Bio‐Voltage Memristors: From Physical Mechanisms to Neuromorphic Interfaces

With the rapid development of emerging artificial intelligence technology, brain–computer interfaces are gradually moving from science fiction to reality, which has broad application prospects in the field of intelligent robots. Looking for devices that can connect and communicate with living biological tissues is expected to realize brain–computer interfaces and biological integration interfaces. Brain‐like neuromorphic devices based on memristors may have profound implications for bridging electronic neuromorphic and biological nervous systems. Ultra‐low working voltage is required if memristors are to be connected directly to biological nerve signals. Therefore, inspired by the high‐efficient computing and low power consumption of biological brain, memristors directly driven by the electrical signaling requirements of biological systems (bio‐voltage) are not only meaningful for low power neuromorphic computing but also very suitable to facilitate the integrated interactions with living biological cells. Herein, attention is focused on a detailed analysis of a rich variety of physical mechanisms underlying the various switching behaviors of bio‐voltage memristors. Next, the development of bio‐voltage memristors, from simulating artificial synaptic and neuronal functions to broad application prospects based on neuromorphic computing and bio‐electronic interfaces, is further reviewed. Furthermore, the challenges and the outlook of bio‐voltage memristors over the research field are discussed.

biological tissues. [12][13][14][15][16][17][18] Thus, bio-voltage memristor is expected to be implemented in low power neuromorphic computing, brain-computer interfaces, and biological integration interfaces. Memristors have a simple sandwich structure consisting of a top electrode (TE), an active layer, and a bottom electrode (BE) (as illustrated in Figure 1d). [19][20][21][22][23][24][25][26] The functional realization of the bio-voltage memristor relies on the reversible switching of the active layer between a high-resistance state (HRS) and a low-resistance state (LRS) under 50-120 mV bio-voltage. [3,[27][28][29][30][31][32][33][34][35] The functional simulation and bio-voltage matching of synapses and neurons can be realized by using the switching properties of bio-voltage memristor, which has great application potential in the field of building brain-inspired neural networks and provides infinite possibilities for the realization of advanced braincomputer interfaces (Figure 1e). Therefore, the exploration of brain-like neuromorphic devices based on bio-voltage memristors can promote the interaction between energy-efficient artificial neuron networks and biological neural networks. [36][37][38] Up to now, there are many reviews that analyze and study neuromorphic memristors from the perspectives of devices, materials, mechanisms, applications, and so on, promoting the development of this emerging field. [2,[39][40][41][42][43][44][45][46][47][48][49] However, unlike the existing reviews, in order to construct devices that can connect and communicate with living biological tissues to function directly in response to signals from the brain, [1] the interaction of biological systems with electronic neuromorphic systems is based on bio-voltage to meet the requirements of energy-efficient biological systems. Therefore, a comprehensive review of recent advances in bio-voltage memristors is urgently needed. In this review, we systematically present the various physical mechanisms and materials of memristors for operating voltages to reach bio-voltages ( Table 2). Next, the developments of bio-voltage memristors are further explored, from the successful simulation of artificial synaptic and neuronal functions to applications based on computing and bio-electronic interfaces. Furthermore, the current challenges and the potential future directions of bio-voltage memristors in realizing long-term stable and efficient bioelectronic interaction research are discussed.

Mechanisms of Bio-Voltage Memristors
To investigate how bio-voltage memristors work, we review the various types of bio-voltage memristors including active layer catalytic type, nanogap type, quantum wires (QWs) type, van In biological synapses, neurotransmitters from the presynaptic membrane into the synaptic cleft are received at the postsynaptic plasma membrane by NMDA and AMPA receptors/ion channels, resulting in the opening or closing of the ion channels, eventual ion influx into the postsynaptic neuron, and establishment of postsynaptic potentials, which suggest that the process plays an important role in regulating rapid changes in the membrane conductance and membrane potential of the postsynaptic cell. c) Schematic of an actual action potential which can be divided into four stages, including resting, depolarization, repolarization, and hyperpolarization. d) Schematic structure of a memristor synapse with a sandwich structure consisting of a TE, an active layer, and a BE, and its typical switching voltages meet 50-120 mV (bio-voltage) in biological systems. e) Output bio-voltage pulse signals of LIF based on memristor.

Active Layer Catalytic Type
The switching behavior of electrochemical metallization memristors is generally associated with active metal cations in the active layer. Its switching dynamics mainly involve three processes: anodic oxidation (active metals M → M + + e − ), M + migration, and cathodic reduction (M + + e − → M). [50] As M are readily oxidized to cations in the surrounding environment, M + migration is generally not a threshold event. Thus, selectively controlling the cathode reduction may become a key factor impacting the switching voltage (ΔV th ) of memristors. To reduce the ΔV th of the memristors, the introduction of catalyst can reduce the reduction overpotential (ΔE), which plays an important role in facilitating the cathodic reduction process; and ultimately, accelerates the formation of conductive filaments (CFs) (Figure 2a).
Inspired by biology, Fu et al. introduced a catalyst based on purified protein nanowires harvested from Geobacter sulfurreducens as an active layer of the memristor to realize switching operations at bio-voltages of 40-100 mV. [10] The protein nanowires contribute to accelerate reduction of Ag, changing the metal ion reactivity and electron transfer properties, promoting cathodic reduction, which ultimately leads to a decrease in switching voltage (Figure 2b). [51][52][53][54][55] As shown in Figure 2c, Ag nanoparticles were distributed between a pair of Ag electrodes of the biovoltage memristor after electroforming ((i) → (ii)). However, the distribution of Ag nanoparticles was not observed ((ii) → (iii) after removal of the protein nanowires with ultrasonication). These results certificated that the Ag conduction channel was completely established in the protein nanowire film extracted from G. sulfurreducens biomaterial. Similar active-layer catalytic memristor with bio-voltage can also be extended to other material systems. [25,56] Therefore, it is feasible to reduce the switching voltage and even meet the requirements of bio-voltage by introducing an active layer with catalytic function in the memristor.

