Review of solid-state proton devices for neuromorphic information processing

This is a review of proton devices for neuromorphic information processing. While solid-state devices utilizing various ions have been widely studied for non-volatile memory, the proton, which is the smallest ion, has been relatively overlooked despite its advantage of being able to move through various solids at RT. With this advantage, it should be possible to control proton kinetics not only for fast analog memory function, but also for real-time neuromorphic information processing in the same time scale as humans. Here, after briefing the neuromorphic concept and the basic proton behavior in solid-state devices, we review the proton devices that have been reported so far, classifying them according to their device structures. The benchmark clearly shows the time scales of proton relaxation ranges from several milliseconds to hundreds of seconds, and completely match the time scales for expected neuromorphic functions. The incorporation of proton degrees of freedom in electronic devices will also facilitate access to electrochemical phenomena and subsequent phase transitions, showing great promise for neuromorphic information processing in the real-time and highly interactive edge devices.


Introduction 1.Neuromorphic information processing
Neural circuits and CMOS digital circuits are the two most sophisticated information processing hardware.Neural circuits excel at processing information from the surrounding environment in real time with low power consumption, while CMOS digital circuits excel at processing large amounts of recorded data at high speed.Although they are complementary in a sense, the understanding of neural circuits is currently limited, and their technology is not yet fully utilized.It is desirable to understand the characteristics of neural circuits in comparison with CMOS digital circuits, and to build technologies that exploit them.
CMOS digital circuits are significantly versatile information processing hardware.The basic constituents are simply PMOS transistors and NMOS transistors, and hence, a largescale circuit can be designed in an automated manner.In fact, transistor integration continues to evolve exponentially, with the number of transistors on a single chip exceeding 10 billion.CMOS digital circuits are fundamentally fast as well since transistors are operated by high-mobility electrons.They can move circuits at a high frequency of clock signals and process large amounts of data at high speeds.Armed with this number and speed of transistors, CMOS digital circuits have achieved great versatility; a given task is converted into a huge number of logical expressions, but CMOS digital circuits can readily handle them due to their number and speed.
Because of the superiority of CMOS digital circuits, it is not considered promising to try to replace what CMOS digital can do with different hardware.In fact, spintronics has pioneered rebooting computers with its non-volatility, which CMOS digital cannot do.Quantum annealers target optimization problems, which CMOS digital is not good at.InGaZnO and organic electronics have been applied to displays with their transparency and flexibility, which CMOS digital cannot achieve.Neuromorphic electronics, which mimic some aspects of neural circuits, will also need to clarify functions that cannot be realized with CMOS digital technology.Fortunately, neural circuits are expected to have many functions that cannot be realized by CMOS digital circuits.Such functions may include the nonlinear dynamics of neurons, synapses, and their networks, 1) the real-time capability to control motor organs while processing the body's myriad sensor information in parallel, 2) and the mechanisms that enable the integration of a wide variety of unreliable devices into a whole system, 3) although we don't have much understanding on each of these issues.Currently, efforts toward hardware design are being made to mimic the characteristics of neurons and synapses with digital or analog circuits 4) and sometimes even material properties, 5) and to construct large-scale networks based on them. 6)As possible clues for such hardware design, here, we will focus on two unique features of neural circuits, which are unavailable in the CMOS digital circuits.
Unlike CMOS digital circuits, neural circuits are slow.This is because neural circuits utilize molecules and ions to process information.This slowness may seem like a disadvantage in terms of CMOS digital circuits, but it is rather necessary for neural circuits.In fact, the slowness allows neural circuits to operate on the same time scale as the surrounding environment, 7) allowing for real-time operation with high efficiency (Fig. 1).Slow motion is hard to be generated by scalable CMOS elements because both resistive and capacitive elements available on a Si chip are severely limited.Slowness may sound primitive but is a unique feature of neural circuits, and is expected to show its true value in edge information processing that heavily interacts with the surrounding environment in real time with minimum power consumption.
Another important feature of neural circuits is its diversity.For example, there is a wide variety of ions and molecules used in neurons and synapses, and the firing patterns of neurons 8) or the learning rules of synapses 9) differ from one part to another in the neural circuit.As opposed to uniform digital circuits, this diversity of neural circuits has been developed bottom-up during evolution, and seems to be the source of the functional emergence in neural circuits.Unfortunately, neither the technology to fabricate devices using a variety of ions and molecules nor the technology to construct large-scale systems using such diverse devices currently exists.However, at the very least, the introduction of an ionic degrees of freedom into electron-based uniform hardware will be important in the design of hardware that mimics neural circuits in the future.In fact, it is expected that various physical phenomena such as electrochemical reactions and subsequent phase transitions will become available by utilizing ionic degrees of freedom.

