Emulating Nociceptive Receptor and LIF Neuron Behavior via ZrOx‐based Threshold Switching Memristor

For the progress of artificial neural networks, the imitation of multiple biological functions is indispensable for processing more tasks in a complex working environment. Memristors, which possess these advantages such as uniformity, high switching speed, and smaller device scale, are the better candidates compared to conventional complementary metal–oxide–semiconductor (CMOS) technology in artificial neural networks. In this work, an Ag/ZrOx/Pt threshold switching memristor (TSM) is designed to overcome the drawback of the large variation in the non‐volatile filament type memristor. The cycle‐to‐cycle and device‐to‐device variations are 5.6% and 4.9%. This device has mimicked the “nociceptive threshold,” “relaxation,” “no adaptation,” and “sensitization” features for the nociceptor which can prevent the artificial intelligence system from dangers in the external environment. Additionally, with the change in the strength of the external stimulus, the artificial neuron is also built by emulating “all‐or‐nothing,” “threshold‐driven‐spiking,” and “strength‐modulated” characteristics. The proposed threshold‐switching memristor allows the simultaneous emulation of the biological nociceptor and leaky integrate‐and‐fire neuron for the first time, which represents an advance in the bioinspired technology adopted in future artificial neural networks and humanoid robots.


Introduction
The von Neumann bottleneck has hindered the development of traditional digital computers because of the high energy consumption and the increased processing time. [1] Inspired by the extensive parallelism and high efficiency, researchers adapt the structures of the human brain and the processing principles into the circuit and put a great effort to emulate the diverse function of the biological system. [2] The sensory nervous system www.advelectronicmat.de responding to change. [14] A sensory system consists of sensory receptors, neural pathways, and parts of the brain involved in sensory perception. [15] Nociceptor is a key sensory receptor that recognizes noxious stimuli such as mechanical stress, chemical molecules, extreme temperatures, etc., through which a warning signal is generated and delivered to the central nervous system to initiate a motor response that minimizes potential physical damage. [16][17][18] An artificial nociceptor must possess four fundamental features: [19][20][21][22] i) nociceptive threshold, meaning that it is triggered to generate an output when the strength of the stimuli exceeds the threshold, ii) relaxation wherein the sensory pain disappears after a certain relaxation time, iii) no adaptation, meaning that a nociceptor cannot adapt to the stimulus it has experienced, iv) sensitization, expressing that the injured nociceptor is more sensitive to the harmful stimuli than the non-injured one. [23] Yoon et al. successfully proposed a diffusive memristor to exhibit all the features of a biological nociceptor, [24] including sensitivity regulation which is an essential property of biological nociceptors. [17] It paves a new way for the application of humanoid robots. Additionally, a good sensitivity could allow the nociceptor to respond to the noxious stimuli of short pulse width, and the work by Yoon et al. had reached the time length of 1 ms.
Most oxide-based memristors suffer from a high leakage current and high cycle-to-cycle variation. To deal with these issues, we chose ZrO x as the dielectric layer. ZrO x is a high-k transition metal oxide with a wide bandgap and large conduction band offset. [25,26] Some reports have suggested that ZrO x is a candidate for the gate insulator of metal oxide TFTs, due to its large capacitance density and a small leakage current density. Meanwhile, according to the DFT calculations, [27] silver is found to be stably and positively charged inside ZrO 2 , but the migration barrier is high (2.48 eV for Ag + interstitial). Therefore, Xue et al. suggested that the Ag cations have a high possibility to be reduced by the electron flow coming from the cathode. [27] This is not beneficial for building a "stable filament" in memory devices but it shall be advantageous for threshold switching. Here, we designed a highly reliable threshold-switching memristor (TSM) with an Ag/ZrO x /Pt structure. First, the artificial nociceptive receptor was built by performing the four fundamental characteristics. With the achievement of those nociceptor functions, the Ag/ZrO x /Pt TSM is proper to emulate another important biological behavior, the LIF neuron. To mimic LIF neuron activity, we changed the duration time of electric signals for better feasibility in neuromorphic computing. Multi-functions of biological neurons, all-or-nothing, threshold-driven spiking, and strength-modulated spiking have been achieved by the ZrO x -based TSM. The simultaneous realization of the functions for nociceptor simulation and neuromorphic computing further provides a novel approach for the development of ultrafast, in situ responsive, and delay-sensitive intelligent systems. [15,28]