Nanogap Type
Reducing the distance between the two metal electrodes of a memristor to 1 nm, we call it a nanogap-type memristor, which works by forming and rupturing metal CFs in the nanogap between a solid-electrolyte electrode and an inert electrode. [34,[57][58][59][60][61][62][63] To start with, the size of unit devices can be greatly reduced, which is beneficial to obtain ultrahigh-density storage. Then, the time for forming CFs is shorter due to nanometers distance, which is conducive to the acquisition of high-speed devices. More importantly, the lower operating voltage required for nanogap-type memristor can better satisfy the energy-efficient requirements of biological system. Takeo et al. successfully constructed Ag/AgS 2 /Pt nanogap type device by using a scanning tunneling microscope (STM) probe to form a nanogap on the AgS 2 film. As AgS 2 is a solid electrolyte electrode, the resistance of the AgS 2 layer is much smaller than gap layer; so, the resistance of the AgS 2 layer in the device can be ignored. Thus, it is believed that the factor that determines the resistance change of the device mainly comes from the gap layer. Furthermore, the formation of nanoscale Ag CFs by the nanogap between the Pt tip of the STM and the Ag 2 S film was confirmed in Figure 3a, which relies on a solid electrochemical reaction of Ag 2 S. Meanwhile, one can see from Figure 3b that the operating mechanisms with formation and annihilation processes of Ag atom bridge of bio-voltage memristor with Ag/Ag 2 S/Pt structure were accomplished by applying 80 and −30 mV voltages. [34] To deeply explore the mechanism of growth and shrinkage of Ag CFs, the STM-controlled phenomenon was explained in Figure 3c. When a forward voltage is applied, the equilibrium state in which the electrochemical potential of Ag + ions in the Ag 2 S electrode is equal to the electrochemical potential of Ag atoms on the surface (the activation energy for reduction [E R, from Ag + to Ag] is equal to oxidation [E O, from Ag to Ag + ] at the Ag 2 S surface) is broken; so that, the Ag + ions diffuse to the sub-surface of the Ag 2 S to increase the concentration. [64] This causes the E R to become smaller than the E O , which promotes the accumulation of  Adv. Electron. Mater. 2023, 9,2200972 www.advelectronicmat.de  Ag atoms to form Ag CFs in the nanogap, leading to the switch from HRS to LRS. On the contrary, E O becomes smaller than E R under the condition of applying negative voltage, which accelerates the shrinkage of the Ag CFs, triggering the switching from LRS to HRS. [57,64] Similar resistive switching operation of electrode nanogap-type bio-voltage memristors can generalize to nanogap-type devices utilizing Cu 2 S [62] and RbAg 4 I 5 [63] as the solid electrolyte electrode.

Quantum Wires Type
The structure of QWs type memristors usually uses QWs as the active layer and active metal (such as Ag) as the TE. Due to the larger diameter of the nanowires (NWs), NWs memristor tends to cause electrons to hop between parallel Ag sub-branches and provides a greater number of short metallic paths for electrons to traverse in the NWs, leading to higher HRS current. Compared with NWs memristor, the smaller diameter of the QWs (≈10 nm) memristor is more advantageous in maintaining the stability of HRS resistance and low-power operation (Figure 4a).
Controlling the length of QWs may become an important factor affecting the switching voltage of memristors. Poddar et al. demonstrated improved memristor operating voltage by tuning the diameter of the QW. [65] Figure 4b shows the switching voltages decrease significantly as the QWs length of the MAPbI 3 QWs devices decreases, indicating the existence of a threshold electric field that triggers the electrical switch. [66] When the QWs length is tuned to 180 nm, the MAPbI 3 QWs devices with the low switching voltages of 100 and −80 mV (Figure 4c), a high selectivity of 10 7 , and ultra-fast switching speed of 193.3 ps (writing speed) and 200 ps (erasing speed), can meet the requirements of bio-voltage systems. The ultra-fast switching speed is mainly attributed to the shorter QWs length; Ag ions and electrons will traverse smaller distances to complete formation and rupture of CFs. The switching mechanism of the MAPbI 3 QWs bio-voltage memristors depends on the formation and rupture of Ag conduction filament ( Figure 4d). [65] The ionic radius of Ag ions (110 pm) is smaller than that of iodide ions (220 pm); so, it is easier to move within the MAPbI 3 QWs. Then, driven by the electric field, Ag + ions are reduced to Ag atoms by fast moving electrons in the monocrystalline QWs. [67][68][69] Therefore, it is beneficial to reduce the working voltage of the memristor by using QWs as the active layer.