Proton devices
In light of these two features of neural circuits, devices that use ions in addition to electrons are considered important for future research on neuromorphic information processing hardware.As such an ion-based device, non-volatile memory has already been extensively researched. 10)Oxide devices using oxygen ions, atomic switches using copper and silver ions, and Li-ion devices derived from batteries are representative examples.However, in general, ions are much larger than electrons, and only a limited number of materials are capable of transporting ions at RT.As a result, only certain materials can be used in these ionic devices, and the basic device design of building functionality through the stacked structure of different materials is not possible.In addition, Li ions are particularly problematic in terms of contamination, making them difficult to mount on a Si chip.
In this sense, the proton are the ions that have been relatively overlooked.Since they have no inner-shell electrons, they are invisible to photoelectron spectroscopy and energy dispersive X-ray spectroscopy.Furthermore, they are often indistinguishable from hydrogen in surface adsorbed water, making them difficult to measure in the first place.However, they are by far the smallest, most invasive ions that can travel at RT in a variety of materials, and can take full advantage of the stacked structures of different materials.In fact, protons are known to be solid-soluble in a number of metals 11) and to diffuse and dope electrons in various oxides. 12)In recent years, more important roles of protons have been disclosed, such as converting vanadate and nickelate from electronic conductors to insulators, 13,14) acting as negative charges as hydride ions, 15) and playing an important role in the stability of InGaZnO. 16)hus, proton devices will provide an important basis for exploiting features of neural circuits in information processing hardware.In fact, by using protons, slow time constants that could not be achieved with electrons are expected to be realized with scaled devices, which will be important for edge information processing.In addition, since protons can penetrate into various materials and change their properties, the concert of protons and electrons in a stacked structure can enable a variety of device functions.Because of the promise of protons, this review summarizes the proton devices that have been reported so far, classifying them according to their device structures.Since most of the literature presented here uses protons to mimic the behavior of synapses and neurons, the functions of biological synapses and neurons are also summarized in the beginning.The way of thinking about protons in stacked structure is also briefly introduced before reviewing proton devices.While the similar concept was previously introduced in the conference abstract by the same authors, 17) this review thoroughly covers the background and the related works on proton devices.Currently, it is not clear how mimicking synaptic and neuronal function with material properties will lead to future information processing 030801-2 © 2024 The Author(s).Published on behalf of The Japan Society of Applied Physics by IOP Publishing Ltd hardware beyond CMOS digital circuits.However, the proton technology developed there will form an important basis for future neuromorphic hardware.

