Device Structures
Two distinct types of HZO-based TFTs were fabricated to implement an ARC system: nonvolatile FeTFT for the synaptic device in the readout network and volatile MPBTFT for the physical reservoir and LIF neuron. While both TFTs utilize HZO thin films as ferroelectric materials, variations in the material composition of the HZO thin films impart the TFTs with either volatile or nonvolatile characteristics, diversifying their functionality within the system. Both TFTs incorporate an IGZO channel within a metal–ferroelectric–metal–insulator–semiconductor (MFMIS) structure, achieving enhanced electrical characteristics (Supplementary Note 3). In particular, an MPBTFT was designed as a DG device, enabling its operation to be modulated not only by the bottom-gate (BG) but also by the top-gate (TG). TG can be utilized to refine the reservoir states of MPBTFTs and execute inhibition operations in LIF neurons. The structural similarities between FeTFT and MPBTFT facilitate the integration of both types of TFTs on a single wafer using self-align etching process. The fabrication process for both TFTs is detailed in the Methods section and Supplementary Fig. 2.
Nonvolatile Synaptic FeTFT for Readout Network
Figure 2a shows a schematic illustration and cross-sectional view of a nonvolatile FeTFT with an MFMIS structure. The cross-sectional transmission electron microscopy (TEM) images of the fabricated FeTFTs are shown in Fig. 2b. The structure of the FeTFT with a Mo/HZO/TiN/ZrO2/IGZO stack was examined using energy-dispersive spectroscopy (EDS) analysis (Supplementary Fig. 3). The crystallinity of the HZO (Hf:Zr = 2:1) thin films were investigated using X-ray diffraction (XRD) analysis (Supplementary Fig. 4a). Figure 2c shows the top optical images of the FeTFT array for the readout network within the ARC system. Each FeTFT functions as a synaptic device within the readout network by leveraging the inherent nonvolatile memory effects of ferroelectric materials.
Capacitors with a metal–ferroelectric–metal (MFM) structure subjected to identical fabrication processes as the FeTFTs were utilized to verify the ferroelectric properties of the FeTFTs. The ferroelectric switching current is distinguished from non-ferroelectric switching currents, such as displacement and leakage currents, through positive-up-negative-down (PUND) measurements–wherein 100 kHz triangular pulses are applied. The switching current and polarization as a function of the applied voltage are depicted in Fig. 2d. The HZO film exhibits the magnitude of remanent polarization (2Pr) of approximately 37.7 µC/cm2 in the sweep range from − 3.8 V to 3.8 V. Figure 2e illustrates the hysteretic transfer characteristics of the FeTFT. The transfer characteristics demonstrate a counterclockwise hysteresis loop with a wide memory window (MW) of approximately 2.2 V at a constant drain current (ID) of 10 nA.
The nonvolatile FeTFTs can emulate the plasticity of biological synapses through the partial polarization switching of the HZO film [31, 32, 55, 56]. The long-term potentiation (LTP) and long-term depression (LTD) characteristics of the FeTFT were investigated by applying program (PGM) and erase (ERS) pulses (Fig. 2f). The pulse widths were consistently maintained at 10 µs, and PGM pulse amplitudes increased from 3.5 V to 4.1 V in steps of 0.04 V, while ERS pulse amplitudes decreased from − 4.35 V to − 5.1 V in steps of − 0.05 V. Multilevel synaptic weights were obtained through the FeTFT exhibiting a highly linear conductance response, characterized by coefficients βp = 0.04 and βd = 1.95 [6]. Figure 2g illustrates the LTP and LTD characteristics of the FeTFT over 20 cycles. Low cycle-to-cycle variation is observed (σ/µ < 0.025), indicating consistency and reliability in the operation of the FeTFT over multiple operational cycles. The retention characteristics of the FeTFT with various memory states are depicted in Fig. 2h, showing its nonvolatile memory characteristics. The FeTFT preserves the stored information for 103 s with a slight degradation, verifying the robustness and reliability of the device. The device-to-device variation and selective PGM/ERS operations within the FeTFT array were investigated (Supplementary Figs. 5 and 6). The results confirm the suitability of the nonvolatile FeTFTs for synaptic devices.
