Tunable Hydrogen Dynamics Under Electrical Bias for Neuromorphic Memory Applications

Abstract

A wide variety of materials and device architectures have been explored for memristor applications targeting neural network simulations, most of which rely on oxide-based structures that exhibit resistive switching driven by oxygen-vacancy–mediated memory effects. In this study, we present a novel approach for modulating resistive and nonvolatile memory behavior in oxide semiconductors through the controlled injection and extraction of hydrogen. The proposed two-terminal device incorporates a hydrogen source layer that facilitates the diffusion of hydrogen ions into the active oxide matrix, where they form hydroxide (OH) bonds and locally modulate the electron concentration. This process induces a stable and reversible memory effect under an applied electric field. Hydrogen exchange predominantly occurs at the interface between the active and insulating layers, with the latter serving as a buffer to maintain an optimal hydrogen concentration. Furthermore, neural network simulations were performed by utilizing the synaptic characteristics controlled via hydrogen modulation, achieving a recognition accuracy of 97.2% on the MNIST data set. The effects of input data resolution and weight quantization on recognition performance were also systematically investigated and discussed.

1. Introduction

Efficiently processing large amounts of data is essential for many artificial intelligence-based technologies. In traditional von Neumann architecture systems, serial data transfer and physical separation of memory and logic cores create bottlenecks during the processing of massive data sets. This separation degrades performance and increases the energy consumption of the system during data processing. In recent semiconductor research, next-generation computing technologies that surpass traditional system architectures have gained significant attention. One notable area is “neuromorphic computing,” which aims to mimic the low power consumption of the human brain, which manages a complex network of 100 million neurons and 100 trillion synapses while operating with remarkable energy efficiency, consuming only approximately 20 W of power. To match the brain’s energy efficiency and performance, neuromorphic computing research is increasingly focused on developing new nonvolatile memory technologies, particularly resistive switching-based technologies known as memristors, as well as in-memory computing paradigms to improve energy efficiency.

Neuromorphic devices must fulfill several key requirements to effectively mimic brain function. First, they must exhibit conductance changes that can represent weights in response to external electrical signals. This hysteresis in electrical conductance mimics the function of artificial synapses by representing weights as binary states (0 and 1) or in multilevel form. Second, achieving fast switching speeds and high stacking densities is crucial to meeting the requirements of 3D integrated circuit architectures. In memristors, the two-terminal structure is emerging as a promising candidate for next-generation nonvolatile memory because it offers high stacking densities and simplified manufacturing processes. Compared to three-terminal devices, two-terminal configurations can reduce cell size, maximizing memory density while lowering manufacturing costs owing to their simpler structure. In addition, two-terminal devices are highly compatible with complementary metal oxide semiconductor (CMOS) processes, enabling seamless integration with existing semiconductor technologies. Their fast switching characteristics further enhance their operating speed, making them suitable for high-speed and low-power applications. These advantages are crucial for advancing memristor technology for future memory and neuromorphic computing applications. Researchers have made attempts to develop artificial synapses that can mimic the human brain with nonvolatile characteristics using next-generation memory devices such as phase change memory, magneto-resistive random access memory, and resistive switching random access (RRAM).

Among the next-generation memory candidates, the development of RRAM based on various materials has been the focus of extensive research because of advantages including fast switching speeds, low power consumption, high endurance, a simple two-terminal structure, and compatibility with CMOS fabrication processes. Most oxide semiconductor-based RRAM studies have been relied on the principle that resistive switching is governed by the migration of oxygen vacancies , under an applied electric field. Forming an oxygen-deficient oxide semiconductor layer is essential to improving memory performance through resistance modulation. This requires designing a material system and device architecture that promote the migration of substantial oxygen vacancies. Consequently, controlling oxygen vacancies has become vital for optimizing the performance of oxide semiconductor-based resistive memory devices. Techniques including the regulation of oxygen partial pressure during deposition, postdeposition annealing in an oxygen atmosphere, and interface or oxygen concentration modulation via plasma treatment have been explored to achieve this objective. , However, this approach poses challenges to the long-term operational stability of the device and exhibits structural weakness in terms of memory characteristics when applied to devices.

Recent study efforts have involved controlling the migration of externally injected elements rather than inducing the migration of conventional oxygen or oxygen defects. Onen et al. reported a switching device using Pd as a hydrogen storage and control gate in a WO₃-based device. The device operates by modulating channel conductance via the electrochemically controlled intercalation of protons (selected for their small radius and light mass) into WO₃. The utilization of hydrogen in oxide semiconductors has been studied through approaches including memory resistors, transistors using oxides, and hydrogen movement in oxide semiconductors. In amorphous indium–gallium-zinc oxide (a-IGZO), which is a representative oxide semiconductor, hydrogen serves as an n-type dopant and induces an increased electron concentration when it is injected. , The hydrogen in a-IGZO serves as a shallow donor when combined with oxygen to form −OH bonds, , which increases the electron density, causing the Fermi level to shift closer to the minimum conduction band. Hydrogen can also affect the role of oxygen vacancies by deactivating or reactivating them, depending on its interaction with the a-IGZO matrix.; , this can change the Fermi level of a-IGZO by enhancing or degrading its resistive switching properties , and affecting switching uniformity and durability. Therefore, precise control of hydrogen incorporation through process optimization is critical to improving the performance and reliability of a-IGZO-based RRAM devices. However, precise incorporation control is challenging because hydrogen is a highly mobile and small element. These inherent difficulties necessitate the development of innovative methods to effectively control the resistive switching characteristics of RRAMs.

