Oxide‐Based Electrolyte‐Gated Transistors with Stable and Tunable Relaxation Responses for Deep Time‐Delayed Reservoir Computing

Time‐delayed reservoir computing with marked strengths of friendly hardware implementation and low training cost is regarded as a promising solution to realize time and energy‐efficient time series information processing and thus receives growing attention. However, achieving a sufficient number of reproducible reservoir states remains a significant challenge, which severely limits its computing performance. Here, an electric‐double‐layer‐coupled oxide‐based electrolyte‐gated transistor with a shared gate and varying channel lengths is developed to construct a deep time‐delayed reservoir computing system. A variety of short‐term synaptic responses related to inherent ion‐electron‐coupled dynamics at the electrolyte/channel interface are demonstrated, reflecting a flexibly regulable channel current. Different stable and tunable relaxation responses corresponding to varying channel lengths are obtained to enrich reservoir states combined with virtual nodes ways. The spoken‐digit classification and Hénon map prediction tasks are implemented with high accuracy (≈92.2%) and ultralow normalized root mean square error (≈0.013), respectively, validating the significant improvement of the computing performance by introducing additional relaxation responses. This work opens a promising pathway in exploiting oxide‐based electrolyte‐gated transistors for realizing temporal information processing hardware systems.


Introduction
The rapid advancements in artificial intelligence (AI) technology, coupled with the proliferation of intelligent edge devices, have DOI: 10.1002/aelm.202300652ushered in an information-driven "big data" era.This progress necessitates the creation of novel computing systems that exhibit lower power consumption and higher processing speeds, as traditional processors are incrementally hindered by the von Neumann bottleneck. [1,2][5] Reservoir computing (RC), as a neuromorphic computing paradigm, originates from recurrent neural networks and is adept at processing temporal information.Significantly, RC has an evident superiority in reducing the computational energy and time consumption owing to its simplified training process, where only the tuning of output weights towards target signals using a straightforward linear regression algorithm is required while the input and reservoir weights are fixed and need no training. [6]This superiority prompts RC to be widely applied to edge computing for Internet of things as a prominent hardware implementation approach.
Traditional RC systems typically require a large number of elements and then face significant volume and energy costs. [7,8]n recent decades, intensive attention has been paid to timedelayed RC, which consists of single nonlinear physical node and a delayed feedback loop containing several virtual nodes segmented through time multiplexing. [9]This contributes to a more friendly hardware implementation by reducing the number of physical devices and thus lowering computing costs.However, the time-delayed RC necessitates devices with rich dynamic characteristics to provide adequate computing resources.To date, the hardware implementation of time-delayed RC has primarily been based on various physical phenomena derived from a diverse range of materials and devices, e.g.nanowires, [10][11][12] conductive organics, [13][14][15][16] memristors, [6,17,18] and transistors, [19][20][21][22][23] which are equipped with nonlinearity and short-term memory characteristics.
To further improve computing performance, deep RC systems, which are composed of multiple layered standard RC systems based on specialized multiplexing, have been proposed to enhance high dimensionality from a device structure design perspective.This strategy introduces diverse outputs to serve as additional computing resources and helps to alleviate the limitation that small numbers of outputs can be collected from the limited output terminals of physical reservoirs.However, it is still a challenge to obtain a sufficient number of stable and tunable reservoir states in the time-delayed RC materials and devices due to the intrinsic stochasticity of physical mechanisms, which severely restricts the computing performance on hardware. [10,24][27][28][29] One notable strength of EGTs is the separation of "write" and "read" operation modes, with the former occurring at the gate and the latter at the source drain, providing tunable and stable channel current responses. [30]A deterministic switching mechanism of EGTs relies on the ensemble behavior of defects within the bulk switching layer to store and process analog resistance states without the probabilistic and stochastic filament formation, thereby offering a more stable current responses compared with filamentary devices. [31]hree main types of EGTs, utilizing H + , Li + , and O 2− ions as mobile charge carriers, have been reported. [30]Among them, highly reactive cationic species (H + and Li + ions) with a low migration energy are prone to intercalate into channel, causing nonvolatile memory.[34] In addition, O 2− ion-based EGTs using solid oxide electrolyte show good cycling stability and compatibility with complementary metal-oxide-semiconductor (CMOS) fabrication technology, [35] which is beneficial to the hardware implementation of RC systems.
In this work, we developed an EGT integrated by three oxidebased EGTs with a shared gate and varying channel lengths by using sub-stoichiometric tantalum oxide (TaO x ) as the electrolyte and amorphous indium-gallium-zinc-oxide (-IGZO) as the channel materials.TaO x has stable amorphous phase and adaptive lattice rearrangements of oxygen vacancies among metal oxides, which is beneficial to the device's stability. [36,37]The -IGZO film serves as an ideal channel layer in thin film transistors due to its high electron mobility. [38]The short-term plasticity and stable, tunable relaxation responses are explored in detail.The varied responses of EGTs with different channel lengths enable the construction of a deep time-delayed RC to enhance computing performance.The spoken-digit classification and Hénon map prediction tasks have been performed to verify its ability to process temporal information.