Material Analysis
To fabricate a new artificial nociceptor that can lower the power dissipation in addition to performing essential neuron functions, we designed a metal-insulator-metal structure consisting of a zirconium oxide layer. The schematic illustration of the TSM is shown in Figure 1a. To build the TSM, we used a stack structure of Ag/ZrO x /Pt with a junction size of 200 × 200 µm 2 . To confirm the atomic ratio of zirconium and oxygen, the X-ray photoelectron spectroscopy (XPS) measurement was performed in Figure 1b. The peaks position at 530.52 and 532.02 eV correspond to O lattice and O non-lattice , respectively. Generally, the O non-lattice peak is considered the signal of oxygen vacancies. According to the area of these two peaks, it is observed that the ratio of oxygen vacancies is about 16.32%. Furthermore, the peaks at 182.7 and 185.0 eV correspond to Zr 4+ 3d 5/2 and Zr 4+ 3d 3/2 , respectively. The peaks at 180.9 and 183.3 eV represent Zr sub-oxide 3d 5/2 and Zr sub-oxide 3d 3/2 . Since the relative peak intensity of Zr 4+ is larger than that of Zr sub-oxide, the deposited ZrO x layer is mainly composited of Zr 4+ . Additionally, we calculated the atomic percent of zirconium and oxygen in XPS with the aim of the stoichiometric ratio of Zr:O. The atomic percent of an element is attained by dividing the peak area by the sensitivity factor. We confirmed that the stoichiometric ratio of Zr:O was 1:1.6. A high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) image is illustrated in Figure 1c to confirm the thickness of all films, in which the junction structure consists of Ag (≈54 nm)/ZrO x (≈23 nm)/Pt (≈98 nm) layers. All the layers were deposited by sputtering, details of the deposited process are explained in the Experimental Section. For high-resolution transmission electron microscopy (HRTEM) images, Ag and ZrO x layer which was fabricated by sputtering both have the polycrystalline structure in Figure 1d,e. We also use the selected area electron diffraction (SAED) pattern to demonstrate the polycrystalline structure in Figure S1a,b, Supporting Information. Additionally, the grazing-incidence X-ray diffraction (GIXRD) pattern on the ZrO x /Pt stack is shown in Figure S1c, Supporting Information. In addition to the peaks pertaining to Pt, there is a peak located at 2θ = 28°. This peak is pertaining to the (111) crystal plane of monoclinic ZrO 2 (ICDD-PDF 00-037-1484), which possesses the highest diffraction intensity and the d-spacing is 0.3165 nm. Therefore, the crystallized ZrO x is the monoclinic ZrO 2 phase although it is oxygen deficient. In previous research, the top electrode Ag was pre-dispersed in the dielectric layer which causes the threshold voltage to change and is hard to control. [29] To verify whether the top electrode Ag pre-disperses into the ZrO x dielectric layer, energy-dispersive X-ray spectroscopy (EDS) elemental analysis was performed in a selected area in cross-sectional HAADF images. Different Ag, Zr, O, and Pt element mapping confirms the presence of multilayer structures as shown in Figure 1f. Furthermore, it is suggested that Ag didn't disperse into the ZrO x layer during the fabricating process because of no observed signal of Ag in the ZrO x layer.

Figure 2a
shows the typical I-V curves of the Ag/ZrO x /Pt TSM for consecutive fifty I-V sweeps. The red curve was measured during the first voltage sweep process, and the gray curves were measured from subsequent 49 I-V sweep cycles. Applying the positive voltage sweeps (0 to 1 to 0 V), a compliance current www.advelectronicmat.de (I cc ) of 10 µA was adopted during the measurements to prevent permanent damage to the device. When the bias voltage exceeded the threshold voltage (V th ) around 0.4 V, the current value experienced an abrupt increase and reached the I cc level, which means that the device switched from a high resistance state (HRS) to a low resistance state (LRS). As the bias voltage was below the hold voltage (V hold ) around 0.02 V, there was a sudden current drop to about 10 pA. This device performs a typical threshold switching behavior at the positive sweep voltage. This threshold phenomenon results from the spontaneous rupture of the conductive filaments. Additionally, the threshold characteristic also shows promising potential in pain sensation and neuromorphic computing, such as the so-called LIF neuron.
