Direct Triboelectric Programming of a Ferroelectric Synaptic Transistor for Neuromorphic Tactile Perception

ABSTRACT

Triboelectric nanogenerators provide a direct electrical interface to mechanical stimuli, yet most triboelectric–neuromorphic integrations rely on volatile conductance modulation or indirect capacitive coupling, limiting memory‐assisted tactile perception. Here, we demonstrate direct triboelectric programming of a MoS₂/HfO₂/PbZr_0.3Ti_0.7O₃ ferroelectric transistor, where rectified triboelectric pulses are coupled to ferroelectric switching to realize cumulative conductance modulation. The spike‐like outputs generated by mechanical contact–separation are translated into history‐dependent conductance evolution through polarization‐controlled electrostatic modulation in PbZr_0.3Ti_0.7O₃ assisted by charge trapping in HfO₂. Systematic correlation between triboelectric pulse amplitude, repetition frequency, and channel conductance reveals that repetitive mechanical inputs progressively accumulate polarization‐mediated states without external signal amplification or continuous waveform processing. The device exhibits fundamental synaptic characteristics under electrical stimulation, while triboelectrically generated inputs enable cumulative tactile programming. By constructing a force–frequency‐dependent conductance map and defining a reference current window, tactile conditions are encoded into stable state regimes, allowing threshold‐based discrimination based on accumulated device states rather than instantaneous signals. Furthermore, the experimentally parameterized encoding framework is extended to electrocardiogram signal classification, demonstrating its capability to transform temporal bio‐signals into device‐relevant conductance features for neuromorphic signal processing. These results establish a physics‐grounded route for analog memory‐assisted tactile state encoding in neuromorphic sensory systems.

Triboelectric signals directly program a ferroelectric synaptic transistor, enabling cumulative polarization‐controlled conductance modulation and neuromorphic tactile perception based on force‐ and frequency‐dependent mechanical stimuli.

1. Introduction

Inspired by the brain's ability to integrate sensing, memory, and decision‐making within distributed neural networks, neuromorphic computing has emerged as a promising paradigm for energy‐efficient information processing [1, 2, 3, 4]. A key feature of neuromorphic systems is the co‐localization of memory and computation, enabling history‐dependent and event‐driven operation without continuous data processing. Extending this concept to physical sensory modalities requires device platforms capable of directly translating external stimuli into nonvolatile internal states. In this context, ferroelectric synaptic devices have attracted significant attention due to their intrinsic nonvolatility, analog weight tunability, and fast polarization switching [5, 6, 7]. In particular, ferroelectric field‐effect transistors (FeFETs) offer a three‐terminal architecture that decouples read and write operations while enabling polarization‐controlled channel conductance modulation. These properties make FeFETs attractive candidates for hardware‐level synaptic implementation [8, 9, 10]. However, most reported ferroelectric synapses are driven exclusively by externally generated electrical pulses, limiting their direct interaction with real‐world mechanical stimuli and requiring additional sensing and signal‐processing circuitry.

Bridging this gap requires neuromorphic devices that can respond directly to external physical stimuli. In recent years, synaptic systems driven by optical [11, 12, 13], thermal [14, 15], or chemical [16, 17, 18] inputs have been actively explored as pathways toward sensory‐integrated neuromorphic computing, demonstrating that external physical signals can be directly translated into device‐level plasticity. However, compared with these modalities, mechanically driven neuromorphic platforms, particularly those based on triboelectric transduction, remain relatively less developed despite the fundamental importance of tactile perception. Mechanical interactions encode environmental information through variations in stimulus amplitude and temporal repetition, forming the basis of tactile sensing. Owing to these properties, tactile sensing plays a critical role in robotics, healthcare monitoring, and human–machine interfaces, where mechanical interactions convey rich and continuous information [19, 20, 21, 22]. Therefore, translating such mechanically encoded information into stable internal states at the device level is crucial for realizing compact neuromorphic sensory systems.

Triboelectric nanogenerators (TENGs), which convert mechanical contact–separation into electrical signals, provide a direct electrical interface to tactile stimuli [23, 24, 25]. Nevertheless, most TENG‐integrated synaptic systems rely on high gate‐equivalent voltages [26, 27], capacitive coupling [28], or volatile conductance modulation mechanisms such as ion migration [29, 30] or transient charge trapping [31, 32]. As a result, memory‐assisted tactile encoding based on intrinsic ferroelectric polarization switching remains underexplored. To address this gap, recent studies have begun to interface TENG outputs with ferroelectric transistor platforms [33], demonstrating nonvolatile state modulation driven by triboelectric signals. For example, Fang et al. demonstrated a TENG‐driven FeFET structure as a nonvolatile ferroelectric transistor memory [26]. Lee et al. reported triboelectric‐driven ferroelectric organic transistor platforms, including a flexible intrinsic‐synaptic tactile sensory organ [34]. More recently, Kim et al. applied triboelectric programming to MoS₂‐based FeFET synapses, enabling tactile sensory memory and pattern recognition at the system level [35]. However, despite these advances, a physics‐grounded understanding of how spike‐like triboelectric pulses directly induce polarization‐mediated cumulative conductance evolution remains lacking. In particular, the relationship between mechanical stimulus parameters (e.g., force amplitude and repetition frequency) and the progressive evolution of ferroelectric states has not been systematically established. Clarifying this correlation is, therefore, essential for moving beyond proof‐of‐concept demonstrations toward cumulative tactile state encoding at the device level.

Motivated by this need, in this work, we demonstrate direct triboelectric programming of a MoS₂/HfO₂/PbZr_0.3Ti_0.7O₃ (PZT) FeFET, in which rectified triboelectric pulses directly modulate ferroelectric polarization to produce cumulative nonvolatile‐conductance modulation. We elucidate the interplay between polarization dynamics in PZT and charge trapping in HfO₂, governing history‐dependent synaptic state evolution under repetitive mechanical inputs. By constructing a force–frequency‐dependent conductance map and defining a reference current window, we show that tactile conditions can be encoded into stable device states without continuous waveform monitoring. We further extend the experimentally parameterized FeFET encoding framework to electrocardiogram (ECG) signal classification, highlighting its potential as a temporal feature‐encoding element for neuromorphic bio‐signal processing. This approach establishes a compact device platform for cumulative tactile state encoding based on direct triboelectric–ferroelectric coupling.