Synapse
A synapse essentially transmits information from the axon of a presynaptic neuron to the dendrite of a postsynaptic neuron [Fig.2(a)].Neurotransmitters are stored in numerous synaptic vesicles at the presynaptic axon tip, and each time a firing signal is transmitted to the axon, the neurotransmitter is released from the cell.Two major synapse-induced phenomena are the altering of the postsynaptic potential of the dendrite, and the altering of the signal transmission to the postsynaptic potential [Fig.2(b)].
The change of the postsynaptic potential is due to the opening and closing of ion channels on the postsynaptic dendrite side in response to neurotransmitters released from the axon side.The increase in postsynaptic potential is called excitatory post-synaptic potential, and the decrease in postsynaptic potential is called inhibitory post-synaptic potential.Depending on the type of ion channel and its mechanism, there are fast post-synaptic potential lasting from several ms to several tens of ms and slow post-synaptic potential lasting from several tens of seconds to several minutes. 17)The fast post-synaptic potential is thought to be used for signal transduction, while the slow potential change is thought to be used for short-term memory etc.
Changes in the synaptic transmission can be divided into short-term plasticity (<1 h) and long-term plasticity (>1 h) as shown in Fig. 2(c).Short-term plasticity is thought to play an important role in signal transduction and temporal information processing, while long-term plasticity is thought to play an important role in learning.Short-term plasticity includes facilitation for less than 1 s and potentiation for longer than several tens of seconds. 17)The short-term plasticity is caused by various reasons such as neurotransmitter depletion and slow recovery of Ca concentration on the presynaptic axon side.The short-term facilitation is responsible for paired pulse facilitation (PPF), in which the effect of a second input spike is amplified when two spike signals are continuously input.On the other hand, longterm plasticity is often caused by the dendritic side and is attributed to Ca channels that only works under high postsynaptic potential as shown in Fig. 2(d) (details are below). 18)This mechanism explains long-term potentiation (LTP), in which the long-term increase in synaptic transmission is induced by HF input spikes, and spike time dependent plasticity (STDP), in which the long-term change in synaptic transmission is induced in response to the relative timings of firing between pre-and postsynaptic neurons.For example, in the case of LTP, continuous input of HF spiking signals results in high postsynaptic potentials and activation of Ca channels in dendrites [bottom left in Fig. 2(d)].As a result, input spike signals from axons cause dendritic Ca concentrations to increase [bottom center in Fig. 2(d)], which creates new ion channels and leads to a long-term increase in the synaptic transmission [bottom right in Fig. 2(d)].In the case of STDP, when dendrites fire, the postsynaptic potentials increase, and Ca channels are activated.Then, spike signals from axons can open the Ca channels and increase Ca concentration, and hence, increases the number of ion channels in dendrites.

Neuron
The fundamental role of neurons is to integrate information, in other words, converging input information in time and space, and diverging it from the axon to hundreds of neurons.Specifically, input signals change the membrane potential, and when the membrane potential reaches a threshold, the neuron generates a spike signal of its own.Unlike synapses, the membrane potential of a neuron is reset when it outputs a spike signal.In this respect, neurons are essentially processing units and do not have a memory function to hold information for short or long periods of time, as is the case with synapses.
The basic behavior of a neuron is described by the cell membrane as an electrical capacitor and the opening and closing of K + and Na + ion channels (Fig. 3). 18)In steadystate neurons, the Ka + concentration inside the membrane is high and the Na + concentration is low due to the Ka + -Na + ion pump.The Ka + channels keep slightly open, and as a result, Ka + ions leak from inside to outside, leading to the inside of the membrane negatively charged [Fig.3(a)].When the membrane potential reaches a threshold under the influence of a synaptic input signal, the Na + channel opens wide and Na + flows into the membrane, resulting in a positive membrane potential [Fig.3(b)].Due to this positive membrane potential, the Ka + channel opens wide, and Ka + flows out of the membrane, resulting in the inside of the membrane negatively charged again [Fig.3(c)].This process of alternating opening of Na + and K + channels generates a spike voltage as shown in Fig. 3(d).This mechanism has often been reproduced in CMOS analog circuits using a PMOS transistor as the Na + channel and a NMOS transistor as the K + channel [Fig.3(e)]. 19)he above is the simplified mechanism by which neurons fire once, but on a longer time scale, individual neurons show their own firing patterns depending on the frequency and intensity of continuous input signals.To begin with, neurons accumulate the input signal as a membrane potential, but as mentioned above, there is also K + leakage, so the input signal is not purely integrated but rather averaged over a short period of time by leaky integration.When the average value of the input signal exceeds the threshold value, neurons start to fire in a characteristic pattern.This firing pattern shows a large diversity such as monotonous regular spiking, sudden bursts of continuous firing, and so on. 8)While regular spiking is explained by a simple leaky integrate fire (LIF) model, understanding other firing patterns requires a neuronal model that accounts for the dynamics of each ion channel. 20)A wellknown neuron model is the Hodgkin Huxley model, 21) which accurately describes ion channel dynamics in terms of fourvariable nonlinear differential equations.However, the firing process is complicated and difficult to simulate, so the Izhikevich model, 8) which is an algorithmic rewriting of the firing process, is often used in simulations.Accurate Hodgkin Huxley models are useful in scientific research to understand the mechanisms of neural circuits.LIF and Izhikevich models, on the other hand, are more practical from an engineering perspective especially when trying to simulate functionality in the network as a whole.