Van der Waals Interfaces Type
Both the BE/channel interface and the TE/channel interface are necessary for the switching behavior of a memristor with a sandwich structure, which greatly affects the working voltage and stability of the memristor. However, the poor stability and yield of most of the current memristors are still limited by the non-ideal interface between the metal electrode and the active channel. Active layer catalytic type bio-voltage memristors. a) Schematic of an active layer with catalytic function in memristor that promotes the cathodic reduction by reducing the ΔE to lower the ΔV th of device. b) The schematic diagram of reduction mechanism of Ag + ion by protein nanowires extracted from G. sulfurreducens bio-material to potentially catalyze bio-voltage memristors. c) The SEM images of G. sulfurreducens protein-nanowire memristor with distributed Ag nanoparticles distributed ((i) → (ii) before and after electroforming) and without distributed Ag nanoparticles ((ii) → (iii) after removal of the protein nanowires), which proved that the Ag CFs were completely established in the G. sulfurreducens protein nanowire film. Reproduced with permission. [10] Copyright 2020, Springer Nature.

www.advelectronicmat.de
Compared with high operating voltage devices fabricated using conventional direct deposition techniques or the buffer layer embedding method, [70][71][72][73] the vdW interfaces can effectively guarantee an ideal electrode/active channel interface, resulting in improved device stability and lower operating voltage. [74] To optimize the device performance, Li et al. fabricated a vdW interfaces type memristor of bio-voltage operating voltage (set/reset voltage of 120 mV/−40 mV) by physically sandwiching prefabricated metal electrodes (Ag TE and Au BE) on both sides of the ultrathin 2D channel material InSe (Figure 5a). [74] The reason for the reduction in the low switching voltage of the memristor can be considered from van der Waals electrodes. On the one hand, the vdW TE minimizes the damage to the channel material caused by conventional metal deposition, thereby enabling an ultrathin channel thickness of 1.6 nm, while preserving the inherent channel properties, which is critical to reduce the voltage required for the formation of Ag CFs. [75] To further investigate the effect of TE on the reproducibility and stability of memristors, schematics and corresponding I-V characteristics of the vdW Ag/3.2 nm Inse/vdW Au and deposited Ag/3.2 nm Inse/vdW Au devices are presented in Figure 5b. The deposited Ag/3.2 nm Inse/vdW Au device has no resistive switching behavior, again demonstrating that conventional TE fabrications induce considerable damage. [76] On the other hand, the root mean square (RMS) roughness of vdW BE/2.4 nm InSe and rough BE/2.4 nm InSe channel are 0.27/0.27 and 1.1/0.6 nm, respectively. Thus, the vdW BE ensures intimate BE/channel contact with low contact resistance compared to the conventional poor contact of rough BE ( Figure 5c). In addition, the contact resistance of vdW metal electrodes shows lower series resistance. Therefore, the several geometric advantages of vdW structure have dramatic impact on device performance. Still, the detailed mechanism that vdW electrodes can reduce the switching voltage needs to be further explored in future work.

Heterostructures Type
Transition metal oxides (TMOs) such as MoO x and WO x can be regarded as semiconductors with native oxygen-deficiencyinduced dopants, which play a vital role in the resistive switching behavior of memristors. The operating voltage of memristors fabricated by layered transition metal dichalcogenides (TMDs) and ultrathin TMOs heterostructures can meet the requirements of energy-efficient biological systems. Schematic of the vertical-stack TMO/TMD memristor and many vacancies point defects present at the TMO and TMD heterointerface are shown in Figure 6a. The Ag/MoO x /MoS 2 /Ag and Ag/WO x /WS 2 /Ag devices exhibit excellent memory behavior at positive and negative bio-voltages (100 mV) in Figure 6b. [77] The X-ray photoelectron spectroscopy (XPS) spectrum indicates that the oxidation of MoS 2 is limited to ≈3 nm of the surface due to the limited oxygen diffusion conditions under the condition of annealing at 200 °C for 3 h (Figure 6c). It is speculated that the MoO x is composed of Mo 6+ and in the lower oxidation state Mo 5+ composition. In addition, as the depth increases, the ratio of Mo 6+ to Mo 5+ decreases gradually, indicating the density of surface oxygen vacancies may be lower than that of the bulk. Moreover, as the oxidation temperature increases, the content of insulating MoO 3 at the surface will be higher ( Figure 6d). The switching voltage of MoO x /MoS 2 memristors with different thicknesses of MoS 2 layers is almost unchanged (Figure 6e), proving that the MoS 2 layer mainly acts as a mechanical support layer and while the resistive switching behavior with bio-voltage is mainly controlled by the ultrathin MoO x layer. Furthermore, to provide a reliable forming-free process and bio-voltage switching, the ultrathin MoO x layer should also be thick enough to ensure a continuous layer and adequate blocking of charge carriers. The resistive switching behavior of the bio-voltage memristor with TMO/TMD heterostructures structure depends on the formation and rupture of the oxygen vacancies CFs in Figure 6f.

Iodine Vacancy Type
Perovskite halides materials have high defect migration speed and low defect migration barrier, which are important materials for studying high-performance bio-voltage memristors. [78][79][80] In Figure 7a, when the thickness of the CH 3 NH 3 PbI 3 film was reduced to 220 nm, the threshold voltage of CH 3 NH 3 PbI 3 -based memristors was reduced to 0.11 V to meet the requirements of biological systems. [78] To investigate the mechanism of the lower threshold voltage observed in devices with Ag anodes, physical mechanism of Ag/CH 3 NH 3 PbI 3 /Ag memristors is shown in Figure 7b. [79] The interfacial reaction  [34] Copyright 2011, Springer Nature. c) The operation mechanisms with formation and disappear processes of Ag atom bridge of nanogap-type memristors by applying positive/negative bias.