Volatile Double-Gate MPBTFT for Physical Reservoir
The distinguishing feature of the ARC system is the reservoir, which is a network of nonlinear dynamic nodes that transform input signals into a high-dimensional space, enabling the system to process temporal information. In an MPBTFT-based ARC system, these nodes are represented by volatile DG MPBTFTs, which are physical reservoirs characterized by nonlinear and dynamic responses to input stimuli. Figure 3a shows a schematic illustration and cross-sectional view of a DG MPBTFT featuring two different gates: BG and TG. The BG side operates as an MPBTFT with an MFMIS structure, while the TG side functions as a TFT with a metal–insulator–semiconductor (MIS) structure. This DG configuration affords versatile control and modulation of the electrical characteristics of the MPBTFT. Figure 3b shows the cross-sectional TEM images of the fabricated DG MPBTFTs. The structure of the DG MPBTFT with a Mo/HZO/Mo/ZrO2/IGZO/ZrO2/Mo stack was examined by EDS analysis (Supplementary Fig. 7). An XRD analysis was conducted for the HZO (Hf:Zr = 3:5) thin films (Supplementary Fig. 4b). Figure 3c shows the top optical images of the DG MPBTFT array for the physical reservoirs within the ARC system. The XRD analyses of both volatile and nonvolatile HZO thin films are detailed in Supplementary Note 2b.
The capacitors featuring an MFM structure, which underwent identical fabrication processes to the DG MPBTFTs, were used to confirm the polarization properties of the DG MPBTFTs. Through PUND measurements using 100 kHz triangular pulses, the switching current and polarization as a function of the applied voltage are depicted in Fig. 3d. The HZO film exhibits the 2Pr of approximately 66.6 µC/cm2 in the sweep range from − 2.0 V to 2.0 V. A high polarization magnitude at a relatively low voltage is obtained through the HZO film near the MPB. The switching current demonstrates a partial hysteresis of the HZO film within the DG MPBTFT. The polarization of the HZO film reverts to a non-polarized state at 0 V, indicating the capability of the HZO film to transition between the polarized and non-polarized states. The symmetric switching current peaks observed near 0 V highlight that the coercive voltage (Vc) and the energy barriers between the phases are sufficiently small to rapidly reset the polarization state. The HZO thin film near the MPB exhibits a high dielectric constant (Supplementary Fig. 8). The polarization within the HZO film is switched when an electric field is applied, whereas it returns to the initial non-polarized state in the absence of an electric field. Figure 3e presents the hysteretic transfer characteristics of the DG MPBTFT under various TG voltage (VTG) conditions. The transfer characteristics exhibit a counterclockwise hysteresis loop with a MW of approximately 0.5 V at a constant ID of 10 nA when VTG = − 3.0 V.
To verify the volatile memory characteristics of the DG MPBTFT, a single PGM pulse was applied to the device, followed by a read operation at a constant voltage for a specific duration. Figure 3f illustrates the dynamic response of the ID in the DG MPBTFT over time, subsequent to the application of a PGM pulse with a width of 100 µs. PGM pulses with different amplitudes were applied at the same initial state. As the amplitude of the PGM pulse (VPGM) increased, the corresponding ID increased. Following the removal of the PGM pulse, IDs decayed toward the initial state in all cases. This relaxation process fitted well with the double exponential function (red lines in Fig. 3f). The dynamic response of IDs in the DG MPBTFT to successive PGM pulses with various VPGMs was investigated (Supplementary Fig. 9). ID increases significantly as VPGM increases, while ID decays back to its initial state in the absence of a PGM pulse. The fitting parameters for these volatile memory characteristics are shown in Supplementary Fig. 10. Note that the volatile DG MPBTFTs exhibit a rapid current relaxation process, ensuring fast data processing capabilities of the ARC system.
The changes in ID when a PGM pulse is applied to the BG varies depending on the TG bias (Supplementary Fig. 11). The formation of an accumulation layer in the IGZO channel, crucial for efficiently coupling the voltage applied to the BG to the ferroelectric layer during PGM operation, is affected by the TG bias [57, 58]. Specifically, applying a positive voltage to the TG can inhibit the formation of an accumulation layer in the IGZO channel, diminishing the PGM efficiency and the amount of change in ID. Figure 3g shows the ID response according to various TG pulse amplitudes when a constant pulse train of 4.0 V amplitude is applied to the BG. As the TG pulse amplitude increases, the amount of change in ID decreases. The dynamic interaction between the BG and TG enables modulation of the sensitivity of the physical reservoir to the input signals. This inherent flexibility in the DG configuration improves physical reservoir characteristics.
