Nonlinear Dynamics in HfO₂/SiO₂-Based Interface Dipole Modulation Field-Effect Transistors for Synaptic Applications

Abstract

In the pursuit of energy-efficient spiking neural network (SNN) hardware, synaptic devices leveraging emerging memory technologies hold significant promise. This study investigates the application of the recently proposed HfO₂/SiO₂-based interface dipole modulation (IDM) memory for synaptic spike timing-dependent plasticity (STDP) learning. Firstly, through pulse measurements of IDM metal–oxide–semiconductor (MOS) capacitors, we demonstrate that IDM exhibits an inherently nonlinear and near-symmetric response. Secondly, we discuss the drain current response of a field-effect transistor (FET) incorporating a multi-stack IDM structure, revealing its nonlinear and asymmetric pulse response, and suggest that the degree of the asymmetry depends on the modulation current ratio. Thirdly, to emulate synaptic STDP behavior, we implement double-pulse-controlled drain current modulation of IDMFET using a simple bipolar rectangular pulse. Additionally, we propose a double-pulse-controlled synaptic depression that is valuable for optimizing STDP-based unsupervised learning. Integrating the pulse response characteristics of IDMFETs into a two-layer SNN system for synaptic weight updates, we assess training and classification performance on handwritten digits. Our results demonstrate that IDMFET-based synaptic devices can achieve classification accuracy comparable to previously reported simulation-based results.

1. Introduction

Research on solid/solid interfaces is intricately connected to the evolution of semiconductor devices. Throughout the extensive history of semiconductor devices, substantial progress has been made in understanding the various interfaces that constitute these devices, such as metal/semiconductor, semiconductor/semiconductor, and oxide/semiconductor interfaces. In the development of MOSFETs, the fundamental building blocks of Si large integrated circuits, the interfacial electronic states, especially at the oxide/Si interface in the gate stack structure, are crucial elements influencing device operation and performance. Historically, the discovery of high-quality SiO₂/Si interfaces formed by the thermal oxidation of silicon substrates in around 1960 paved the way for the mass production manufacturing of CMOS integrated circuits [1,2]. In the 2000s, research institutions worldwide actively pursued the development of high-k gate dielectrics, accumulating knowledge about gate stacks containing various metal oxides [3,4]. The adoption of HfO₂-based dielectrics in current state-of-the-art MOSFETs is a technological fruition of these research and development efforts [5]. In the 2010s, research on non-volatile memories based on MOSFETs incorporating high-k dielectrics experienced a significant surge [6,7]. Ferroelectric field-effect transistor (FeFET) memory, integrating ferroelectric materials into the gate stack, and conventional flash memory with high-k dielectrics garnered considerable attention [8,9]. The discovery of ferroelectric HfO₂ in 2021 captured researchers’ interest due to its excellent material compatibility with silicon semiconductor technology [10], driving continued active research and development for memory applications [11,12,13,14]. A notable recent research trend in FeFETs, similar to the purpose of this study, is the development of analog memories for neuromorphic applications [15,16].