www.advelectronicmat.de
at the anode of devices using Ag can be expressed as: → Ag + I − + e − ↔ AgI + e − . The formed AgI x region becomes an effective I − reservoir, thereby inhibiting the diffusion of I − from the anode electrode region and the recombination process with V I s in the MAPbI 3 active layer. Thus, the switching behavior of Ag/CH 3 NH 3 PbI 3 /Ag devices depends on the increase or decrease of V I concentration corresponding to the increase or decrease in device conductivity. The SEM images of Ag/CH 3 NH 3 PbI 3 /Ag memristor at pristine state and LRS and energy dispersive spectroscopy (EDS) spectra collected at labeled positions 1 and 2 reveal that Ag clusters are only aggregated near the anode but not form the CFs throughout the CH 3 NH 3 PbI 3 film (Figure 7c,d). Moreover, typical SEM and EDS measurements of Ag/CsPbI 3 /Ag memristor at a threshold voltage of 80 mV suggest a decrease in the iodine concentration in the CsPbI 3 film (Figure 7e,f), [80] which further proves that the low voltage switching behavior may originate from the generation of V I s under an applied electric field. However, the application of memristors in BCI is restricted due to the toxicity of lead halide perovskite materials. In this review, lead halide perovskite materials mainly provide a method or idea for the realization of bio-voltage memristor. Thus, it is necessary to further investigate the lead-free low bio-voltage memristors (such as Au/Rb 3 Bi 2 I 9 /Pt and Au/Cs 3 Bi 2 I 9 /Pt devices). [81]

Phase Change Type
Phase change type memristors can utilize the different conductive states of material phases to achieve memory. 2D phase change materials represented by MoS 2 have broad application prospects in many fields such as nanoelectronic and optoelectronics devices due to special energy band structure, semiconducting, or superconducting properties. MoS 2 has a stable 2H phase and a metastable 1T metallic phase. Cheng et al. successfully fabricated a MoS 2 nanosheet phase change type memristor with a bio-voltage of 100 mV by an intercalation-assisted exfoliation method, mainly owing to the existence of a metastable 1T metallic phase. Figure 7a shows bulk MoS 2 device with an ohmic feature. The inset of Figure 8a shows schematic diagram of Ag/bulk MoS 2 /Ag device and 2H phase with two layers per unit cell stack in the hexagonal symmetry with trigonal prismatic coordination of bulk MoS 2 . [82] Unlike the more stable and essentially dominant 2H phase of bulk MoS 2 , the metastable 1T metallic phase with monolayer unit cells in tetragonal symmetry with octahedral coordination of the exfoliated MoS 2 nanosheets device is observed in Figure 8b. [33] As can be seen in Figure 8c, the MoS 2 nanosheets device has resistive switching behavior. Figure 8d further reveals the Raman spectra of MoS 2 bulk and nanosheet. The Raman shift of E1g peak is 280 cm −1 ,  [83] Thus, these results demonstrate the existence of the 1T phase in the MoS 2 nanosheets.
As the 1T metallic phase MoS 2 is a metastable structure, only ultra-low voltages are required to induce lattice distortion of the Mo and S atoms. The octahedral coordination distortion structure of Mo atoms of 1T phase is proved by Cheng et al. through density functional theory calculations, which leads to the strong interaction of Mo d z 2 , d xy , and p z orbitals with S p z and p x orbitals to create the orbital hybridization, leading to valence and conduction bands overlap without gaps. [84] The electrons delocalize to stabilize the structure of 1T phase in MoS 2 nanosheets, resulting in its metallic character. Owing to the different metastable structures of the 1T phase in MoS 2 nanosheets, lattice distortion occurs when the Mo and S ions in the 1T phase are induced to be displaced by a positive voltage with 66 mV. This electric field-induced lattice distortion can www.advelectronicmat.de enhance the delocalization of electrons, leading to a significant increase in electrical conductivity, [33] so the device changes from LRS to HRS. Conversely, when the reverse voltage (−98 mV) is applied, Mo and S ions are redisplayed from the on state to the off state; the device transitions from LRS to HRS. Thus, the realization of the bio-voltage memristor is mainly attributed to the lattice distortion of the metastable 1T phase induced.

Artificial Synapse and Neuron Based on Bio-Voltage Memristors
Synapse and neuron are the basic units of how the brain learns and processes multi-dimensional information. Artificial synapse and neuron based on bio-voltage memristor can not only imitate the brain's efficient and low-power neural learning process but also facilitate the integrated interactions with living biological cells, which paves the way for the combination of artificial neuromorphic systems and biological neural networks.

Artificial Synapse
The bio-synapse that has the ability to adjust its own strength (synaptic weight) according to the information contained in the input action potentials, and maintain this change in strength even after the input action potentials have disappeared (synaptic plasticity), plays an indispensable role in the transmission of information between neurons (as illustrated in Figure 9a). [85][86][87] Thus, the reason why the brain can realize advanced functions such as learning and multiple memory is inseparable from synaptic plasticity. [3] In addition, synaptic devices based on biovoltage memristors can closely emulate the Ca 2+ dynamics for presynaptic and postsynaptic terminals of bio-synapse, which are the key elements for constructing efficient artificial neural networks for novel neuromorphic computing systems. [10,[88][89][90] Meanwhile, compared with previously reported artificial synapses, the characteristics of bio-voltage memristors are closer to the parameters of biological synapses in signal amplitude. [10] From the perspective of time, the changes in the connection strength of synaptic plasticity can be divided into short-term synaptic plasticity (STP) and long-term synaptic plasticity (LTP). [3,32]

STP
STP reflects that synaptic plasticity occurs in a short period of time, usually in the order of microseconds to a few minutes, and can be recovered rapidly in a short period of time, mainly including paired-pulse facilitation (PPF) and pairedpulse depression (PPD). Fu et al. demonstrated the bio-voltage www.advelectronicmat.de memristor with protein nanowire achieved PPF (conductance increase temporarily) and PPD (conductance decrease temporarily) behaviors at high frequency (900 Hz) and low frequency (10 Hz) with an amplitude of 100 mV, respectively (as show in Figure 9b). [10] The combined effect of PPF and PPD behaviors plays a crucial role in the transmission of information in the nervous systems. Therefore, it is of great significance to research STP in short-term memory (STM).

