Enhancement of Synaptic Performance through Synergistic Indium Tungsten Oxide-Based Electric-Double-Layer and Electrochemical Doping Mechanisms

Abstract

This study investigated the potential of indium tungsten oxide (IWO) channel-based inorganic electrolyte transistors as synaptic devices. We comparatively analyzed the electrical characteristics of indium gallium zinc oxide (IGZO) and IWO channels using phosphosilicate glass (PSG)-based electrolyte transistors, focusing on the effects of electric-double-layer (EDL) and electrochemical doping. The results showed the superior current retention characteristics of the IWO channel compared to the IGZO channel. To validate these findings, we compared the DC bias characteristics of SiO₂-based field-effect transistors (FETs) with IGZO and IWO channels. Furthermore, by examining the transfer curve characteristics under various gate voltage (V_G) sweep ranges for PSG transistors based on IGZO and IWO channels, we confirmed the reliability of the proposed mechanisms. Our results demonstrated the superior short-term plasticity of the IWO channel at V_G = 1 V due to EDL operation, as confirmed by excitatory post-synaptic current measurements under pre-synaptic conditions. Additionally, we observed superior long-term plasticity at V_G ≥ 2 V due to proton doping. Finally, the IWO channel-based FETs achieved a 92% recognition rate in pattern recognition simulations at V_G = 4 V. IWO channel-based inorganic electrolyte transistors, therefore, have remarkable applicability in neuromorphic devices.

1. Introduction

The advancement of artificial intelligence (AI) technology has drawn significant inspiration from the structure and functionality of the human brain. The human brain, comprising approximately 10¹¹ neurons and approximately 10¹⁵ synapses, forms a complex neural network that efficiently performs tasks such as problem-solving, learning, and memory while being highly energy-efficient [1,2,3,4,5,6]. Elucidation of biological systems can aid the development of neuromorphic computing systems, contributing to the progress of smart AI technologies capable of learning and memory functions akin to those of the human brain [7,8]. Recent studies have emphasized the application of neuromorphic technology, which mirrors human cognitive functions, to enhance the efficiency and adaptability of such systems [9,10,11]. These advancements highlight the potential of neuromorphic technologies across diverse domains such as image recognition, natural language processing, autonomous driving, and human–machine interfaces. They form a pivotal framework for developing artificial perceptual learning systems [12,13,14].

In this paper, we explore methods for constructing energy-efficient and highly parallelized neuromorphic systems by mimicking the synaptic functions of the human brain. Elucidating the operational mechanisms of the human brain, where synapses regulate the strength of neuron connections based on the intensity of stimulation to control learning and memory, is crucial for reliably implementing short-term and long-term memories in neuromorphic devices [15,16].

Recently, synaptic operations utilizing three-terminal devices based on charge-based field effects, electric double layers, electrochemical effects, and ferroelectric field effects have been demonstrated [17,18,19,20,21]. In particular, solid-state electrolyte-based electric-double-layer (EDL) and electrochemical effects have attracted attention because of their relatively stable chemical properties and cost-efficient processing [22,23]. Depending on the materials of the solid electrolyte and channels, as well as the engineering of device fabrication, these transistors could exhibit changes in channel operation mechanisms and electrical characteristics based on specific threshold voltages [24]. Specifically, synaptic characteristics based on the EDL mechanism were suitable for short-term synaptic plasticity, while synaptic characteristics based on the electrochemical effects mechanism were reported to be suitable for long-term synaptic plasticity [25,26]. Therefore, this study aimed to precisely control the EDL and electrochemical doping (ECD) processes in single synaptic devices based on solid-state electrolyte films responding to stimulus intensity and demonstrate the selective implementation of short-term plasticity (STP) and long-term plasticity (LTP) characteristics.

Indium gallium zinc oxide (IGZO) amorphous oxide semiconductors (AOSs) have been extensively utilized as channels in synaptic devices employing solid-state electrolyte films. However, IGZO is susceptible to moisture, leading to reduced device reliability [27]. To address this challenge, recent research has focused on indium tungsten oxide (IWO) without Ga/Zn components [28]. IWO exhibits high stability because of its resilience against moisture and acids, coupled with excellent electrical performance in thin-film transistors (TFTs), owing to its high electron mobility [29,30]. Consequently, the potential applicability of IWO channels in synaptic devices has been recognized.