Research on the dipole layers induced at dielectric/dielectric interfaces has primarily advanced for the purpose of controlling the threshold voltage of high-k MOSFETs [17,18]. It has been reported that a small dipole layer is formed at the HfO₂/SiO₂ interface used in this study [19,20]. Interfacial dipole modulation (IDM) memory was conceived based on the studies of such high-k gate stack structures, with the objective of inducing MOS threshold voltage shifts similar to FeFETs. The first reported IDM results were observed in a HfO₂/SiO₂ stack structure incorporating a 1 monolayer (ML) of TiO₂, as depicted in Figure 1a [21]. A MOS capacitor with such a stack structure exhibits a high-frequency C-V curve with small hysteresis, as illustrated in Figure 1b. This phenomenon was elucidated by the alteration in the bonding state around the Ti atoms at the SiO₂/HfO₂ interface, causing a change in the potential difference (interfacial dipole) between SiO₂ and HfO₂. Subsequently, hard X-ray photoelectron spectroscopy studies of IDM metal–insulator–metal structures have revealed changes in the Ti oxidation states synchronized with memory operation, supporting the notion that the origin of IDM is the structural change around the interfacial Ti atoms [22]. As IDM operates at amorphous dielectric/dielectric interfaces, it does not necessitate high-temperature crystallization annealing, in contrast to ferroelectric HfO₂, and has been verified even with low-temperature annealing at around 300 °C [23]. Furthermore, as shown in Figure 1c, the modulation operation exhibits a gradual change with respect to the applied voltage. This feature is also thought to originate from atomic-scale disorder at the amorphous dielectric/dielectric interface. The gradual threshold change is anticipated to be beneficial for analog operation in synaptic devices, as discussed later. On the other hand, the modulation range of HfO₂/SiO₂-based IDM is relatively small, up to about 0.33 V [21,23]. To extend the modulation range, a multilayered HfO₂/SiO₂ structure has been proposed. For example, an FET incorporating a multilayered IDM with six TiO₂ modulation layers achieves a threshold voltage shift of over 1 V and a current change of over six orders of magnitude [21]. The pulse response characteristics of IDMFETs have been investigated, demonstrating stable and repetitive modulation. The authors believe that these characteristics make it suitable for use as a synaptic device in spiking neural networks and have conducted detailed measurements. This report presents the pulse response measurement of IDMMOS capacitors, explains the physical origin of their characteristics, and describes the current modulation of IDMFETs to verify their potential as synaptic devices in spiking neural networks.

Energy-efficient neuromorphic computing hardware is a recent hot topic in the electronic memory device research field, and a variety of emerging-memory-based technologies have been proposed [24,25,26]. In particular, research on analog memory aimed at emulating biological synapses is active, and the realization of highly integrated synaptic devices is expected. One criterion for evaluating synaptic devices is the linearity of synaptic weight update characteristics. For example, in the current mainstream of deep neural networks, good linearity in weight updates is preferred and increased nonlinearity leads to poor learning and inference performance [27,28]. Also, regarding the next generation AI technology, spiking neural networks (SNNs), most studies show that linear weight updates are advantageous [29,30]. However, there is also the possibility of intentionally introducing nonlinear update dynamics to improve the performance of unsupervised SNN learning [31,32,33]. In SNN systems, as illustrated in Figure 2a, spikes transmitted from presynaptic neurons undergo weighting through synapses before being input into the postsynaptic neuron [34,35]. In post-synaptic neurons, input spikes contribute to the rise in membrane potential, leading to output spikes when the membrane potential reaches a specific threshold according to leaky integrate-and-fire (LIF) dynamics. Spike timing-dependent plasticity (STDP) serves as a fundamental synaptic learning rule, updating synaptic weights based on the temporal difference between pre-synaptic and post-synaptic spikes. Numerous studies have explored STDP learning employing emerging memory devices, with a predominant focus on two-terminal resistive-switching memory due to its benefits in high-density integration [36,37]. Meanwhile, three-terminal memory devices like FET-type memory have garnered attention as synaptic devices capable of achieving concurrent STDP learning within a single device [38,39,40]. This study aims to achieve concurrent STDP in IDMFETs utilizing a straightforward spike waveform and operation approach. We apply the experimentally observed STDP-like responses of IDMFET to a simple SNN network and assess its suitability for unsupervised SNN learning.

2. Materials and Methods

2.1. Oxide Deposition and Device Fabrication

The HfO₂, TiO₂, and SiO₂ constituting the IDM stack structure were deposited in the same chamber via the high-vacuum electron-beam (EB) evaporation method [41]. The thicknesses of the HfO₂ and SiO₂ were estimated from the calibrated deposition rates using transmission electron microscopy (TEM) and in situ X-ray photoelectron spectroscopy (XPS, Vacuum Science Instruments, Bad Schwalbach, Germany). The modulation layer of 1-ML TiO₂ was determined based on in situ XPS measurements of TiO₂ deposition on the Si surface. Following the oxide deposition of the IDM stack, a post-deposition annealing (PDA) process was performed at 400 °C for 30 min in an O₂/Ar (~20%) atmosphere. Subsequently, a 50 nm-thick Ir layer was deposited onto the sample surface to serve as the gate electrode, using an EB deposition method.

