Flexible Sensory Platform Based on Oxide-based Neuromorphic Transistors

Inspired by the dendritic integration and spiking operation of a biological neuron, flexible oxide-based neuromorphic transistors with multiple input gates are fabricated on flexible plastic substrates for pH sensor applications. When such device is operated in a quasi-static dual-gate synergic sensing mode, it shows a high pH sensitivity of ~105 mV/pH. Our results also demonstrate that single-spike dynamic mode can remarkably improve pH sensitivity and reduce response/recover time and power consumption. Moreover, we find that an appropriate negative bias applied on the sensing gate electrode can further enhance the pH sensitivity and reduce the power consumption. Our flexible neuromorphic transistors provide a new-concept sensory platform for biochemical detection with high sensitivity, rapid response and ultralow power consumption.

Inspired by the dendritic integration and spiking operation of a biological neuron, flexible oxide-based neuromorphic transistors with multiple input gates are fabricated on flexible plastic substrates for pH sensor applications. When such device is operated in a quasi-static dual-gate synergic sensing mode, it shows a high pH sensitivity of ~105 mV/pH. Our results also demonstrate that single-spike dynamic mode can remarkably improve pH sensitivity and reduce response/recover time and power consumption. Moreover, we find that an appropriate negative bias applied on the sensing gate electrode can further enhance the pH sensitivity and reduce the power consumption. Our flexible neuromorphic transistors provide a new-concept sensory platform for biochemical detection with high sensitivity, rapid response and ultralow power consumption.
With the recent interest in brain/computer interfaces 1 , soft robotics 2 , wearable electronics 3 and skin-like sensory systems 4 , flexible devices have attracted growing attention. These emerging devices require new fabrication schemes that enable integration with soft, curvilinear and time-dynamic human tissues. Among these devices, flexible sensors are becoming increasingly significant in a wide-variety of novel applications such as in vivo monitoring 5 , delivery of advanced therapies 6 , artificial sense organs 7 , etc. As a fundamental component for sensor application, field-effect transistors (FETs) based sensors have been intensively investigated due to their inherent advantages of miniaturization, facilitated integration, direct transduction and label-free detection [8][9][10][11] . The classical sensing mechanism of the FET-based sensor is attributed to a charge-dependent interfacial potential due to the adsorption of potential-determining species at sensing membrane/electrolyte interface 12 . The sensitivity is limited to ~59.2 mV/decade (Nernst limit) at room temperature when the threshold voltage (V th ) is recorded as the output signal. It should be noted here that above mentioned measurements are based on the quasi-static electrostatic coupling mode, which potentially increases the time consumption and energy dissipation. But in smart sensory platforms, such as implantable devices and wearable sensory systems, low power consumption is one of the most important pre-requisites.
Synergic integration of presynaptic inputs from the dendrites plays an important role for sensory information process and cognitive computation, and the idea of building bio-inspired solid-state devices has been around for decades 13,14 . In 1992, Shibata et al. proposed Si-based neuron transistors with multiple input gates that are capacitively coupled to a floating gate 15 . The "on" or "off " state of the neuron transistors depends on the integrated effect of the multiple input gates. One of the unique features of the neuron transistors is the ultralow power dissipation during calculation due to the gate-level sum operation in a voltage mode. From then on, Si-based neuron transistors have attracted much attention for chemical and biological detection due to the easy adjustment of threshold voltage [16][17][18][19][20] . When an asymmetric gate capacitor structure is adopted, magnification of V th shift can be observed in the neuron transistor when the sensing gate experiences a load from electrolyte. This device concept scales up the surface potential shift by the capacitance ratio between the sensing gate and the control gate [21][22][23] . But, up to now, flexible electrolyte-gated neuron transistors with amorphous oxide channel layers for biochemical sensing applications have not been reported.