On this basis, we fabricated synaptic transistors utilizing phosphosilicate glass (PSG) electrolyte films spin-coated with IWO AOS. To evaluate the superior electrical and synaptic characteristics of PSG-IWO transistors based on the EDL and ECD mechanisms, we conducted a comparative analysis with PSG-IGZO transistors. First, through gate bias tests, we observed that PSG-IWO transistors formed residual currents by doping protons in the channel at the ECD threshold voltage (U_th = 2 V) [31]. By fabricating IWO and IGZO over thermal SiO₂ films and applying U_th for proton doping and interface charge trapping to PSG-AOS and SiO₂-AOS, respectively, we confirmed that the IWO channel exhibited excellent residual current characteristics induced by proton doping [32]. Furthermore, by applying various gate voltage sweep ranges to the PSG-AOSs and analyzing the transfer curve operation characteristics based on the EDL and ECD processes, the advantages of IWO related to the doping and de-doping processes were confirmed.

To assess the synaptic characteristics of the proposed PSG-AOSs, we measured the excitatory post-synaptic current (EPSC) after single- and multiple-spike neural stimulation with spike amplitudes of 1 V (EDL) and 3 V (ECD). The measured results were compared using quantified data, such as the EPSC change ratio, EPSC gain, and paired-pulse facilitation (PPF) [33,34]. Ultimately, the successful emulation of STP and LTP was confirmed owing to the relatively superior EDL and ECD characteristics of PSG-IWO. Additionally, based on the normalized potentiation and depression (P/D) results from measurements at spike amplitudes of 4 V and 30 pulses, the nonlinearity factors, asymmetry ratio (AR), and dynamic range (DR) were extracted, and certain trends were identified [35,36]. Finally, using the Modified National Institute of Standards and Technology (MNIST) dataset, handwritten digit recognition rates were obtained through artificial neural network (ANN) simulations [37]. As a result, PSG-IWO achieved a recognition rate of 92% for the digit ‘9’ at epoch 20, demonstrating its suitability for successful artificial synaptic hardware network configurations. We anticipate that the electrical analysis-based evaluation method proposed for AOS synaptic devices in this study will be widely adopted in neuromorphic computing studies. Furthermore, the IWO channel may significantly contribute to the advancement of synaptic devices owing to its exceptional EDL and ECD characteristics.

2. Materials and Methods

2.1. Material Specifications

In this study, the transistors and metal–oxide–semiconductor (MOS) capacitors were fabricated using a common p-type (100) Si substrate with a resistivity of 1–10 Ω·cm (LG Siltron Inc., Gumi, Republic of Korea) for both devices. For the fabrication of the PSG electrolyte film, 20B spin-on glass (SOG) and P509 spin-on dopant (SOD) (both from Filmtronics Inc., Butler, PA, USA) were employed. Aluminum (Al) pellets (purity > 99.999%; THIFINE Corp., Incheon, Republic of Korea) were used to fabricate the transistor source/drain (S/D) and capacitor electrodes. For the transistor channel fabrication, IWO sputter targets (In₂O₃:WO₃ = 99:1 wt%, THIFINE Co., Ltd., Incheon, Republic of Korea) and IGZO sputter targets (In₂O₃:Ga₂O₃:ZnO = 4:2:4.1 mol%, THIFINE Co., Ltd., Incheon, Republic of Korea) were utilized.

2.2. Fabrication of PSG-Based AOS Channel Transistors

We fabricated the transistors using a PSG film with two different types of channels: IWO and IGZO. Initially, a PSG solution, composed of a blend of SOG and SOD solutions, was spin-coated onto p-type Si substrates. These substrates were cleaned using the wet-chemistry-based standard Radio Corporation of America (RCA) cleaning process. The spin coating was performed at 6000 rpm for 30 s. After the coating, an optimized series of annealing processes was carried out: pre-baking at 70 °C for 3 min, followed by baking at 100 °C, 150 °C, and 200 °C for 2 min each, and post-annealing at 650 °C for 1 h in forming gas (5% H₂ + 95% N₂). To ensure a reliable comparison of the characteristics of PSG film-based IWO and IGZO transistors, each with a film thickness of 300 nm, IWO and IGZO were sequentially deposited with a thickness of 30 nm on the same substrate (p-Si/PSG, 1 × 1 cm²) using RF magnetron sputtering. Finally, an Al film with a thickness of 100 nm was deposited using electron-beam (E-beam) evaporation and was patterned via lift-off to form the S/D electrodes. The dimensions (width × length) of the channel and S/D electrodes were 120 µm × 60 µm and 150 µm × 120 µm, respectively.