Results and Discussion
Figure 1a shows the optical images of EGT (left) and enlarged channels (right).It is obvious that as-prepared EGT possesses a common gate structure with varying channel lengths.The atomic force microscopy (AFM) images of the surfaces of -IGZO and TaO x films are presented in Figure 1b,c, respectively.The surface roughness of -IGZO and TaO x films measures 0.217 and 0.269 nm, respectively, indicating high surface quality.Figure 1d exhibits the cross-sectional transmission electron microscopy (TEM) image of a single EGT.A distinct contrast difference among the various layers is visible, reflecting diverse elemental compositions in different layers.The energy-dispersive X-ray spectroscopy (EDS) measurement is conducted to identify the composition, as depicted in Figure 1e.The elemental mapping of the cross-section indicates an Au/Ti/TaO x /-IGZO/SiO 2 hetero-structure from the top gate to the substrate.To confirm the crystallinity of as-fabricated -IGZO and TaO x films, electron beam diffraction analysis is performed, as illustrated in Figure S1 a,b (Supporting Information), respectively.The corresponding diffraction spectrum of both films exhibits multiple diffuse diffraction rings without distinct diffraction spots, which is a typical characteristic of amorphous structure.Besides, the Xray diffraction (XRD) spectra, devoid of characteristic peaks, further substantiate the amorphous nature of the -IGZO and TaO x films, as shown in Figure S1c,d (Supporting Information).
The schematic image of electrical measurement setup is shown in Figure S2 (Supporting Information), in which the gate voltage (V g ) and drain voltage (V d ) are directly applied to the gate and drain electrodes, respectively.The channel current (I d ) is monitored by applying a small DC read voltage (≈0.1 V) between the drain and source electrodes.Figure 2a shows the curves of the gate leakage current (I g ) as a function of V g at different sweep rates for the EGT with a channel length of 2 μm.The V g value is ≈3 orders of magnitude smaller than I d value, manifesting that the effect of I g on the performance is negligible.
In Figure 2b, typical I d -V d curves of an n-type field-effect transistor are depicted.I d first exhibits a linear increase at low V d , revealing good ohmic contact, and then gradually approaches a saturated value.Also, the I d value varies significantly under different V g levels, demonstrating strong gate control ability.Figure 2c displays the I d -V g curves at different V g sweep rates ranging from 0.67 to 3.33 V s −1 , illustrating an inherently nonlinear response of I d to V g , which allows the EGT-based reservoir to map inputs to high-dimensional states space for feature extraction.As V g sweeps from −5 to 5 V and then back to −5 V, the EGT switches between high resistance and low resistance states.Owing to relatively slow migration speed of oxygen ions and vacancies, a delay in response to the gate voltage sweep exists. [27,39]With increasing the gate voltage sweep rate, less oxygen ions, and vacancies are driven, resulting in a reduction of the on-state current.Moreover, a larger anticlockwise hysteresis is observed in the I d -V g curves with a lower V g sweep rate, attributed to the delay of O 2− ions transport in the TaO x layer in response to the V g sweep, and serves as the origin of the short-term memory. [27]he switching and hysteretic behavior observed in this study can be elucidated through the formation and disappearance process of ion-electron coupled EDL, [29,40] as shown in Figure 2d-f.At the initial high-resistance state (OFF state), the O 2− ions and oxygen vacancies are randomly distributed in the TaO x film.When applying a positive V g , an external electrical field (E ex ) is generated (step 1, Supporting Information), causing the oxygen vacancies to drift along the E ex and accumulate at the electrolyte/channel interface.Concurrently, the O 2− ions drift in the opposite direction and accumulate at the electrolyte/gate interface, resulting in the formation of an internal electrical field (E in ) built by the concentration gradient of O 2− ions in the electrolyte.And the magnitude of E in increases with V g .At the electrolyte/channel interface, the accumulated oxygen vacancies induce electrons in the channel, which drift from source to drain electrodes driven by the V d .When V g approaches its upper limit (5 V) which refers to the maximum value of applied V g during the test process (step 2, Supporting Information), E ex and E in reach equal maximum values, leading to the formation of a stable EDL at the electrolyte/channel interface.Consequently, the EGT switches to low-resistance state (ON state).After that, as V g reduces from its upper limit (step 3, Supporting Information), E in grows to be dominant, driving the diffusion of O 2− ions and oxygen vacancies back to their initial states while the EDL gradually dissipates.This spontaneous diffusion process is generally thought to be origin of the delay and short-term memory characteristics in the system.