The variation of the device was also studied, which will cause a large deviation in the artificial neural networks. [30] Both the cycle-to-cycle and device-to-device variations were tested in Figure 2b,c. In Figure 2b, the threshold switching voltages were counted from the 50 I-V cycles in Figure 2a. It is clear that the distribution of the threshold voltages ranges from 0.3 to 0.45 V. The distribution of the threshold voltages was fitted by Gaussian functions. The average value (µ) and the standard deviation (σ) are 0.36 and 0.02 V, respectively. In our Ag/ZrO x / Pt device, the value of σ/µ is only 5.6%. For the device-to-device variation, 25 devices were randomly selected from those fabricated TSMs and tested for 5 cycles of DC I-V sweep to calculate the average value of threshold voltage for each device. The details of measured I-V curves are shown in Figure S2, Supporting Information. The distribution of the threshold voltage of different devices is shown in Figure 2c. The average value (µ) and the standard deviation (σ) are 0.35 and 0.017 V, respectively.

www.advelectronicmat.de
Furthermore, the value of σ/µ is only 4.9%. To verify the lower cycle-to-cycle and device-to-device variation of the TSM, the value of σ/µ of our device and other devices are shown in Table  S1, Supporting Information. It is worth noting that our TSM performs better uniformity compared to the others, which possess more advantages in neuromorphic computing. [31][32][33] The typical I-V curves of the ZrO x -based TSM under different compliance currents from 10 nA to 100 µA are shown in Figure 2d. The on/off ratio was more than 10 5 under the compliance current of 100 µA, which is an indispensable advantage in the application of selectors. [34] It is worth noting that the device would transfer from threshold switching to memory switching under 1 mA compliance current in Figure S3, Supporting Information. Furthermore, it was observed that the threshold voltages (V th ) under different compliance currents were similar obviously. However, the hold voltage (V hold ) decreased with the increase of the compliance current, as shown in Figure 2e. When the compliance current increase, the diameter of the conductive filament would also increase, which means a stronger and more stable conductive path. [35] Therefore, the weaker www.advelectronicmat.de conductive filament needs a higher hold voltage (V hold ) to prevent the rupture. Furthermore, the current response of the device under AC was also investigated by applying consecutive voltage pulses. Repeatable threshold resistive switching behavior was confirmed in Figure 2f. A programming pulse (4 V, 100 µs) was used to switch the resistance state from HRS to LRS. Then, the pulse interval was 2 ms to make the device fully relax back to its original state. A read pulse (0.05 V, 100 µs) was used to read the HRS state. As shown in Figure 2f, the current value for HRS and LRS were both 100 µA after applying 220 pulse cycles, which means that the device was breakdown. Therefore, the endurance for the Ag/ZrO x /Pt device is approximately 220 pulse cycles.
The conduction mechanism in this TSM device was based on the formation and rupture of the conductive filament shown in Figure 2g. During the voltage sweep process, we can divide it into three stages: i) the applied voltage First, the Ag atoms from the top electrode are oxidized to Ag ions due to the applied electric field (Ag → Ag + + e − ). Owing to the application of positive bias, the oxidized Ag ions can move toward the Pt bottom electrode. Once the Ag ions reach the bottom electrode, they could be reduced to Ag atoms and precipitate at the interface of the ZrO x dielectric layer and the bottom electrode. When the V appl reaches the value of V th (V appl > V th ), these reduced Ag atoms could form a weak conductive filament because the compliance current limits the further growth of Ag filament. Finally, as V appl decreases below the value of V hold (V appl < V hold ), the conductive filament had a spontaneous rupture resulting from the surface energy minimization, which is related to the Gibbs-Thomson effect. To explain the Gibbs-Thomson effect in detail, we need to know the relationship between chemical potential and surface free energy following the Gibbs-Thomson equation in Equation (1): [36] 2 GT Ag macro Ag where γ is the surface free energy, r is the radius of the particle, V m is the molar volume, Ag macro µ is the chemical potential of macrocrystalline Ag electrode, Ag nano µ is the chemical potential of nanosize filament, and V GT is the Gibbs-Thomson potential which is a positive value. In previous research, it displayed that the Gibbs-Thomson potential increases with the change of resistance state from ON state to OFF state, which means that the surface free energy γ would decrease to the minimum value. Because of the minimization of surface free energy, the chemical potential of Ag conductive filament is larger than the Ag electrode. In thermodynamics, the hosted atoms would migrate from high chemical potential to low chemical potential. Therefore, these Ag atoms constituting the conductive filament would migrate to the side of the Ag electrode, resulting in the rupture of the conductive filament spontaneously.
