-SCO matrix. High-resolution HAADF-STEM image of the vertical interfaces for an individual MgO nanopillar (Fig. 1D) shows an alternate stacking originating from fully oxygenated octahedral and oxygen-deficient tetrahedral sublayers in BM-SCO matrix, indicative of brownmillerite structure with horizontal oxygen vacancy channels. The clear boundaries between BM-SCO and MgO phases were formed along the OOP direction. Oxygen sublattices were further determined by annular bright-field (ABF)–STEM image (Fig. 1E). The fast Fourier transform (FFT) electron diffraction pattern (inset of Fig. 1E) of BM-SCO also confirms the brownmillerite phase of BM-SCO (52, 53). On the basis of the above results, a crystallographic modeling at the vertical interface between BM-SCO and MgO is constructed in Fig. 1F. The formation mechanism underlying the VAN architecture is detailed in section S1.

Artificial synaptic simulation

To investigate resistive switching, we compared the switching characteristics of BM-SCO/LSMO/STO and S₅₀M₅₀/LSMO/STO devices. The S₅₀M₅₀ memristor displays improved resistance switching characteristics, including electroforming-free operation, a higher on/off ratio, extended cycling endurance, and enhanced uniformity across multiple devices (section S2 and figs. S7 to S9). Further increasing the fraction of BM-SCO in the nanocomposite films leads to a degradation in resistive switching performance (fig. S10). Therefore, the S₅₀M₅₀ composition is considered as an optimal ratio for enhancing the resistive switching behavior of BM-SCO. The S₅₀M₅₀ memristor maintains stable performance in I-V measurements at 85°C (fig. S11), which is typical of practical environments (46, 54, 55), with a slight decrease in on/off ratio and an influence on retention performance. Its switching behavior (fig. S12) and low power consumption (≤ 0.8 nJ) further support its potential for practical applications.

In human brain system, synapses are crucial structures that are responsible for the transmission of stimuli between neurons (Fig. 2A). Five continuously positive and negative voltage sweeps were applied to investigate the possibility that the resistance of the S₅₀M₅₀ memristor can be progressively modified under dc model (Fig. 2, B and C). Compared to bipolar voltage sweeps (fig. S7), lower voltage ranges of 0 → ±1.5 V → 0 V were selected to prevent the device from hard breakdown. It is seen that the current gradually rises (reduces) with the positive (negative) sweep cycles, and the slope of the I-V curve for each subsequent scan increases/decreases compared to the previous scan. The gradual adjustment in resistance of the device with the voltage accumulation suggests the excellent consecutively regulable memory resistance characteristics. This behavior closely resembles the dynamic strengthening and weakening of biological synapses. The ability of the S₅₀M₅₀ device to modulate current in a tunable and continuous manner allows the memristor to mimic the functionalities of biological synapses. The S₅₀M₅₀ device also presents analog history–dependent resistive switching behavior (56), as the gradual increase (decrease) in current can be observed with continuous positive (negative) voltage sweep (fig. S13).

Fig. 2. Synaptic behavior simulation.

(A) Schematic representation of an artificial synapse. Consecutive voltage scans with (B) 0 V → +1.5 V → 0 V positive bias and (C) 0 V → −1.5 V → 0 V negative bias for S₅₀M₅₀ memristor. Evolutions of (D) PPF and (E) PPD with pulse interval for S₅₀M₅₀ memristor. The insets show the corresponding pulse waveforms. (F) The LTP and LTD of S₅₀M₅₀ memristor. The insets show the applied write pulses. Variation of current as a function of the amplitude (G), width (H), and interval (I) of the write pulse.

Synaptic plasticity is categorized into short-term and long-term plasticity, both of which are crucial for adaptive neural function (section S3). As shown in Fig. 2D, applying two voltage pulses (4 V, 30 ms) successively results in the second pulse eliciting a higher stimulation current than the first. As the interval between the pulses increases, the facilitative effect diminishes. Conversely, two negative voltage pulses (−1 V, 30 ms) lead to an inhibitory effect (Fig. 2E). The fitting results of relaxation factors τ₁ and τ₂ for PPF are 17.53 and 476.58 ms, and those for PPD are 5.97 and 134.17 ms, respectively. Both are similar to biological synapses, indicating that the S₅₀M₅₀ device has good short-term synaptic plasticity.