In this study, a novel approach to regulate the resistive switching characteristics of a-IGZO-based RRAM devices through the controlled introduction of externally supplied hydrogen was proposed. In this approach, an external electric field induces hydrogen ions to drift into the a-IGZO active layer, thereby modulating the resistive switching behavior. The hydrogen supply layer comprises a silicon nitride (SiNx) film deposited via plasma-enhanced chemical vapor deposition (PECVD), ensuring compatibility with CMOS fabrication processes. The externally applied electric field facilitates the drift of the injected hydrogen into and out of the a-IGZO layer. The proposed hydrogen-controlled a-IGZO device demonstrates precise tunability of its resistive switching characteristics in response to the applied voltage and exhibits stable switching performance even after prolonged operation.

2. Results and Discussion

2.1. Uniform Resistive Switching of the Hydrogen-Injected Memristor Structure

In this study, a hydrogen injection-controlled (HIC) structure was designed to measure the electrical and structural properties of hydrogen-controlled resistive memories. Figure a illustrates that both terminal electrodes are composed of Ti/Au (5/50 nm). The a-IGZO active layer in which resistance is modulated is deposited using a sputtering system under an oxygen partial pressure of 10%, resulting in a 40 nm thin film. A 10 nm-thick SiNx thin film is used as the hydrogen source layer to facilitate the forced injection of hydrogen into the a-IGZO layer. SiNx is widely used in CMOS semiconductor processes and is conventionally deposited via PECVD using silane (SiH₄) and ammonia (NH₃) as precursor gases. During PECVD, these precursors undergo plasma-induced reactions, forming Si–H or N–H bonds, which can be readily dissociated within the material, thereby enhancing the mobility of hydrogen and enabling its transformation into different chemical states. Consequently, thin SiNx films contain substantial mobile hydrogen. Studies have demonstrated that the hydrogen concentration within these films depends on the precursor ratio during deposition. When SiNx thin films with a high mobile hydrogen content are directly deposited onto an a-IGZO layer, the resistance of the layer changes readily because of modifications in the bonding state of the a-IGZO surface. , This phenomenon is primarily attributed to the hydrogen plasma bombardment that occurs during the deposition process. A substantial quantity of hydrogen ions is injected into the underlying material owing to plasma bombardment and forced injection during the hydrogen plasma process. Given their low mass and high mobility, these hydrogen ions gradually diffuse into the a-IGZO layer. This migration influences the resistive properties of the a-IGZO layer and considerably impacts the operational characteristics and retention stability of the resistive memory device.

1.

(a) Schematic of the proposed stacked structure of the memory device with hydrogen injection and control (Au (top)/SiNx/SiOx/IGZO/Au (bottom)), (b) cross-sectional TEM image and EDS color mapping showing the elemental distribution of each layer in the device, (c) current–voltage characteristic curve of the top electrode, with SET proceeding at a positive voltage and a counterclockwise hysteresis loop, (d) cumulative probability distribution of the high and low resistance states of the device, (e) input pulse width and voltage used for the potential application and depressurization in the HIC device, and (f) electrical observations of the gradual rise and fall of conductance with the applied pulse.

To effectively regulate uncontrolled hydrogen migration, a 10 nm thin silicon oxide (SiOx) film is incorporated between the a-IGZO and SiNx layers, functioning as a hydrogen buffer and reservoir layer. Hydrogen ions (H⁺) drift in the direction of an applied external electric field, promoting hydrogen diffusion within the SiOx layer via the Grotthuss mechanism. , The hydrogen ions form Si–H or Si–OH bonds through this proton-hopping process, establishing distinct draft pathways through the SiOx layer and enabling their relocation within the material. , These mechanisms are crucial in regulating hydrogen diffusion and efficiently managing hydrogen transport. Additionally, SiOx possesses excellent insulating properties and a wide bandgap, which aid in controlling hydrogen migration in the absence of an external electric field, thereby contributing to enhanced device retention properties. The fabricated samples (Figure a) were analyzed using transmission electron microscope (TEM) cross-section imaging to determine layer thickness and energy dispersive X-ray spectroscopy (EDS) color mapping to assess material distribution (Figure b). Electrical characterization was conducted after confirming that all constituent materials in each layer were deposited under well-controlled conditions.