To further substantiate the aforementioned ion-electron coupled EDL working mechanism, X-ray photoelectron spectroscopy (XPS) analysis of TaO x films is performed.The O 1s spectra of TaO x films deposited under different oxygen environments (1.2, 2.4, and 3.6 sccm) are shown in Figure 2g-i, respectively.Two distinct peaks are observed in the O 1s spectra, corresponding to O 2− ions (O I ) at approximately 529 eV and oxygen vacancies (O V ) at around 530.2 eV, respectively. [41,42]The area ratio of O V /O I , denoting the number of oxygen vacancies in the sample, decreases with the increase of oxygen content, reflecting the more complete oxidation of Ta.Given that the TaO x film utilized in this work is deposited at 2.4 sccm oxygen environment and possesses a higher oxygen vacancy concentration than that of the film deposited under a 3.6 sccm oxygen environment, the presence of oxygen vacancies in the former is confirmed, regardless of the latter's oxygen vacancy content.Because the XPS measurement is a semi-quantitative test method to weigh the oxygen vacancy concentration, the variation tendency of the O V /O I area ratio cannot represent the actual oxygen vacancy concentration of TaO x films.Figure 2g-i reflects that the variation tendency of O V /O I ratio is consistent with proposed oxygen vacancy-electron coupled EDL working mechanism.Hence, the proposed operation mechanism involving O 2− ions, oxygen vacancies, and electrons is deemed plausible.Besides, it is noted that the oxygen vacancy content of TaO x film should be controlled within an appropriate range to obtain the typical transfer curves and the required relaxation characteristics for RC implementation.Hence, an oxygen condition of 2.4 sccm is used to sputter the TaO x film.
The TaO x -based EGT, featuring reversible charging and discharging process, is closely analog to biological synapses in terms of structure and neurotransmitters delivery, as depicted in Figure 3a.The gate imitates the axon of presynaptic neurons, which generates pre-spikes.The mobile O 2− ions in the TaO x electrolyte serve as neurotransmitters within the synaptic cleft.The -IGZO channel including a drain simulates the dendrite of postsynaptic neurons.And the channel current can be viewed as the synaptic weight denoting the synaptic connection strength. [43,44]erein, such a TaO x -based EGT is capable of mimicking the synaptic short-term plasticity.Figure 3b shows a typical excitatory postsynaptic current (EPSC) generated by the TaO x -based EGT in response to a pre-spike (3 V, 30 s).It is noted that the EPSC intensity can be tuned by the amplitudes and durations of presynaptic pulses, as shown in Figure 3c,d, respectively.As pulse amplitude and duration increase, the EPSC progressively intensifies due to an accumulation of oxygen vacancies at the EDL interface driven by presynaptic pulses.Besides EPSC response, pairedpulse facilitation (PPF) is another important short-term plasticity, which describes the channel current change when two consecutive presynaptic pulses are applied (see Figure 3e).The EPSC induced by the second presynaptic pulse is larger than that of the first one on the condition of the pulse interval (Δt) is smaller than the EPSC decay time stimulated by the former pulse.The PPF index, defined as A 2 /A 1 , where A 1 and A 2 represent the first and second EPSC values, presents a gradual decrease as a function of the increasing Δt in the Figure 3f.This behavior can be well fitted by the following double exponential relaxation function: [45,46] PPF index = 1 + C 1 exp(−Δt∕ 1 ) + C 2 exp(−Δt∕ 2 ) ( 1 ) where C 1 and C 2 are the initial facilitation magnitudes,  1 and  2 are the characteristic relaxation time.The fitted  1 and  2 are 22 and 313 ms respectively, which are comparable to those of biological synapse. [47]These findings imply that the prepared EGT has a highly adaptable channel current, which is particularly valuable for information processing.The prepared oxygen ion-based EGT can generate marked excitatory postsynaptic current and decay current responses in a shorter scale, as shown in Figure S3a (Supporting Information).With further decreasing the pulse width, the decay current response is hard to measure (Figure S3b, Supporting Information), which might be caused by insufficient time for ion migration.
To explore the cycling stability of TaO x -based EGT, the consecutive current response to identical voltage pulse is measured over 200 cycles (see Figure 4a).The decay curves are normalized concerning the point where the voltage decreases to 0, and are displayed in Figure 4b.The corresponding normalized current at t = 5 s is plotted in Figure 4c.As observed, the curves show a highly small variation.To quantify this variation, the ratio ΔI d /I d, average is employed, where ΔI d and I d, average are the differential of the maximum and minimum and the average of the extracted normalized current, respectively.The calculated ratio is ≈4%, which is notably smaller than the reported value of other reservoir devices such as Cs 2 AgBiBr 6 -based optoelectronic device [22] and ferroelectric diodes, [48] indicating an exceptionally stable relaxation property of the as-fabricated EGT.The current responses of randomly selected EGTs with a channel length of 2 μm exhibit good uniformity under the same voltage pulse, and the calculated device-to-device variation is 0.081 (Figure S4, Supporting Information).This remarkably low variation is essential to ensure the separation property of reservoir states, and thus effective feature extraction.