In the context of long-term synaptic plasticity, LTP underlies the processes of learning and memory, while LTD helps filter and forget outdated information. The collaboration between these mechanisms allows the brain to flexibly process complex information, which is achieved through pulse modulation (Fig. 2F). Specifically, when 100 positive pulses (5 V, 30 ms) are applied to the S₅₀M₅₀ device, the current gradually increases, exhibiting multiple conductance states. Applying 100 negative pulses (−3 V, 30 ms) results in a gradual decrease in current from the previously established level. Five successive LTP and LTD sequences (fig. S14) illustrate that the measured long-term synaptic plasticity is reproducible and reliable. The degree of change in the synaptic weight can also be controlled by modulating the pulse amplitude (fig. S15) or pulse numbers (fig. S16) in LTP and LTD. The variation in the current for both LTP and LTD is larger when the pulse voltage or pulse number is increased, providing flexibility for controlling the synaptic plasticity. This memory capability (fig. S17) allows the device to sustain multiple conductance states under external stimulation, facilitating neuromorphic computing applications. Furthermore, by varying the amplitude, pulse width, and interval of the pulses, further regulation of the device is achieved (Fig. 2, G to I). The response current increases with greater pulse amplitude and width, while it also rises as the pulse interval decreases. These experimental results demonstrate that the S₅₀M₅₀ device exhibits remarkable synaptic plasticity, making it a suitable component for constructing computational arrays that more closely resemble biological neural networks.

Neuromorphic computing

The multiple resistance states and strong synaptic plasticity of the S₅₀M₅₀ device highlight its potential for neuromorphic computing applications. To further evaluate this, we constructed a convolutional neural network (CNN) to identify various traffic signs, using data from the German Traffic Sign Recognition Benchmark dataset. As illustrated in Fig. 3A, the CNN architecture comprises three convolutional layers, three pooling layers, and a fully connected layer, ultimately classifying the signs into 43 distinct categories. The network uses the rectified linear unit (ReLU) activation function, with a 3 × 3 convolution kernel size, and uses max pooling for dimensionality reduction. Training of the network is conducted online via forward and back propagation, with the detailed procedure depicted in fig. S18. Taking advantage of the fact that memristors can simplify matrix operations, a crossbar array as shown in Fig. 3B was constructed to provide a hardware framework for the neural network. In this array, the conductance values of the memristors are mapped to the neural network’s weights, which are updated on the basis of the LTP and LTD characteristics. To fully reflect the good brain-like computing performance of the S₅₀M₅₀ device, we enhanced the dataset to a certain extent, including adding a certain range of random noise (15%), random rotation (15°), and random brightness change (0.2), as shown in Fig. 3C. The dataset comprises 39,210 images, with 80% allocated for training and 20% for testing. As illustrated in Fig. 3D, with an increasing number of iterations, the recognition accuracy of the neural network improves progressively under ideal conditions, S₅₀M₅₀ devices, and data augmentation scenarios. Correspondingly, the loss value decreases gradually, as shown in Fig. 3E. After 40 iterations, the accuracy under ideal conditions reached 99.4%, while the S₅₀M₅₀ device achieved a recognition accuracy of 97.9%. When combined with data augmentation, the S₅₀M₅₀ device attained an accuracy of 96.3%, approaching the ideal benchmark and surpassing 95% (Fig. 3F). The recognition accuracy for each category and the corresponding confusion matrix are detailed in figs. S19 and S20. The above results show that neuromorphic computing based on the S₅₀M₅₀ devices can handle complex computing tasks and show excellent performance in large-scale data processing and pattern recognition.

Fig. 3. Neuromorphic computing.

(A) Schematic of CNN-based traffic sign recognition. (B) The circuit diagram for simulating neural networks based on memristor crossbar arrays. (C) Dataset enhanced with noise, rotation, and brightness adjustments. (D) Recognition accuracy as a function of number of iterations. (E) Variation of loss value with iterations. (F) Comparison of accuracy rate under three training methods.