The electrical characteristics were assessed by applying a positive voltage to the top electrode while grounding the bottom electrode. As shown in Figure c, the HIC device with a SiNx layer functioning as the hydrogen supply layer exhibited memory characteristics. Each data set includes measurements obtained at the same voltage over 50 iterations. An increase in current was observed upon applying a positive voltage to the upper part of the device, as shown in Figure c. An approximately 15-fold increase in current was observed at 0.5 V. In the SET state, the device demonstrated electrical characteristics similar to those of previously reported oxide-based memristors. In addition, the low resistance state (LRS) was retained after removing the voltage, indicating stable retention. This switching was restored to the initial high resistance state (HRS) when the opposite voltage was applied to the device, indicating that reversible resistive switching is possible. The current exhibited counterclockwise hysteresis as the voltage switched from positive to negative, as indicated by the directional arrows in Figure c. This hysteretic behavior is characteristic of resistive switching devices, which switch between different resistance states (e.g., HRS and LRS) when a voltage is applied. The current response follows a loop, indicating the device’s ability to retain and switch between these resistance states depending on the voltage applied. To evaluate the device-to-device and batch-to-batch uniformity, a total of 60 devices (20 devices from each of three independent batches) were analyzed under identical bias conditions. The corresponding statistical distributions (Figure S1) exhibit narrow variability at each voltage, confirming the excellent reproducibility and reliability of the proposed device architecture. Figure d illustrates the stability of the HIC device’s resistance measured at 0.5 V over 30 sweep cycles as a cumulative probability. Additionally, Figure d illustrates the stable resistance states (LRS/HRS) of the HIC device, measured at 0.5 V over 30 sweep cycles. This reveals the device’s ability to reliably switch between high and low resistance states over multiple cycles, indicating its stability and potential for nonvolatile memory applications. The device’s ability to maintain a stable resistance value over multiple cycles suggests its suitability for applications that require repetitive switching, such as neuromorphic computing and memory storage. To clarify the role of hydrogen supplied from the SiNx layer, electrical characteristics were also investigated for a control device without the SiNx layer (Au/SiOx/a-IGZO/Au) (Figure S2). In the device incorporating SiNx, hydrogen generated as a byproduct during SiNx deposition can be readily injected into the SiOx and IGZO interfaces under a positive bias applied to the top electrode, resulting in the SET process. In contrast, in the control device where only a SiOx insulator is formed on top of IGZO, a SET process is observed only under a negative bias applied to the top electrode. This behavior corresponds to a typical oxygen-vacancy-driven interfacial gradual resistance switching mechanism.

Programming the memristor element is critical to using it as an operator that can set a desired resistance or conductance (weights) value through a potential rise and erase or modify this value through a depression. Figure eillustrates the voltage and width of the pulse shapes used for the potentiation and depression, employing identical pulse trains. During potentiation, 5.5 V was applied with a pulse width of 6 μs, and current sensing was performed at 0.5 V. During depression, the same readout voltage of 0.5 V was maintained and a pulse width of 9 μs was used. As shown in Figure f, the electrical conductivity of the hydrogen-controlled RRAM device changed gradually in response to the applied pulse when the voltage was fixed at 0.5 V. Based on the synaptic behavior, the linearity factor (α) was calculated using the following equation

$$G=\left\{\begin{array}{l}{[(G_{\max }^{\alpha }-G_{\min }^{\alpha })\omega +G_{\min }^{\alpha }]}^{1/\alpha }(\alpha \ne 0)\\ {G_{\min }(G_{\max }/G_{\min })}^{\omega }(\alpha =0)\end{array}\right\}$$ 1

where G _max and G _min represent the proportionally increased synaptic weights until the maximum and minimum synaptic weight values are reached, respectively. In an ideal synaptic behavior, the internal variable, ω, ranges from 0 to 1, and α = 1. Under 50 repeated pulses, the linearity of potentiation (α_pot) and depression (α_pot) on the hydrogen-controlled RRAM were 0.33 and 0.48, respectively. Additionally, a proportional but moderate conductance modulation was observed, with, a conductance ratio between G _max and G _min of 3.4, indicating stable analog weight updates. The synaptic characteristics shown in Figure f demonstrate a highly linear evolution of conductance even under identical pulse input conditions. This behavior is clearly distinct from the barrier-modulation-driven gradual switching typically observed in oxygen-vacancy-based interfacial switching mechanisms. In such interfacial switching, charge transport is governed by modulation of the interfacial barrier height, resulting in an exponential dependence of current on the barrier properties and, consequently, inherently nonlinear conductance changes. In contrast, in the hydrogen-injection-based switching proposed in this work, hydrogen is introduced at the interface and subsequently diffuses into the bulk of the active layer. As a result, the conductivity of the entire active layer is gradually modulated, leading to continuous and uniform conductance updates and, thereby, highly linear synaptic characteristics.