LTP
Conversely, LTP is generally thought to be the memory storage of information at synapses for hours or even days, involving long-term potentiation (LTPot) with increased synaptic weight and long-term depression (LTD) with decreased synaptic weight. To demonstrate LTPot and LTD functions, pulses response characteristics in bio-voltage memristor of Ag/BiOI/Pt device with different polarity were performed in Figure 9c. [91] Under a series of positive write pulses with an amplitude of 800 mV for 1 s, the current continuously increased, indicating a potentiating process. While a series of erase pulses was applied (−800 mV, 1 s), the current continued to decrease and the device was depressed. LTPot is essential for new learning and memory, while LTD is associated with the removal of memory. Thus, it is necessary to further study LTP in long-term memory (LTM).

Transition From STM to LTM
More interestingly, bio-voltage memristors can not only simulate the behaviors of bio-synapses but also implement model related to human memory and learning behaviors, proposed by Atkinson and Shiffrin in 1968. [34] Figure 9d shows the model consists of sensor memory (SM), STM, and LTM. Sensory receptors exposed to external stimuli convert the different forms of physical stimuli they detect into encodings of information. Information is first used as SM which is stored in sensory registers for a short period of time. STM is a temporary enhancement in response to external lowfrequency stimuli. The parts of information will be quickly forgotten after the stimulus is removed, but the selected information will be stored for LTM under the action of highfrequency stimulation, which is important for learning and memory. [34,86,87] The bio-voltage memristor with Ag/AgS 2 / Pt structure designed by Ohno et al. reveals the multistore process. [34] Before the formation of the STM, the SM first forms that conductance slightly increases during the pulse stimulation and then returns to its initial value immediately after pulse withdrawal. Next, the conversion of STM to LTM is by adjusting the inter-stimulus time (T = 20 and 2 s) and repeating stimulation with the same input pulse amplitude and width (V = 80 mV, W = 5 s), establishing a strong dependence on the learning frequency (interval time of input pulses), as shown in Figure 9e,f. [34] Figure 7. Iodine vacancy type bio-voltage memristors. a) I-V curves of the resistive switching behavior of Ag/CH 3 NH 3 PbI 3 /Pt memristors. Reproduced with permission. [78] Copyright 2016, Wiley-VCH. b) Physical mechanisms of Ag/CH 3 NH 3 PbI 3 /Ag memristors. c) SEM images of Ag/CH 3 NH 3 PbI 3 /Ag memristor at pristine state and LRS show the locations of the Ag electrodes and the formed Ag clusters by red and black dashed lines, respectively. d) Energy dispersive spectroscopy (EDS) spectra collected at labeled positions 1 and 2. Reproduced with permission. [79] Copyright 2017, Wiley-VCH. e) Schematic diagram and typical I-V curves of Ag/CsPbI 3 /Ag memristor. f) EDS of the CsPbI 3 -based memristor shows the intensity change of the iodine X-ray peak before and after SET. Reproduced with permission. [80] Copyright 2020, Springer Nature. www.advelectronicmat.de

Artificial Neuron
Biological neurons receive information from the pre-neuron through a large number of dendrites, and integrate the information to output to the post-neuron through axons to complete the integration and transmission of spatial information. The main electrical operations of neurons include integration and firing. [6] Neuronal membrane acts as capacitor dielectric layers that can hold charges generated by the summation of spatiotemporal currents within the cell body. [10] When the membrane potential (V m ) increases to be greater than the excitation threshold, the membrane potential will gradually decrease and return to the initial state after the neuron discharges. According to Q = C m × V m (Equation (1), C m is the membrane capacitance and Q is the net charge in a given cytosolic volume), we can deduce to m m m m (Equation (2), I denotes injection current and g m V m is the leaky current). Figure 9g shows schematic of a biological neuron integrating excitatory postsynaptic current (EPSC), indicating excitatory postsynaptic current approach linear and sub-linear postsynaptic temporal integration at high frequency (short-interval input spikes, Δt < C m /g m ) and low frequency (long-interval spikes, Δt > C m /g m ) in biological neurons, respectively. Temporal integration of protein nanowire-based bio-voltage memristors in artificial neurons is investigated by using biological action-potential-like pulse spikes (100 mV, 1 ms) and varying the frequency (or pulse interval) from low-frequency (50 Hz) to high-frequency (900 Hz) (Figure 9h,i). At certain frequencies, neurons spontaneously repolarize to HRS after integrating a certain number of spikes before firing. It is also observed that the spike number needed for neural firing had a stochastic distribution, similar to a biological neuron (Figure 9j). Moreover, Figure 9k reveals correlation between the average pulse number of excitatory postsynaptic potential (EPSP) signals required for neural firing and different frequencies. The number of pulses showed a nearly linear temporal integration at high frequencies (200-900 Hz), while at low frequencies (≤100 Hz), this trend deviated as a sub-linear summation. Such integration is similar to postsynaptic temporal integration in biological neurons. Therefore, the artificial neuron constructed by the bio-voltage memristor not only acts on the biological action potential but also exhibits a temporal integration similar to that of the biological neuron.