In addition, this stable relaxation characteristic can be regulated by altering the pulse parameter.Figure 4d exhibits the spike number-dependent tunability of the relaxation process.The corresponding decay curves (Figure 4e) can be well fitted by Kohlrausch stretched exponential function as follows: [49] where  is the characteristic relaxation time of the decay process.Figure 4f illustrates that the relaxation time increases with the rise of spike number, which can be attributable to the fact that more O 2− ions driven by stronger inputs require a longer time to spontaneously diffuse to their initial state.Similar results with pulse amplitude, width, and frequency are also observed in Figure S5 (Supporting Information).Consequently, the TaO x -based EGT exhibits a tunable relaxation time, rendering it a potential candidate for temporal and sequential information processing.
Considering that information is typically input to the reservoir and processed in the form of voltage pulse trains, the current response of the device to different pulse trains is examined based on this stable and tunable relaxation property.Figure 4g displays the current response to distinct pulse trains, each consisting of 10 consecutive pulses with varying pulse amplitudes but fixed pulse width and pulse interval.A gradual accumulation of response currents under each pulse train is observed, and the response currents differ from one another with varying pulse amplitudes.Similar phenomena are also evident in the cases of different pulse widths and intervals (see Figure 4h,i) respectively).Such results reveal that the device could respond differently to different inputs, indicating that it possesses rich and distinguishable reservoir states, which are essential for implementing RC.
In addition to pulse parameters, channel current responses can also be regulated by modifying channel lengths.The channel current value can be calculated by the following equation: [50] Where W is the width of channel layer, L is the length of channel layer, μ is the mobility, C 1 is the capacitance of dielectric layer, V th is the threshold voltage.Figure 5a displays varied current responses of EGTs with different channel lengths to voltage pulse (3 V, 20 s).With the increase of the channel length, the EPSC gradually decreases, which is consistent with Equation (3).In order to explore the difference of relaxation mechanisms for the EGTs with different channel lengths, corresponding decay curves are normalized with respect to the point at which the voltage decreases to 0 (refer to the inset).These normalized decay curves are then fitted using Equation ( 2), and the obtained relaxation To perform RC, it is essential to establish a model that correlates the present current response (I(t)) with the current response at previous time step (I(t-1)) and the present input voltage (V g ).Here, the EPSC process is divided into a charging process during pulse width and a decay process after removing the pulse, resulting from the accumulation at the channel/electrolyte interface and the relaxation behavior of ions, respectively (see Note S1 Supporting Information).The charging and decay processes can be described by following equations respectively [29] : where k,  c , T,  d , , T D , and T I are coefficient, time constant of charging process, presynaptic pulse duration, time constant of decay process, and stretch index, simulated pulse width and interval, respectively.Besides, k values related to pulse amplitude can be fitted by a linear formula of a•V g -b. Figure 6a presents the current responses caused by different gate voltage pulses with V g ranging from 1 to 2.5 V in 0.5 V increments and a fixed pulse width of 15 s for the EGT with a 5 μm channel.With the increase of the gate voltage amplitude, the current value gradually increases because more oxygen vacancies are driven to the electrolyte/channel interface, inducing more extra electrons at the channel surface.Subsequently, the charging processes are fitted using Equation ( 4), and corresponding coefficient (k) values are fitted using a linear formula, as depicted in Figure 6b,c, respectively.Due to the limited oxygen vacancy concentration in the electrolyte layer and driving ability of gate voltage, the current value gradually saturates over time.Regarding the decay process under 2.5 V, it is approximately fitted using Equation (5) (see Figure 6d).Similar fitted results of EGTs with 2 and 10 μm channels are illustrated in