2.2. The Behavior of Hydrogen and Oxygen at the Interface

Under an electric field, HIC devices exhibit interesting electrical properties, most likely as a result of hydrogen drift. Time-of-flight secondary ion mass spectrometry (TOF-SIMS) and X-ray photoelectron spectroscopy (XPS) were conducted to confirm the introduction of hydrogen into the a-IGZO active layer as a result of electric field application. The application of these techniques enabled the analysis of hydrogen content, elemental distribution, and chemical states within each layer, providing evidence of hydrogen incorporation and validating the hydrogen doping process of the material. Figure a presents the SIMS results, which quantify hydrogen concentrations across the layers of the HIC device. Furthermore, SiNx fabricated via PECVD contains a high hydrogen concentration, with measurements exceeding 10²⁰ atoms/cm³ obtained in this study. The hydrogen concentrations in the SiOx and a-IGZO layers were approximately 10¹⁹ and 10¹⁸ atoms/cm³, respectively. Although the interfaces between SiNx and SiOx were well-defined, the hydrogen distribution at the SiOx/a-IGZO interface exhibited distinct variations between the SET and RESET states. The interface between SiOx and a-IGZO was defined based on the distribution of indium (In) within the a-IGZO layer to precisely determine the boundary. When a positive voltage was applied to the top electrode of the HIC device, the hydrogen drift became evident upon switching to the SET state, as shown in Figure a, with hydrogen accumulating at the SiOx/a-IGZO interface. The hydrogen ions that penetrated the SiOx/a-IGZO interface diffused into the a-IGZO layer, slightly increasing the overall hydrogen concentration in this layer.

2.

Compositional analysis to determine the hydrogen content and oxygenation state. (a) Analysis of the hydrogen depth profile at the interface and in the bulk of the device in the SET and RESET states using SIMS, (b) analysis of the oxygen bonding states in the SET and RESET states of the HIC device using XPS, and (c) percentage of hydroxide bonded to hydrogen at the interface.

Conversely, during the RESET state, in which the opposite voltage is applied, a significant reduction in hydrogen concentration was observed at the SiOx/a-IGZO interface, indicating hydrogen ion expulsion from the a-IGZO layer. Notably, the expelled hydrogen ions did not return to their source (supply layer), i.e., the SiNx layer or the SiNx/SiOx interface, but instead remained localized at the SiOx/a-IGZO interface. This behavior suggests that hydrogen ion migration was driven by the concentration gradient between the two interfaces, with ions released from one interface, traversing the oxide layer, and subsequently becoming trapped near the opposite interface. Consequently, hydrogen ions remained concentrated at the interface throughout the process. These findings indicate that SiOx functions as a buffer layer and hydrogen reservoir and is crucial in regulating hydrogen ion migration and retention within the device.

SIMS further confirmed the pronounced activity of hydrogen at the SiOx/a-IGZO interface. XPS depth profile analysis was performed to investigate changes in the oxygen states at the interface, the corresponding bonding states, and the distribution of oxygen and hydrogen. The XPS measurements were conducted using a depth profiling technique in which the signals and bonding states of each layer from the top Au electrode, through the SiNx/SiOx/a-IGZO stack, to the bottom Au electrode (Figure b) were analyzed. The depth position was assigned as a notation for the depth direction in this profiling analysis. The SiOx and a-IGZO interface was identified based on the Si 2p and In 3d signals, which are plotted as a function of depth position on a logarithmic scale in Figure c. The interface was centered around depth position #31, spanning positions #30 to #32. XPS revealed a significant variation in the oxygen binding state across the SiOx/a-IGZO interface. The oxygen signal obtained from the XPS can be categorized into three regions: approximately 530.2, 531.3, and 532 eV. These regions correspond to metal–oxygen (M–O) bonds, oxygen bonds weakened by oxygen deficiency, and hydroxyl (−OH) groups, respectively. The blue deconvolved XPS peaks are primarily attributed to oxygen deficiency, which alters bonding configurations owing to changes in coordination number. The XPS peak associated with oxygen deficiency was most pronounced at depth position #32, corresponding to the bulk a-IGZO layer. However, at depth position #31, which represents the interface between a-IGZO and SiOx, a notable difference between the SET and RESET states was observed. The relatively lower XPS peak intensity in the SET state suggests that hydrogen accumulation at the a-IGZO interface effectively passivates oxygen deficiencies. This phenomenon, which is commonly observed in oxides, is known as “multicenter bonding”. In this phenomenon, hydrogen interacts with its surrounding environment in a manner similar to oxygen in oxygen vacancies. The enhanced multicenter bonding effect in the SET state relative to that seen in the RESET state at depth position #31 explains the reduced XPS peak associated with oxygen deficiency.