For a physical reservoir within an ARC system to function effectively, the reservoir state, manifested as the ID of the volatile DG MPBTFT, must evolve dynamically over time in response to an input signal. Distinct reservoir states for different temporal inputs are essential for transforming and expanding input signals across the high-dimensional state space of the reservoir [5, 18–40]. Figure 3h demonstrates the ID evolutions of the DG MPBTFTs in response to 16 different input pulse trains, each comprising 4 timeframes. For each timeframe, a pulse with an amplitude of 4.0 V and a width of 100 µs was applied to the BG, while a fixed voltage of 0 V was applied to the TG. The time interval between successive pulses is 3 ms (see Supplementary Fig. 12 for ID evolutions over various time intervals). The device-to-device variations of the DG MPBTFTs were further investigated (Supplementary Fig. 13). The results verify the capability of the DG MPBTFT to effectively distinguish input pulse trains. Nevertheless, some input pulse train configurations exhibit similar final conductance states. To address this states overlap issue, the TG of DG MPBTFT is utilized in this study.
Figure 3i illustrates the ID evolutions of the DG MPBTFTs in response to various input pulse trains with TG utilization. The proposed pulse scheme for TG is depicted in the upper panel of Fig. 3i. The utilization of TG enhances the distinctness between reservoir states compared to the case in which TG is not employed and clearly distinguishes each state. The enhanced distinction between reservoir states is attributed to PGM efficiency affected by VTG. Gradually decreasing VTG across 4 timeframes augments the influence of the input data in the latter timeframe (see Supplementary Note 4 for details). Supplementary Fig. 14 contrasts the normalized difference between the reservoir states of the DG MPBTFT with and without TG utilization. Note that the final reservoir states can be further optimized by adjusting the pulse scheme for TG (see Supplementary Fig. 15 for ID evolutions with different TG pulse scheme). The utilization of TG in DG MPBTFTs not only clearly distinguishes reservoir states but also enables the expansion of reservoir states. A demonstration of 32 distinct reservoir states (5 bits) in response to input pulse trains with 5 timeframes is shown in Fig. 3j. The proposed strategy effectively leverages the TG to modulate the electrical characteristics of DG MPBTFTs, thereby enhancing the performance and area/energy efficiency of the ARC system by ensuring sufficiently distinct reservoir states or expanding the states.
Volatile Double-Gate MPBTFT for Leaky Integrate-and-Fire Neuron
Within the ARC system, the versatility of volatile DG MPBTFTs extends to their application as LIF neurons, facilitating the seamless integration of physical reservoirs, readout networks, and LIF neurons. DG MPBTFTs utilize the inherent polarization volatility and partial polarization switching capabilities of the HZO film to implement the leaky effect and integrate-and-fire function of LIF neurons. Figure 4a shows a schematic of the DG MPBTFT-based LIF neuron, which eliminates the need for capacitors and complex circuits that are typically associated with larger footprints. The DG MPBTFT-based LIF neuron exhibits high area efficiency and functional versatility. Moreover, the DG MPBTFT can process both excitatory and inhibitory pulses through BG and TG, respectively, within a single device. This characteristic negates the necessity for additional devices or connections, streamlining the composition of the neuron [53–55]. The pulse schemes for the BG and TG are shown at the bottom of Fig. 4a.
Figure 4b illustrates the ID response of the DG MPBTFT when excitatory pulses are applied to the BG. During this time, a fixed voltage of − 3.0 V is applied to the TG. The DG MPBTFT-based LIF neuron demonstrates an integrate-and-fire function. The application of excitatory pulses to the BG induces gradual polarization switching of the HZO film, increasing ID. When ID exceeds a specific threshold current (Ith) required for the neuron to fire, the neuron self-resets to its initial state, which is attributed to the inherent leaky effect. This characteristic eliminates the necessity for an external reset circuit. Neuronal behaviors with various reset times are shown in Supplementary Fig. 16. The roles of base voltage (Vbase) and high voltage (Vhigh) in modulating neuronal behavior are further investigated. As Vbase increases, the polarization of the HZO film is preserved, thereby decelerating the leaky effect (Fig. 4c). A high Vbase maintains the polarization of the HZO film during the delay between pulses, thus strengthening the integration function of the neuron. In addition, even if the same reset time is given, the neuronal behavior varies distinctly with Vbase. The neuron is completely reset to its initial state under low Vbase conditions. By contrast, under higher Vbase conditions, the neuron cannot be reset to its initial state within the same reset time, thus requiring an extended reset time for complete initialization. Figure 4d shows the neuronal behavior as a function of Vhigh. An increase in Vhigh results in more significant polarization switching in the HZO film and increases the ID change. However, an insufficient Vhigh (e.g., Vhigh = 1.0 V) barely switches the polarization of the HZO film, thus impairing the integration function of the neuron. These results highlight the critical roles of Vbase and Vhigh in modulating neuronal behavior. The impact of the delay between pulses on neuronal behavior is depicted in Supplementary Fig. 17.