2. Results and Discussion

2.1. Device Architecture

The overall device concept is illustrated in Figure 1. In the biological sensory–synaptic–decision pathway (Figure 1a), tactile stimuli are converted into neural spike signals, integrated through synaptic transmission, and ultimately interpreted by the brain. This process highlights the cumulative and history‐dependent nature of biological tactile perception. Inspired by this mechanism, the corresponding device architecture is designed in Figure 1b. In the device implementation, mechanical stimulation is converted into spike‐like electrical signals by a TENG. Then, the rectified TENG signals are directly applied to the gate of a MoS₂/HfO₂/PZT FeFET, enabling polarization‐controlled channel conductance modulation. Through repetitive stimulation, the conductance accumulates in a history‐dependent manner, analogous to synaptic weight modulation. The accumulated conductance state is then interpreted within a predefined current window to classify tactile conditions into distinct regimes (e.g., weak/slow, normal, strong/fast).

FIGURE 1.

Conceptual illustration of triboelectrically driven neuromorphic tactile perception. (a) Biological tactile perception process mediated by synaptic signal transmission and cumulative weight modulation. (b) Device‐level implementation using a TENG‐driven synaptic FeFET, where mechanical stimuli are converted into triboelectric pulses that program cumulative conductance states for tactile perception and threshold‐based decision making.

2.2. Electrical Characteristics of TENG

The electrical output characteristics of the fabricated TENG were systematically investigated under periodic mechanical contact–separation actuation. As shown in Figure 2a, the TENG device generates spike‐like voltage (V _TENG) waveforms, whose peak amplitude increases monotonically with increasing applied mechanical force. When the applied force is varied from 1.3 to 2.1 N, the peak V _TENG increases correspondingly, while the overall waveform profile remains consistent. This behavior indicates that the stimulus intensity is primarily encoded in the voltage amplitude. The spike‐like nature of the TENG output originates from the dynamic electrostatic potential variation during the contact and separation transitions. As illustrated in Figure 2b,c, the TENG operation can be divided into four stages: fully separated, contacting, fully contacted, and separating. In the fully separated and fully contacted states, the charge distribution remains quasi‐static, resulting in negligible output voltage. In contrast, rapid charge redistribution during the contacting and separating transitions induces transient current flow in the external circuit, producing sharp voltage spikes [36]. This transient behavior is characteristic of contact–separation‐mode TENGs and directly reflects interfacial charge transfer dynamics.

FIGURE 2.

Electrical characterization of the PTFE‐based TENG. (a) Output V _TENG waveforms generated under different applied mechanical forces. (b) Spike‐like V _TENG waveform originating from the separation‐contact motion. (c) Working mechanism based on contact electrification and electrostatic induction during the contact–separation cycle. (d) Rectified unipolar V _TENG pulses. (e) Dependence of V _TENG on applied force, demonstrating force‐sensitive electrical generation with error bars indicating cycle‐to‐cycle variation. (f) Stability and (g) reproducibility of the TENG output over repeated mechanical cycles.

To enable gate‐driven FeFET operation using TENG signals, the intrinsically bipolar TENG output was rectified using diode circuitry to produce unipolar voltage pulses (Figure 2d). After rectification, the polarity becomes uniform while preserving the force‐dependent amplitude variation, enabling direct coupling to the transistor gate without additional amplification circuitry. The quantitative relationship between mechanical force and output voltage is summarized in Figure 2e. The peak V _TENG exhibits a monotonic increase with applied force, confirming stable and reproducible force‐to‐voltage transduction. At a fixed operating frequency of 1.0 Hz, rectified pulse trains generated under different force conditions exhibit consistent waveform shapes with varying amplitudes, as shown in Figure 2f. The repetition rate of the pulses is governed by the mechanical actuation frequency, while the pulse amplitude is dictated by the applied force magnitude. These characteristics demonstrate that mechanical stimulus parameters such as force and frequency can be independently encoded into the amplitude and the repetition rate of electrical spikes. Such well‐defined electrical outputs provide suitable programming signals for investigating direct triboelectric modulation of the FeFET synapse in the following sections.

Furthermore, the triboelectric output exhibits excellent stability and reproducibility, as confirmed by monitoring the peak voltages over repeated mechanical cycles (Figure 2g). Both the positive (V _TENG+) and negative (V _TENG−) peak voltages remain nearly constant over 10³ operation cycles, exhibiting only minor fluctuations without noticeable degradation or drift. The positive peaks stabilize around approximately 1.8–2.2 V, while the negative peaks remain near −2.6 to −3.0 V, demonstrating stable charge generation and release during repetitive contact–separation processes. The absence of significant cycle‐dependent variation confirms the mechanical robustness of the stacked Cu/polytetrafluoroethylene (PTFE)/Cu structure and the reliability of the triboelectric signal for repetitive gate programming. This stability is particularly important for cumulative ferroelectric modulation, where consistent pulse amplitude is required to ensure controlled state evolution in the FeFET.

2.3. Polarization‐Controlled Electrical Characteristics of FeFET

The structural and ferroelectric properties of the MoS₂/HfO₂/PZT FeFET were first examined to verify polarization‐controlled channel modulation. The Raman spectrum of the transferred MoS₂ channel is shown in Figure 3a. The characteristic E¹ _2g and A_1g modes are observed at approximately 384 cm⁻¹ and 404 cm⁻¹, respectively, with a peak separation of ∼20 cm⁻¹, corresponding to a 2–3 layer MoS₂ film [37]. This thickness provides an appropriate balance between electrostatic gate control and carrier transport stability.

FIGURE 3.

Structural and electrical characteristics of the MoS₂/HfO₂/PZT FeFET. (a) Raman spectrum of the MoS₂ channel. (b) P–V hysteresis loops of the HfO₂/PZT gate stack. (c) PUND‐extracted P–V hysteresis loops of HfO₂/PZT stack. (d) Switching current response of the HfO₂/PZT stack. (e) PUND extracted P–V hysteresis loops of MoS₂/HfO₂/PZT stack. (f) Switching current response of the MoS₂/HfO₂/PZT stack. (g) I _D–V _G transfer characteristics of the MoS₂/HfO₂/PZT FeFET (L = 5 µm, W = 200 µm) showing clockwise hysteresis. (h) I _D–V _D output characteristics. (i) Multi‐level memory states programmed by different V _pulse.