Analysis of resistive switching mechanism

To understand resistive switching behaviors of Pt/BM-SCO/LSMO/STO and Pt/ S₅₀M₅₀/LSMO/STO devices, we analyzed the charge carrier transport by replotting the I-V curves in the double logarithm coordinate (fig. S21 and section S4). Compared with the BM-SCO device, the space charge–limited current (SCLC) becomes prevalent at low-voltage range for the S₅₀M₅₀ device. Overall, the I-V slopes of the regions in the S₅₀M₅₀ device are commonly larger than the corresponding parts of the pure BM-SCO phase. This discrepancy may be correlated with the different number of traps (oxygen vacancies) in the films (47). The number of oxygen vacancies generated in the S₅₀M₅₀ VAN film (as evidenced in fig. S1A) is larger than that in the BM-SCO film, resulting in the dominance of SCLC and trap-filled behaviors in the S₅₀M₅₀ memristor.

To directly unveil the microscopic nature of oxygen ions movement in S₅₀M₅₀ memristor, we further conducted in situ STEM characterizations for the device under high-resistance state (HRS) and low-resistance state (LRS). During the measurement, the tungsten tip was electrical biased and made a contact with the top Pt layer in the device (fig. S23). The LSMO bottom electrode was connected to the TEM grid and grounded. Figure 4 (A and B) displays the close-up HAADF-STEM and ABF-STEM images at HRS state near a single MgO nanopillar, respectively. We could only observe the BM-SCO phase in the matrix and the vertical interfaces between the nanopillar and the matrix, as indicated by the alternating dark and bright stripes. After applying a positive voltage of +3.3 V to trigger the LRS state in the device, the stripe contrast disappeared near the vertical interface region, while BM-SCO structure remains away from the MgO nanopillar, as shown more clearly on the right side of the images in Fig. 4 [C (HAADF) and D (ABF)]. The area along the nanocolumn is different from brownmillerite structure region and is denoted by an orange dashed line. Corresponding intensity line profiles (fig. S24) show that the relative intensity of oxygen ions in the LRS state is stronger than that in the HRS state, suggesting that the oxygen concentration in the LRS state is higher than that in the HRS state. Figure 4E recorded the voltage application process during resistance switching process. A voltage of +3 V was applied to the device initially, and the current increased gradually. As the voltage reached +3.3 V, a sudden increase in the current can be observed, indicating that a phase change occurred along the vertical interfaces. Then, this voltage was maintained, and the current increased at a higher rate than at the beginning, corresponding to the filling of oxygen ions on the left side of the nanopillar, away from the interface. The FFT image of the dashed line area (Fig. 4F) displays the characteristic of a perovskite structure. Therefore, we concluded that in the vertical interface region, BM-SCO transforms into a perovskite structure (PV-SCO) with the chemical formula of SCO_3−δ. Thus, the preferential movement of oxygen ions within the vertical interfaces of the VAN film was directly revealed using in situ STEM. The migration of oxygen ions along the vertical interface is also supported by Co L-edge shift in electron energy-loss spectroscopy (EELS) studies that performed at both the vertical and nonvertical interface areas at the LRS state. As shown in Fig. 4G, the shift of the Co L₂ and L₃ peaks to higher energy levels indicates a higher Co valence state in the vertical interface region than in the nonvertical interface region (57). This finding provides direct evidence for the oxygen migration–induced Co valence state alteration at the vertical interface, which is consistent with the structural results unveiled by STEM.

Fig. 4. In situ STEM and mechanism analysis.

HAADF-STEM (A and C) and ABF-STEM (B and D) images for HRS and LRS states of S₅₀M₅₀ memristor, respectively. The contrast correspondence is illustrated by the structural model overlapping in images (A) and (C). The green arrows in (D) indicate the oxygen tetrahedral layers. (E) The relationships between current and voltage and current and time during the process of applying voltage. (F) The FFT pattern of the dashed line area in (C). (G) Comparison of Co L-edge EELS spectra of the vertical interface and nonvertical interface regions under LRS. (H) COMSOL simulation of an electric field under 0 V (left) and +5 V (right) for BM-SCO:MgO memristor. (I) Schematic drawings of HRS and LRS of BM-SCO:MgO memristor, where the octahedral and tetrahedral layers of SCO are shown in rose red and light blue, respectively. The cyan pillar denotes MgO.