A 2-IDMMOS capacitor containing two TiO₂ modulation layers was prepared for pulse response characterization. Using the high-vacuum EB evaporation method, the IDM oxide stack, comprising a 3.5 nm-thick top HfO₂ layer, a 1.8 nm-thick inner SiO₂ layer, a 1.8 nm-thick inner HfO₂ layer, and two TiO₂ modulation layers, was formed on an n-type Si(100) substrate covered with a thermally grown SiO₂ layer approximately 5 nm thick. After the above PDA, Ir gate electrodes with a diameter of 200 μm were fabricated using a stencil mask method.

The IDMFET was fabricated through the following steps [42]. Initially, n⁺ source/drain (S/D) regions were formed on a p-type Si(100) substrate using the ion implantation method, followed by the formation of an approximately 10 nm-thick thermally grown SiO₂ layer. After etching the SiO₂ layer to 5 nm using diluted hydrofluoric acid, a multilayered HfO₂/SiO₂ IDM stack, comprising 6 TiO₂ modulation layers, was deposited using the high-vacuum EB evaporation method. This 6-IDM stack is composed of a 3.5 nm-thick top HfO₂ layer, 1.8 nm-thick inner SiO₂ layers, 1.8 nm-thick inner HfO₂ layers, and six TiO₂ modulation layers. After the above PDA, a 50 nm-thick Ir layer was deposited and, subsequently, gate electrode patterns with a gate length of 1 μm and a gate width of 100 μm were formed using lithography and reactive-ion etching.

2.2. Pulse Response Measurements of IDM Devices

The pulse response characteristics of the IDMMOS capacitor were observed through a repeated sequence of a voltage pulse stimulus and a C-V measurement at 1 MHz. Negative voltage pulses to the gate electrode in this experiment failed to generate adequate minority carriers (holes) to create an inversion layer on the Si surface. Consequently, we could not obtain a sufficient oxide electric field compared to positive voltage pulses. To address this limitation, we intentionally generated holes near the Si surface by exposing the sample surface to light [21,23]. In such cases, it becomes imperative to maintain a relatively wide pulse width, considering the hole diffusion process from the electrode-uncovered Si region to the Si region beneath the gate electrodes. In this experiment, a pulse width of 20 msec was employed.

The pulse response measurement of the IDMFET was conducted using a repeated sequence of a voltage pulse stimulus and drain current (I_d) measurements with V_ds = 50 mV [21]. For comparison, the DC drain current–gate voltage (I_d-V_g) curves were also measured with V_ds = 50 mV. Furthermore, to emulate the synaptic STDP behavior, two voltage waveforms with a time difference were applied to the gate and drain electrodes of the IDMFET. A repeated sequence of this double-pulse-controlled stimulus and I_d measurement was performed. The details of the voltage waveforms are described below.

2.3. SNN Architecture for Pattern Recognition

A simple two-layer feedforward network with 784 neurons in the input layer, N neurons in the hidden layer (where N is a variable), and 10 neurons in the output layer was utilized to perform training and classification tasks on the MNIST handwritten digits dataset [43], as shown in Figure 2b. Spike-based temporal processing optimized for the experimentally observed responses of the IDMFETs was implemented, referencing a computationally efficient simplified model proposed by Iakymchuk et al. [44]. In the input neurons, 28 × 28 pixels with 256 gray levels were transformed into 784 spike trains with a frequency range from 200 Hz to 1.25 kHz. These spikes were then transmitted to the hidden neurons through feature synapses, thereby raising the membrane potentials of the neurons. The hidden neurons operate on a LIF model with dynamic threshold adjustments and a 10 msec refractory period, generating output spikes when their membrane potential exceeds a certain threshold. Additionally, a winner-take-all (WTA) competitive algorithm was implemented through lateral inhibition in the hidden layer neurons.

The STDP weight update of the feature synapse was performed based on the time difference between the pre-synaptic spike and the post-synaptic spike, with the time delay from post-synaptic spike firing to reaching the synapse set to 80 μs. Furthermore, to stabilize unsupervised learning, an additional weight depression function, independent of the input neuron’s frequency and referred to as frequency-independent depression (FID), was introduced for all synapses undergoing STDP. This FID operation simply produces the effect of reducing the synaptic weight distribution. Through the above procedure, the feature synapses acquire characteristics of the input images through unsupervised learning. Additionally, the classification synapses update weights through supervised learning based on STDP to associate the hidden neurons with the output neurons. These synaptic learning processes enable the entire network to function as a pattern classifier. A total of 60,000 MNIST images were used for training feature and classification synapses, and 10,000 images were used for the performance test.