In the HfO₂/PZT gate stack, the polarization–voltage (P–V) characteristic curves clearly display well‐defined hysteresis loops with substantial remanent polarization (Figure 3b), confirming robust ferroelectric switching in the PZT layer. The incorporation of HfO₂ in the present MoS₂/HfO₂/PZT was intentionally adopted to enable polarization‐controlled channel modulation, while suppressing direct carrier injection into the ferroelectric layer during gate programming. As summarized in Figure S1, a capacitor composed of a single PZT layer (i.e., Au/PZT/Pt) exhibits typical abrupt ferroelectric switching behavior in both P–V and capacitance–voltage (C–V) measurements, where nearly full polarization is achieved once the applied voltage exceeds the coercive voltage, largely independent of further increases in sweep amplitude. In contrast, the Au/HfO₂/PZT/Pt stack shows expanded hysteresis windows with an applied voltage‐dependent polarization evolution. This voltage‐tunable polarization behavior originates from the insertion of the HfO₂ interlayer, which modifies the internal electric field distribution across the HfO₂/PZT ferroelectric/insulator bilayer stack. The high‐k HfO₂ interlayer electrically isolates the semiconductor channel from the ferroelectric film while maintaining polarization‐driven electrostatic coupling, thereby suppressing leakage currents during polarization switching [38, 39]. This metal–ferroelectric–insulator–semiconductor (MFIS) structure is particularly critical for triboelectric programming, where transient voltage spikes must modulate polarization without inducing excessive leakage currents.

To further examine the influence of the MoS₂ channel on ferroelectric switching behavior, positive–up–negative–down (PUND) measurements were performed on the HfO₂/PZT stack both with and without the MoS₂ layer (Figure 3c–f). The PUND measurement scheme enables separation of the pure switching polarization from non‐switching contributions, such as leakage current and capacitive charging, thereby allowing direct evaluation of intrinsic ferroelectric polarization reversal [40, 41]. Figure 3c,d show the PUND‐extracted P–V hysteresis loops and corresponding transient current responses of the HfO₂/PZT stack without MoS₂ channel, respectively. The polarization and switching current responses increase with pulse amplitude, confirming voltage‐dependent ferroelectric switching in the gate stack. Figure 3e,f present the PUND‐extracted P–V hysteresis loops and transient current responses after MoS₂ channel integration, respectively. The preserved hysteresis and switching current responses indicate that ferroelectric polarization switching remains active in the MoS₂/HfO₂/PZT MFIS structure. The PUND measurements were conducted using voltage pulses of ±0.5 to ±5 V with a pulse width of 1 ms and an interval time of 2 ms. From the integrated charge of each pulse, the switching charge (Q _sw) was extracted according to [42, 43]:

$$Q_{sw}=\frac{\left(Q_P-Q_U\right)+\left(Q_N-Q_D\right)}{2}$$ (1)

where Q _P and Q _N correspond to the polarization switching charges measured during the positive and negative switching pulses, respectively, and Q _U and Q _D represent the non‐switching reference pulses. The extracted Q _sw values confirm that ferroelectric polarization‐switching responses are maintained even after integration of the MoS₂ channel. Although slight asymmetry is observed between positive and negative switching branches, this behavior is likely caused by the asymmetric MFIS electrostatic boundary conditions [44], including interfacial trap states and work‐function mismatch between the Pt metal electrode and the MoS₂ semiconductor channel in the MoS₂/HfO₂/PZT/Pt stack. Nevertheless, the overall switching behavior remains stable, confirming that the MFIS configuration preserves effective ferroelectric field coupling to the MoS₂ channel. Furthermore, low‐voltage PUND measurements ranging from ±0.5 to ±5 V show detectable sub‐coercive polarization‐switching components, supporting that the rectified TENG pulse of approximately 1.95 V can progressively induce partial polarization modulation under repeated stimulation.

With the ferroelectric switching behavior verified, the transfer characteristics (I _D–V _G) of the FeFET were further investigated to assess polarization‐controlled channel conductance modulation. Figure 3g shows the I _D–V _G curves of the MoS₂/HfO₂/PZT FeFET measured at V _D = 100 mV. Clear hysteresis windows are observed during forward and reverse gate‐voltage sweeps, confirming polarization‐dependent electrostatic modulation of charge distribution in the MoS₂ channel. Through a comparative analysis of three reference samples including a typical FeFET, a charge‐trap FET, and an MFIS FeFET, the observed hysteretic behavior was confirmed to arise from ferroelectric polarization switching in HfO₂/PZT rather than the charge‐trapping effect in HfO₂ (Figure S2). Notably, the hysteresis direction is clockwise (i.e., antihysteresis), which deviates from the ideal counterclockwise loop expected from purely ferroelectric switching [45, 46, 47]. This clockwise hysteresis can be understood in terms of polarization (in PZT)–trap (in HfO₂) coupling in the MFIS FeFET structure. Under positive gate bias, electrons accumulated in the MoS₂ channel are partially trapped at native point‐defect states within the HfO₂ interlayer. The trapped electrons lower the local electrochemical potential inside HfO₂ and effectively shift the flat‐band condition, resulting in a positive shift of the threshold voltage (V _th). Consequently, during the backward sweep under the same positive gate bias, the modified electrostatic environment gives rise to a clockwise hysteresis loop. Importantly, the trapped charges are not freely relaxed because they are electrostatically confined by the strong ferroelectric polarization field of the underlying PZT layer. The polarization‐induced electric field stabilizes the trapped electrons and sustains the modified flat‐band condition, leading to sweep‐direction‐dependent threshold modulation. This polarization–trap coupling mechanism accounts for the observed antihysteresis behavior in the MoS₂‐based MFIS FeFET. A more detailed discussion of the polarization–trap coupling mechanism, including the schematic band modulation model, is provided in Figure S3.