Figures 4e–g show the neuronal behavior when inhibitory pulses are applied after a sequence of excitatory pulse trains. Notably, even with a shorter reset time (1.0 ms) compared to Figs. 4b–d, the application of sufficient inhibitory pulses facilitates the complete reset of the neuron to its initial state. This is particularly evident as the inhibitory voltage (Vinh) intensifies, enhancing the inhibitory efficiency of the neuron and requiring a greater number of excitatory pulses for the neuron to fire (Fig. 4e). These results highlight the potent inhibitory efficiency of higher Vinh levels. Furthermore, modulation of the width or quantity of inhibitory pulses significantly influences the inhibitory efficiency of neurons. Inhibitory pulses with extended pulse widths inhibit the neuron (Fig. 4f). The application of numerous inhibitory pulses also inhibits the neuron, leading to a complete reset to its initial state (Fig. 4g). These results verify the suitability of volatile DG MPBTFTs as LIF neurons, emphasizing their ability to finely modulate neuronal behavior in response to varying excitatory or inhibitory pulses (see Supplementary Note 5 for details).
MPBTFT-based Analog Reservoir Computing System for Digit Recognition
An ARC system utilizing volatile DG MPBTFTs and nonvolatile FeTFTs is shown in Fig. 5a. Volatile DG MPBTFTs are employed as physical reservoirs and LIF neurons, while nonvolatile FeTFTs constitute the readout network. The MPBTFT-based ARC system operates through unique interactions between the physical reservoir, readout network, and LIF neurons, as detailed in Supplementary Note 6.
A handwritten digit recognition task was performed using the MNIST dataset to evaluate the performance of the MPBTFT-based ARC system. The images are preprocessed prior to their introduction into the reservoir (see Fig. 5b and Methods section). The encoded input pulse trains are applied to the volatile DG MPBTFT-based physical reservoir. The state of each physical reservoir evolves over time, and the state in the last timeframe is transmitted to the readout network. The reservoir states in the last timeframe are shown for the physical reservoirs based on DG MPBTFTs and linear resistors (right side of Fig. 5b). The response of the linear resistor is solely contingent on the magnitude of the input pulse applied at any given instant, thereby lacking the capacity to retain information regarding previous states. The absence of memory effects in linear resistors leads to an inability to distinguish between similar images. By contrast, the DG MPBTFTs exhibit well-distinguishable reservoir states. This attribute significantly enhances the ability of the system to discern subtle differences in inputs, thereby facilitating accurate pattern recognition. Note that the configuration of the size and number of sections in the input encoding process can be adjusted. A comprehensive explanation of the input encoding process is detailed in the Methods section.
Figure 5c shows the results of performing a digit recognition task utilizing several different device-based physical reservoirs: volatile DG MPBTFTs, nonvolatile FeTFTs, and linear resistors. Linear resistor-based physical reservoirs, limited by their capacity to retain information solely from the last timeframe of the input signal, demonstrate the lowest level of accuracy (~ 83.63%). When FeTFTs are utilized as physical reservoirs, the system lacks the capacity to process temporal information. However, these FeTFT-based physical reservoirs exhibit different responses depending on the number of input pulses applied. This characteristic results in a slight enhancement in accuracy (~ 85.45%) over linear resistors, which is attributable to the increased number of reservoir states (five states per reservoir). In contrast, DG MPBTFT-based physical reservoirs facilitate temporal information processing. This allows 16 reservoir states (4 bits) for each physical reservoir, significantly increasing the accuracy (~ 89.42%). In particular, the highest accuracy (~ 90.23%) is achieved when utilizing TG, which effectively differentiates between the reservoir states without any overlap. This distinct advantage demonstrates the suitability of the DG MPBTFT as a physical reservoir, highlighting its potential for accurately handling complex computing tasks through enhanced reservoir state discrimination. The system utilizing 32 reservoir states (5 bits) significantly enhances the area efficiency (~ 20% enhancement) with a slight degradation in accuracy (~ 88.25%). This enhancement is attributed to the reduction in the number of physical reservoirs required for input encoding and the subsequent number of synaptic devices that constitute the readout network. The flexibility of the MPBTFT-based ARC system enables for a significant reduction and optimization of hardware resources with slight accuracy adjustments. The average accuracies achieved for the various physical reservoirs are shown in Fig. 5d. Figure 5e shows the confusion matrix obtained using the MPBTFT-based ARC system. The system accurately distinguishes ten types of handwritten digits. These results substantiate the proficiency of DG MPBTFTs utilizing TG to distinguish each reservoir state with remarkable distinction and even expand reservoir states. The MPBTFT-based physical reservoir consumes ~ 22.5 pJ per input, while the FeTFT-based synaptic device consumes ~ 0.2 pJ per input (Supplementary Note 7). Benchmarking analysis against previous studies is presented in Supplementary Table 1.