Figures S7 and S8 (Supporting Information).The extracted parameter values are shown in Table S1 (Supporting Information) in supporting information.The general process of RC is described in Note S2 (Supporting Information).To assess the memory capacity (MC), a critical factor closely related to the computational performance of RC sys-tems, a short-term memory task is performed at first (see Note S3 Supporting Information).The input signal is initially subjected by a masking process to generate 10 virtual nodes for each reservoir.Subsequently, distinct reservoir states are obtained and utilized for learning in conjunction with the training data.Both correlation (Cor 2 ) and MC are appraised through linear regression training involving 2000-time steps, followed by testing with 500 time steps in this study.Figure 7a illustrates a decrease in Cor 2 as T delay increases, which is a typical feature of short-term memory.Furthermore, as depicted in Figure 7b, the calculated MC value escalates from 2.04 to 3.05 when the reservoir count rises from 1 to 3.This signifies that incorporating diverse channel lengths to generate varying relaxation responses contributes to not only enhanced dimensionality but also increased MC in RC systems, thereby assisting in the improvement of computational performance.
A spoken-digit classification task is undertaken to demonstrate the potential of the constructed deep time-delayed RC system for time-dependent information processing, (see Note S4 Supporting Information).The auditory waveform of spoken-digit signals, as displayed in Figure 7c, initially undergoes filtering and multiplication by a 64 × M (M = 64) mask matrix of 1 for time multiplexing within each interval , resulting in the conversion to input voltage pulse trains with a time step  which is equivalent to the pulse period ( = /M = 0.06 s)(see Figure S10, Supporting Information).These converted pulse trains possess a pulse width of 0.04 s and a pulse interval of 0.02 s, with the pulse amplitude linearly transformed to a range of 0 to 3 V.Afterwards, the preprocessed inputs are applied to the reservoir, and the varying current responses are recorded as reservoir states, as demonstrated in Figure 7d. Figure 7e presents the classification outcomes of different digits with an average accuracy of ≈92.2% in the as-prepared RC system integrated with three reservoirs.This is comparable to the accuracy achieved in other RC systems, which typically employ dozens of parallel reservoirs. [6,29,51]igure 7f displays the classification accuracy of the RC system with varying reservoir counts, highlighting an increase in accuracy from 89.8% to 92.2% as the reservoir count augments from 1 to 3. The accuracy of tasks is closely related to the decay characteristics of channel current.Different decay current responses can improve the MC and enrich reservoir states by increasing the number of reservoir node states.In this work, we develop an EGT with a shared gate and varying channel length to obtain different current responses.By adding different output current responses from different channel lengths to serve as additional computing resources, the accuracy is markedly improved.At present, the relationship between the accuracy and the physical properties of devices is not very clear and lacks theoretical basis, which needs further research.
The Hénon map prediction task, a classic discrete-time dynamic system exhibiting chaotic behavior and serving as a benchmark task for RC, has been implemented as well (see Note S5 Supporting Information).The input x(n) is transformed into pulse trains via a masking process employing a mask matrix with a mask length of 30, where the width and interval measure 0.04 s and 0.02 s respectively, and the pulse amplitude is linearly mapped to the voltage range of [0, 3 V].Then, the pulse trains are applied to the reservoirs to generate 30 reservoir states for each reservoir.The combined 90 reservoir states, provided by 3 reservoirs with distinct channel lengths, are utilized to predict the x(n+1) value.Figure 8a,b showcases the predicted outputs generated by the as-prepared RC system during the training and testing phases for the initial 200 time steps, respectively.It is clear that both predictions exhibit strong alignment with the ideal targets.The corresponding 2D plots of Hénon map during these two processes are presented in Figure 8c,d, demonstrating a successful reconstruction of the strange attractor.The computed normalized root mean square error (NRMSE) value during the testing process is ≈0.013, which is substantially lower than values reported in other RC systems, [6,48,52] as shown in Table S2 (Supporting Information).Figure 8e illustrates the NRMSE values of the RC system with varying reservoir counts, indicating a gradual decrease in NRMSE values as the number of integrated reservoirs increases.This observation aligns with the results observed in the spoken-digit classification task, further substantiating the potential for time series information processing.