2.3. Characterization Method

To prevent changes in the EDL capacitance characteristics of the PSG film due to external humidity, the film was stored in an environment with a relative humidity of approximately 40%. Additionally, to minimize the variables caused by optical noise, all measurements were conducted inside a dark box. First, we measured the thicknesses of the PSG film formed by spin coating during the device fabrication process, as well as the Al deposited by E-beam evaporation and the IWO and IGZO materials deposited by sputtering. These measurements were performed using a Dektak XT Bruker stylus profiler (Bruker, Hamburg, Germany). Next, for the analysis of MOS-structure capacitor devices, we used an Agilent 4284A precision LCR meter (Hewlett-Packard Corporation, Palo Alto, CA, USA). To verify the EDL and ECD characteristic mechanisms of the fabricated transistors, a 2231A-30-3 precision DC power supply (Keithley Instruments, Cleveland, OH, USA) and an Agilent 4156B precision semiconductor parameter analyzer (Hewlett-Packard Co., Palo Alto, CA, USA) were used in conjunction. Finally, to implement the synaptic characteristics, two identical Agilent 8110A pulse generators (Hewlett-Packard Co., Palo Alto, CA, USA) were also utilized.

3. Results and Discussion

3.1. EDL and ECD Mechanisms of IWO and IGZO

To mimic the synaptic characteristics based on the EDL and ECD mechanisms applied to the proposed device, elucidating the proton characteristics within the electrolyte film through the frequency-dependent capacitance behaviors of the PSG film is essential [38]. Protons form an EDL at the interface between the PSG film and the AOS channel when an electric field is applied. In certain cases, protons penetrate the AOS channel from the electrolyte film, enhancing the current retention characteristics of the channel. Specifically, when an electric field is applied, the O-H bonds in the P-OH groups within the PSG dissociate, and protons migrate to the interface via continuous hopping. Furthermore, at a certain threshold voltage (U_th), the protons accumulated at the interface dope into the channel, acting as electron donors and forming a current path in the n-type AOS channel [39,40,41]. Consequently, elucidation of the channel formation criteria based on these two mechanisms, according to the strength of the gate voltage (V_G), allows for precise implementation of synaptic behavior.

First, the threshold voltage (V_th) was defined from the transfer curve of the device, and the proton-doping voltage (U_th) was defined through the DC bias test [42]. We then analyzed the drain current (I_ds) characteristics of the IWO and IGZO channels formed on the same PSG thin film in response to the gate voltage (V_G) and the influence of protons on the channel.

Figure 1a is a schematic of a three-terminal transistor composed of p-Si/PSG/AOSs. The characteristics of the IWO and IGZO channels formed on the same PSG thin film were analyzed from both TFT and synaptic perspectives. Figure 1b shows the capacitance–frequency (C–f) curve measured for the MOS capacitor with a p-Si/PSG/Al structure. Aluminum was deposited in a circular shape with a diameter of 200 µm and a thickness of 100 nm using a shadow mask. Generally, the evaluation of synaptic characteristics is performed in the low-frequency range [43,44]. The fabricated PSG demonstrated excellent capacitance characteristics, reaching up to 16.9 µF/cm² at 1 Hz and maintaining a value of 0.34 µF/cm² even at 1 kHz, indicating the feasibility of mimicking synaptic behavior across various frequency ranges. Figure 1c illustrates the I_ds mechanism of the PSG-AOS channel when V_th < V_G < U_th. The protons accumulated at the PSG/channel interface induce the EDL mechanism, promoting the movement of electrons within the channel. Figure 1d shows the I_ds mechanism of the PSG-AOS channel when U_th < V_G. The protons at the interface penetrate into the channel, doping the AOS channel. Figure 1e depicts the I_ds mechanism of the PSG-AOS channel at V_G = V_th. While the protons forming the EDL quickly relocate or recombine within the PSG film, the protons involved in the ECD process take relatively longer to de-dope [45]. Additionally, IWO demonstrates a superior ECD process compared to IGZO, indicating that a greater number of protons remain doped within the channel at V_G = V_th. Figure 1f,g illustrate the I_ds characteristics over time based on Figure 1c–e. Initially, when V_D = 1 V and V_G = 0 V (0–5 s), no current flowed in all cases. However, with V_D = 1 V and V_G (0.5/1/2/4 V, 5–35 s), the channel current varied according to the V_G. At 35 s, the I_ds values for IWO and IGZO were 4.1/7.3/11.7/17.1 µA and 3/6/9.7/14.3 µA, respectively. The IWO channel, dominated by an indium composition, exhibited higher conductivity and superior current characteristics because of the engineered oxygen vacancies through an optimal annealing process [46]. Subsequently, when V_D = 1 V and V_G = off (35–60 s), different residual current (I_R) characteristics appeared, depending on the V_G conditions (0.5/1/2/4 V, 5–35 s). When V_G was 0.5/1 V (5–35 s), the I_R value for a relatively short period existed because of the EDL characteristics, and the I_R difference between the two channels was attributed to the I_ds value at 35 s. When V_G was 2 V (5–35 s), the I_ds values at 60 s were 3 and 2.1 µA, respectively, showing a significant difference from the I_ds values under the V_G conditions (0.5/1 V, 5–35 s). Consequently, around V_G = 2 V, the ECD mechanism (Figure 1d) was operating; in this study, we defined U_th as 2 V [47]. Additionally, for V_G (4 V, 5–35 s), the I_ds values at 60 s were 13.3 and 6.1 µA, respectively. In conclusion, IWO exhibited better I_R characteristics than IGZO when the same U_th voltage was applied. This directly caused the difference in synaptic characteristics between IWO- and IGZO-based transistors.