3. Results and Discussion

3.1. Pulse Response of IDM MOS Capacitors

The pulse response characteristic of the IDMMOS capacitor (illustrated in Figure 3a) indicates that the flat-band voltage (V_fb) undergoes stable positive and negative shifts in response to the voltage polarity switch every 50 pulses. Here, the V_fb shift from the initial V_fb (ΔV_fb) were plotted. Both positive and negative V_fb shifts display obvious nonlinear responses, exhibiting substantial changes immediately after the polarity switch and gradual suppression in the amount of change as the pulse count increases. Utilizing the approximation formula for nonlinear characteristics [45,46], the estimated nonlinear parameter ν for both positive and negative V_fb shifts is approximately six, indicating a nearly symmetrical response. The pulse voltage (V_p) dependence illustrated in Figure 3b indicates an increase in modulation amplitude and ν with the rise in V_p, as depicted in Figure 3c. As a result, IDMs exhibit inherently nonlinear and near-symmetric responses and the degree of nonlinearity varies depending on the operating conditions, making it necessary to consider these specific characteristics in synaptic applications.

Next, we delve into the reasons behind the nonlinear response. The V_p-dependent V_fb shift in Figure 3b incorporates information about both the interface dipole state and response characteristics, which is useful for analyzing their relationship. ΔV_fb on the y-axis corresponds to the strength of the interface dipoles, as shown in Figure 1a. Here, we assumed that the unit dipole switches between two states: large and small. In this scenario, the maximum V_fb shift occurs when all the unit dipoles at the HfO₂/SiO₂ interface switch due to the electric field; under the opposite electric field, the opposite maximum V_fb shift occurs when all the unit dipoles switch to the opposite state. Additionally, the maximum modulation width of the 2-IDM structure is 0.66 V, as previously reported. In this context, the ratio of un-switched unit dipoles, that is, switchable unit dipoles, is defined as θ_D. In the following discussion, θ_D was estimated from the experimentally obtained ΔV_fb based on the above assumptions. On the other hand, the modulation rate, dV_fb/dt (V/sec), can be estimated from the ΔV_fb shift per pulse and the oxide electric field E_ox (V/cm) can be estimated from the relationship between the ideal C-V curve of the MOS structure and V_p [47]. Consequently, we can establish the relationship between dV_fb/dt and E_ox as shown in Figure 4a. It is essential to note that even if the switching rate of the unit dipole is constant, the modulation rate varies depending on θ_D. For instance, a change in θ_D from θ_D = 0.5 to θ_D = 0.65 or θ_D = 0.35 predicts the characteristics (I) and (II) in Figure 4a. However, the experimental results indicate more significant changes that cannot be explained by a simple θ_D difference.

The experimentally obtained dV_fb/dt is considered to be proportional to the number of switchable unit dipoles. Therefore, the following relationship can be predicted: dV/dt = ΔV_max·k·θ_D, where ΔV_max is the maximum V_fb modulation of 0.66 V and k (s⁻¹) represents the reaction rate associated with the structural change responsible for dipole modulation. As mentioned earlier, IDM originates from alterations in the chemical bond state around the Ti atom. Previous studies have indicated that this phenomenon can be elucidated by the electric-field-induced chemical bond breakage model [21]. In our study, we continue to employ the following equation, based on the bond-breaking model:

$$k=v_{0}exp\left(-\frac{∆H_0^{*}-p_{eff}E}{k_{B}T}\right)$$

(1)