Basic proton characteristics
The basic behavior of protons in solid thin films is introduced in this chapter by taking anatase TiO 2 polycrystalline films on Si substrates for example. 22)First of all, protons can be doped to TiO 2 thin films in various ways: thermal annealing in a hydrogen atmosphere, spillover using noble metal catalysts, and electrochemical reduction in an aqueous solution.The electrochemical reduction is especially simple.When an Al electrode is deposited at the edge of a TiO 2 thin film and immersed in an acid or alkaline solution at RT for several tens of seconds, the Al is oxidized and the TiO 2 is reduced, resulting in the entire TiO 2 film being doped with protons [Fig.4(a)].
Proton doping changes a TiO 2 thin film from an insulator to a conductor.This is because hydrogen dopes electrons into TiO 2 as a donor impurity.Secondary ion mass spectrometry (SIMS) shows that the doped hydrogen is mainly located near the lower interface with SiO 2 as shown in Fig. 4(b).As with electrons, the proton distribution is strongly affected by the interface.
Due to the high mobility of protons in crystals, unlike other ions, grain boundaries are not necessarily a major diffusion pathway of protons.In fact, when protons are electrochemically doped into TiO 2 thin films, the proton doping area can be defined well by the resist mask [Figs.4(c) and 4(d)].Here, the grain boundaries were found to have no significant impact on the in-plane distribution of doped protons.This means, when protons are doped in aqueous solution, most of them enter from the grain surface and diffuse inside the grain toward the interface.Protons of about 6 × 10 13 cm −2 sheet concentration around the interface are relatively stable.In fact, they are slowly released from the TiO 2 surface with a time constant in the order of 100 days at RT. On the other hand, it was also found that the bulk part of the TiO 2 thin film (40 nm thickness) can store ten times as many protons as the interface in sheet concentration. 23)This large number of protons can be observed only when we dope protons through a few-nm SiO 2 cap on the surface of the TiO 2 film to temporarily prevent the release of protons from the surface [Fig.4(e)].These protons inside TiO 2 slowly leaves the SiO 2 cap in about 1000 s where the diffusion inside the SiO 2 cap is the rate-limiting process [Fig.4(f)].Finally, only protons around the interface are left.
Thus, TiO 2 film is a container for protons, SiO 2 is a proton cap, and the interface works as a proton trap.By clarifying the role of each material and structure in this way, it will be possible to design more complex proton devices based on stacked structures.

Structure of protonic synaptic devices
Broadly, protonic synaptic device can be categorized into two types depending on number of terminals either three or two terminals.Three terminal devices are mostly studied owing to the advantage of additional terminal to realize learning operation and signal transmission process simultaneously through the gate terminal and channel respectively.On the other hand, fabrication of three terminal devices is complicated and less scalable than that for two terminal devices.In this perspective, two-terminal devices are advantageous owing to their simple structure and scalability.
For a conventional three terminal device, either a bottomgate or a top-gate configuration is always adopted. Figure 5(a) shows a schematic image of a top-gate transistor, where the source and drain electrodes are buried below the dielectric and the semiconductor channel.][34][35][36][37][38] As shown in Fig. 5 On the other hand, two-terminal devices are simpler, where two electrodes are coupled to each other either vertically or laterally through a proton conducting layer.In a typical twoterminal vertical device [Fig.][41][42] In this case, either top electrode or a dedicated layer (placed between top electrode and proton conductor) may act as a source for proton reservoir.By applying electric field between two terminals, migration of protons can be controlled inside proton conducting layer.As for the proton conducting layer, a mixed conductor (electronic and ionic) is sometimes used.In this case, the proton motion modifies the electronic conductance between the two terminals.Similarly, in case of two-terminal lateral structure [Fig.44]