The output characteristics (I _D–V _D) of the FeFET, obtained by sweeping V _D from 0 to 1 V under different gate biases, are shown in Figure 3h. The I _D increases monotonically with increasing V _G, confirming n‐type conduction in the MoS₂ channel. These results verify stable field‐effect modulation of the channel current under gating operation. Furthermore, consistent I _D–V _D and I _D–V _G characteristics were reproducibly obtained from multiple devices fabricated under identical process conditions (Figures S4 and S5), demonstrating robust process reliability and uniform polarization–trap coupling behavior in the MFIS structure. All measured devices showed comparable output characteristics, with monotonic I _D increase under increasing V _G and no evidence of abnormal leakage or device degradation. In addition, the transfer curves consistently exhibited clockwise hysteresis with similar memory windows and V _th modulation trends as the sweep voltage increased. The progressive broadening of the memory window with enhanced programming strength is observed across different samples, confirming stable polarization‐controlled modulation. The limited variation in hysteresis window and drain current levels further indicates good fabrication uniformity and reliable integration of the MoS₂/HfO₂/PZT stack. Collectively, these results demonstrate that the polarization‐controlled modulation (i.e., bias voltage‐dependent memory state tunability) is an intrinsic and reproducible feature of the MoS₂/HfO₂/PZT MFIS device scheme.

Building upon this stable transistor operation, Figure 3i demonstrates the multi‐level conductance tunability of the FeFET under different programming conditions (V _pulse = ±1 to ±8 V). The drain current exhibits a gradual and monotonic change as the gate programming strength is incrementally enhanced. This enables multiple distinguishable conductance states rather than abrupt binary switching. Importantly, the programmed states remain clearly separated without significant overlap, indicating stable and reproducible state formation. The persistence of these intermediate conductance levels confirms that the polarization‐controlled on‐state current modulation is sufficiently robust to support analog‐like multi‐level operation. Such stable multi‐state behavior provides the essential electrical foundation for synaptic weight representation and cumulative tactile state encoding in the subsequent triboelectric‐driven experiments.

In addition, the endurance characteristics confirm the stable reversible conductance modulation of the MoS₂/HfO₂/PZT MFIS FeFET. As shown in Figure S6a,b, the program and erase current states remain clearly distinguishable over 1000 cycles, indicating that repeated operation does not significantly degrade the memory window or the programmed conductance states. Furthermore, the PUND hysteresis loops measured during repeated cycling show no noticeable collapse or distortion (Figure S6c,d), confirming that the ferroelectric polarization of the PZT layer is preserved without significant fatigue. These results indicate that the repeated conductance modulation in the MFIS FeFET is not dominated by irreversible trap generation or polarization degradation, supporting its reliability for cumulative synaptic operation.

2.4. Electrical Synaptic Characteristics

To evaluate the synaptic functionality of the FeFET under electrical stimulation, gate voltage pulses were applied while monitoring the I _D response at a constant V _D (= 100 mV). Figure 4a shows the excitatory postsynaptic current (EPSC) responses induced by a single negative gate pulse with varying pulse amplitudes (V _pulse = −2 to −5 V) and pulse widths (t _pulse = 10 µs–1 ms). Upon pulse application, the drain current increases abruptly, followed by gradual relaxation after pulse removal. The peak change in postsynaptic current (ΔPSC) increases systematically with both pulse amplitude and pulse width, indicating that stronger or longer stimuli induce a larger degree of polarization‐mediated conductance potentiation. This pulse amplitude‐ and width‐dependent potentiation behavior is quantitatively summarized in Figure 4b, where ΔPSC increases monotonically with increasing stimulation strength. In contrast, the positive V _pulse (+2 to +5 V) generates inhibitory postsynaptic current (IPSC) responses, as shown in Figure 4c. The ΔPSC decreases incrementally with increasing V _pulse and t _pulse, corresponding to synaptic depression behavior. The extracted ΔPSC values in Figure 4d confirm a controllable and polarity‐dependent depression process. These EPSC/IPSC behaviors are consistent with the V _th modulation mechanism described in Section 2.3, arising from the polarization–trap coupling mechanism governing polarity‐dependent conductance shifts.

FIGURE 4.

Electrical synaptic characteristics of the fabricated FeFET. (a) EPSC responses induced by negative gate pulses with varying V _pulse and t _pulse. (b) ΔPSC in EPSC as a function of V _pulse for different t _pulse. (c) IPSC responses induced by positive gate pulses with varying V _pulse and t _pulse. (d) ΔPSC in IPSC as a function of V _pulse for different t _pulse. (e) PPF behavior under different t _inter. (f) PPF index as a function of t _inter. (g) PPD behavior under different t _inter. (h) PPD index as a function of t _inter. (i) LTP/LTD characteristics under pulse trains with different amplitudes. Error bars in (b), (d), (f), and (h) represent the standard deviation obtained from 5‐times repeated measurements.

Short‐term synaptic plasticity was further evaluated through paired‐pulse facilitation/depression (PPF/PPD) measurements. Figure 4e,g present the PPF and PPD characteristics, respectively, measured under pulse‐to‐pulse intervals (t _inter) ranging from 2 to 150 ms. When two identical pulses (V _pulse = ±3 V, t _pulse = 10 µs) are applied with a short interval, the second response (A ₂) is enhanced (for PPF) or suppressed (for PPD) relative to the first response (A ₁), owing to incomplete relaxation of polarization charges and trapped carriers. In biological synapses, such facilitation and depression arise from transient modulation of neurotransmitter release probability [48, 49], whereas in this device, they originate from residual polarization–trap coupling. As t _inter increases, the facilitation or depression effect gradually diminishes, reflecting the temporal decay of ΔPSC. The PPF and PPD indices plotted in Figure 4f,h, respectively, exhibit exponential relaxation behavior. The decay was fitted using a double‐exponential function [50, 51]:

$$PPF\left(orPPD\right)Index=y_0+A_{1}exp\left(\frac{-t}{{\tau }_1}\right)+A_{2}exp\left(\frac{-t}{{\tau }_2}\right)$$ (2)

which captures the coexistence of fast and slow relaxation processes. For PPF, the extracted time constants were τ ₁ = 3.36 ms and τ ₂ = 54.60 ms, whereas for PPD, τ ₁ = 0.23 ms and τ ₂ = 53.88 ms. The coexistence of fast and slow relaxation components within a few to tens of milliseconds indicates multi‐level charge dynamics in the fabricated MFIS FeFET, closely resembling the temporal short‐term memory characteristics observed in biological synapses [52, 53] and neuromorphic devices exhibiting multi‐timescale relaxation dynamics [54, 55].