To further explore the conduction mechanism of Ag/ZrO x /Pt memristors, the I-V characteristic curve was transformed into linearized voltage-dependent models of different carrier transport or conduction mechanisms for curve fitting in Figure S4, Supporting Information. For the data in the nonlinear region (HRS) of the I-V curves, the voltage-dependent relationship used was described as ln(I)−√V, log(I)−log(V), ln(I/V 2 )−1/V, ln(I/V)−√V, and ln(I)−1/V as shown in Figure S4b-f, Supporting Information, corresponding to the five possible carrier transport mechanisms in HRS which are Schottky emission, Ohmic conduction, Fowler-Nordheim tunneling, Poole-Frenkel emission, trap-assisted tunneling. [37][38][39] The R-square for the abovementioned five mechanisms are 0.98, 0.95, 0.48, 0.49, and 0.23. The R-square for Schottky emission meets the scientific community requirement (>98%). However, the R-square for the other four conduction mechanisms does not meet the requirement. Therefore, we can concern the latter four fitting methods unreasonable. In contrast, the curve of ln(I)−√V calculated based on the experimental data reveals better agreement due to the linear trendline, confirming the dominance of Schottky emission mechanisms in HRS. For the sake of confirming the Schottky emission, we also measured the I-V characteristic under different temperatures (298, 323, 348, 373, and 398K) and transformed it into the temperature-dependent plot (lnI/T 2 vs 1/T) in Figure S5, Supporting Information. [40] The linear relationship identified the thermionic (Schottky) emission well. In addition, the Schottky barrier height can be extracted according to Schottky emission conduction equation, as shown in Equation (2). [41] exp /4 where J is the current density, A* is the effective Richardson constant, T is the absolute temperature, q is the electronic charge, ϕ B is the Schottky barrier height, E is the electric field, ε is dielectric permittivity, and k is the Boltzmann constant. From the slope of ln(I/T 2 ) versus 1/T plot ( Figure S5, Supporting Information) and taking the dielectric constant of ZrO x as 15.8 (extracted from C-V measurement), the Schottky barrier height is calculated to be 0.17 eV. Since the Ag filament protrudes into the ZrO x layer, the Schottky barrier shall arise from the Ag/ZrO x interface. We have identified the electron affinity of ZrO x is about 3.7 eV in our previous work. [42] The Ag filament can be reasoned in form of nanocrystals and the work function of Ag nanocrystals is about 3.89 eV. [43] The Ag/ZrO x Schottky barrier thus is around 0.19 eV. Therefore, the extracted Schottky barrier is proper and it could further confirm the conductive mechanism in HRS.

Nociceptor
Nociceptive receptor, namely nociceptor, which can sense the stimulus from the external environment, is an indispensable part of the nervous system. Figure 3a is the working principle of the artificial nociceptor. When the external stimulus is received by the neuron, the electric signal is generated and transmitted to the nociceptor which determines whether the stimulus exceeds the nociceptive threshold. If the stimulus exceeds the nociceptive threshold, an action potential will be sent to the spinal cord. [17,44] Therefore, the brain receives the signal, and the feeling of the pain is produced to warn us. When the input pulse that exceeds the nociceptive threshold is applied to the TSM, the resistance state would be changed www.advelectronicmat.de from HRS to LRS. Due to the above-mentioned switching behavior, the current response will be detected at the output terminal. There are four features of nociceptor: nociceptive threshold, relaxation, no adaptation, and sensitization successfully demonstrated by applying voltage pulse stimuli. First, a series of pulses (with 100 µs pulse width) was applied to the TSM to emulate the nociceptive threshold characteristics. The pulse amplitude gradually increased from 4 to 7 V, and the current response is shown in Figure 3b. It can be observed that the current value did not increase until the pulse amplitude increases to 6 V. By increasing the pulse amplitude to 7 V gradually, the current output was enhanced further. Therefore, the TSM would output a signal if the electric stimuli exceed the nociceptive threshold (6 V). A more serious pain sensation is attributed to the larger stimulus intensity. In Figure 3c, a series of pulses with different pulse widths are applied to the memristor, and the pulse amplitude is fixed at 6 V. It shows that a sufficient long pulse width (100 µs) pulse is needed to generate the current response. Intending to protect the human body, it is essential to detect the noxiously shorter stimulus with high sensitivity right after the artificial nociceptor receives the stimulus. Compared with the reported artificial nociceptors in Table S2, Supporting Information, the ZrO x -based TSM could detect with a shorter stimulus (100 µs). The nociceptive threshold feature was fully demonstrated owing to the rupture and formation of the Ag filament in the TSM device. When the applied pulse is not vigorous enough, the disconnection of conductive filament results in a lower current. On the contrary, a higher current is observed if the applied pulse amplitude exceeds the nociceptive threshold.