Amorphous oxide-based transistors were proposed as promising fundamental unit in sensory platform due to their low process temperature, superior electrical properties, high reliability and easy reproducibility [24][25][26] . To date, remarkable sensing performances have been demonstrated in these oxide-based transistors [27][28][29]  as potential candidates with a low operation voltage due to the strong EDL modulation at the electrolyte/channel interface 30,31 . Recently, oxide-based EDL transistors gated by solid-state inorganic electrolytes were proposed by our group 32,33 . At the same time, artificial synapses and neuromorphic transistors with low power consumption and fundamental biological functions were mimicked in these devices [34][35][36] . In the present work, flexible sensory platform based on individual protonic/electronic coupled indium-zinc-oxide (IZO) neuromorphic transistor was fabricated on plastic substrates. Such neuromorphic transistor exhibited a high sensitivity when a quasi-static dual-gate synergic modulation mode was adopted. Most importantly, single-spike dynamic sensing of such flexible neuromorphic transistor was also investigated, and pH sensing with ultra-high sensitivity, very quick response/ recover time, and extremely low power consumption were realized. Figure 1a shows the schematic diagram of a flexible IZO-based neuromorphic transistor with multiple in-plane gate electrodes for pH sensing application. A miniature Ag/AgCl reference electrode immersed into a 5.0 μ L pH buffer solution droplet on the nanogranular SiO 2 (n-SiO 2 ) electrolyte film acts as a sensing gate (G 1 ). In-plane aluminum (Al) electrodes (G 2 , G 3 …G n ) are used as control gates. The distinctive feature of our device is that sensing gate and all control gates are located at the same plane. The capacitive network of the neuromorphic transistor is plotted in Fig. 1b. The carrier density of the IZO channel layer can be electrostatic modulated by the weighted sum of all inputs from the sensing and control gates. The weight for each gate is directly proportional to the capacitive factor normalized by the total capacitance of the floating gate 5 . Figure 1c displays the top-view optical image of the IZO-based neuromorphic transistor sensor. The channel width (W) and length (L) is 1000 and 80 μ m, respectively. As a proof of concept, only one sensing gate and one control gate electrode is used in the present work. The distance between the control gate electrode and the drain electrode is 300 μ m. Figure 1d shows a picture of the IZO-based neuromorphic transistor array on PET plastic substrate, exhibiting its flexible nature under external force. Figure 2a shows the transfer characteristics of the IZO-based EDL transistor at a constant V DS of 1.5 V. Gate voltage applied on lateral Al gate electrode is swept from − 1.5 V to 1.5 V and then back. A clear anticlockwise hysteresis window of ~0.4 V is observed, which is likely due to the mobile protons within the nanogranular SiO 2 electrolyte 37 . Subthreshold swings (SS), current on/off ratio (I on /I off ) and V th are estimated to be ~175 mV/decade, ~6.4 × 10 5 , and − 0.3 V, respectively. In addition, field-effect electron mobility (μ FE ) at the saturation region is estimated to be ~12 cm 2 /V.s by the following equation:

Results
where C i (~2.7 μ F/cm 2 ) is the specific capacitance of the SiO 2 electrolyte measured from two in-plane Al gate electrodes at 1.0 Hz (Supporting Figure S1). For practical flexible electronics application, flexible devices should be bendable without sacrificing their electrical properties. The influence of mechanical bending on the electrical characteristics of our devices was investigated. Figure 2b shows the transfer curves recorded before, during and after bending by a cylinder with a radius of 1.0 cm. The images of the measurement process are shown in the insets of Fig. 2b. Good reproducibility is obtained on different test conditions. Moreover, mechanical stress tests have also been performed by bending the sample repeatedly. Figure 2c shows the transfer curves recorded at repetitive bending cycles. The flexible neuromorphic transistors survive after more than 1000 flex/flat cycles