Beyond short‐term plasticity, the long‐term plasticity characteristics were also investigated by long‐term potentiation/depression (LTP/LTD) measurements. As shown in Figure 4i, consecutive pulse trains were applied in stepwise‐increasing amplitude blocks, resulting in a progressive evolution of synaptic weight with pulse number. Within each pulse block, ΔPSC gradually increases in LTP and decreases in LTD, demonstrating analog‐like and nonvolatile modulation behavior. Notably, as the pulse amplitude increases from one block to the next, the achievable ΔPSC window expands correspondingly, leading to an enlarged dynamic range of the synaptic weight. This stepwise expansion of ΔPSC indicates that stronger programming stimuli enable access to higher polarization states and deeper electrostatic modulation levels. The weight update remains continuous and distinguishable across different amplitude regimes, confirming stable multi‐level state formation. This gradual and amplitude‐dependent evolution originates from the cumulative polarization–trap coupling mechanism, where successive pulses progressively modulate the internal electrostatic potential landscape. Such controllable dynamic range expansion provides a robust electrical foundation for analog neuromorphic learning and cumulative tactile state encoding, as discussed in the following section.

2.5. Direct Triboelectric Programming and Cumulative Tactile State Encoding

To extend the device‐level synaptic characteristics toward a system‐level tactile perception framework, the triboelectrically driven FeFET was evaluated under mechanical stimuli with varying force and frequency. As schematically illustrated in Figure 5a, external mechanical inputs applied to the TENG generate electrical pulses whose amplitude and repetition rate depend on the applied force and actuation frequency. After rectification, these triboelectric pulses are directly applied to the FeFET gate without an external gate power supply. The resulting I _D can be further processed using a comparator circuit for threshold‐based event detection, enabling conceptual integration of sensing, memory, and decision functions within a compact hardware architecture analogous to the biological perception–memory–decision pathway [56]. The detailed circuit configuration is shown in Figure 5b, where the rectified V _TENG serves as the presynaptic gate stimulus. Owing to the nonvolatile characteristics of the FeFET, repetitive triboelectric pulses induce cumulative ΔPSC modulation through polarization–trap coupling, resulting in progressive ΔPSC evolution.

FIGURE 5.

Triboelectrically driven neuromorphic tactile system. (a) Schematic illustration of the TENG–FeFET integrated tactile perception system, where mechanical force is converted into electrical gate stimuli. (b) Circuit configuration including TENG rectification and FeFET‐based threshold detection. (c) Force‐ and frequency‐dependent ΔPSC responses under rectified TENG stimulation with window‐based state classification, where the shaded region denotes the representative normal‐state operating window defined near the average value of the total PSC dynamic range of approximately 0.47 µA. (d) Force–frequency state map constructed from the maximum ΔPSC, showing window‐based classification into low, normal, and high regimes.

Figure 5c presents the time evolution of I _D under different combinations of stimulation frequency (0.5–2.0 Hz) and applied force (1.3–2.1 N). The triboelectric input pulses generated from the TENG device and directly applied to the FeFET gate are shown in Figure S7. For a fixed force level, increasing the stimulation frequency leads to a faster and more pronounced conductance accumulation. At low frequencies, sufficient relaxation of polarization–trap coupling occurs between successive pulses, limiting cumulative modulation. In contrast, higher frequencies reduce the relaxation interval, resulting in incomplete charge detrapping and enhanced cumulative polarization screening, thereby accelerating the increase in ΔPSC. This frequency‐dependent accumulation behavior is analogous to spike‐rate‐dependent facilitation in biological synapses [57]. In addition to frequency dependence, a clear force‐dependent response is observed. At a given frequency, increasing the applied force generates higher V _TENG, enabling access to larger polarization switching fractions and deeper electrostatic modulation. Consequently, stronger mechanical stimuli produce a wider ΔPSC excursion and a faster transition toward higher synaptic weights. In contrast, weak stimuli result in limited accumulation and remain confined to lower conductance regimes. This force‐selective behavior reflects the intensity discrimination observed in biological tactile receptors [58, 59], where stimulus strength is encoded through firing rate and synaptic weight modulation [60].

To confirm device‐to‐device reliability and reproducibility, the force‐ and frequency‐dependent ∆PSC measurements were performed using multiple devices under the same experimental conditions as those used in Figure 5c. A representative result from another device is provided in Supporting Information (Figure S8), showing qualitatively consistent cumulative conductance modulation. The shaded region in Figure 5c defines a reference ΔPSC window for state classification based on accumulated ΔPSC. If ΔPSC remains below the window, the input is classified as a low‐level tactile state. Once the accumulated ΔPSC exceeds the upper boundary, the system transitions into a high‐level state. Importantly, due to the nonvolatile nature of ferroelectric polarization, the classified state is retained even after mechanical stimulation ceases, enabling memory‐assisted decision‐making based on stimulus history rather than instantaneous signal amplitude alone. Overall, the triboelectrically driven FeFET encodes stimulus frequency and force into analog, cumulative conductance states through the same polarization–trap coupling mechanism responsible for the anti‐hysteresis and synaptic plasticity described earlier. This unified physical mechanism enables direct mechanical‐to‐memory transduction, supporting compact and low‐energy tactile perception systems.

The force–frequency map in Figure 5d summarizes the steady‐state ΔPSC levels obtained after repeated mechanical stimulation. To reflect device‐to‐device reliability and reproducibility, the map was constructed using the average ΔPSC values obtained from multiple devices. A defined operating current window is used as a reference for state classification. Conditions in which the accumulated ΔPSC remains within this window are interpreted as normal states, whereas deviations beyond the upper or lower boundaries indicate abnormal operating conditions. The labeled regions in the map correspond to stimulus combinations that drive the device outside the defined conductance range, reflecting either excessive or insufficient cumulative modulation.

Importantly, the emergence of abnormal states is governed by cumulative ΔPSC accumulation rather than a simple binary threshold applied to the amplitude or frequency alone. High‐force inputs can induce significant polarization switching even at low frequencies, while frequency‐dependent accumulation becomes more pronounced when the stimulus amplitude approaches the boundary of the operating window. This indicates that the device response reflects a synergistic interaction between stimulus intensity and temporal repetition, with stimulus amplitude dominating under extreme conditions. Because state classification is based on accumulated ΔPSC relative to a reference window, rather than explicit waveform decoding, the same framework can be extended to pulse‐like mechanical bio‐signals such as heartbeats. In such scenarios, repetitive mechanical signals can progressively shift the device's conductance across the decision boundary, enabling threshold‐based sensing and synaptic memory‐assisted decision‐making without continuous signal processing. Consequently, the force–frequency mapping demonstrates that the triboelectrically driven FeFET encodes tactile stimuli into cumulative analog ΔPSC states, enabling event‐driven perception based on intrinsic device memory. By integrating sensing, memory, and state discrimination within a single hardware element, this approach provides a compact and energy‐efficient platform for neuromorphic tactile systems.