3.2. Comparison of Residual Current Based on ECD and Trapping

To provide a comprehensive understanding of the I_R characteristics of IWO and IGZO based on the ECD mechanism, a thorough analysis was conducted. First, to support the differences in ECD characteristics due to proton doping in IWO and IGZO, we compared p-Si/SiO₂/AOSs transistors utilizing high-quality thermal oxide (SiO₂) with excellent interface characteristics against p-Si/PSG/AOSs. Specifically, the superiority of the ECD process in PSG–IWO was verified by comparing the I_R characteristics according to the U_th extracted in Figure 1f,g for the PSG-AOSs with the I_R characteristics according to U_th, where electron trapping occurred at the SiO₂-AOSs interface [48].

Figure 2a illustrates the mechanisms of proton doping and electron trapping for p-Si/PSG/AOSs and p-Si/SiO₂/AOSs-based TFTs when U_th < V_G. Figure 2b demonstrates the de-doping and de-trapping operations when V_G ≤ V_th. First, a comparison of I_R characteristics was conducted for the PSG/IWO, PSG/IGZO, SiO₂/IWO, and SiO₂/IGZO TFTs, considering the two different mechanisms [49,50]. The period from 0 to 75 s was divided into three segments based on the DC bias voltage applied to the gate: 0 to 5 s (V_off), 5 to 20 s (U_th), and 20 to 75 s (V_th, V_th − 1 V, V_th − 2 V, V_th − 3 V, and V_th − 4 V). The V_th for each TFT was −0.81 V for PSG/IWO, −0.63 V for PSG/IGZO, −0.50 V for SiO₂/IWO, and 0.53 V for SiO₂/IGZO. Figure 2c shows the case of PSG/IGZO, where the I_R value at 20 s is 7.9 µA, and the I_R value at 75 s is 3.5 µA (V_th)/1.4 µA (V_th − 1 V), depending on the V_G values. The I_R values at V_th − 2 V, V_th − 3 V, and V_th − 4 V were considered to be in the off-state as they were below 1 nA. Figure 2d shows the case of PSG/IWO, where the I_R value at 20 s is 14.2 µA, and the I_R value at 75 s is 13.6 µA (V_th)/13.5 µA (V_th − 1 V)/12.6 µA (V_th − 2 V)/2.8 µA (V_th − 3 V), depending on the V_G values. The I_R value at V_th − 4 V was considered to be in the off-state, as it was below 1 nA. The retention characteristics based on I_R were extracted using the equation (I_R at 75 s/I_R at 20 s) × 100%. Consequently, PSG/IWO and PSG/IGZO showed values of 95.6/95.1/88.7/19.7/0.1% and 44.3/17.7/0.1/0.1/0.1%, respectively, depending on the V_G values. We observed that the retention characteristics of IWO were significantly higher than those of IGZO from V_th − 2 V onwards, indicating that the ECD process, according to U_th, was exceptionally superior in IWO. Figure 2e,f represent the cases of SiO₂/IGZO and SiO₂/IWO, respectively, and the values of the retention characteristics according to V_th, V_th − 1 V, V_th − 2 V, V_th − 3 V, and V_th − 4 V are similar. Additionally, the retention characteristics of both IWO and IGZO, considering I_R at 75 s/I_R at 20 s, were confirmed to be below 0.1%. Figure 2g shows the fitting results of the time-dependent retention characteristics extracted using Equation (1) for the I_R values at V_th − 1 V from 20 to 50 s. The retention results at 30 s were 96.8%, 23.6%, 0.1%, and 0.1% for PSG/IWO, PSG/IGZO, SiO₂/IWO, and SiO₂/IGZO, respectively.