with negligible change in the transfer characteristics. The variations in V th and μ FE with the repetitive bending cycles are extracted, as shown in Fig. 2d. After 1000 cycles of bending and recovery, a small positive shift of ~0.1 V in V th and only ~10% reduction in μ FE are measured. The results indicate that the flexible neuromorphic transistors have good mechanical reproducibility and durability. We will next study the pH sensing performance of the devices operated in the quasi-static mode. Figure 3a shows the transfer curves of the neuromorphic transistor based sensor operated at the linear region (V DS = 0.1 V) with the sensing gate immersed into solution droplets with different pH values. The inset in Fig. 3a shows the layout of this normal pH sensing measurement. Clear negative shift of the transfer curve is observed when pH value decreases from 10 to 4. It has been reported that acidic solution can give rise to a more positive surface potential due to the ionic interaction at the solution/SiO 2 interface 38,39 . In our case, positive surface potential will make protons within SiO 2 electrolyte migrate to the electrolyte/IZO channel interface, which will induce excess electrons in the IZO channel and a negative shift of transfer curve. When the gate voltage at a drain current of 10 nA is defined as the responsive voltages (V R ), a sensitivity of ~37.4 mV/pH is realized, as shown in Fig. 3b. This value is comparable to the reported FET sensors using SiO 2 as a sensing material 40 . In order to improve the sensing performance of the IZO-based neuromorphic transistor, dual-gate synergic modulation mode is investigated. During the measurements, G 1 is biased at different fixed voltages and G 2 is swept from -2.0 to 1.0 V. The measuring schematic is shown in Fig. 4a. Figure 4b shows the transfer curves (I DS -V G2 ) curves measured at V DS = 0.1 V with pH value changed from 10 to 4 and V G1 fixed at 0.3 V and − 0.6 V, respectively. Similarly, the transfer curves shift to the negative direction when the pH value decreases at a fixed V G1 . Here, we should point out that more obvious shifts in the transfer curve are induced by pH variation when V G1 = − 0.6 V. The sensitivity in terms of V R shift is plotted as a function of V G1 (Fig. 4d). The pH sensitivity increases when V G1 shifts from a positive value to a negative value. A maximal pH sensitivity of ~105 mV/pH is obtained when V G1 = − 0.6 V. The improved sensitivity obtained at a negative V G1 is attributed to amplified capacitive coupling factor between these two gates (G 1 and G 2 ). Asymmetric dual-gate capacitive coupling can result in intrinsic amplification of the measured surface potential shifts. Theoretical analysis of the quasi-static pH sensing process can be found in Supporting Note 1. Jayant et al. also reported that this technique merely scaled the surface potential shift, but did not signify any change in the intrinsic properties of the electrolyte interface 21 . Figure 4c shows real-time responses of I DS of the IZO-based neuromorphic transistor sensor in different pH solutions for 180 s at fixed V DS = 0.1V, V G1 = 0 V and V G2 = 0.2V. It is observed that I DS increases gradually to a stable value. The steady I DS increases stepwise with discrete changes in pH value from 10 to 4. The sensitivity S of a sensor can also be defined as the relative change in channel conductance, S = (|G− G 0 |)/(G 0 ) = Δ G/G 0 41 . In our case, the response conductance to pH = 10 is defined as G 0 . Therefore, the sensitivity (∆G/G 0 ) is estimated to be ~ 2.2 for pH = 4 at equilibrium state. We also find that the sensitivity (∆G/G 0 ) can be improved by a negative bias applied on sensing gate (G 1 ), as shown in the right axis of Fig. 4d. A highest sensitivity of ~38.3 is obtained at a fixed V G1 of − 0.6 V. This value is much higher than those reported in nanoscale transistor sensors 42,43 . This is because that an appropriate negative voltage applied on the sensing gate (G1) can make the neuromorphic transistor operated in the subthreshold regime, in which the sensitivity in terms of current variation can be exponentially enhanced due to the most effective gating effect of charges bound on a surface 44 .