Additionally, the XPS signal corresponding to hydroxyl groups diminished as the analysis progressed toward the bulk a-IGZO layer. This trend is illustrated in Figure c, in which the hydroxyl ratio is represented by a color map, with red indicating a high hydroxyl content and white representing a lower ratio. In the SET state, the hydroxyl ratios at depth positions #31 and #32 were significantly higher than that at depth position #30, suggesting that hydrogen-related species diffuse from the SiOx layer into the deeper a-IGZO regions. However, in the RESET state, the hydroxyl ratio remained high only at depth positions #28 and #29, which are located before depth position #30. The color distinction in Figure c clearly differentiates the SiOx and a-IGZO regions based on hydroxyl distribution. These findings align with the previously discussed SIMS analysis, confirming that hydrogen accumulates at the SiOx/a-IGZO interface and migrates into the a-IGZO layer or across the SiOx interface, depending on the SET/RESET state.

2.3. Hydrogen Movement and Resistance Mechanism Governed by Potential Difference

The compositional analysis of each layer revealed that the distribution of hydrogen atoms and hydroxyl molecules varies across layers, materials, and electrical states of the device. The active layer, a-IGZO, was deposited through radio frequency (rf) sputtering using only argon and oxygen plasma. However, the hydrogen content in this material is within 10¹⁷–10¹⁸ atoms/cm³, which is consistent with the values observed in this study. SiOx and SiNx have relatively high hydrogen content compared to the active layer; nonetheless, studies have confirmed that it is challenging for hydrogen to enter the active layer by deposition alone. The mobile hydrogen capable of migrating into the active layer is primarily generated by the hydrogen plasma during SiNx formation, which accumulates at the SiNx/SiOx interface. Subsequently, a substantial quantity of the mobile hydrogen in SiOx is transported toward the active layer or the SiOx interface under the influence of an electric field. Figure illustrates the hydrogen drift mechanism in the HIC device. The SiOx layer, which functions as a buffer and reservoir for hydrogen, facilitates hydrogen migration into the a-IGZO layer upon the application of an electric field. This drafting predominantly occurs at the SiOx/a-IGZO interface. Furthermore, the electric field drives the hydrogen introduced into the a-IGZO layer deeper into it. When a negative voltage is applied to the HIC top electrode, the injected hydrogen moves to the SiOx/a-IGZO interface or into the SiOx layer. This hydrogen modulation is crucial in altering the electrical properties of the device through two primary mechanisms. Hydrogen injection into IGZO not only increases the electrical conductivity, but also reduce the concentration of pre-existing oxygen vacancies through the formation of multicenter bond s. This interaction stabilizes cation dangling bonds below the valence band, effectively eliminating defect states within the bandgap.

3.

Schematic representation of the voltage-dependent hydrogen drift in an HIC device and the role of each layer. Hydrogen is injected with the creation of the SiNx layer and, as hydrogen is injected into the SiOx, a substantial amount is distributed at the interface. The voltage-mobilized hydrogen migrates into the oxide semiconductor layer, changing the electrical conductivity of the oxide semiconductor layer.

The first mechanism is the hydrogen-induced modulation of the carrier concentration in a-IGZO. Hydrogen increase the electron density in the active layer, leading to enhanced electrical conductivity and partial passivation of intrinsic oxygen-related defects. The second effect arises from the movement of charged hydrogen atoms, which generates an image charge across the SiOx potential barrier, resulting in an image-force-induced barrier-lowering effect at the potential barrier; this effect maximizes the memory effect of HIC devices.

The current conduction in a memristor device is primarily determined by the dielectric layers, material properties of the oxide semiconductor, and interfacial properties between layers. These factors collectively determine the conduction mechanism, which can be broadly classified into two categories: interface- and bulk-limited conduction. The dominant conduction mechanism in a device depends on the resistive state of the memristor, magnitude of the applied electric field, and the operating temperature. In addition, multiple conduction mechanisms can coexist within the same device, with the relative contribution of each mechanism varying with the operating conditions and the resistive state of the memristor. Based on these complexities, the unique material properties and external operating parameters must be considered in the design and optimization of memristor devices for specific applications.

In this section, we explore different conduction mechanism models for a-IGZO-based memristors by matching their current–voltage (I–V) characteristics. Figure a illustrates the HRS and LRS of the I–V curve in the positive bias section, which can be divided into respective regions owing to the substantial difference in slope in the curve. In the presence of a Schottky barrier, the primary conduction mechanism is often considered to be Schottky emission (thermionic emission) (Region I in Figure a). In this mechanism, electrons are thermally excited and subsequently injected into the conduction band of the dielectric material, overcoming the Schottky barrier, as shown in Figure b. During this process, the image-force barrier-lowering phenomenon occurs as a result of movement of the charged hydrogen injected by the electric field; this effect is significant because the amount of hydrogen injected increases as the electric field becomes stronger. The emitted electron current density (J) can be described by

$$J=A{T}^2\exp \left(\frac{-q{\mathrm{\Phi }}_{\mathrm{B}}}{\mathit{kT}}\right)$$ 2