Applications of Bio-Voltage Memristors
Bio-voltage memristor can be used to simulate the functions of biological synapses and neurons. It has broad application prospects due to its own advantages such as low power consumption, high efficiency, and easy integration. Bio-voltage memristors can not only be used to build novel computing systems including neuromorphic computing and logic circuits but also to realize bioelectronic interface for bio-signal processing and wearable neuromorphic interface application. Reproduced with permission. [33] Copyright 2016, American Chemical Society. www.advelectronicmat.de

Neuromorphic Computing
As an indispensable member of constructing efficient neuromorphic computing systems, memristors have made important progress in multilayer perceptron such as convolutional neural network and recurrent neural networks (RNNs). [92,93] It realizes the computing in the applications of image processing, speech recognition, face recognition, and detection. Interestingly, bio-voltage memristors can build RNNs, enabling low-power reservoir computing (RC). RC derived from RNNs has been successfully used to realize applications such as dynamic system recognition, time series forecasting, and pattern generation. [94][95][96] Figure 10a shows schematic diagram of a bio-voltage memristorbased RC system for neural activity analysis, consisting of input Figure 9. The artificial synapse and neuron successfully deduced by bio-voltage memristor. a) The dynamic behavior of Ca + plays an indispensable role in the synaptic plasticity. b) STP with temporary changes of synaptic weight, including PPF and PPD. Reproduced with permission. [10] Copyright 2020, Springer Nature. c) LTP, involving LTPot with increased synaptic weight and LTD with decreased synaptic weight. Reproduced with permission. [91] Copyright 2022, Wiley-VCH. d) Synaptic operation-based learning and memory model proposed by Atkinson and Shiffrin. e,f) the conversion of STM to LTM can be realized by adjusting the inter-stimulus time. Reproduced with permission. [34] Copyright 2011, Springer Nature. g) Schematic of a biological neuron integrating EPSC. h,i) Temporal integration of protein nanowire-based bio-voltage memristors in LIF artificial neurons is investigated by using biological action-potential-like pulse spikes (100 mV, 1 ms) and varying the frequency from low-frequency (50 Hz) to high-frequency (900 Hz). j) The spike number needed for neural firing had a stochastic distribution, similar to a biological neuron. k) A correlation between pulse number and spike frequency. Reproduced with permission. [10] Copyright 2020, Springer Nature.
www.advelectronicmat.de layer, reservoir layer, and readout layer. [80] To start with, the input layer is responsible for directly firing memristors from neural spikes collected from firing neurons. Second, the concept of virtual nodes extends the reservoir space to facilitate handling complex temporal inputs. In the end, a simple ANN used as readout layer is responsible for producing the final output (recognizing four common neural firing patterns, including "Tonic", which corresponds to low-frequency spikes with a constant interval; "Bursting", which corresponds to groups of high-frequency spikes with a constant inter-group interval; "Irregular," which corresponds to irregularly fired spikes; and "Adapting," which corresponds to spikes with increased intervals). More importantly, memristor as reservoir plays a key role in RC systems for neural activity analysis. Zhu et al. also constructed a RC system with the possibility of real-time neural data analysis by utilizing dynamic bio-voltage CsPbI 3 memristors with extremely low threshold voltage and inherent short-term memory with shortterm memory effect. [80] To begin with, evolution of read current of four neural firing patterns for four different types of spike trains is illustrated in Figure 10b, suggesting that the bio-voltage CsPbI 3 memristors can be used as reservoirs for analyzing the spike trains. Then, a series of simulation and testing experiments is designed to test the ability of the RC system to analyze the neural firing patterns. 31 virtual nodes generated from a single memristor device are taken together to form the memory state and fed to a simple fully connected neural network. Thus, we only need to train the readout layer to achieve pattern recognition, mainly owing to the fact that temporal features in the input spike train can be transformed into features in the stored state. Experimentally obtained correlated recognition results of four neural firing patterns depicted in a false color confusion map, correspond to an overall recognition accuracy of ≈87.0% (as illustrated in Figure 10c,d). In addition, the memristor-based reservoir system has significantly higher recognition accuracy of simulated neural discharge patterns at different readout layer sizes than the integrated system that integrates input data over a Figure 10. Application of bio-voltage memristor for neuromorphic computing and logic circuit, respectively. a) Schematic diagram of a bio-voltage memristor-based RC system for neural activity analysis (four neural firing patterns including "Tonic," "Bursting," "Irregular," and "Adapting"). b,c) Evolution of read current of four different types of spike trains and experimentally obtained recognition results depicted in a false color confusion map of four neural firing patterns. d) The memristor-based reservoir system has significantly higher recognition accuracy of simulated neural discharge patterns at different readout layer sizes than the integrated system. e) Utilizing the bilayer CNN as the readout layer to analyze the firing pattern recognition. f,g) Transitional trains of four neural firing patterns, and the corresponding experimentally measured current responses and detected moments of the pattern transitions. h) Represented by color maps of experimental results from the RC system versus the manually labeled ground truth. Reproduced with permission. [80] Copyright 2020, Springer Nature. I-k) The circuit for the implementation of a SR latch of bio-voltage memristor. Reproduced with permission. [25] Copyright 2018, American Chemical Society.
www.advelectronicmat.de fixed time period (in Figure 10d). In addition to pattern recognition, real-time detection of emission pattern changes in stream pulse trains requires further research. [97] Utilizing a bilayer convolutional neural network (CNN) as the readout layer for firing pattern recognition is displayed in Figure 10e. In addition, the five output neurons contain the four firing patterns and a "transition" neuron that identifies transitions between any two of the four firing patterns. Figure 10f-h shows transitional trains of four neural firing patterns ("Tonic → Bursting," "Bursting → Irregular," and "Irregular → Adapting"), and the corresponding experimentally measured current responses and detected moments of the pattern transition. The experimental prediction results of the entire spike train from the RC system are basically consistent with the manually labeled ground truth results. These results the bio-voltage CsPbI 3 memristor-based RC system can monitor the evolving activity of neural activity in real time. However, the application of bio-voltage memristors to a large biological neural network still requires the development of new neural computing and optimized devices. Besides RC-based RNNs, biovoltage memristors can also play more important role by interacting with other biological neural networks.