$$\mathrm{R}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(\%\right)=\left({\mathrm{I}}_{\mathrm{R},20\mathrm{s}\mathrm{e}\mathrm{c}+\mathrm{t}}/{\mathrm{I}}_{\mathrm{R},20\mathrm{s}\mathrm{e}\mathrm{c}}\right)\times 100, t=0 \mathrm{t}\mathrm{o} 30 \mathrm{s} (\mathrm{i}\mathrm{n}\mathrm{c}\mathrm{r}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s} 2.5 \mathrm{s}, 13 \mathrm{s}\mathrm{t}\mathrm{e}\mathrm{p}\mathrm{s})$$

(1)

3.3. Comparison of Electrical Characteristics of PSG-Based IWO and IGZO

Considering the EDL and ECD characteristics of IWO and IGZO based on the PSG electrolyte film, we evaluated the operational characteristics of the device through the transfer curve of the transistor. Generally, when only EDL characteristics are considered, applying V_G in a double-sweep mode results in dynamic characteristics where protons within the electrolyte film accumulate at the interface during the forward sweep, leading to a slight I_R due to the accumulated protons in the backward sweep. This characteristic results in a counterclockwise hysteresis window in the transfer curve [51]. Additionally, owing to the ECD characteristics, protons are doped into the channel during the forward sweep according to the magnitude of the V_G, and I_R is formed in the channel until the de-doping process is completed in the backward sweep, also resulting in a counterclockwise hysteresis window [52]. Consequently, we analyzed the difference in hysteresis window characteristics due to the EDL operating under the same electrolyte film, as well as the difference in hysteresis window characteristics during the proton-doping and de-doping processes of IWO and IGZO.

Figure 3a,b show the transfer curve characteristics of IGZO and IWO, respectively, with a V_G sweep range from −6 V to a maximum (Max.) V_G (0 V to 4 V in 0.5 V increments) applied in a double-sweep mode at V_D = 1 V. The V_th and hysteresis window for each transfer curve were fitted in Figure S1 (Supplementary Materials). The average V_ths for IGZO and IWO were −0.63 V and −0.81 V, respectively, extracted using the constant current method [53]. Additionally, as the Max. V_G increased, the hysteresis window of IGZO increased linearly from 0.24 V to 1.43 V with a linearity (R²) of 99.80%, while the hysteresis window of IWO increased linearly from 0.63 V to 2.5 V with an R² value of 99.44%. R² is a measure of the goodness-of-fit between a linear model and the variability in experimental data. Figure 3c shows the hysteresis windows for IGZO and IWO, considering their different Vths, plotted against Max. V_G − V_th. For both IWO and IGZO, the common Max. V_G − V_th range was from 0.86 V to 4.73 V, with the difference in hysteresis window increasing from a minimum of 0.39 V to a maximum of 1.07 V. Therefore, IWO exhibited a larger hysteresis window than IGZO, which could be attributed to the differences in electrochemical process characteristics, suggesting an advantage of the IWO channel in proton doping to the channel. Figure S2 (Supplementary Materials) shows the transfer curve characteristics of IGZO and IWO, with a V_G sweep range from the minimum (min.) V_G (−8 V to −3 V in 1 V increments) applied in a double-sweep mode at V_D = 1 V. Unlike the hysteresis window changes caused by the Max. V_G, the hysteresis window caused by the minimum V_G, showed no variation [54]; this supported the operational reliability of the fabricated devices. Additionally, in the case of IWO, I_R in the forward sweep of the double-sweep could be observed when the minimum V_G was −3 V. This indicated that the on/off operation of the channel was limited because the electric field strength required for the de-doping of protons doped into the channel was not met. Therefore, an analysis was conducted to evaluate the specific transfer curve operation characteristics of IGZO and IWO from the perspective of the de-doping process.