Protonic devices in the past literature
As explained in the last chapter, the protonic devices in the past literature can be categorized into four: three-terminal vertical, three-terminal lateral, two-terminal vertical, and twoterminal lateral.The choice of material for electrodes, dielectric and channel strongly affect the output results.Here we will cite some of the cases studied so far.The more detailed literature based on three-terminal and twoterminal devices are summarized in Tables I and II respectively.

Three terminal vertical devices
Yao et al. 24) demonstrated a three-terminal vertical protonic solid-state electrochemical synapse for physical neural networks as shown in Fig. 6(a).In this case Pd was used as top gate and gold was used as source and drain terminal.Nafion electrolyte of 300-400 nm thickness is used as the proton source and the proton conductor.A 50 nm thick WO 3 layer was used as a proton-sensitive channel.The synapse design relies on a charge-controlled mechanism, modulated electrochemically in solid-state.The device operates by shuffling protons in a three-terminal configuration and demonstrates a continuum of resistance states over seven orders of magnitude.As represented in Fig. 6 24) is a current controlled one, Burgt et al. 25) developed a voltage controlled three-terminal non-volatile electrochemical neuromorphic organic device (ENODe), which works at a low-voltage.The basic structure of the device [Fig.6(d)] consists of a poly (3,4ethylenedioxythiophene): polystyrene sulfonate (PEDOT:PSS) film partially reduced with poly(ethylenimine) (PEI) as a postsynapse (channel) and a PEDOT:PSS layer as pre-synapse (gate).Both layers form a sandwich structure having a central electrolyte layer.Cations (protons) flow from the presynaptic electrode into the postsynaptic electrode through the electrolyte, resulting in protonation of the PEI, at the same time electrons flow through the external circuit.The device has the capability to switch at low voltage and energy, displaying over 500 distinct, non-volatile conductance states within 1 V range.It achieves high classification accuracy when implemented in neural network simulations.The plastic ENODes are also fabricated on flexible substrates, making them compatible with three-dimensional architectures, and enabling the integration of neuromorphic functionality in stretchable electronic systems.Presynaptic voltage pulse could result in five discrete post synaptic non-volatile conductance states which was represented in Fig. 6(e).The application of 500 voltage pulses could result 500 distinct conductance states [as shown in Fig. 6(f)] which is promising for high density nonvolatile applications.In contrast to the non-volatile property, the volatile property was also observed for sub-threshold

PROGRESS REVIEW
potentiation, where the function of PPF is established in this ENODe device as demonstrated in Fig. 6(g).

Three terminal lateral devices
An example of three terminal in-plane device is demonstrated by Liu et al., 34) as shown in Figs.7(a) and 7(b).A chitosan membrane of 100 μm thickness is used as a dielectric proton source and Indium-zinc-oxide (IZO) as a channel material having 20 nm thickness and 30 μm width.Excitatory postsynaptic current (EPSC) triggered by different numbers of gate pulses (4.0 V, 10 ms) has been demonstrated in Fig. 7(c).
A transition from a short-term memory to long-term memory was mimicked in this freestanding artificial synapse.Proton migration and EDL electrostatic modulation are responsible for the rise in channel current when the gate pulse voltage is low.Reversing the electrostatic modulation process is possible by removing the gate voltage.Protons in the chitosan can enter the IZO channel when the gate pulse voltage is high enough, allowing for the electrochemical doping of the IZO channel layer.Thus, the synaptic transistors appear to have a non-volatile memory.The schematic diagram for PPF measurement is displayed in Fig. 7(d).The EPSC caused by two presynaptic spikes with a pulse interval (Δt) of 50 ms is seen in Fig. 7(e).On the presynaptic (IZO gate) input, two consecutive presynaptic spikes (1.5 V, 50 ms) were applied in quick succession.
When compared to the first presynaptic spike, the second one causes a bigger EPSC.It is not possible for some of the protons that were activated by the first spike to diffuse back to their equilibrium position when the interval is less than the mobile proton's relaxation duration.Next, at the channel/ chitosan contact, the EDL modulation of the remaining protons facilitates the channel current.The PPF is plotted against the time interval between the paired pulses in Fig. 7(f).A maximum PPF of around 222% is achieved when Δt = 10 ms.As Δt increases, the PPF steadily diminishes.