MPBTFT-based Analog Reservoir Computing System for Waveform Classification and Time-Series Prediction
In addition to the handwritten digit recognition task, where the original static image data are converted to temporal data, waveform classification and time-series prediction tasks were performed to further evaluate the capabilities of the MPBTFT-based ARC system for processing temporal signals [59]. The quintessential applications of ARC systems lie in their proficiency in processing temporal data [59, 60]. In an MPBTFT-based ARC system, the implementation of the waveform classification and time-series prediction tasks relies heavily on masking techniques. This technique augments input information and increases the number of virtual nodes within the temporal domain, enhancing the computational capabilities [18, 19, 21, 28–30]. A comprehensive explanation of the masking process is detailed in the Methods section.
Through the masking process, each DG MPBTFT creates M virtual nodes in response to the application of an M-timeframe input pulse train. The state of each virtual node is characterized by the ID response of the DG MPBTFT within each corresponding timeframe. These virtual nodes of the DG MPBTFTs effectively expand the reservoir size from N to N×M, significantly enhancing the computational density without requiring additional hardware. The expanded reservoir states are subsequently fed into an (N×M) × 1 fully connected single-layer readout network. The readout network is trained using a linear regression method and outputs the classification or prediction results through a linear combination of all the reservoir states. A detailed description of the linear regression method can be found in the Methods section. An MPBTFT-based ARC system that utilizes the masking process is illustrated in Fig. 6a.
In the waveform classification task, the input data contain randomly arranged sine and square waveforms (Fig. 6b). The target output is a binary sequence of 0 and 1, representing sine and square waveforms, respectively. The MPBTFT-based ARC system classified the waveforms with a normalized root mean squared error (NRMSE) of 0.0044 (Fig. 6c). This indicates a sufficiently low classification error, demonstrating the proficiency of the system in handling classification tasks.
In addition to the waveform classification task, a time-series prediction task was performed to comprehensively assess the performance of the MPBTFT-based ARC system in handling temporal signals. A benchmark task is focused on Hénon map prediction, which is a typical discrete-time dynamic chaotic system prediction [61]. A detailed description of the Hénon map prediction task can be found in the Methods section. Figure 6d shows the target values (black line) and outputs of the MPBTFT-based ARC system after training (red line). The ARC system predicts a time-series dataset and achieves a low NRMSE of 0.035. Figure 6e shows a two-dimensional (2D) representation of the Hénon map, which exhibits excellent consistency between the target and predicted values. This result demonstrates that the ARC system can effectively reconstruct strange attractors of the Hénon map.
In this study, complex real-world data that extend beyond theoretical models are addressed to further evaluate the applicability and effectiveness of the MPBTFT-based ARC system in practical scenarios. The number of confirmed coronavirus disease 2019 (COVID-19) cases from January 2020 to March 2023, provided by Johns Hopkins University, was used for the evaluation [62]. The MPBTFT-based ARC system forecasts the future number of confirmed COVID-19 cases based on the historical data of previously confirmed cases. Figure 6f shows the actual number of confirmed cases (black line) and the predicted number of confirmed cases (red line). The remarkable concordance between the predicted and actual numbers of confirmed cases verifies the exceptional precision of the system in forecasting outbreak progression. Despite the complexity of real-world predictive tasks, a low NRMSE of 0.28 is achieved. The impressive system performance is primarily attributed to the polarization dynamics inherent in DG MPBTFTs, which result in a robust ability to process temporal features. This result demonstrates the potential of the MPBTFT-based ARC system in public health analytics and response strategies, providing a valuable tool for predictive modeling in epidemiology by leveraging its sophisticated temporal data processing capabilities.