Figure 4a,d show the linear scale I_D–V_G curves of IGZO and IWO according to the V_G sweep range (from −n V to n V). Continuous measurements were conducted while gradually increasing the V_G sweep range. We observed that the I_R increased in the backward sweep because of the degree of doping as V_G increased. Additionally, we confirmed that at n = 10 for IGZO and n = 6 for IWO, a current of 2 µA and 6 µA, respectively, was formed at the initial negative voltage in the forward sweep. This indicated that the current flowed even at the initial negative voltage because the protons strongly doped into the channel were not de-doped. Figure 4b,e show the logarithmic scale double-sweep I_D–V_G curves for n = 10 for IGZO and n = 6 for IWO. Strong channel doping was observed in IWO within the V_G range from −5 V to 5 V, whereas for IGZO, despite the V_G range from −9 V to 9 V, the channel doping was relatively weak. Particularly for IWO at n = 6, considering that the minimum current in the forward sweep was above 10 nA, the proton-doping characteristics were far superior compared to IGZO. Figure 4c,f show the double-sweep I_D–V_G curves measured in the V_G range from −n V to 4 V (with n increments of 1) to analyze the de-doping characteristics of IGZO and IWO in the doped state, as shown in Figure 4b,e. As a result, the complete de-doping of IGZO was confirmed when the start voltage of the forward sweep reached −9 V. For IWO, de-doping was not observed even at −15 V but occurred at the start voltage of −16 V. Ultimately, the doping process in IWO occurred more strongly at lower voltages, and the strongly doped protons within IWO required more energy for de-doping compared to IGZO.

3.4. Synaptic Characteristics Based on EDL and ECD Processes

We clarified the correlation between protons within the PSG film and the channel current with respect to U_th = 2 V based on the EDL and ECD mechanisms of PSG-IWO and PSG-IGZO. Subsequently, we evaluated the applicability of these synapse devices from the perspectives of STP and LTP. Figure 5a illustrates the principle of biological synapses in the brain. When a pre-synaptic spike stimulus is applied, a complex interaction between neurotransmitters and receptors occurs, resulting in an EPSC as a post-synaptic response [55,56]. In the fabricated device, the electrical pulse applied to the gate is considered pre-synaptic, and the current characteristics formed in the channel, based on the EDL and ECD mechanisms, are considered to correspond to the EPSC. Electrical pulse amplitudes of 1 V (V_G < U_th, EDL) and 3 V (V_G > U_th, ECD) were applied, and the EPSC was measured according to the variables of spike duration and number. First, we analyzed the synaptic characteristics of IWO and IGZO based on the EDL mechanism. Figure 5b shows the PPF index for a spike amplitude of 1 V (V_G < U_th, EDL), a duration of 100 ms, and two spikes. The PPF index is depicted as a function of the spike number while the interval between the first and second spikes (Δt_inter) varies from 10 ms to 4.4 s (applied 20 times). The second spike, following the first spike that induces an EPSC (A₁), results in an amplified EPSC (A₂) based on the EDL mechanism. The PPF index can be calculated using the following double-exponential decay relationship [57]:

$$PPF index={A+C}_1\mathit{exp}(-\Delta t/{\tau }_1)+C_2\mathit{exp}(-\Delta t/{\tau }_2),$$

(2)

where C₁ and C₂ denote the initial facilitation magnitudes, while τ₁ and τ₂ indicate the characteristic relaxation times. The PPF indexes for IWO and IGZO obtained values of 147.7 and 136.3 at Δt_inter of 10 ms and 109.7 and 100 at Δt_inter of 4.4 s, respectively, confirming the relatively superior synaptic plasticity of IWO. The higher conductivity characteristics of the IWO channel resulted in higher EPSC values and amplification for a 1 V spike amplitude and two spikes, suggesting that IWO was more suitable as a synaptic application material based on the same EDL mechanism [58]. Figure S3 (Supplementary Materials) shows a comparison among the PPF indexes of the proposed device and various IWO channel-based synaptic transistors [27,28,29,30]. The Δt_inter values ranged from 10 ms to 25 ms. Our PSG-IWO synaptic transistor demonstrated relatively superior performance in terms of the PPF index. Figure S4 (Supplementary Materials) shows the EPSC for a spike amplitude of 1 V (V_G < U_th, EDL), duration of 100 ms, and a single spike, as well as the EPSC for spike amplitudes of 1 V (V_G < U_th, EDL), durations of 100/300/500/700/900 ms, and a single spike. Figure S5 (Supplementary Materials) shows the EPSC for a spike amplitude of 1 V (V_G < U_th, EDL), duration of 100 ms, and 10 spikes at frequencies of Δt_inter = 1/3/5/7/9 Hz. Figure 5c shows the EPSCs of IWO and IGZO according to the EDL and ECD mechanisms in a single spike.