Finally, an ECG classification task was performed using experimentally extracted synaptic characteristics to further demonstrate the applicability of the proposed FeFET beyond tactile signal processing (Figure 6). ECG signals are temporally structured physiological waveforms in which diagnostic information is encoded not only in instantaneous signal amplitudes but also in the temporal evolution of characteristic waveform components. Therefore, the experimentally measured synaptic dynamics of the FeFET were utilized as device‐level temporal encoding functions to transform ECG waveforms into conductance‐domain representations prior to neural‐network classification. The ECG data used in this simulation were obtained from the MIT‐BIH Arrhythmia Database [61], containing five representative heartbeat classes: normal beat (N), left bundle branch block beat (L), right bundle branch block beat (R), premature ventricular contraction beat (V), and atrial premature beat (A). Each ECG segment was normalized and resampled to 200 points before FeFET‐based encoding. As schematically illustrated in Figure 6a, the encoded feature matrix was subsequently flattened and used as the input to a lightweight multilayer‐perceptron classifier consisting of a fully connected dense layer with 32 hidden units and ReLU activation, followed by a dropout layer with a dropout rate of 0.15 and a final dense layer containing five softmax output neurons. The first dense layer projects the FeFET‐encoded 400‐dimensional feature vector into a compact latent representation, allowing the classifier to combine temporal conductance features distributed across the ECG segment. The ReLU activation introduces nonlinearity into the classifier, while the dropout layer randomly deactivates a fraction of the hidden units during training to reduce overfitting and improve generalization. The final softmax layer converts the learned representation into class probabilities for the five ECG categories. Because only a single hidden layer was employed, the classification performance mainly reflects the effectiveness of the FeFET‐based temporal encoding rather than the complexity of the classifier architecture [62].

FIGURE 6.

FeFET parameter‐based ECG signal encoding and classification. (a) Schematic of ECG feature encoding using G _norm and ΔG _norm, followed by multilayer perceptron‐based five‐class classification. (b) Class‐averaged ECG waveforms for the heartbeat categories, where the solid lines represent the mean waveform, and the shaded regions indicate the standard deviation for each class. (c) Normalized LTP/LTD characteristics obtained under identical‐pulse and modulated‐pulse operation. (d) Training accuracy comparison between identical‐pulse‐based nonlinear encoding and modulated‐pulse‐based linear encoding. Confusion matrices obtained using (e) identical‐pulse‐ and (f) modulated‐pulse‐based encoding.

Figure 6b presents the class‐averaged ECG waveforms for the five heartbeat categories. Although all heartbeat classes exhibit the characteristic QRS complex associated with ventricular depolarization, noticeable differences exist in waveform morphology, peak amplitude, pulse width, and temporal evolution before and after the QRS event. These subtle temporal variations contain diagnostically relevant information and therefore provide an appropriate benchmark for evaluating the temporal feature‐extraction capability of the FeFET‐based encoding framework. To emulate the experimentally observed synaptic dynamics of the FeFET, the measured LTP/LTD characteristics were normalized and fitted using a widely adopted nonlinear synaptic update model [63]:

$$G_{\mathrm{norm}}=\frac{G-G_{\min }}{G_{\max }-G_{\min }}$$ (3)

$$G_{\mathrm{LTP},\mathrm{norm}}=\frac{1-\exp \left(-{\beta }_{\mathrm{LTP}}\frac{n}{N}\right)}{1-\exp \left(-{\beta }_{\mathrm{LTP}}\right)}$$ (4)

$$G_{\mathrm{LTD},\mathrm{norm}}=\frac{\exp \left(-{\beta }_{\mathrm{LTD}}\frac{n}{N}\right)-\exp \left(-{\beta }_{\mathrm{LTD}}\right)}{1-\exp \left(-{\beta }_{\mathrm{LTD}}\right)}$$ (5)

where β _LTP and β _LTD denote the nonlinearity factors for potentiation and depression, respectively, while n and N represent the cumulative pulse number and total pulse number, respectively. As shown in Figure 6c, two experimentally measured synaptic‐update conditions were considered. The identical‐pulse condition was obtained using fixed‐amplitude potentiation and depression pulses of −2.5 V and +3.4 V, respectively, whereas the modulated‐pulse condition employed amplitude‐incremental pulse schemes ranging from −1.5 to −2.25 V for potentiation and +0.5 to +4.3 V for depression. For both conditions, 100 consecutive pulses with a pulse width of 100 µs were applied during the LTP/LTD measurements. The extracted fitting parameters included the dynamic range (i.e., G _max/G _min) and the potentiation/depression nonlinearity factors (i.e., β _LTP and β _LTD). The obtained values were G _max/G _min = 1.21, β _LTP = 3.72, and β _LTD = 4.80 for the identical‐pulse condition, and G _max/G _min = 1.19, β _LTP = −0.31, and β _LTD = 0.56 for the modulated‐pulse condition (see also Figure S9).

Using these experimentally extracted parameters, each ECG waveform was transformed into conductance‐domain temporal features. Specifically, two conductance‐related features were generated at each time point: the normalized conductance (G _norm) and the normalized conductance variation (ΔG _norm). The G _norm feature represents the accumulated FeFET synaptic state generated by sequential ECG‐derived inputs, while ΔG _norm indicates the local temporal variation of the conductance state between adjacent time points. Accordingly, the resulting 200 × 2 feature matrix preserves both memory‐dependent accumulated responses and transient waveform variations in the conductance domain. This feature matrix was then flattened into a 400‐dimensional vector and used as the input to a lightweight multilayer perceptron. Figure 6d compares the classification accuracy obtained using the identical‐pulse and modulated‐pulse encoding schemes. The identical‐pulse‐based encoding achieved a classification accuracy of 91.7%, whereas the modulated‐pulse‐based encoding achieved 88.5%. This result is noteworthy because linear and symmetric conductance updates are generally considered advantageous when synaptic devices are directly used as trainable weight elements in artificial neural networks. In the present framework, however, the FeFET conductance evolution does not serve as a trainable synaptic weight array but rather as a front‐end temporal encoding kernel. Therefore, the stronger nonlinear conductance evolution observed under the identical‐pulse condition can effectively enhance the separation of subtle ECG waveform features in the conductance domain, resulting in improved classification performance. This observation suggests that conductance nonlinearity, although often considered undesirable for hardware neural‐network training, can become advantageous when the device is employed as a temporal feature‐encoding element.

