Realization of nociceptive receptors based on Mott memristors

Inspired by the human brain, researchers are exploring more parallel and efficient computational models and hardware architectures to break the von Neumann bottleneck. In the biological nervous system, each individual neuron exhibits the remarkable ability to process thousands of different synaptic inputs unequally. ¹⁾ The sensory nervous system (SNS) is an integral component of the complex neural network responsible for processing sensory information which converts the detected stimuli into initial biochemical signals, ultimately reaching sensory centers in the brain. ²⁾ Nociceptors are key sensory receptors that perceive various types of harmful stimuli and relay these harmful stimuli to exceed a certain threshold to the central nervous system and minimize potential physical damage. ³⁾ Therefore, the SNS is crucial for artificial intelligence systems to gather external data, adapt to real-world changes, and aid robots in efficiently performing tasks with high accuracy and robustness. ⁴⁾

CMOS neuron circuits have been widely used for the construction of synapses and neurons. ^5,6) However, the constructed circuits require numerous transistors, resulting in complexity and limitations in area efficiencies and low energy efficiency. ^7–9) Hence, developing novel electronic devices to address these issues has become the primary focus of researchers. Mousam et al. considered an Ag/TiO₂/Pt/SiO₂/Si memristor that exhibits very low power consumption and has the ability to imitate synaptic and nociceptive functions. ¹⁰⁾ Similarly, Saransh et al. fabricated a photoelectric synaptic memristor based on Zn₂SnO₄/(Cu₂O) heterostructure, which exhibits primary synaptic functions and mimics the essential features of a nociceptive receptor. ¹¹⁾ Recently emerging devices, especially the Mott memristors, are expected to be promising candidates to implement the biological synapse and various neuronal functions. Mott memristors, possessing a threshold property and volatile resistive switching, exhibit rich dynamics and high bio-mimetic capabilities, enabling the construction of high-density and highly efficient spiking using a single device. This leads to a significant reduction in circuit complexity. ¹²⁾

In this paper, we have analyzed the TS mechanisms behavior in the Pt/Ag/NbO_x /W Mott memristor. By conducting a comparative analysis with the memristor of inert electrodes, the threshold characteristics of Ag-dominated devices are revealed. Then, a single memristor made of NbO_x materials is used to emulate the LIF artificial neuron by changing the input signal, which exhibits leaky integration, threshold-driven fire, self-relaxation features, and strength-modulated spiking. Furthermore, the device can be explored to build the artificial nociceptive receptor, including allodynia, and hyperalgesia. The three fundamental and sensitization characteristics of nociceptors have been imitated, which demonstrates the possibility of NbO_x -based nociceptors for the application of humanoid robots.

To fabricate Pt/Ag/NbO_x /W memristors, we use a silicon wafer as a substrate. First, 40 nm W was deposited as the bottom electrode (BE) on the substrate by magnetron sputtering. Next, a 40 nm NbO_x film was fabricated using the magnetron sputtering technique on the W BE. After that, a thin Ag metal layer was deposited on top of the NbO_x film by magnetron sputtering. Finally, a 30 nm Pt was deposited by magnetron sputtering as the top electrode (TE). Electrical characteristics of devices were conducted by a Keithley 4200A-SCS semiconductor analyzer at room temperature. During all electrical measurements, the voltage was biased on the TE (Pt), and the BE (W) was grounded.

In this work, two terminals Pt/Ag/NbO_x /W devices were fabricated to imitate an artificial neuron and nociceptors. Figure 1(a) shows the cross-sectional scanning electron microscope (SEM) image of the device. In the human cognitive system, external pain information is perceived through the sensory receptors in the skin. Once the injury signal exceeds the threshold of nociceptors, it is transmitted to the central nervous system via connected neurons to provide a warning, minimize potential damage, and perform data computation, as illustrated in Fig. 1(b). ^13,14) The designed devices schematic diagram is shown in Fig. 1(c). The application of an electrical signal to the device serves as the external stimulus. When the amplitude of the pulse exceeds the threshold voltage of the memristor, the device transitions into an "on" state, leading to the detection of a current signal at the output terminal of the device.