The duration was consistently 500 ms, and spike amplitudes of 1 V (V_G < U_th, EDL) and 3 V (V_G > Uth, ECD) were applied. Figure S6 (Supplementary Materials) shows the EPSC at 3 V (V_G > U_th, ECD), durations of 300/500 ms, and a single spike. Figure 5d represents the EPSC change ratio ((I − I₀)/I₀ × 100%) according to the retention time for an EPSC induced by a single spike. Here, I refers to the residual EPSC after a certain retention time following a spike, and I₀ is the resting current before electrical spike stimulation [59]. For 1 V (V_G < U_th, EDL), the retention times of 1 and 4 s resulted in IWO values of 466% and 112%, respectively, and I₀ values below 1 nA at 12 and 20 s, respectively. IGZO obtained values of 235% and 36% at 1 and 4 s, respectively, and I₀ values below 1 nA at 12 and 20 s, respectively. For 3 V (V_G > U_th, ECD), the retention times of 1, 4, 12, and 20 s resulted in IWO and IGZO values of 31,328%, 2556%, 21,739%, and 20,016%/14,597%, 10,719%, 7147%, and 5966%, respectively. Figure 5e shows the EPSCs of IWO and IGZO according to the EDL and ECD mechanisms in 20 spikes. The duration was consistently 100 ms, the interval between spikes was 10 ms, and spike amplitudes of 1 V (V_G < U_th, EDL) and 3 V (V_G > U_th, ECD) were applied. Figure 5f shows the fitting results of the EPSC gains (A₂₀/A₁) extracted from an EPSC induced by 20 spikes. A₂₀ refers to the EPSC induced by the twentieth spike, and A₁ refers to the EPSC induced by the first spike. For 1 V (V_G < U_th, EDL), the EPSC gains of IWO and IGZO were 4.1 and 3, respectively, and for 3 V (V_G > U_th, ECD), the gains were 5.3 and 3, respectively. Consequently, from the perspective of residual EPSC characteristics through the EPSC change ratio in terms of the LTP, the order of superiority is IWO (ECD) > IGZO (ECD) > IWO (EDL) > IGZO (EDL). From the perspective of EPSC characteristics in terms of the synaptic amplification rate through an EPSC gain, the order of superiority is IWO (ECD) > IWO (EDL) > IGZO (ECD) > IGZO (EDL).

3.5. Recognition Rate in MNIST ANN Simulations

To validate the neuromorphic computing capabilities of the proposed PSG-IWO synaptic transistor, P/D measurements were conducted using PSG-IWO and PSG-IGZO devices at spike amplitudes of 4 V and −4 V [60]. Based on the P/D characteristics, parameters such as nonlinearity factors, AR, and DR were extracted, leading to the calculation of normalized conductance. Subsequently, ANN technology was employed to simulate the training of the MNIST handwritten digit dataset, thus evaluating the recognition rates for specific digits based on the normalized conductance from the PSG-IWO and PSG-IGZO synaptic transistors. Figure 6b shows the endurance of IWO and IGZO over five cycles of P/D. Commonly, the gate was applied with a spike amplitude of 4 V (potentiation)/−4 V (depression), a duration of 100 ms, and an interval between spikes of 300 ms for 30 spikes each. The V_D read-spike was applied with a spike amplitude of 1 V, a duration of 200 ms, and an interval between read-spikes of 200 ms for 60 spikes. The maximum and minimum conductance values for each cycle of IWO were 6.54/6.59/6.61/6.63/6.63 µS and 2.85/2.85/2.86/2.86/2.86 µS, respectively, and for IGZO, 4.10/4.13/4.19/4.20/4.25 µS and 1.98/1.98/2.01/2.01/2.02 µS, respectively. These deviations in maximum and minimum conductance over the five cycles were within an acceptable range, ensuring the reliability of the fabricated devices. Additionally, the average maximum and minimum conductance values for IWO were approximately 2.43 µS and 0.86 µS higher, respectively, compared to IGZO, indicating the superior amplification rate and long-term potentiation characteristics of IWO. Figure 6a shows a schematic representation of a four-layer fully connected ANN with input (784 neurons), hidden (256 neurons, 128 neurons), and output (10 neurons) layers for recognizing the MNIST handwritten digits (‘9’). The two hidden layers enhanced the reliability of learning and processing between the input and output layers, positively impacting the recognition rate. Figure 6c shows the conductance calculated by normalizing each conductance to the minimum conductance (G_#/G₁). DR, representing the P/D characteristics, is defined as the ratio of the maximum to minimum conductance values, with a larger value being ideal (Equation (3)) [61].