Logic Circuit
Brain-inspired neural networks can provide a general architecture to efficiently implement various logical functions. Inspired by neural networks, through matching the circuit structure, memristor with excellent resistive switching behavior and nonvolatile memory characteristics has realized a variety of logical computing functions, including AND, OR, NOT, NOR, and IMP, which lays a foundation for the effective realization of memory and computing fusion. In particular, bio-voltage memristor has been successfully used to construct sequential logic circuits (SR latches), which is an important component of digital information processing systems. Biomaterial-based rDnaJ-6 bio-voltage memristor exhibits excellent bipolar nonvolatile switching properties at 120 mV or −80 mV in Figure 10i. [25] To research the SR latch circuit of NAND gate, two rDnaJ-6 bio-voltage memristors were connected to a common node, which connected a load resistor of 50 KΩ to ground, as depicted in Figure 10j. The SR latch included two input signals (S [set] and R [reset]) and two corresponding output signals (Q and Q ). As can be seen from the truth table, when both input signals S and R are low, the rDnaJ-6 latch output maintains its original value (defined as 0). Moreover, when one of the input signals is high but the other is low, the output Q is forced to assume a high (defined as 1) or low state, and then Q remains fixed until the signal applied to input signal S or R changes. Figure 10k confirms the use of rDnaJ-6 memristors in SR latches, promising to form a fundamental building block for future ultra-low power digital electronic technology.

Bio-Signal Processing
As electronic devices such as wearable sensors and intracellular biological probes typically record very small signals (sub-100 mV range), they need to be connected to communication amplifier circuits for amplification prior to signal processing. This pre-processing process will increase the power requirements for closed-loop bioelectronic systems. Unlike electronic devices that record physiological signals with high power consumption and complex integration, [98][99][100] bio-voltage memristors can directly provide the possibility of processing bio-signals, thereby satisfying bio-energy saving system requirements and reduced integration. Bio-signal processed by circuit of leaky integrate-and-fire (LIF) neuron with tunable integrateand-fire response integrates the bio-voltage memristor in an RC circuit with a parallel capacitor (C = 100 µA) and series resistors (R = 10 KΩ) are illustrated in Figure 11a. [10] The memristor that mimics post-neuron can integrate signals from pre-neuron by a capacitor. Once the capacitor's potential accumulates to a certain value, the memristor turns on and transitions to low resistance (R LRS ). When R LRS is considerably smaller than R, it will discharge the capacitor to lower V m . After a period of time similar to the "refractory period" in biological neuron after triggering, the next round of input pulses starts to be integrated. Therefore, it can be considered that the artificial bioreporter can monitor the changes of bio-signals according to the input frequency-dependent firing of artificial neuron. The pulse frequency of normal heart rate (1.16 Hz) failed to trigger artificial neuron firing, but the pulse frequency of abnormal heart rate (3 Hz) increased the charging rate and generated a membrane potential greater than the threshold, which successfully triggered neuronal firing (Figure 11b,c). Moreover, the pulse number required to trigger neuron firing at different frequencies is studied in Figure 11d. Thus, these results reveal that memristor-based artificial neurons at biological action potentials with the ability to process bio-signals on-site hold great potential for next-generation bioelectronic interactions.

Wearable Neuromorphic Interface
Integrated sensor and memristor-based artificial neuron to constitute front-tend afferent circuit and back-end execution system, forming wearable neuromorphic interfaces that can efficiently and intelligently process bio-stimulation to achieve intelligent responses. [101,102] Notably, the inherent amplitude mismatch between the conventional sensing signal and the computational signal makes it unsuitable at the biological level for wearable integrated interfaces. [55] Thus, bio-voltage signal processing is fundamental to unifying sensing and computing functions in biological systems. The resistance of LRS and HRS of a high-performance G. sulfurreducens protein nanowires device on a polyimide substrate did not cause considerable degradation at the different bending radius and after being bent 10 000 times, [55] which proves the potential of biovoltage memristors with high mechanical flexibility in wearable electronics due to thin device thicknesses that reduce bending stiffness and active layer of ideal flexibility. Meanwhile, Fu et al. reported that bio-voltage memristors could be integrated with a planar protein nanowire sensor to detect respiratory signals, a capacitor to modulate the membrane potential (V m ), and an LED to serve as a visual alert to abnormal respiratory rate to successfully construct a wearable neuromorphic interface, as www.advelectronicmat.de indicated in Figure 11e. [55] To begin with, simulation circuit of the artificial neuron in response to breath emulates normal or fast breath rate with low or high frequency (0.3 or 1 Hz) spiking input in Figure 11f. The V m balances at a low value of ≈15 mV below V th at normal breathing rates, thereby not triggering neuronal firing, and conversely, the breathing interval is shortened at fast breathing rates, resulting in a decrease in firing, thereby increasing the equilibrium V m . Neuron firing is triggered when the V m approaches V th . Neuron firing rapidly releases the V m below V th ; so, if the spiking input continues, the neuron can continue to fire. Then, in the experiment, the artificial neurons remained silent (V m of the artificial neurons was below 30 mV) at normal respiratory rates (0.3 Hz) in Figure 11g. In contrast, neuronal firings were triggered to cause LED illumination (V m was increased to ≈40 mV) upon abnormal respiration rates (1 Hz) in Figure 11h. As we can see, the experimental results about dynamic response in the neuromorphic interface are consistent with the predictions from the simulations, demonstrating the great potential of bio-voltage memristors in future self-sustained wearable neuromorphic interface applications.