Two terminal vertical devices
A two terminal vertical protonic memristive device was reported by Wang et al. 39) as shown in Fig. 8 illustrates memory cycling.The I SD spike that results from the RESET V SD pulse has the same period of time and size as the I SD spike that results from the ON V SD pulse.This indicates that a certain amount of hydrogen is transferred between the source and drain contacts resulting a H + current in the Nafion (I SD ).Applying a 2 Hz, 1 V sine wave to the protonic device and cycling it confirms the device's features, demonstrating a distinct hysteresis between the ON and OFF states [Fig.9(f)].

Timescales of proton devices
From the benchmarks in Tables I and II, it can be seen that proton relaxation generally has a timescale of milliseconds to several hundred seconds.Figure 1 summarizes these time scales and shows that different active materials yield different proton relaxation time scales.It can be seen that amorphous WO 3 and amorphous MoO 3 have long time constants of several seconds or longer, (In)ZnO 2 have time constants of several seconds to tens of ms, and PEDOT-based materials have time constants of several tens of ms to several ms, approximately speaking.And most importantly, the time domains covered by these proton devices are perfectly consistent with the time scales of human activity.On the other hand, electrons are overwhelmingly faster, and even with the full use of sub-threshold currents in transistors, the time scale that can be realized with on-chip scalable electronic devices is 100 ms or shorter. 46)Thus, it can be said that the proton is inherently more suitable than the electron in terms of time scale for the information processing required in the human living environment.
It should be noted that some past literature on threeterminal vertical devices has used proton control with pulse voltages much shorter than the time scale shown in the benchmark of Tables I and II.For example, as shown in Fig. 6(c), 24) the pulse voltage used was 5 ms wide while the proton relaxation time constant was 50-100 s.In Ref. 47, an ultrashort pulse voltage of 5 ns is applied to the Pd/PSG (Phosphorous Silicate Glass) gate to modulate the electronic conductivity of the polycrystalline WO 3 channel.Where did this discrepancy in time scales come from?The pulse widths in these past three-terminal devices correspond to the delay time of the electrical conductivity modulation, which is different from the proton relaxation time scales as shown in Fig. 10.The proton relaxation time is the time constant for the relaxation to the final state at the applied voltage, which is long as summarized in the benchmark.Also, in many cases, conductance changes on this time scale are volatile, as is the plasticity of biological synapses.On the other hand, the delay time for conductance modulation is the time required to modulate conductance over a small range compared to the relaxation phenomena described above, and is much shorter than the relaxation time.Also, on short time scales, proton  relaxation is almost negligible, so it is considered a nonvolatile memory, and the relaxation phenomenon is referred to as the retention property.Thus, the relaxation time is long because it corresponds to the entire relaxation phenomenon to the final state, and the delay time is short because it corresponds to only a small amount of time at the beginning of the relaxation phenomenon (Fig. 10).There are also several other reasons for the short pulse widths in the past literature: 24,47) only a small change in conductivity is needed in a single pulse, only the channel surface needs to be modulated in three-terminal devices, and the pulse voltage used is relatively high (10 V for Ref. 47).When considering the proton function in information processing, delay time and relaxation time have very different meanings.As for the delay time in non-volatile memory operated with protons, the advantage over electronic devices is the analog nonvolatility of the protons, not the delay time.Even if we try to use this proton delay time for temporal information processing, the time scale of this delay overlaps with the time scale of the electronic device, and has no practical value.On the other hand, when the proton relaxation phenomenon is used for real-time information processing, the long relaxation time itself becomes an advantage over electronic devices.As shown in Fig. 1, long relaxation time is important for human-related information processing, and in fact, the neural circuits of living organisms utilize the relaxation phenomenon of synapse plasticity, as shown in Fig. 2(c), in various types of information processing.In the field of engineering, it is also known that such a relaxation phenomenon can be utilized as a short-term memory for realtime information processing using echo state networks. 48) Conclusions Some of the ions, including protons, are in the unique position of having high mobility and penetration ability in solid materials, but not as much as electrons.The moderate timescales of these ions allow the biological neural circuits to efficiently perform the information processing necessary in their living environment.Therefore, it is thought that useful edge information processing can be realized by combining protons, the smallest ions, with neuromorphic technology.In fact, as shown in this review, the time constant of proton relaxation can be modulated from a few milliseconds to several hundred seconds by changing the active material, and these timescales completely match those of the human living environment.In the future, proton devices may functionalize various electrochemical reactions and phase transitions, which were inaccessible in conventional electronic devices, and the mimicry of neural circuits will be an important guideline in connecting those functions to useful information processing.Fig. 10.The long timescale of proton relaxation, which is the focus of this paper, contrasts with the short timescale of proton modulation, which is mentioned in previous papers.24,47)