where A is Richardson’s constant, T represents the absolute temperature, q is the electric charge, Φ_B denotes the junction barrier, and k is the Boltzmann constant. According to eq , the relationship $\ln \left(\frac{J}{T^2}\right)\propto \sqrt{E}$ is observed at constant temperature, which is confirmed by the linear relationship between ln­(J) and √E, as shown in Figure b. These results suggest a consistent correlation between the current density and the applied electric field and temperature. An effective barrier height of 0.83 ± 0.13 eV, as calculated from the experimental data, occurred in the low-field region (<∼0.3 MV/cm). When the electric field exceeds approximately 0.3 MV/cm, the current conduction in the hydrogen-controlled RRAM cannot be explained using the thermionic emission mechanism alone. At higher electric fields, the conduction mechanism transitions from thermionic emission to thermionic field emission (region II in Figure a). Hydrogen doping in a-IGZO increases the electronic conductivity of the material while pushing the Fermi level up near the conduction band. The I–V characteristics were interpreted using thermionic field emission, as shown in Figure c, to confirm this behavior, with the barrier height found to be 0.31 ± 0.11 eV. This decrease in barrier height allows more electrons to move, with the interaction between hydrogen doping and the applied strong electric field promoting electron mobility, thereby changing the conduction mechanism.

4.

Analysis and schematic of electron transport in HIC devices. (a) Electric field applied to the HIC device; the resulting current density and distribution of the current density as functions of the electric field strength, divided into three regions (I, II, and III) based on electron conduction mechanisms. (b) In Region I, electrons are transported through thermionic emission (inset illustrates barrier height as a function of temperature). (c) In Region II, electrons are transported across the barrier through thermionic field emission. (d) In region III, electrons are transported by the interfaces or traps in SiOx through Poole-Frankel conduction. The inset illustrates the second derivative I–V characteristics of a tunneling device that uses SiOx (Au/SiOx/Au).

Any material used as a dielectric layer in a memristor device can generate traps, which often contain defects and impurities that can contribute to electron conduction. These defects and impurities can be inherent properties of the material or can be caused by the high electric fields applied during the device’s initialization process through a phenomenon called “forming”. Traps in dielectrics and at interfaces can be caused by the electrical stress generated during the forming process prior to device operation. This high electrical stress can create various types of subgap states within the bandgap of the dielectric, interfering with electron travel paths or creating unintended paths. Curves I–V in Figure d align with the Poole–Frankel (P–F) emission model at high electric fields (>0.5 MV/cm) over a constant temperature range (303–373 K). Poole-Frankel emission is the dominant conduction mechanism under high electric fields, explaining the electron transfer between traps inside the bulk. The applied strong electric field drives the electrons trapped within SiOx to move between traps and cross the potential barrier. The inset in Figure d illustrates the subgap state that occurs during the forming process and is measured by a tunneling device (metal/SiOx/metal) used to observe the tunneling of inelastic electrons.

The energy level and physical location of the corresponding trap can be estimated by observing the positive polarity of the second derivative I–V signal, which indicates the soft breakdown of the dielectric that occurs during the forming of the device. The shape of these tunneling signals can be related to trap-assisted tunneling or charge-trapping processes. A distinctive feature of the second derivative I–V at low voltages is the valley-to-peak shape signal, which presumably is generated due to the shallow traps associated with P–F emission. The positions of the peak-to-valley and valley-to-peak signals reveal a similar attraction to the energy level of the trap in the dielectric material. As shown in Figure S3, temperature-dependent I–V characteristics further comfirm this transition behavior. The conduction mechanism evolves from thermionic field emission at low electric fields to P–F conduction at higher fields, supporting the filed -driven nature of charge transport in the hydrogen-controlled RRAM device.