It is known that the designed vertical interfaces using two structurally incompatible oxides could generate higher concentration of oxygen vacancies than the regions at nonvertical interfaces (34). On the basis of this consideration, moderate charges were added at the vertical interfaces between BM-SCO and MgO to perform the COMSOL electric field simulation. As the voltage applied to the top electrode increases from 0 to +5 V, the electric field intensity at the vertical interfaces becomes increasingly higher (movie S1). Therefore, those localized electrical fields along the vertical interfaces under positive bias (Fig. 4H) will strongly attract the negatively charged oxygen ions from LSMO bottom electrode. Ultimately, the oxygen ion distribution can be well controlled by the localized electrical field at the vertical interfaces. For the BM-SCO device (fig. S25A), the (001)-oriented film has an ordered oxygen vacancy channels running IP, but no oxygen vacancy channels open to the film surface. The easy oxygen ion migration path between top and bottom electrodes is blocked by the alternative stacking of octahedral and tetrahedral layers along the OOP direction. Consequently, oxygen ions extracted from the bottom electrode migrate along the IP direction of the BM-SCO device, and the formation of conducting filaments along the OOP direction is therefore impeded. This accounts for the parameter instability observed in BM-SCO memristor (fig. S8). With the introduction of MgO nanopillars into the BM-SCO film (Fig. 4I), a high density (approximately terabit per square inch) of the vertical oxygen vacancy channels is created by the vertical interfaces in the VAN film. These vertical oxygen vacancy channels are effective guided filament regions, enabling precision resistive switching performance engineering, by eliminating the randomness and dimensionality of the conducting filaments (metallic PV-SCO_3−δ) in the BM-SCO phase. Thus, the preferential migration of oxygen ions extracted from LSMO bottom electrode can be manipulated along the vertical interfaces, connecting directly with top electrode and bottom electrode (fig. S25B). The schematic diagram of VAN structure under LRS state (right of Fig. 4I) shows that the PV-SCO_3-δ phase is formed along the vertical interfaces between BM-SCO and MgO. Therefore, much improved memristive performance in S₅₀M₅₀ VAN memristor is achieved. The model system of S₅₀M₅₀ successfully reveals the oxygen ions dynamics in VAN-based memristors.

DISCUSSION

In summary, epitaxial BM-SCO and BM-SCO:MgO films were designed to investigate the microscopic mechanism of resistive switching in memristors. We found that S₅₀M₅₀ VAN memristor exhibits electroforming-free, more endurable, and stable performance and various synaptic behaviors compared with BM-SCO memristor. A CNN was constructed on the basis of the S₅₀M₅₀ device for the realization of high-precision traffic sign recognition. The preferential oxygen ion migration along the vertical channels created by self-assembled vertical interfaces was directly visualized using in situ STEM. As the conduction channels are spatially confined at the vertical interfaces, the vertical and low-dimensional preferential path helps ensure uniformity in the switching parameters, thus exhibiting a better control over the device performance. This work significantly deepens the understanding of resistive switching mechanism in VAN structures and also demonstrates that S₅₀M₅₀ VAN is suitable for application as memory material in the recent future.