where ${\upsilon }_0$ is the molecular vibrational frequency, typically on the order of ~10¹³ (s⁻¹), and T and k_B are the temperature (K) and Boltzmann’s constant, respectively. From these relationships and the experimentally obtained E_ox dependence, we can estimate the zero-field activation energy ${∆H}_0^{*}$ (eV) and the effective dipole moment p_eff (eÅ) for each θ_D. The θ_D dependence of ${∆H}_0^{*}$ and p_eff for positive ΔV_fb shifts is summarized in Figure 4b, and the results estimated by the same analysis for negative ΔV_fb shifts are shown in Figure 4c. We can find that, for both cases, both ${∆H}_0^{*}$ and p_eff increase when θ_D falls below 0.5. Studies on the dielectric breakdown of gate dielectrics have reported that ${∆H}_0^{*}$ and p_eff are sensitive to local bonding configuration [48,49]. In addition, in the previous report on the IDM response, it has been suggested that the observed ${∆H}_0^{*}$ of about 0.7 eV for large θ_D is close to the breakage of the Ti-O bond [21]. Namely, the observed θ_D dependence in this study likely exhibits that the Ti bonding configuration contributing to IDM varies depending on θ_D. Since IDM occurs at an amorphous oxide interface, it is natural that there are variations in the bond length and bond angle in the chemical bonds of interfacial Ti atoms. Therefore, it is reasonable to assume that the initial structural change starts from the bonding with low ${∆H}_0^{*}$. In addition, there is a possibility that the structural change itself affects approximate bonding, that is, IDM itself leads to structural variations with higher ${∆H}_0^{*}$ values. From the above experimental results and considerations, we conclude that the nonlinearity in the IDM response is an unavoidable feature caused by the amorphous oxide interface.

3.2. Pulse Response of IDMFETs

We can easily predict that converting the threshold voltage (V_th) shift induced by the IDM into a change in the channel current of the FET will result in a response characteristic that is different from the IDM response, since the channel current–gate voltage relationship of the FET is not ideally linear, i.e., general I_d-V_g characteristics include at least a linear region and a sub-threshold region [50]. Before describing the synaptic characteristics of the IDMFET, we will briefly discuss the fundamental DC I_d-V_g curve and pulse-induced I_d change. The DC I_d-V_g curves shown in Figure 5a indicate that approximately 1 V hysteresis takes place with a sweeping voltage range of ±4.5 V. To convert the IDM-induced V_th shift into I_d change, it is suitable to use the read V_g within this hysteresis range. Here, the sub-threshold swing was estimated to be approximately 100 mV/decade, suggesting that the I_d change caused by a 0.1-V V_th shift is expected to be an order of magnitude current change.

The amplitudes of the I_d modulations marked as (I), (II), and (III) in Figure 5a represent the pulse-induced I_d changes observed under different readout V_g voltages and the same pulse conditions. Here, the pulse voltage (V_p) and pulse width (t_p) were set to ±5.4 V and 800 μs, respectively, and the V_p polarity was switched every 300 pulses. The changes in the pulse response characteristics (I), (II), and (III) shown in Figure 5b indicate that the I_d increase and decrease exhibit opposite behavior regarding nonlinearity. As for the I_d increase, (I) exhibits a nonlinear response, (II) approaches a linear response, and (III) shows an inverted nonlinear response, exhibiting that the nonlinear coefficient (ν₊) changes from positive to negative. Regarding the I_d decrease, the nonlinear coefficient (ν₋) is always positive and the nonlinearity becomes stronger in the order of (I), (II), and (III). On the other hand, even with the same read V_g, the nonlinearity changes significantly depending on the pulse voltage V_p (Figure 5c,d). In the lower graph of Figure 5d, we present the ratio of the nonlinear parameters for I_d increase and decrease (ν₊/ν₋) as an indicator of asymmetry. Here, ν₊/ν₋ values approaching 1 indicate proximity to symmetric response and smaller V_p values have better symmetry. In summary, the nonlinearity and asymmetry of the IDMFET exhibit complex behavior dependent on the read and pulse conditions. A summary of the ν₊/ν₋ ratios measured under various conditions (Figure 5e) shows that the general tendency is that the asymmetry becomes stronger when aiming for a large current ratio (I_max/I_min). This implies that, simultaneously, the nonlinearity of the I_d decrease becomes stronger.