After the noxious stimulus is removed, the "relaxation" process begins to dissipate the signal. As shown in Figure 3d, the first pulse (7 V 100 µs) is applied to switch the memristor on. After relaxing in different pulse intervals (5, 10, and 100 µs), the device received the second pulse (4 V 100 µs) which cannot switch the device in general. The relaxation time of shorter pulse intervals (5 and 10 µs) return the device to HRS when the second pulse was applied as shown in Figure 3e. However, it shows no current response to the second pulse if the pulse interval increases to 100 µs. It is worth noting that the recovery of the memristor after receiving the pre-noxious stimulus is faster than that in the previous research because of the shorter full relaxation time (100 µs). [24,45] Another important characteristic of the nociceptor is "no adaptation." If the pulse width of the successive stimuli is prolonged, the biological nociceptor could not adapt to the stimuli and in turn it generates the signal continuously to prevent the noxious stimuli. For the purpose of mimicking this characteristic, Figure 3f reveals that the device reached a constant current at 4.2, 3.4, and 2 ms after subjecting to pulses with amplitude of 3.5, 4, Figure 3. Nociceptor characteristics. a)The nociceptor system in an artificial nociceptor circuit consists of a ZrO x -based TSM. b) A series of 100 µs input voltage pulses (orange curve) of variable pulse amplitude (4 to 7 V) and the corresponding output current (magenta curve). c) A train of input voltage pulses (orange curve) is composed of a variable range of pulse widths from 5 to 400 µs, with a 6 V pulse amplitude, and the output current (magenta curve). d) Input voltages of 7 and 4 V with time intervals of 5, 10, and 100 µs, respectively. e) The corresponding output current for the 7 and 4 V input pulses. At an interval of 100 µs, the device is relaxed with no output response. f) The current response of pulse trains with different amplitudes (3.5, 4.0, and 4.5 V) applied before the 5 ms, and followed by the low voltage pulse train with 0.01 V amplitude in purpose of reading current response after the repeated noxious stimuli. All pulses are in width of 50 µs.
www.advelectronicmat.de and 4.5 V (width = 50 µs), respectively. After that, the response current did not change with the repeated input of the same stimuli. Therefore, the device shows "no adaptation" characteristics. When the stimulating pulse train was completed at 5 ms, a subsequent low-voltage pulse train (0.01 V/50 µs) was executed. It can be noticed that the device returned to the low-current state after applying the 3.5 V/50 µs and 4 V/50 µs pulse trains. However, after executing the 4.5 V/50 µs pulse train, the device remained at the high-current state even though the read voltage was only 0.01 V, implying that the Ag filament in the device became permanently connected after the 4.5 V/50 µs pulse train. This result well emulates that severe pain could cause irrecoverable damage, in addition to the "no adaptation" characteristic.
The last feature of nociceptors is "sensitization" which is important to protect the human body that had already been injured by the increase in pain sensation. As shown in Figure 4a, sensitization can be further divided into "allodynia" and "hyperalgesia" characteristics. Allodynia means abnormal pain induced by a stimulus that does not normally provoke pain, and hyperalgesia means an increased pain sensation to a normally painful stimulus. [46] To emulate the sensitization feature, we first applied the pulse of higher pulse amplitude than the threshold (7 and 8 V, 100 µs pulse width) to mimic the injury state of the artificial nociceptor and the device subjected to 0 V is considered as the reference in Figure 4b. Then, a pulse series of different pulse amplitudes (4, 4.5, 5, 5.5, and 6 V, 100 µs pulse width) are applied to the device to record the current change under the three different states of the nociceptor as shown in Figure 4c. Obviously, we observe that the current response of the "injured" (7 and 8 V) nociceptor is higher than that of the "non-injured" (0 V) nociceptor under a pulse series. The differences in the output current under three different states result from the 7 and 8 V pulses causing partial filament formation. To verify the difference between the "injured" (7 and 8 V) state and the "non-injured" (0 V) state, the output currents under a pulse series of varying pulse amplitudes are plotted in logarithmic and linear scale in Figure 4d,e. It is clear that the threshold shifts to lower voltage under the "injured" (7 and 8 V) state, corresponding to the "allodynia" property in Figure 4d. In Figure 4e, the higher output currents under the "injured" (7 and 8 V) state are also observed, demonstrating the "hyperalgesia" phenomenon.