MATERIALS AND METHODS

Film fabrication and characterization

Fifty-nanometer-thick BM-SCO and BM-SCO:MgO nanocomposite thin films with 19-nm LSMO bottom electrodes were epitaxially grown on (001)-oriented STO substrates via pulsed laser deposition (PLD) system. A KrF excimer laser (λ = 248 nm) with the energy density of 1.1 J/cm² was used for the PLD fabrication of the BM-SCO, BM-SCO:MgO, and LSMO films at a 1-Hz repetition rate. The LSMO films were deposited at a substrate temperature of 750°C under an oxygen environment of 100 mtorr. The deposition of BM-SCO or BM-SCO:MgO was subsequently carried out on the LSMO layers under the same growth condition. The PLD chamber was evacuated to a base pressure of 1 × 10⁻⁴ Pa before introducing the pure oxygen. After deposition, the samples were cooled down with a 10°C/min cooling rate to room temperature while maintaining the identical oxygen pressure. Circular Pt top electrodes with the size of 30 μm by 30 μm were deposited ex situ on BM-SCO and BM-SCO:MgO films through magnetron sputtering. The crystalline structure and strain states of the films were examined by high-resolution XRD from a PANalytical Empyrean diffractometer with Cu Kα source (λ = 1.5406 Å). An Asylum Research MFP-3D atomic force microscope was used to analyze the surface morphology of the films with tapping mode. The x-ray photoelectron spectroscopy analysis was performed on a Thermo Fisher Scientific Nexsa instrument equipped with an Al Kα source (hν = 1486.6 eV). The binding energy of the O 1s, Co 2p, Mg 1s, and Sr 3d of the BM-SCO and S₅₀M₅₀ thin films was determined and corrected by referencing the C 1s peak to 284.6 eV. We further explored the influence of the orientation of oxygen vacancy channels on the resistive switching behavior of the VAN structure, with the corresponding structural and electrical characterizations presented in figs. S26 and S27, respectively.

Scanning transmission electron microscopy

The TEM samples were prepared using focused ion beam technique (Helios G4, FEI, USA). The atomic images of the nanocomposite film were characterized using a JEOL ARM300 STEM (Japan), equipped with a double spherical aberration (Cs) corrector and an x-ray EDS (JED-2300 Series) with two 158-mm² silicon drift detectors. The HAADF image was acquired with a probe convergence angle of 24 mrad and a collection angle of ~64 mrad. The EELS measurement was carried out using a Gatan K2 Summit camera. The EELS spectrum imaging was performed with a dispersion of 0.1 eV per channel and a 500-eV drift tube energy using a 4000-pixel-wide detector for the acquisition of Co L_2,3-edge signals. The atomic-scale in situ biasing STEM analyses were performed on the above Cs-TEM with a PicoFemto double tile biasing TEM holder (ZEPTools Technology Company) at I-t mode. A tungsten tip was used as the mobile electrode, which was precisely controlled by a piezoelectric system.

Electrical characterization

The electrical characteristics were recorded on a home-built probe station using Keysight B2902A, Keithley 2602B and Keithley 2636B source/measure units. During the measurement, the LSMO bottom electrode was grounded, while the Pt top electrode was biased. The switching characteristics and synaptic plasticity of the devices were measured by applying different scanning voltages and pulses. The conducting mechanisms of the BM-SCO and S₅₀M₅₀ memristors were analyzed by plotting log I versus log V curves, using ohmic and SCLC mechanisms.

Neuromorphic computing

A CNN was implemented using the PyTorch deep learning framework for feature extraction and classification. The model consists of three convolutional layers (3 × 3 kernels) with 32, 64, and 128 output channels, each followed by 2 × 2 max pooling. The extracted features are flattened and passed through three fully connected layers (512, 128, and 43 nodes) with ReLU activations, while the final layer uses softmax for classification. Data preprocessing included resizing images to 32 × 32 and applying augmentations such as Poisson and Gaussian noise, brightness adjustments, and rotations (±15°). Normalization was performed to ensure numerical stability. These operations were implemented via “torchvision.transforms” and integrated into the training pipeline. The dataset was split into 80% training and 20% testing subsets. The model was trained for 40 epochs using cross-entropy loss and the Adam optimizer, with a batch size of 64. Training performance was monitored through real-time recording of training loss and test accuracy after each epoch.

COMSOL simulation

A finite element method was used to map the electrical field distribution in the VAN device using the COMSOL Multiphysics 6.2 program. The dimensions of the BM-SCO matrix are 40 nm in width and 50 nm in height, while the MgO nanopillar takes 20 nm in width and 50 nm in height. We adapted the numerical model approach from previously reported works on resistive switching in oxides (58, 59).

Supplementary Materials

The PDF file includes:

Figs. S1 to S27

Sections S1 to S4

Legend for movie S1

References

Other Supplementary Material for this manuscript includes the following:

Movie S1