Next, inspired by the spiking operation mode of a biological neuron, we have investigated the single-spike pH sensing performance of our neuromorphic transistor sensors. Due to the distinctive dynamic characteristics of the proton migration, our device presents a unique time dependent transient property. During the measurement, equilibrium is disturbed by a small voltage pulse applied on the control gate (G 2 ). The dynamic spike current response to such a disturbation contains the pH sensing information. After the detection, the device will quickly recover to the original equilibrium state. Moreover, during the single-spike sensing process, the energy consumption is extremely low, which is preferred for portable and wearable sensory applications. The single-spike pH sensing measurement of the detection is schematically illustrated in Fig. 5a. At first, a disturbing spike V G2 (0.2 V, 10 ms) was applied on control gate (G 2 ), and a synchronous reading spike V D (0.02 V, 10 ms) was applied on drain electrode to measure the output current. As shown in Fig. 5b, when the pH value is changed from 4 to 10, the response current (I DS ) decreases from 512 to 80 nA. We also find that the logarithm of I DS decreases linearly with increasing pH value, and a high sensitivity (∆G/G 0 ) of ~5.6 is estimated, as shown in Fig. 5c. The characteristic time of the dynamic process of proton migration in the nanogranular SiO 2 electrolyte is in the order of few milliseconds. The response/ recover time is estimated to be ~5.0 milliseconds, which is much shorter than that operated in quasi-static mode. The reproducibility of the single-spike pH sensing measurement is also investigated. Figure 5d shows the response currents stimulated by repeated voltage pulse spikes with pH = 6. The results indicate a good reproducibility of single-spike detection of pH values. If we define the value of σ (I p )/Ave(I p ) as the noise factor, where σ (I p ) is the standard variation of the repeated spike current peaks, and Ave(I p ) is the average value of repeated spike current peaks. The value of σ (I p )/Ave(I p ) is calculated to be only ~1.7% for pH = 6. Detailed analysis of the reproducibility and noise of the spike sensing can be found in Supporting Note 2. The power consumption of our system can be estimated by multiplying the reading voltage, the channel current and the spike duration time 45 . Figure 5e shows the average energy dissipation for single-spike pH detection in each pH value from 10 to 4 with a spike duration time of 10 ms. The power consumption reduces from 103 pJ/spike to 15.6 pJ/spike when the pH value increases from 4 to 10. Of course, the power consumption can be further reduced by reducing the spike voltage and spike duration time. The influence of bending on the sensitivity is also investigated. As shown in Fig. 5f, after 1000 bending cycles, the sensitivity reduction is less than 10% for both quasi-static and single-spike sensing modes.
The single-spike sensing performance implemented with an asynchronous reading spike is also investigated. Figure 6a shows the sensitivity as a function of the inter-spike interval (Δ t) between V D and V G2 . If Δ t < 0, the reading spike V D is applied before V G2 . In this case, the protonic disturbance does not happen in the sensing process, thus the sensitivities are close to the equilibrium state and a sensitivity (∆G/G 0 ) of ~2.7 is obtained. When Δ t ≥ 0, the measured sensitivity is time interval (Δ t) dependent. A highest pH sensitivity (∆G/G 0 ) of ~5.6 is obtained when Δ t = 0 and it gradually reduces to 2.7 with increasing Δ t. Figure 6b shows the sensitivity as a function of spike duration time. At present, in order to accurately measure a low current in the nA scale by semiconductor analyzer, the shortest spike duration we can used is 10 ms. In this case, the maximal pH sensitivity (∆G/G 0 ) is measured to be ~5.6. We can anticipate that the sensitivity can be improved further when the spike duration time is reduced further. Detailed theoretical analysis of the influence of spike duration on the sensitivity can be found in Supporting Note 3. We also find that the sensitivity decreases gradually to ~ 2.4 when the spike duration increases to 2 s. These results indicate that the neuromorphic transistor sensor tends to arrive at equilibrium state with the increase of spike duration. The sensitivity of single-spike pH sensing performances of our neuromorphic transistor sensor can be improved further by additional gate synergic modulation. Figure 6c shows the influence of voltage bias applied on G 1 on the sensitivity when the device is operated in single-spike mode. The pH sensitivity increases when V G1 shifts from positive to negative. A maximum sensitivity (∆G/G 0 ) of ~63 can be obtained when a negative voltage of − 0.2 V is applied on G 1 . We also investigated the influence of voltage bias applied on V G1 on the energy dissipation of single-spike sensing measurement. Our results indicate that the energy dissipation can be gradually reduced when the V G1 is changed from 0.2V to − 0.2 V. As shown in Fig. 6d, an ultra-low energy dissipation of ~0.6 pJ/spike is estimated for pH = 10 at V G1 = − 0.2 V with the spike duration of 10 ms. Similar to the quasi-static synergic mode, an appropriate negative V G1 can make the device operate in the subthreshold regime. Thus, an enhanced sensitivity can be obtained. At the same time, negative bias can reduce the spike sensing current, which is critical for energy dissipation reduction.