LIF Neuron
As shown in Figure 5a, the schematic of a biological neuron illustrates the basics. The membrane potential is referred to as the potential difference between the exterior and the interior of the neuron. It is manipulated by the movement of Na + , K + , Ca 2+ , and other ions near the ion channel. [47] After receiving the external stimuli from these dendrites, the membrane potential will arise and result in the permeable membrane through which K + ions could pass. An action potential would be generated at the axon hillock as a result of depolarization. During depolarization, voltage-gated Na + channels open due to an electrical stimulus. As the Na + ions infiltrate into the cell in a hurry, their positive charges change the membrane potential

www.advelectronicmat.de
inside the cell from negative to positive. [48,49] If the membrane potential exceeds the neuronal threshold value (≈−70 mV), an action potential would be "fired", which means that an output spike was generated and passed to the postsynaptic neuron. [50,51] After generating the output spike, the membrane potential decrease to its pristine state during the relaxation period. The whole neuron activity is a fast depolarization-hyperpolarization process in biology, also called LIF behavior.
The LIF neuron can be implemented by a simple circuit, where the ZrO x -based TSM is connected with a capacitor in parallel. The capacitor could be either an external capacitor or intrinsic capacitance of the Ag/ZrO x /Pt device. [2,52] To demonstrate whether the intrinsic capacitance exists in the TSM, we carried out impedance spectroscopy measurements on ZrO x -based TSM. The applied frequency range is 10 3 to 10 6 Hz with different DC bias voltages (0-0.5 V) as shown in Figure S6, Supporting Information. It was found that the phase angle θ was near −90° as the DC bias voltages were 0-0.3 V, which proved the existence of intrinsic capacitance in the TSM. [53] In an artificial neuron, the all-or-nothing spiking is thought to be a fundamental feature to perform the spiking neural networks. It means that the output spike only has a binary value. [54] To mimic this characteristic, the four 1 µs pulses with different pulse amplitudes (5.5, 7, 8.5, and 10 V) were applied to the ZrO x -based TSM. The current response is shown in Figure 5b. When applying the weaker pulses (5.5 and 7 V), there is no obvious current response observed. However, with the increase of the pulse amplitude to 8.5 and 10 V, the TSM was triggered to generate an output current spike. It is worth noting that the strength of output spikes was the same when applying stronger pulses of varying pulse amplitude, which fully demonstrated the all-or-nothing spiking property.
For LIF behavior, it is essential to apply a sufficient input signal which exceeds the neuronal threshold to generate the output spike, is called neuronal-threshold-driven spiking. When applying consecutive identical pulses to the device, the Ag ions would flow in the dielectric layer to form the weak conductive filament resulting from the electric field. As the conductive filament formed, the current increased abruptly, which is the fire behavior. Subsequently, the conductive filament would rupture during the pulse interval, which is identified as the "leaky" behavior. Figure 5c shows the experiment results. Initially, the successive 500 pulses (5 V, 1 µs) were applied to the Ag top electrode, and the pulse interval was 1 ms. The artificial neuron could not fire because of insufficient stimuli lower than the threshold. Then, we increased the pulse amplitude to 6 V, and it induced a firing event. From these results, we conclude that a low pulse amplitude is not sufficient to generate an output spike despite applying a large number of identical pulses.