2.4. Operational Reliability Based on Pulse Signal Input

Endurance characteristics are one of the essential parameters of RRAM memory devices. Changes in current or resistance values owing to repeated device operation cause long-term operation reliability problems in a device. A pulse-type input was used to sequentially switch the SET and RESET states of the HIC device to control the injection and expulsion of hydrogen. To quantitatively evaluate cycle-to-cycle variability, we measured the current distribution of the HRS and LRS states over more than 10⁴ switching cycles, as shown in Figure a. The statistical results demonstrate that both states are well separated with narrow variation, confirming stable and reproducible switching behavior. During more than 10⁴switching cycles, the memory characteristics remained stable, as shown in Figure a. The LRS exhibited a consistent current without a noticeable deviation, whereas the HRS demonstrated a slight variation. This change in current was minimal. Furthermore, even when the cycle-to-cycle variability was analyzed across the entire endurance test window, no significant degradation or overlap between HRS and LRS was observed. Extrapolation of the current trend indicates that more than 10⁹ pulse cycle iterations would be required to reduce the current gap value, which is a memory characteristic, by less than half. The long-term memory capability of an RRAM device is an essential functionality in performing cognitive operations to retain computed weights. In memristor applications, the device plays two main roles: as an operator, and in remembering and recognizing calculated weights in order to utilize them. Retention, which is the ability to maintain long-term memory, is used in the operator function but is more crucial in the recognizer role. The proposed HIC device demonstrated excellent retention capability, as the current (or resistance) gap, a key measure of memory performance, did not exhibit a decreasing trend (Figure b). In particular, direct retention measurements up to 10⁶ seconds confirmed that both the LRS and HRS states remained clearly separated with negligible degradation. This experimental evidence supports the device’s ability to sustain long-term memory without relying on extrapolation. The stable resistance demonstrated during pulse application and the sustained retention of memory characteristics indicate that the hydrogen injected during the SET process remained immobilized and did not migrate or escape during repeated read operations. This is one of the main functions of a SiOx layer, as mentioned earlier: to act as a buffer and as a barrier that can prevent hydrogen ions from returning to the buffer layer. This high retention capability is a typical characteristic of nonvolatile memory, which retains information even when the electricity supply has been disconnected.

5.

Long-term operation reliability and cognitive computation simulation: (a) durability measurements of the HIC device and (b) retention measurements. (c) Recognition rate based on image resolution when using five weight levels through weight quantization. (d) Recognition rate variation based on the number of quantized weight levels used for learning and cognitive operations and the resolution size of the input signal.

Finally, based on the potentiation and depression characteristics shown in Figure f, a multilayer neural network was constructed using a backpropagation algorithm. The network consisted of an input layer, two hidden layers, and an output layer, following a conventional multilayer perceptron (MLP) architecture. The MNIST data set, Modified National Institute of Standards and Technology, was employed as the input data. Each input image in the MNIST data set is composed of a 28 × 28 pixel matrix. To evaluate the impact of input resolution on image recognition performance, the original MNIST images were downsampled to resolutions of 28 × 28, 15 × 15, 10 × 10, 8 × 8, 5 × 5, 4 × 4, and 3 × 3. Furthermore, weight quantizationa fundamental technique in machine learningwas applied to implement training using quantized synaptic weights derived from the 50-pulse potentiation/depression data shown in Figure f. Only five discrete weight levels were used during the training process for handwritten digit recognition. Figure c presents a comparison of the recognition accuracy as a function of input resolution when five quantized weight levels were adopted. The number of training epochs was limited to ten for all cases. Even with only five quantized weight levels, the network achieved a recognition accuracy exceeding 90% when the input resolution was maintained at 8 × 8 or higher, indicating that such resolution is sufficient for a relatively simple data set like MNIST. Figure d illustrates the relationship between the number of quantized weight levels, input resolution, and recognition accuracy. The results suggest that for simple image data sets such as MNIST, a large number of memristor weight levels is not necessary. The difference in recognition accuracy between a memristor system capable of representing 50 weight states and one limited to five states is negligible. However, the input resolution plays a crucial role in improving recognition accuracy. To achieve a recognition rate above 80%, at least a 5 × 5 input resolution (i.e., 25 input nodes) is required.

3. Conclusions

This study presents a novel memristor device designed for weight computation and memory functions in neural network hardware accelerators. The device employs a new resistive switching strategy in which hydrogen injection into the a-IGZO active layer is precisely regulated through a SiOx interface. This mechanism provides a reliable means of tuning resistance states, essential for efficient memory and computation. A distinctive feature of the design is the dual use of SiNx and SiOx layers, serving as a hydrogen supplier and buffer, respectively. Hydrogen primarily resides at the SiOx/a-IGZO interface and is injected or expelled under electrical bias. The applied electric field promotes hydrogen migration, enabling stable and reversible switching. As a result, the device functions as a nonvolatile memory element highly compatible with CMOS fabrication techniques. Electronic conduction is governed by thermionic emission and electron transfer between traps, with SiOx traps playing a critical role in modulating charge transport and enhancing stability. The durability of the hydrogen-based memory state was verified through repeated SET/RESET cycles and long-term retention tests, which demonstrated stable programmability and persistence of the SET state. In contrast to conventional memristors, which rely on oxygen-vacancy-driven switching, the HIC device operates via controlled hydrogen migration. The HIC-based device exhibited stable retention characteristics even after more than 10⁶ seconds, while also demonstrating symmetric and reliable synaptic behavior. Based on these synaptic properties of the HIC, numerical simulations for handwritten digit recognition using the MNIST data set were performed. The influence of input image resolution and the limited number of available weight levels on recognition accuracy was systematically analyzed and discussed. These findings are expected to provide valuable insights for the development of various hydrogen-based semiconductor systems and next-generation memory devices.