The confusion matrices in Figure 6e,f further support the above interpretation. Compared with the modulated‐pulse‐based encoding, the identical‐pulse‐based encoding exhibits improved classification consistency across all heartbeat categories. In particular, reduced inter‐class confusion is observed for heartbeat classes exhibiting similar waveform morphologies, indicating that the nonlinear polarization–trap‐coupled conductance dynamics generate more distinguishable conductance‐domain trajectories during temporal encoding. These results suggest that the intrinsic nonlinear synaptic behavior of the FeFET should not be regarded solely as a nonideal weight‐update characteristic. Instead, when employed as a signal‐encoding element, the experimentally measured polarization–trap‐coupled conductance dynamics can function as an effective temporal feature‐extraction kernel for physiological signal classification. The demonstrated ECG classification further highlights that the triboelectrically programmable FeFET can operate not only as a tactile perception device but also as a temporal signal encoder for physiological data, thereby expanding its applicability toward bio‐signal processing and neuromorphic healthcare electronics.

To place the present platform in the context of previously reported TENG‐driven neuromorphic synaptic devices, the representative device architectures, operating mechanisms, neuromorphic functionalities, operating voltages, and synaptic energy consumption reported in the literature are summarized in Table 1. In many previously reported systems, triboelectric signals are utilized mainly as gate‐equivalent stimuli that modulate channel conductance through floating‐gate charge trapping, ion migration, or capacitive tribotronic gating mechanisms. These approaches often require relatively high triboelectric output voltages ranging from several tens to hundreds of volts or rely on additional coupling structures to translate mechanical signals into effective gate programming. In contrast, the present work demonstrates direct triboelectric programming of a ferroelectric MFIS FeFET using rectified TENG pulses with a comparatively low operating voltage of approximately 1.95 V. In addition, the proposed device exhibits low synaptic‐event energy consumption (∼2 pJ), estimated from the postsynaptic/read‐side operation, further supporting its suitability for energy‐efficient neuromorphic operation. The cumulative conductance evolution originates from polarization–trap coupling within the HfO₂/PZT gate stack, enabling nonvolatile synaptic weight modulation directly driven by mechanical stimuli without external amplification or complex signal‐conditioning circuitry. Furthermore, while most previous studies primarily demonstrate fundamental synaptic responses such as EPSC, PPF, or LTP under triboelectric stimulation, the proposed device establishes a force–frequency‐dependent conductance map that enables cumulative tactile state encoding based on self‐updated synaptic states. Together with the FeFET‐parameter‐based ECG classification simulation, this comparison indicates that the proposed TENG–MFIS FeFET platform provides not only a direct mechanical‐to‐memory transduction mechanism for tactile perception but also an experimentally grounded synaptic encoding framework for broader neuromorphic sensory processing.

TABLE 1.

Comparison of the key features for various TENG‐driven synaptic devices.

Device Scheme	TENG		Synaptic device			Synaptic Functions	Neuromorphic Application	Synaptic Energy	Refs.
	TENG Materials	Output Voltage	Channel	Active Layer	Driving Mechanism
TENG–MFIS FeFET	PVDF/PS	∼220 V	Pentacene	PMN‐PT (ferroelectric)	Ferroelectric polarization switching	—	—	∼6 µJ	[26]
TENG–FGFET	Al/PTFE	∼±90 V (gate‐equivalent)	MoS2	Graphene floating gate	Triboelectric gating + Photo‐assisted charge trapping	—	—	∼40 µJ	[31]
TENG–vdw FGFET	Cu/PTFE	∼±50 V (gate‐equivalent)	MoS2	Graphene floating gate	FN tunneling‐based charge trapping in FG	—	—	∼60 nJ	[32]
TENG–Fe‐OFET	Human skin/PI	∼±10 V	Pentacene	BT/P(VDF–TrFE)	Ferroelectric polarization switching	—	—	∼45 nJ	[34]
TENG–DGFET	Cu/FEP	−20–60 V (gate‐equivalent)	IGZO	Air dielectric	PPC from oxygen vacancies	—	—	∼84 nJ	[70]
TENG–vdw semi‐FGFET	Cu/FEP	∼±65 V (gate‐equivalent)	WSe2	Graphene semifloating gate	FN tunneling‐based charge trapping reconfiguring pn junctions	EPSC/IPSC, PPF/PPD, LTP	—	∼74.2 fJ	[71]
TENG–vdw FGFET	Al/Kapton/Silicone rubber	Tens of volts (gate‐equivalent)	InSe	Graphene floating gate	FN tunneling‐based charge trapping in FG	EPSC/IPSC, PPF/PPD, LTP/LTD	—	∼165 aJ	[72]
TENG–DGFET	Cu/PTFE	∼±60 V (gate‐equivalent)	Graphene / MoS2	Graphene/MoS2 heterostructure	Triboelectric gating + Photonic excitation	EPSC/IPSC, STP/LTP	Artificial neural network for image recognition	∼1 µJ	[27]
TENG–EGFET	Silk‐fibroin/ITO	∼5 V	IGZO	Chitosan electrolyte	EDL formation via proton migration	EPSC, PPF, LTP	Biological tactile perception	∼37.5 nJ	[29]
TENG–IGT	Al/PTFE	0.1–0.5 V (gate‐equivalent)	MoS2	Ion gel (EDL)	EDL formation	EPSC, PPF, LTP	Contact‐electrification‐activated artificial afferents	∼11.9 fJ	[30]
TENG–FeFET	Al/PDMS	Tens of volts (gate‐equivalent)	MoS2	P(VDF–TrFE) ferroelectric polymer	Ferroelectric polarization switching	EPSC/IPSC, LTP/LTD	Human tactile peripheral system	∼30 pJ	[35]
TENG–IGT	Au/FEP/Kapton	∼2 V	PDVT‐10	Ion‐gel (EDL)	EDL formation via ion migration	EPSC, PPF, STM/LTM	Artificial auditory pathway	∼30 pJ	[73]
TENG–MFIS FeFET	Cu/PTFE	∼1.95 V (rectified TENG output)	MoS2	HfO2/PZT (ferroelectric)	Ferroelectric polarization switching + Charge trapping	EPSC/IPSC, PPF/PPD, LTP/LTD	Cumulative tactile state encoding	∼2 pJ	This work