Fig. 1 .
Fig. 1.Summary of time scales in various physical phenomena, and the types of computing (digital or neuromorphic) that are suitable for handling them.

Fig. 2 .
Fig. 2. (a) A schematic illustration of synapse between the axon and the dendrite.(b) Two major synapse-induced phenomena: the altering of the postsynaptic potential of the dendrite, and the altering of the signal transmission to the postsynaptic potential.(c) The long-term plasticity and the short-term plasticity (potentiation and facilitation), and their time scale difference.(d) The mechanism of LTP and STDP based on the Ca channel activation in the dendrite.

Fig. 3 .
Fig. 3. (a)-(c) The equivalent circuits of a neuron (a) during the leaky-integrate phase, (b) at the threshold voltage, and (c) after initialized.The capacitor corresponds to the postsynaptic membrane, the current source corresponds to the input signal from the presynaptic axon, the voltage source corresponds to the concentration cell for each ion, and the resistance corresponds to the ionic resistance for each ion channel.(d) The corresponding membrane potential (V M ).(e) An analog neuron circuit where K + and Na + channels are implemented by NMOS and PMOS FETs, respectively.

Fig. 4 .
Fig. 4. (a) A schematic illustration of proton doping to TiO 2 via electrochemical reduction in an aqueous solution.(b) A SIMS profile of the TiO 2 thin film on the SiO 2 /Si substrate, where the sample was doped in D 2 O and rinsed in H 2 O to distinguish doped D from surface adsorption water.(c), (d) Mappings of current and corresponding morphology on the TiO 2 surface, where the proton is doped in a limited area via photolithography.White arrows in (d) show the grain boundaries.(e) Sheet carrier density (N 2D ) of the 40 nm doped TiO 2 thin films as a function of SiO 2 cap thickness.The dopes samples were dried, exposed to the air, and Hall measurements were performed at three different times after doping (10 min, 3 h, and 6 h).(f) The corresponding sheet conductance as a function of time after proton doping.Panels (a) and (b) are adapted from Ref. 22.The other panels are adapted from Ref. 23 (license 5678521059654).
(b), the gate voltage can be directly coupled to the semiconductor channel laterally by only one or more lateral electric-double-layer (EDL) capacitor.In this way, the in-plane drain current can be varied by the variation of gate voltage.In three-terminal structures, either gate or dielectric material acts as the source of proton reservoir.Upon application of gate voltage, protons start to migrate towards channel material and make electrochemical changes of the material resulting in the large change of electronic conductance in the channel.The change of polarity of bias at the gate electrode could result in the change of direction of motion of protons inside the proton conducting layer resulting in the reverse change of electronic conductance.