Conclusion and Outlook
This review summarizes the progress of representative biovoltage memristors from three aspects: working mechanisms, artificial synapse and neuron functions, as well as novel applications. We focused on the detailed analysis of physical mechanisms of various types of bio-voltage memristors including active layer catalytic type, nanogap type, QWs type, and vdW interfaces type, heterostructures type, iodine vacancy type, and phase change type, which provides ideas and directions for the adjustment/reduction of the functional voltage of the memristor in future practical applications of biological interfaces. In addition, memristors successfully emulate typical artificial synaptic functions with STP, LTP, and the transition from STM to Figure 11. Application of bio-voltage memristor for bio-electronic interface with bio-signal processing and wearable neuromorphic interface. a) Bio-signal processed by circuit of LIF neuron integrates the bio-voltage memristor in an RC circuit with a parallel capacitor and series resistors. b,c) The pulse frequency of normal heart rate (1.16 Hz) failed to trigger artificial neuron firing, but the neuronal firing was successfully triggered when the pulse frequency was of abnormal heart rate (3 Hz). d) The pulse number required to trigger neuron firing at different frequencies. Reproduced with permission. [10] Copyright 2020, Springer Nature. e) Schematic and circuit diagram of integrated wearable interface, including G. sulfurreducens protein nanowires sensor, G. sulfurreducens protein nanowires memristor, and backend execution. f) Simulated evolution in the membrane potentials and currents from the artificial neuron when normal and abnormal respiration sensing signals are received. g) Membrane potential and corresponding current from artificial neurons when the G. sulfurreducens protein nanowires sensor exposed to normal breathing occurs. h) Membrane potential and corresponding current from artificial neurons to trigger a LED to light up as visual warning when abnormal breathing occurs. Reproduced with permission. [55] Copyright 2021, Springer Nature.
www.advelectronicmat.de LTM, which are the foundations for building efficient artificial neural networks for novel neuromorphic computing systems. Last, the applications of bio-voltage memristors in computing and bioelectronic neuromorphic interface application prospects are introduced. Despite research on biological voltage memristors having made phased progress, there are still some obstacles to be overcome to realize long-term stable and effective bioelectronic interactions. The relevant hurdles are as follows: 1. In terms of device mechanism, the corresponding relationship between the physical mechanisms and performances of bio-voltage devices is still unclear enough to realize their most suitable practical applications, which requires more indepth research. In addition, bio-voltage memristors also have severely non-ideal characteristics such as fluctuation, and noise and drift between devices and operation cycles, which limit the standard operating efficiency reduction of the entire artificial neural network. [103] Thus, it is necessary to improve the performances of the devices from cycle-to-cycle and device-to-device. 2. Artificial synapses and neurons are critical for future interactions of electronic devices with biological neural networks. The realization of large-scale bio-compatible artificial neural networks requires a comprehensive consideration of the relevant factors involved to obtain low-power artificial synapses (spike time-dependent plasticity and spike rate-dependent plasticity) and neurons (Hodgkin-Huxley neurons, stochastic neurons, oscillatory neurons, and LIF neurons) that match the bio-voltage signal. In addition, the actual memristors in terms of synaptic conductance modulation are usually nonlinear and asymmetric. [103,104] At present, there are a few studies on the retention characteristics of bio-voltage memristor, which are worthy of further study. Bio-voltage memristor is faced with the basic dilemma of the barrier for retention (non-ideal characteristics) when the switching voltage is very low. These problems can be studied from device level to circuit system level and algorithm level. First, improving device reliability can be the option of introducing optimized programming pulses or additional layers/dopants that allow rapid linear update of conductance in the same pulse driver. [104] Then, memristor and transistor are integrated to form a 1-transistor-1-memristor (1T1R) structure. [103] Ultimately, optimization algorithms (for example, the hybrid training techniques used to emphasize training the weight of the fully connected layer) can compensate for the non-ideal characteristics of existing device imperfection. [93] 3. In bio-systems, a large number of neurons will be apoptosis under the influence of various inducing factors such as pathological changes or physical damage, which in turn leads to synaptic damage and loss of connections, ultimately causing the loss of memory ability in the brain. Therefore, it is very necessary to study the research of bio-voltage memristors in physical transient electronics that successfully emulates the death phenomenon of biological neurons and synapses, which greatly promotes the application of bio-voltage memristors in defense security information storage technology and integration in biological interface-intelligence augmentation/medicine. In addition, how to balance the contradiction between the stability and degradation characteristics of physical transient bio-voltage memristors is also urgently needed to be explored. 4. Large-scale integration of memristors is the basis of applications. [103] Crosstalk in the cross-bar array is the vital factor that restricts the scale of integration. Choosing transistor to integrate with memristor is an effective way to suppress sneak path issue of the cross-bar array and improve the integration. However, relatively low set and reset voltages seem to be detrimental to memristor applications. Therefore, how to realize the integration of bio-voltage memristor and bio-voltage transistor to meet the voltage matching is the key of future work. 5. How to realize the integrated interaction between neuromorphic devices and biological interfaces is the key to the development of biologically integrated neuromorphic systems, such as biocompatibility. If the biocompatibility of the brain electrodes is not compatible, the collection intensity of the collected bio-signals will be affected. To improve biocompatibility, it is important to seek suitable materials.
With the rapid development of emerging information technologies such as artificial intelligence and brain-computer interfaces, the demand for data storage and computing continues to increase. Memristors will play an increasingly important role in various research fields in the future. Although the application development of bio-voltage memristors is still in its infancy, if the above problems can be improved, our research efforts will make an important contribution to the fusion of low-power artificial neuromorphic and biological neural networks.