3. Conclusions

A triboelectrically driven MoS₂/HfO₂/PZT FeFET was demonstrated as a neuromorphic tactile perceptron device that integrates mechanical sensing, nonvolatile memory, and state‐level interpretation within a single hardware platform. Triboelectric signals generated by mechanical stimuli were directly coupled to the FeFET gate, enabling cumulative analog ΔPSC modulation governed by polarization–trap coupling. The device exhibited both short‐ and long‐term synaptic plasticity, while triboelectric inputs induced history‐dependent conductance evolution without external amplification or continuous waveform processing. Through systematic force‐ and frequency‐dependent measurements, a predefined ΔPSC window‐based perception map was established, allowing tactile conditions to be classified according to accumulated conductance states rather than instantaneous signal amplitude. In addition, ECG classification using experimentally extracted FeFET LTP/LTD parameters verified that the same polarization–trap‐coupled conductance dynamics can serve as a temporal encoding function for neuromorphic bio‐signal processing. These results highlight a direct mechanical‐to‐memory transduction mechanism and suggest a compact, low‐energy FeFET‐based framework for neuromorphic sensory systems beyond tactile perception.

4. Experimental Section

4.1. TENG Fabrication

The triboelectric sensing unit was fabricated in a vertical contact–separation configuration (see Figure S10a). A PTFE sheet layer (2 × 2 cm², 1 mm) was employed as the negative triboelectric material and sandwiched between Cu electrodes, forming a Cu/PTFE/Cu stacked structure. The PTFE–Cu material pair was selected based on their large separation in the triboelectric series. PTFE is positioned near the negative end of the triboelectric series due to its strong electron‐withdrawing fluorinated surface groups and high electron affinity (≈2.2 eV [64]), whereas Cu resides toward the positive side with a work function of approximately 4.6–4.8 eV [65, 66]. This substantial difference in surface electronic properties promotes efficient charge transfer upon contact–separation [67, 68]. During mechanical actuation, periodic contact and separation between the PTFE and Cu layers induce interfacial charge exchange, leading to surface charge accumulation of opposite polarity on the two materials. The resulting potential difference drives transient electron flow through the external circuit, generating spike‐like voltage outputs characteristic of contact–separation‐mode TENGs. The vertical stacked configuration ensures stable contact area and reproducible electrical output under repeated mechanical stimulation [69], making it suitable for direct electrical coupling to the FeFET gate.

4.2. FeFET Fabrication

The FeFET was fabricated in an MFIS configuration (see Figure S10b). First, a Pt (150 nm)/Ti (10 nm) bilayer was deposited on a SiO₂/Si substrate to serve as the bottom gate electrode, where Ti improves adhesion, and Pt provides a chemically stable and conductive base for subsequent high‐temperature processing. A 300 nm‐thick PZT layer was then deposited via a sol–gel process and crystallized at 650°C to obtain a dense polycrystalline ferroelectric phase exhibiting stable polarization switching. The relatively thick PZT layer ensures robust remanent polarization and mitigates leakage during repetitive triboelectric stimulation. Next, a 10 nm HfO₂ interlayer was deposited by atomic layer deposition at 250°C, followed by post‐annealing at 500°C to improve film quality and reduce interfacial defect states. Subsequently, the MoS₂ film grown by chemical vapor deposition was transferred onto the MoS₂/PZT stack to form the semiconductor channel. Rapid thermal annealing at 300°C was performed to remove transfer residues and improve interfacial conformity. Then, the MoS₂ channel (L = 5–20 µm, W = 200 µm) was patterned via photolithography and reactive ion etching to define the active region. Source and drain electrodes (Au/Cr, 100/5 nm) were deposited by electron‐beam evaporation, where the Cr adhesion layer ensures stable contact formation, and Au provides low‐resistance electrical conduction. Finally, a 20 nm Al₂O₃ passivation layer was deposited by atomic layer deposition at 200°C to suppress ambient‐induced adsorption and enhance electrical stability during repeated mechanical stimulation and long‐term retention measurements.

4.3. Measurements and Characterization

The P–V hysteresis measurements of the HfO₂/PZT stack were performed using a ferroelectric tester (Precision RT66C, Radiant Technologies Inc.). PUND measurements were conducted to separate switching and non‐switching polarization components, using voltage pulses of ±0.5 to ±5 V with a pulse width of 1 ms and an interval time of 2 ms. The electrical characteristics of the FeFET devices were measured using a semiconductor parameter analyzer (Keysight B1500A/B1530A, Keysight Technologies Inc.) under ambient laboratory conditions. The I _D–V _G transfer characteristics were obtained by sweeping V _G while maintaining a V _D at 100 mV, unless otherwise specified. The I _D–V _D output characteristics were measured by sweeping V _D under fixed V _G conditions.

For synaptic characterization, electrical pulses with amplitudes from ±1.5 V to ±5 V and pulse widths from 10 µs to 1 ms were applied to the FeFET gate, depending on the experiment. The drain current response was continuously monitored for EPSC/IPSC, PPF/PPD, and LTP/LTD characteristics. Between measurement sequences, the FeFET conductance state was reset toward the initial baseline by applying a gate bias of V _RESET = −5 V. For triboelectric‐driven synapse analysis, the output of the TENG was directly used as the gate stimulus applied to the FeFET. The open‐circuit voltage generated by the TENG under periodic mechanical contact–separation was first characterized using a digital oscilloscope (Keysight DSOX1202G, Keysight Technologies Inc.) to confirm the spike‐like waveform and peak amplitude under controlled force and frequency conditions. The bipolar triboelectric output was then rectified using a diode rectifier circuit to produce unipolar voltage pulses compatible with gate‐driven FeFET operation. No additional amplification or waveform shaping circuitry was employed. The rectified TENG signal was directly connected to the gate terminal of the FeFET, while a constant external drain bias was supplied to monitor the channel current response. In this configuration, mechanical stimuli were translated into cumulative ferroelectric programming solely through triboelectrically generated gate modulation. All measurements were conducted at room temperature in ambient air unless otherwise specified.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: smll74524‐sup‐0001‐SuppMat.docx.

Data Availability Statement

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