Zoom In Zoom Out Reset image size

The memristor we fabricated is a unipolar device that transitions its resistance state through the application of a positive voltage. To investigate the electrical characteristics of NbO_x memristors, a direct current (DC) sweeping voltage of 1.0 V was applied to the device. The Pt/Ag/NbO_x /W memristor exhibits a threshold-switching behavior. Figure 2(b) shows the typical I–V curves of the Pt/Ag/NbO_x /W devices for consecutive 100 I–V sweeps with a compliance current (I_CC) of 8 μA implemented to prevent permanent damage to the device. The red curve represents the measurements obtained during the initial voltage sweep, while the gray curve represents the measurements obtained during the subsequent 99 I–V sweep cycles. When the sweeping voltage exceeded the threshold voltage (V_th), the device turned on, and the current rapidly reached the I_CC level. This indicates the transition from the high resistance state (HRS) to the low resistance state (LRS). As the sweeping voltage falls below the hold voltage (V_hold), which is approximately 0.14 V, the current drops from 8 μA to about 0.1 μA. This indicates the transition from the LRS to the HRS. To investigate the impact of I_CC on the switching behavior of the device, Fig. 2(c) shows the typical I–V curves of the NbO_x -based memristor under different I_CC from 5 μA to 10 μA. It can be seen that the device undergoes a transition from threshold switching to memory switching at the I_CC of 10 μA among the six trials, indicating that the NbO_x -based memristor exhibits TS behavior under low I_CC and resistive switching behavior under high I_CC.

Zoom In Zoom Out Reset image size

In order to demonstrate the repeatability of the TS phenomenon, consecutive voltage pulses were applied to the device. Each switching cycle comprised a pulse sequence, including red-marked pulses (1 V, 10 ms) to turn the device on, and a blue-marked pulse (0.02 V, 10 ms) for reading the HRS state. An interval time of 400 ms was used to ensure that the device fully returned to its resting state. It was observed that the endurance of the Pt/Ag/NbO_x /W device exceeds 1600 pulse cycles, as shown in Fig. 2(d). This threshold mechanism is attributed to the spontaneous rupture of conductive filaments (CFs). Moreover, this unique threshold characteristic holds the potential for emulating biological neural behaviors and neuromorphic computing. ¹⁵⁾

Pt/Ag/NbO_x /W is a structure based on the Mott memristor. The most widely accepted mechanism for the Mott memristor is the transition between metallic and insulating phases induced by temperature and/or electric fields. ¹⁶⁾ Because the active metal Ag is selected as the TE, two different physical processes occur within the NbO_x matrix including the metal–insulator transition (MIT) of the NbO₂ nanochannel and the migration of Ag ions. Under the influence of the positive voltage, the NbO_x layer forms a crystalline NbO₂ channel, in which field-driven native oxygen ions drift leading to the programming of a suboxide NbO₂ nanochannel that underlies threshold-switching behavior. ^17–19) To further investigate the conduction threshold mechanism, the Pt/NbO_x /W memristors were fabricated for comparison. Figure 2(a) illustrates the typical threshold-switching I–V curves during 100 sequential cycles. The voltage is swept from 0 to 2.3 V, while maintaining an I_CC of 0.55 μA. It is worth noting that the Pt/NbO_x /W memristor structure exhibits a slower opening process with a higher threshold voltage and exhibits a significant negative differential resistance effect in comparison to the device with the TE Ag. Therefore, between the two conduction mechanisms, it is possible that the formation and rupture of Ag CFs dominate the threshold behavior of the device. Under positive electrical stimulation, Ag ions are initially generated from the TE and move toward the BE, where they are reduced to form Ag atoms as shown in Fig. 3(a). Subsequently, Ag atoms gradually accumulate to form weak Ag CFs between the two electrodes as shown in Fig. 3(b). The growth of Ag filaments is constrained when the I_CC is set at a low level. The filaments spontaneously rupture due to the Joule heat effect and the minimization of surface energy when the applied voltage is not strong enough as shown in Fig. 3(c). ^20,21) Moreover, a higher I_CC tends to generate stronger Ag filaments, which holds the promise of improving the retention capability of the device as shown in Fig. 3(d). ^22–24) In conclusion, these features are crucial for constructing a biological emulator of neural behaviors and simulating nociceptors.

Zoom In Zoom Out Reset image size

Integration and firing are fundamental functions of neurons that allow for the processing and transmission of information in the nervous system. ²⁵⁾ As shown in Fig. 4(a), biological neurons receive input spikes and undergo continuous integration, persistently accumulating information until the membrane potential reaches a V_th. Then the action potential is triggering and outputting. In this work, we utilized Pt/Ag/NbO_x /W memristors to simulate LIF neurons. Accordingly, Fig. 4(b) shows the experimental realization of a LIF neuron utilizing NbO_x memristors. Under consecutive voltage pulses with a constant amplitude of 0.8 V and width of 200 ms, the devices electrically emulated neuronal behaviors of the constructed LIF neuron. Initially, the neuron unit is in the integration period, wherein the current is accumulated. Once an artificial dendrite is activated, there is a sudden surge in the current, closely resembling the action potential behavior observed in biological neurons. Then, the input signals are replaced by reading pulses with an amplitude of 0.05 V to investigate the neural function of relaxation. It has been observed that the current gradually decreases during the simulation of reading pulses, resembling the relaxation process in LIF neurons. This finding indicates that the Pt/Ag/NbO_x /W memristors possess the capability to simulate the "leaky" integration function observed in biological neurons, where leak channels exist, allowing ions to influx and outflow simultaneously as neurons initiate information processing. ²⁶⁾