$$DR=G_{max}/G_{min}$$

(3)

AR is an indicator of asymmetry in the conductance characteristics; for high learning accuracy, the ideal AR value is 0. AR was estimated using Equation (4) [62]:

$$AR={\displaystyle \frac{MAX\left|G_p\left(n\right)-G_d\left(n\right)\right|}{G_p\left(30\right)-G_d\left(30\right)}} for n=spike numbers$$

(4)

Here, G_p(n) and G_d(n)) refer to the conductance of the channel corresponding to the nth spike stimulus. Finally, nonlinearity is a critical indicator of the relationship between input and output in long-term potentiation and depression, serving as a key parameter in neural network learning and pattern recognition simulations. The nonlinearity factor is quantified using Equation (5) [63]:

$$G=\left\{\begin{array}{c}{\left\{\right({G_{max}}^{\mathsf{\alpha }}-{G_{min}}^{\mathsf{\alpha }})\times \mathrm{w}+{G_{min}}^{\mathsf{\alpha }}\}}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\mathsf{\alpha }$}\right.}\\ {G_{min}\times {(G}_{max}/G_{min})}^{\mathrm{w}}\end{array}{\displaystyle \genfrac{}{}{0pt}{}{ if \mathsf{\alpha }\ne 0,}{ if \mathsf{\alpha }=0.}}\right..$$

(5)

where w is an internal variable ranging from 0 to 1. The nonlinearity coefficient α governs either potentiation (α_p) or depression (α_d), with the ideal value for the nonlinearity factor being 1. The α_p, α_d, AR, and DR of IWO and IGZO were 3.2, −2.3, 0.48, 2.3/3.9, −2.1, 0.49, and 2.07, respectively, indicating that IWO acquired more optimal values for α_p, AR, and DR. Figure 6e presents the simulation results of the recognition rates for the digit ‘9’ across varying numbers of epochs for IWO and IGZO. As the number of epochs increased, both IWO and IGZO exhibited an increase in recognition rates, with a notable decline in the rate of increase starting from epoch 4. Specifically, at epoch 4, IWO and IGZO achieved recognition rates of 86.29% and 76.19%, respectively; by epoch 20, these rates had risen to 91.68% and 83.01%, respectively. Consequently, IWO demonstrated a relatively superior achievement in the recognition rates.

4. Conclusions

In this study, we demonstrated the potential of IWO channel-based PSG inorganic electrolyte transistors for synaptic mimicking applications, focusing on STP and LTP, by comparing them with IGZO channels. Specifically, we confirmed the superiority of the IWO channel based on the characteristics of EDL and ECD, arising from the complex interactions between protons within the PSG film and IWO and IGZO under varying electric field strengths. Using PSG and thermal oxide films, we compared the residual current characteristics of AOSs under DC bias gate voltages, thus verifying the excellent advantages of proton doping in IWO. Additionally, by analyzing the electrical characteristics based on transfer curves under various gate voltage sweep ranges and conditions, we identified the specific hysteresis window of IWO due to EDL and ECD and confirmed the characteristics of IWO during the proton-doping and de-doping processes. To measure the synaptic response of the device as an EPSC, stimuli with spike amplitudes of 1 V (EDL) and 3 V (ECD) were applied based on U_th = 2 V. As a result, IWO exhibited notable differences compared to IGZO, with superior STP characteristics due to EDL and superior LTP characteristics due to ECD. Finally, based on the normalized parameters from the potentiation/depression (P/D) measurements, we conducted recognition simulations for the digit ‘9’ using the MNIST dataset. At epoch 20, IWO achieved a recognition rate of 91.68%, while IGZO acquired 83.01%. In conclusion, these results suggest that the electrolyte film-based IWO channel material is highly suitable for EDL and ECD processes, indicating its potential as a highly applicable oxide semiconductor channel for mimicking short-term and long-term synaptic behaviors of the human brain.

In future research, we intend to perform in-depth follow-up studies focusing on the ECD mechanism of IWO channel-based synaptic devices. Such studies would entail specific enhancements aimed at improving the LTP characteristics.