4. Experimental Section

4.1. Device Fabrication

All devices were fabricated at the Institute of Next-generation Semiconductor Convergence Technology fabrication facility at Daegu Gyeonbuk Institute of Science & Technology in a clean room maintained below 1000 classes. The bottom gold electrode was sequentially stacked with Ti (adhesion layer) using e-beam evaporation (SORONA, SRN-200). The bottom electrode was patterned using photolithography, and a 40 nm-thick a-IGZO layer was deposited using rf-sputter (SORONA, SRN-110) at a 9:1 ratio of Ar and oxygen gas at room temperature. The SiOx (10 nm) and SiOx (10 nm) layers were deposited at 350 °C using PECVD (KCT, KCT-800). The SiOx and SiNx precursors were SiH₄ and N₂O applied at a ratio of 15:2000 sccm and SiH₄ and NH₃ applied at a ratio of 30:15 sccm, respectively. Finally, the top electrode used an e-beam evaporator (SORONA, SRN-200) to deposit the Ti/Au electrode.

4.2. Device Measurement

The I–V characteristics of the device were measured using a Keithley 2636b and 707A switching box, with a TRM250 (Nanoscope systems) probe station and optical scope. An Agilent 81150 arbitrary pulse generator and a Keithley 4200 PMU unit were used for pulse measurements. Unless otherwise noted, all electrical measurements were performed on device with a cross-point active area 50 μm × 50 μm. Experiments on HIC devices were performed at 50 °C for I–V and pulse injection and metrology. To examine the dependence on interfacial area, additional electrical measurements were separately carried out on devices with four different active areas. As shown in Figure S4, the resistance values in both the HRS and the LRS exhibit a systematic and sequential variation as a function of device area.

4.3. Material Characterization

Cross-sectional high-resolution TEM and EDS were performed using a field emission gun TEM (Thermo Scientific, Titan Themis Z). Prior to TEM characterization, the sample was pretreated using a focused ion beam to ablate the cross-section. Depth profile analysis was conducted for hydrogen and oxygen using SIMS (IONTOF, TOF-SIS 5–100) and XPS (Thermo Scientific, ESCALAB 250xi), respectively, for the entire depth of the sample. The components were analyzed using SIMS with a Bi ion source at 30 keV. Ion sputtering was performed to analyze the depth profile in SIMS using O₂ for In, Ga, and Zn, and Cs atom for hydrogen and oxygen. The hydrogen signal in SIMS was quantified at the Evans Analytical Group, and reference samples were used to quantify hydrogen in SiOx, SiNx, and a-IGZO. XPS was used to measure elemental binding energies using Al Kα at 1486.6 eV. Depth profile analysis was performed at 38 depths along the depth direction of the sample, and Ar ions were accelerated to 3 kV and used in the etch process. An optical microscopy image of the fabricated device, showing the top and bottom electrode pattern and the defined active area, is provided in the Supporting Information (Figure S5) to complement the schematic illustration and TEM image in Figure a,b.

4.4. Neural Network Simulation

Handwritten digit images obtained from the MNIST data set were divided into 28 × 28 matrices, yielding 784 data points per image. To investigate the effect of image resolution on recognition performance, the original images were further downscaled into matrices of 15 × 15, 10 × 10, 8 × 8, 5 × 5, 4 × 4, and 3 × 3, thereby generating data sets with varying spatial resolutions of identical digit patterns. For weight quantization, potentiation and depression characteristics were extracted from pulse measurements (50 rising and 50 falling pulses), and the resulting conductance states were quantized into 25, 10, 8, and 5 discrete weight levels, respectively, to evaluate the impact of reduced quantization granularity on learning performance. Hardware-based training simulations were conducted using the IBM Analog Hardware Acceleration Kit, enabling an accurate assessment of device-level characteristics within neuromorphic computing frameworks.

The data that support the findings of this study are available from the corresponding author upon reasonable request.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsami.5c21475.

• Experimental method details (fabrication, measurement, characterization, neural network methodological details), electrical measurement data for comparison groups, electrical characteristic data by batch, electrical characteristics vs temperature, electrical characteristics vs area trend, actual device images and optical microscope photographs (PDF)

⊥.

J.W. contributed equally to this work. This project was supervised and directed by H.-J.L. H.N. conceived the study. H.N. and H.-J.L. designed the experiments. G.R.H. analyzed the transport properties, and J.W. and C.L. conducted the simulations. H.N. carried out the electrical and compositional analysis measurements. S.L. and M.L. provided guidance on interpreting the memristor. All authors contributed to discussions and result analyses. H.N. and H.-J.L. wrote the manuscript.

This research was supported in part by the National Research Foundation of Korea (NRF) grant funded by the Korean government (Nos. RS-2023-NR076512 and RS-2024–00442165) and in part by the Commercialization Promotion Agency for R&D Outcomes (COMPA) funded by the Ministry of Science and ICT (MSIT) (RS-2024–00417600), and in part by the DGIST R&D Program of the Ministry of Science, ICT, and Future Planning (25-ET-02).

The authors declare no competing financial interest.