Zoom In Zoom Out Reset image size

Normal nociceptors exhibit three fundamental characteristics: "relaxation," "threshold," and "no adaptation." ²⁷⁾ As shown in Fig. 5(a), the application of voltage pulses (0.4 V) below the threshold does not trigger the device. Subsequently, the voltage pulse (0.8 V) accompanied by a wide pulse width, is applied to switch the memristor on, simulating a strong stimulus that induces an injury-like response. Following this, the subsequent pulses of 0.4 V can result in an observable current response. Over time, the nociceptor starts the "relaxation" process that the current gradually decreases and the device returns to HRS. This can be attributed to the volatility of the memristor, and the subsequent voltage of 0.4 V is not strong enough to maintain the connection of the CFs for a long time. Therefore, after a period of time, the CFs spontaneously rupture. The "threshold" feature was simulated by applying a series of pulses ranging from 0.42 V to 0.49 V as shown in Fig. 5(b). When the amplitude is below 0.46 V, the probability of spiking is 0. However, when the input pulse amplitude exceeds 0.47 V, the probability of spiking in the device reaches 100%. Hence, when the stimulus exceeds the nociceptive threshold, the device will generate an output pain signal. The observed "threshold" characteristic can be attributed to the formation of conduction paths within the switching layer of the device. In Fig. 5(c), it is observed that the response current remains invariant with repeated additional input of identical stimuli. This characteristic displayed by the devices can be attributed to the constrained growth of Ag conductive filaments, as there is a limited number of Ag ions that can be excited when the same voltage is applied, or the TE contains a limited number of Ag atoms. This phenomenon is similar to the "no adaptation" properties of nociceptors, which are essential for organisms to defend themselves against repetitive noxious stimuli.

Zoom In Zoom Out Reset image size

High sensitivity is also an essential function of the nociceptors, which can be further expressed by "allodynia" and "hyperalgesia" characteristics. ²⁸⁾ Allodynia refers to the abnormal perception of pain in response to a typically non-painful stimulus, representing an abnormal pain response and hyperalgesia is a physiological state characterized by an enhanced sensitivity to painful stimuli. ^29,30) These features of the nociceptors can protect the human body from further injury by increasing sensitivity. To demonstrate the "sensitization" feature, we applied the high amplitude voltage pulse (1 V and 1.1 V, 1 s width) to the device, simulating strong damage that can lead to an injury as shown in Fig. 6(a). To better monitor the changes in current under the "injured" and "normal" cases of the nociceptors, a pulse series of varying amplitudes (0.3, 0.4, 0.5, 0.6, and 0.7 V, 1 s width) are applied to the device to monitor the current changes as shown in Fig. 6(a). Notably, in Fig. 6(b), it can be observed that the "injured" nociceptors exhibit higher currents than the "normal" nociceptors, and the current increases as the "damage" increases. In order to more clearly visualize the distinction between "injured" and "normal" nociceptors, we employed linear and logarithmic scales to show the variation in output current at different input voltages. In Fig. 6(c), it is clear that the threshold voltage decreased in the "injured" state, demonstrating the "hyperalgesia" phenomenon. The current response to voltage stimulations shows a gradual increase, corresponding to the "allodynia" property in Fig. 6(d). These phenomena suggest that the Pt/Ag/NbO_x /W Mott memristor can successfully simulate the allodynia and hyperalgesia characteristics.

Zoom In Zoom Out Reset image size

In conclusion, we have perfectly realized the synaptic and nociceptive behavior simulations using a simple Pt/Ag/NbO_x /W sandwich structure. The Pt/Ag/NbO_x /W Mott memristors exhibit TS characteristics under a low I_CC of 10 μA and it is possible that the formation and rupture of Ag CFs dominate the threshold behavior of the device. Subsequently, the designed memristors are employed to emulate LIF neurons. Finally, we built an artificial nociceptor with three fundamental characteristics, which include "threshold," "relaxation," and "no adaptation." Following exposure to a painful stimulus, the nociceptors have demonstrated "hyperalgesia" and "allodynia" behaviors. All these functions are due to the unique threshold-switching behavior and rich dynamics of the Pt/Ag/NbO_x /W Mott memristor. These results show that the proposed Mott memristors provide an alternative approach for constructing a configurable neuron and thus have great potential for building artificial neural networks and robots.

Acknowledgments