Conclusion
In the study, we have developed a deep time-delayed RC system utilizing the EGT with a common gate and varying

Experimental Section
Device Fabrication: The devices with common gates and different channel lengths (2 μm, 5 μm, 10 μm) were fabricated.The device fabrication process started with the patterning of source and drain electrodes by lithography, followed by depositing Ti/Au (10 nm/35 nm) on pre-cleaned SiO 2 /p-Si (300 nm/500 μm) substrate using electron beam evaporation in which the Ti serving as an adhesion layer were used to form an ohmic contact with channel, and then the source and drain electrodes were formed through a lift-off process.After the patterning of channel layer, the deposition of -IGZO (7 nm) and TaO x (50 nm) were conducted in series using magnetron sputtering, acting as channel and electrolyte layers respectively.Besides, the lift-off operation was performed to remove the extra part.Finally, the gate electrodes were patterned by lithography process and then fabricated by depositing Ti/Au (10 nm/35 nm) followed by a lift-off process.
Measurement Methods: The surface morphology of films was analyzed by atomic force microscopy (AFM, Bruker MultiMode).The crystallinity of TaO x and -IGZO thin films was characterized by X-ray diffraction (XRD, PANalytical X'Pert Pro MRD).And the atomic defect was analyzed utilizing X-ray photoelectron spectroscopy (XPS, PHI X-tool).The planar structure of the devices was characterized by optical microscopy (OM, OLYM-PUS BX53M).Besides, the samples were preprocessed by focused ion beam (FIB, FEI Helios G5 HX dual beam).Then, the cross-sectional morphology of as-fabricated EGTs was characterized by transmission electron microscopy (TEM, FEI Titan Themis 200).And the elemental composition was measured by energy-dispersive X-ray spectroscopy (EDS, Bruker super-X).The electrical measurement of prepared devices was performed by an Agilent B1500 semiconductor parameter analyzer which was connected with a waveform generator/fast measurement unit (WGFMU) module.