Another important characteristic of LIF neurons is the strength-modulated frequency which means that the spiking frequency could be controlled by the strength of the input. To change the strength of the input, the pulse amplitude varied from 6 to 7.5 V with an increment of 0.5 V while the pulse width remained the same at 1 µs, and the pulse interval are 1 ms. In Figure 5d, it was observed that the spiking number also increased when the pulse amplitude gradually increased. Furthermore, the statistical results of the relationship between the spiking number and the pulse amplitude are shown in Figure 5e. The measurement for each pulse amplitude was conducted for 10 different devices to demonstrate the uniformity as shown in Figure S7, Supporting Information. Besides, the pulse width is also a deterministic factor to control the strength of the input. Here, the consecutive 500 pulses of four different pulse widths (1, 1.5, 2, and 2.5 µs) were applied to the device. The pulse amplitude was fixed at 6 V, and the pulse interval was 1 ms. Similar to the previous experiment result, the spiking number varied from the distinct pulse widths in Figure 5f. We also used 10 different devices to perform this measurement. Figure 5g displays the relationship between the spiking number and the pulse width, which demonstrated the strength-modulated frequency well. Interestingly, we found that a large variation of the spiking number occurred when applying stronger input to the device. Given that the stronger input would have a higher probability to form the filament, the stochasticity increases when the probability of generating a spike increases. Similar stochastic neuron characters have been reported. [55,56] Although the firing behavior seems to be stochastic, the firing probability (i.e., spiking number) can be modulated by changing the pulse amplitude or pulse width, as presented in Figure 5e,g. The stochastic neuron devices have been applied to neuromorphic computing for the classification of handwritten digits. [55,56] Particularly, the devices with stochastic dynamics were implemented to add the dropout function in neuronal units to prevent the overfitting issue. [55] This is certainly a potential application for our Ag/ZrO x -based TSM and will be further investigated.

Conclusion
We have successfully fabricated an Ag/ZrO x /Pt TSM to emulate the nociceptive receptor and LIF neuron simultaneously. For the basic electrical characteristics, the volatile switching behavior was observed by repeatedly I-V sweeps. This can be ascribed to the formation and rupture of the weak conductive filament. By the way, the low cycle-to-cycle and deviceto-device variation exhibit the advantage in neuromorphic networks. Realizing the threshold switching behavior, the artificial nociceptor successfully emulates the four important features of nociceptive threshold, relaxation, no adaptation, and sensitization by applying different kinds of input signals. Adjusting the strength of these applied pulses, the LIF neuron can be also mimicked by performing all-or-nothing, neuronal-threshold-driven spiking, and strength-modulated characteristics. All these functions are due to the formation and dissolution of Ag conductive filament in the ZrO x film. It is a pioneering work for emulating multiple biological functions at the same time, which demonstrated the enormous potential of future artificial neural networks and humanoid robots.

Experimental Section
Fabrication of Ag/ZrO x /Pt TSM: Before fabricating the Ag/ZrO x /Pt TSM, the SiO 2 /Si substrate was cleaned by the RCA process. First, the 98 nm Pt bottom electrodes were deposited on SiO 2 /Si substrate via DC magnetron sputtering using a Pt target (99.99%). The sputtering process was performed in an Ar atmosphere, the radio frequency power was 100 W, and the working pressure was 4 mTorr. Then, ZrO x films with a thickness of 23 nm were deposited by radio frequency magnetron sputtering using a zirconium target (99.99%) around Ar/O 2 atmosphere. The radio frequency power was 80 W, and the working pressure was 10 mtorr. Finally, the silver top electrodes of ≈54 nm thickness were deposited on the surface of the ZrO x film by DC magnetron sputtering with the help of a copper shadow mask with the size of 200 × 200 µm 2 . During Ag electrode fabrication, the adopted power was 20 W, and the working pressure was 7 mtorr.
Electrical Measurements: The DC current-voltage (I-V) curve of this device was measured by the Keysight B1500A, and the pulse electrical properties were measured by the Keysight B1530A Waveform Generator/ Fast Measurement Unit (WGFMU) at room temperature. Impedance measurements were performed with a 4294A precision impedance analyzer. The impedance (Z) spectra were recorded in the 1 kHz to 1 MHz frequency range by applying an ac level = 25 mV sine-wave voltage. During the electric test, the voltage bias was applied on Ag top electrode, while the Pt bottom electrode was grounded.
Characterization: The morphologies of the fabricated sample were observed using a TEM (JEOL-2100F CS STEM). The TEM sample was prepared by a focused ion beam. Dark-field STEM images were also acquired, and EDS was employed for further analysis. XPS was performed using monochromatic Al Kα radiation in a JEOL, JAMP-9500F Auger electron spectroscopy instrument. The grazing-incidence X-ray diffraction (GIXRD) analysis was measured by D8 Discover with GADDS (Bruker AXS GmbH, Karlsruhe, Germany).

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.