The above behavior regarding the nonlinearity and asymmetric response can be roughly understood in terms of basic FET operation as follows. We can easily understand that when the I_d modulation is in the linear region or sub-threshold region with a sufficiently small I_max/I_min ratio, the nonlinear and near-symmetric IDM characteristics are directly reflected in the I_d response, as the I_d dynamics of (I) in Figure 5b show. On the other hand, when I_max/I_min is large and the device is operating in the sub-threshold region, even if the V_th shift is constant, the smaller the current, the smaller the absolute I_d change will be exponentially. In other words, in the characteristic of the I_d increase, the current is insensitive to V_th shifts in the initial stage and gradually becomes sensitive, so the nonlinear characteristics are weakened. Conversely, in the characteristic of the I_d decrease, the current is sensitive to V_th shifts in the initial stage and gradually becomes insensitive, so the nonlinearity of FET operation is further superimposed on the nonlinear IDM response. It is easy to predict that a similar effect will occur even when I_max/I_min is large and the I_d modulation straddles the linear and sub-threshold regions. This emphasized asymmetry can be found in the I_d dynamics in (II) of Figure 5b. As I_max/I_min increases further, the current becomes more insensitive to V_th shifts during the initial stage of I_d increase, resulting in an opposite nonlinear response. This can be seen in the I_d dynamics in (III) of Figure 5b. The ultimate goal of this study is to verify whether the nonlinear and asymmetric IDMFET response can be applied to STDP learning.

3.3. Double-Pulse-Controlled Synaptic Operation of IDMFETs

To update the I_d of IDMFETs based on the time difference between pre- and post-synaptic spikes, akin to synaptic weight (w) updates in biological STDPs, it is crucial to carefully choose the pre-spike and post-spike waveforms. However, for compatibility with the digital circuits responsible for neuron information processing, it is preferable to avoid complex waveforms as much as possible. We adopted a simple bipolar rectangular waveform, as shown in Figure 6a. Pre- and post-synaptic spikes have waveforms of the same voltages (V_STDP) and pulse widths (t_STDP) with a time difference Δt. Assuming that a superimposed waveform of pre-synaptic and post-synaptic spikes is applied to the gate stack structure, I_d modulation is expected to depend on Δt, because the period during which a voltage twice V_STDP is applied coincides with Δt. Here, the application period of V_STDP also changes, but since IDM has an exponential response to E_ox, it is expected that it can be ignored by setting an appropriate V_STDP. Figure 6b shows the measurement results in which the sign of Δt alternates every 500 spikes. An increase in I_d is observed at +Δt and a decrease in I_d is observed at −Δt, indicating the expected STDP-like response. This means that synaptic potentiation occurs when a post spike is input after a pre spike is input and synaptic depression occurs at the opposite timing. Furthermore, as Δt approaches 200 μs of t_STDP, the amplitude of the I_d modulation increases, which is a characteristic predicted from the above waveform superposition. On the other hand, we also find that STDP operation exhibits obvious nonlinear and asymmetric potentiation/depression properties. For example, at Δt = ±200 μs, the ν₊/ν₋ ratio was estimated to be 0.2, showing similar asymmetry to the previously discussed single-pulse IDMFET response.