Figure 1 .
Figure 1.a) Optical images of EGT (left) and enlarged channel (right).AFM images of the surface of b) -IGZO and c) TaO x films.d) Cross-sectional TEM image.e) EDS map of distribution of corresponding elements of the cross-section.

Figure 2 .
Figure 2. a) The gate leakage current curves (I g -V g ) at different V g sweep rates.b)The output curves (I d -V d ) at different V g varying from 0 V to 5 V with 1 V steps.c) The transfer curves (I d -V g ) at different V g sweep rates.d-f) Schematic images of ionic switching mechanism for as-fabricated TaO x -based EGT.The XPS O 1s spectrum of TaO x films deposited with different oxygen environment: g) 1.2 sccm, h) 2.4 sccm, and i) 3.6 sccm.

Figure 3 .
Figure 3. a) Schematic diagram of device structure and biological synapse.b) An EPSC triggered by a voltage pulse (3 V, 30 s). c) EPSC stimulated by a series of presynaptic pulses with a same duration time (1 s) and different amplitudes.d) EPSC stimulated by a series of presynaptic pulses with a same amplitude (3 V) and different pulse width.e) The EPSC is stimulated by a pair of gate voltage pulses (2 V, 1 s, with an interval of 2.2 s).f) PPF index plotted as a function of pulse interval between two successive presynaptic pulses.

Figure 4 .
Figure 4. a) Consecutive current response to the voltage pulse (2 V, 2 s), b) the spontaneous decay process of normalized current, and c) the cycling variation of as-fabricated EGT in 200 cycles.d) Spike number dependent tunability of the relaxation process triggered by different numbers of voltage pulse (2 V, 3 s, with an interval of 2 s).e) Decay curves and f) corresponding relaxation time in (d).g) The current response to different pulse trains consisted of 10 consecutive pulses with different pulse amplitude but fixed pulse width of 2 s and pulse interval of 1 s.h) The current response to different pulse trains consisted of 10 consecutive pulses with different pulse widths but fixed pulse amplitude of 2 V and pulse interval of 1 s.i) The current response to different pulse trains consisted of 10 consecutive pulses with different pulse intervals but fixed pulse amplitude of 2 V and pulse width of 2 s.

Figure 5 .
Figure 5. a) The current responses of EGTs with different channel lengths to the voltage pulse (3 V, 20 s).The inset shows the corresponding normalized decay current curves.b) The fitted relaxation time at different channel lengths.c) Various current responses of EGTs with different channel lengths to an identical pulse train.d) Schematic image of deep time-delayed RC system utilizing three different relaxation responses.

Figure 6 .
Figure 6.a) Current responses caused by different gate voltage pulses with V g from 1 to 2.5 V in 0.5 V steps and fixed pulse width of 15 s for the EGT with a 5 μm channel.b) The experimental data (symbols) and corresponding fitting curves (solid lines) in charging process.c) k values as a function of V g and corresponding fitting curve.d) The experimental data (symbols) and approximate fitting curves (solid lines) in decay process.The inset shows the enlarged decay process.

Figure 7 .
Figure 7. a) Fading memory curves of the reservoirs integrated with different reservoir numbers and b) corresponding memory capacity.c) Audio waveform of digit nine pronounced by a speaker.d) Output current responses of the reservoir to input voltage pulse trains.e) Classification results obtained from the as-constructed deep time-delayed RC system based on O 2− ions-based EGTs.f) The classification accuracy of the RC system integrated with different reservoir numbers.

Figure 8 .
Figure 8.The experimental output from the RC system versus ideal target in a) training process and b) testing process.The corresponding 2D plots are in c) training process and d) testing process.e) NRMSE of the RC system integrated with different reservoir numbers.