Fig. 5 .
Fig. 5. Schematic illustration of (a) three terminal vertical (b) three terminal lateral (c) two terminal vertical and (d) two terminal lateral protonic device structures.

Fig. 6 .
Fig. 6.(a) Diagram showing the structure of the protonic electrochemical synapse device.Where, R represents the layer that acts as a solid-state hydrogen reservoir and gate electrode (PdH x ).E is the solid-state electrolyte layer that conducts protons (nafion).A is the conducting channel, or active switching layer (WO 3 ).G is the conductance of the channel between the drain and source electrodes.(b) Electrical conductivity and open circuit potential of channel WO 3 versus gate PdH x as a function of the hydrogen content (x) in H x WO 3 .(c) Channel conductance modulation by positive and negative current pulses (0.5 μA, 5 ms), which lead to potentiation (protonation) and depression (deprotonation), respectively.The different curves correspond to different pulse numbers during potentiation or depression (10, 20, 50, and 100).(d) Another schematic representation of three terminal vertical device structure, where pre-synaptic layer (PEDOT:PSS) and postsynaptic layer (PEDOT:PSS/PEI) are separated by a nafion electrolyte layer transporting cations (protons).(e) The postsynaptic electrode's conductance G displays consistent, non-volatile switching between five distinct states by varying the presynaptic voltage V pre .(f) When the device is regulated with voltage pulses, LTP and depression display 500 distinct states over the operational range.A zoomed-in view highlighting each state is included in the inset.(g) Short-term potentiation and PPF.The interval between two brief pulses determines how much the synaptic weight is momentarily altered.Two distinctive timeframes are obtained by applying an exponential fit.A schematic of how such biasing is usually applied is included in the inset.Panels (a), (b), and (c) of the figure are adapted from Ref. 24.Other panels (d), (e), (f), (g) are adapted from Ref. 25 (License Number: 5693671405524).

Fig. 7 .
Fig. 7. (a)-(b) Schematic images of three-terminal laterally coupled flexible synaptic transistors on freestanding chitosan membranes with IZO as the channel material (c) EPSC measurement with variation of number of gate pulses (d) the schematic diagram of paired-pulse facilitation (PPF) measurement (e) two presynaptic spikes with Δt pre = 50 ms and an IZO channel current reading voltage (V DS ) of 0.5 V caused the EPSCs.The amplitudes of the first and second EPSCs are denoted as A1 and A2, respectively.(f) The PPF index, defined as 100% × A2/A1, is displayed as a function of the inter-spike interval, Δt pre , which is the time interval between two consecutive presynaptic spikes.All panels in the figure are adapted from Ref. 34 (Order License ID:1429940-2).

Fig. 9 .
Fig. 9. (a) Schematic representation of signal transmission across biological synapses.The binding of released neurotransmitters causes the ion channels of orange neurotransmitter receptors to open.(b) Schematic representation of two-terminal protonic device laterally coupled between two PdH x contacts which were separated by proton conducting Nafion.When a voltage (V SD ) is applied an H + current (I SD ) flows.The protonic device is in depression when this H + current depletes the PdH x hydrogen source, forming Pd (which is not proton conducting).(c) The two terminal protonic device having source and drain of 30 μm width with 1 μm gap was seen under a microscope (d) output current (I SD ) and input voltage (V SD ) were plotted against time, where a current spike behavior at V SD = 1 V was evidenced.(e) An equivalent two terminal lateral device was used to demonstrate ON and OFF switching.Three positive SET pulses (V SD = 1.25 V, 0.25 s) and one negative RESET pulse (V SD = -1.25 V, 0.25 s) were applied.(i) Hysteresis behavior can be noticed in the I-V curves for this two-terminal device of PdH x -Nafion system.All panels in the figure are adapted from Ref. 45 (Order License ID: 1429940-3).

Table I .
Summary of three-terminal protonic neuromorphic devices in terms of device type, materials, doping method, proton relaxation time constant, and volatility.

Table II .
Summary of two-terminal protonic neuromorphic devices in terms of device type, materials, doping method, proton relaxation time constant, and volatility.