In order to determine whether the pulse-timing-dependent I_d modulation obtained from the IDMFET can be applied to STDP learning, we need to investigate the different Δt responses acquired within the same I_d range. Therefore, we performed a similar double-pulse measurement that restricted the I_d range, where the sign of Δt is reversed and when I_d exceeds the range of 0.8 to 3.0 μA. Figure 7a presents the comparison of response characteristics for Δt = ±200 μs and ±100 μs. We can see that for the latter, more pulses are required for Δt sign reversal compared to the former. Both results exhibit asymmetric response characteristics and Δt does not approximately affect the ν₊/ν₋ ratio. The ΔI_d−I_d characteristics in Figure 7b can be obtained by converting the measured pulse-induced I_d change into an I_d change for each pulse (ΔI_d). Here, we can find the impact of the asymmetry response. Regarding an I_d increase, a slight ΔI_d value persists even as I_d approaches 3 μA. However, in the case of an I_d decrease, ΔI_d approaches zero more closely as I_d approaches 0.8 μA. The experimentally obtained ΔI_d−I_d data were fitted with an approximate equation: ΔI_d = α(I_d−I₀) + β(I_d − I₀)γ, where α, β, γ, and I₀ are constants. In the simulation study described later, the approximate equation of ΔI_d−I_d data was converted to the synaptic weight, w, in the w range of 0–0.8. The Δw−Δt characteristics of STDP shown in Figure 7c are the result converted from the experimentally obtained ΔI_d−I_d data. Here, in the double pulse scheme shown in Figure 6a, when Δt is ±200 μs, which is equal to the pulse width, the time during which the voltage of ±2 × V_p is applied is maximum of 200 μs, and if it deviates from this, the application time becomes shorter by the maximum one. If it is outside the ±400 μs range or 0 sec, the application time is expected to be zero. The Δw−Δt characteristics of the experimental results show the expected characteristics, with the maximum modulation at 200 μs and the minimum modulation at −200 μs. On the other hand, the effects of IDMFET nonlinearity and asymmetry are manifested as w-dependent potentiation/depression asymmetry. Under conditions where w is close to zero, potentiation is larger than depression, reaching an equilibrium of potentiation/depression around w = 0.4. As w increases further, depression becomes more prominent. In the following simulations, these nonlinear and asymmetric STDP characteristics are applied to unsupervised pattern learning.

On the other hand, an obvious variation is observed in the experimental ΔI_d−I_d data in Figure 7b. Figure 7d illustrates the difference between the approximation curve and measured data across the entire I_d range for the Δt = ±200 μs measurement. The origin of this variation contains the fluctuations of the IDM device itself and measurement system noise. Regarding the former, the fluctuation of the IDM response itself and other V_th fluctuations such as the oxide carrier trap may contribute. In the subsequent simulations, STDP incorporating the distribution of observed variations is applied.

In general, SNN learning requires an additional w update function that differs from STDP, for example, to set initial w values and to optimize and adjust the synaptic learning conditions. In this study, an additional w update of FID is applied to adjust the STDP-based unsupervised learning, as described above. We propose a two-pulse-controlled modulation, as shown in the inset of Figure 7e, which is highly compatible with our STDP operation. Positive and negative voltage pulses, serving as pre- and post-synaptic spikes, are inputted into the IDMFET, inducing the w depression as shown in Figure 7e. The depression effect becomes stronger with the increase in pulse voltage (V_FID) across all w ranges. This depression characteristic is incorporated into SNN simulations using the same approximate equation as the STDP characteristics.

3.4. Unsupervised Synaptic Learning Based on IDMFET Characteristics

First, let us examine how unsupervised learning, combining STDP and FID, operates using a network with N = 100 as an example. In this simulation, when a hidden layer neuron fires, the synapses connected to it are updated by STDP and, subsequently, FID is applied to all synapses that underwent STDP (100% FID). For STDP, we utilized the approximate curve obtained from measurements at V_STDP = 3.5 V, and for FID, the approximate curve with V_FID varied in the range of 3.15 to 3.5 V was employed. Random variations from the distributions estimated by the measurements were incorporated into both STDP and FID. The training dynamics in Figure 8a show the average test accuracy over 10 training/classification cycles, with the shaded area indicating the spread between the maximum and minimum values. In comparison to the result at V_FID = 3.2 V, a higher V_FID of 3.5 V reaches maximum accuracy faster but subsequently experiences more significant accuracy degradation and fluctuates. Here, the number of training images required to reach 90% of the maximum average accuracy is defined as learning efficiency (η). Although V_FID does not significantly affect the maximum accuracy (Figure 8b), a noticeably larger V_FID is advantageous for learning efficiency (Figure 8c). This is presumed to be due to a larger V_FID enhancing the WTA effect, suppressing the probability of overlapping different digit patterns. However, as V_FID increases, the robustness deteriorates after reaching the maximum accuracy, as shown in Figure 8a, suggesting that a large V_FID degrades the information of the pattern once learned. Based on the characteristics of the IDMFET obtained in this experiment, a V_FID of around 3.2 V is considered a balanced and favorable condition.