The use of EDL electrolyte as gate dielectrics in flexible neuromorphic transistors can obviously reduce the operation voltage down to 2.0 V. Our results also demonstrate that spiking operation could greatly reduce the power consumption because the device is usually biased at zero voltage and only low voltage spikes with very short duration time are applied. Such neuromorphic transistors are favorable for flexible and portable sensor applications. Inspired by biological neuron, our neuromorphic transistor is designed with multiple in-plane gates. At present, we only investigate the influence of the second gate on the sensing performances in both quasi-static and spiking modes. Such devices can also be proposed as multi-functional sensors, where one in-plane gate acts as modulation terminal, one in-plane gate acts as calibration terminal, and other in-plane gates act as sensing input terminals. In the future, multiple-gate stochastic resonance effects may also be explored for further sensitivity improvements and power consumption reductions when such multiple-gate devices are operated in neuromorphic sensing mode. In summary, flexible oxide-based neuromorphic transistors were fabricated on plastic substrates. A pH sensitivity of ~105 mV/pH was obtained for quasi-static dual-gate synergic sensing mode. Our results demonstrated that single-spike dynamic sensing mode could remarkably improve the pH sensitivity and reduce response/recover time and power consumption. We also found that appropriate depression voltage applied on the sensing gate could further enhance the pH sensitivity and reduce the power consumption. Our results provided a novel strategy for fabricating biochemical sensors with high sensitivity, rapid response and ultralow power consumption.

Methods
Fabrication of flexible oxide-based neuromorphic transistors. First, 500-nm-thick nanogranular SiO 2 electrolyte films were deposited on ITO-coated PET substrates by plasma enhanced chemical vapor deposition (PECVD) at room temperature. SiH 4 (95% SiH 4 + 5% PH 3 ) and pure O 2 were used as reactive gases. Then, 30-nm-thick IZO channel layer was sputtered on the SiO 2 electrolyte films by using a nickel shadow mask. The sputtering was performed at a RF power of 100 W and a working pressure of 0.5 Pa using an IZO target. The channel width and length were 1000 and 80 μ m, respectively. Finally, 100-nm-thick Al source/drain electrodes and in-plane gate electrodes were deposited by thermal evaporation patterned by another shadow mask.
Preparation of pH solution. pH solutions were prepared by titrating 10 mM phosphate solution with dilute hydrochloric acid or potassium hydroxide solutions. The pH value of the solutions was monitored by a commercial pH meter. All chemicals were purchased from Sinopharm Chemical Reagent Co., Ltd. (China). Such phosphate buffered solutions were used for measurements due to their strong buffer capacity to the influence of external environment. Thus, the changes in pH signal due to relaxation of charges in oxide can be ignored.
Electrical and sensing performance characterizations. The sensing area of the device was immersed in deionized water for 24 hours before measurements. Frequency-dependent capacitances of the SiO 2 electrolyte films were characterized by a Solartron 1260A Impedance Analyzer in air ambient with a relative humidity of ~ 55%. Transistor characteristics and pH sensing performance were recorded by a semiconductor parameter characterization system (Keithley 4200 SCS) at room temperature. After each pH value test, the solution droplet was removed and the sensing area was rinsed two times in deionized water.