From the perspective of reducing calculation costs, it is advantageous to minimize the number of FIDs. Figure 8a illustrates the training dynamics of randomly inducing FID pulses with a 50% probability, demonstrating that both the maximum accuracy and learning efficiency are degraded compared to those of the 100% FID. As depicted in Figure 8b,c, no clear benefit was found from the results of V_FID dependence either. We also investigated various FID probabilities and concluded that FID is always required after STDP. This result suggests that FID is effective for properly operating WTA and accumulating training pattern information in appropriate synapses. It is worth mentioning that previously reported studies on STDP-based unsupervised learning did not incorporate additional pulses such as FID [31,32,33,36,37,38,39,40]. This difference is presumed to be due to the difference in spike waveforms. Generally, more complex spike waveforms are employed to balance potentiation and depression during STDP, for example, spike waveforms include triangle waves and different positive/negative shapes, voltages, and widths. In this study, emphasis was placed on the simplicity of spike waveforms and concurrent STDP learning. An important result of this study is that we were able to achieve efficient unsupervised learning by combining additional FIDs within these constraints.

Next, we briefly mention the impact of the variation of IDMFETs. The training and classification calculation without the variation was also performed, but there are no significant differences in classification accuracy and learning efficiency (Figure 8b,c). We performed similar calculations with a wider distribution than the experimentally observed variation of IDMFETs and found a decline in learning performance. For example, if the variation is 10 times wider than that of the IDMFETs, the maximum accuracy drops to 70%. This means that whereas the current level of variation is acceptable, devices with excessive variation should be treated with caution.

Finally, let us discuss the impact of the feature neuron size. Figure 8e illustrates the training dynamics for different values of n, calculated at V_FID = 3.25 V. Increasing n results in a decrease in learning efficiency due to the increased number of synapses to be learned, where we found a proportional relationship of η = 68 × n. On the other hand, increasing n can improve classification accuracy, as shown in Figure 8e, in which the previously reported accuracy data deduced by similar networks with STDP-based unsupervised learning were compared [31,32,33]. It is important to note that previous studies were not related to the device characteristics or are not based on the actual device dynamics. It is evident that even with the STDP characteristics of the IDMFET, introducing suitable FID operations can achieve accuracy equivalent to conventional SNN. Based on these results, IDMFET is considered a promising candidate as a synaptic device for unsupervised SNN learning. Particularly noteworthy is the fact that, in typical SNN systems, the number of synapses is orders of magnitude larger than that of neurons; therefore, the implementation of high-density synaptic devices using IDMFETs is expected to be highly effective. Additionally, it is expected that not only IDMFETs but also other three-terminal devices with similar nonlinear response characteristics can be applicable to unsupervised SNN learning; for example, ferroelectric HfO₂ FeFETs and flash memory devices are promising candidates [15,16,38]. Achieving high-performance integrated synaptic devices based on these Si-based FETs, including IDMFETs, requires a meticulous investigation of the effects of device scaling, endurance, and reliability during STDP operation, among other considerations. Furthermore, nonlinear responses are frequently observed in non-Si-based three-terminal memory devices, including two-dimensional materials, organic materials, and so on [39,40,51,52]. We expect that such emerging devices can be extended to unsupervised learning, following a similar approach to the one adopted in this study.

4. Conclusions

The pulse response characteristics of HfO₂/SiO₂-based IDM devices were investigated in detail, with the aim of applying them to synaptic STDP learning in SNN systems. The pulse measurements of IDM MOS capacitors reveal that IDM has an essentially nonlinear and near-symmetric response. The cause of the nonlinearity was presumed to be due to a variety of Ti-O chemical bonding states at the HfO₂/SiO₂ interface. The IDMFETs exhibit nonlinear and asymmetric dynamics of pulse-induced I_d modulation that are associated with the superimposition of an inherent nonlinear and near-symmetric IDM response and FET operation characteristics. STDP-like spike-timing-dependent double-pulse-controlled I_d modulation of IDMFETs can be demonstrated, showing nonlinear and asymmetric weight update dynamics. We propose a timing-independent double-pulse-controlled I_d depression of IDMFETs as an additional FID weight update for SNN to optimize unsupervised STDP learning. The training and classification of handwritten digits with a two-layer SNN, leveraging the experimentally observed STDP and FID characteristics, reveals the effectiveness of IDMFET synaptic devices for unsupervised learning in SNNs.