Ion‐specific effects in confined nanochannels and neural network

Since the discovery of biological ion channels and their role in physiology, scientists have been trying to create artificial structures that mimic biological ion channels. Nanochannels such as biological/solid‐state nanopores and porous nanostructures can have fine‐tuned ion behaviors in a confined space where ions are “aggregated”. Thus, one of the most exciting applications of nanochannels is artificial neurons and neural network based on the ion‐specific effects and nanofluidics. In this review article, it is the first time that the ion aggregation behaviors inside the confined nanochannels are summarized, and their connection to neuroscience, especially the artificial neurons and neural network, is explored and envisioned.


INTRODUCTION
Nanochannels are the basis of material and energy exchange in organisms, and they play a vital role in the process of life. [1] Since the discovery of biological ion channels and their role in physiology, scientists have been trying to create artificial structures that mimic biological ion channels. [2][3][4] Nanopores can have fine-tuned ion-specific behaviors, just like biological nanochannels that allow neurons to discharge. In biological ion channels (Figure 1), the most remarkable features are responding to external stimuli and the ability to differentiate ions of the same charge, such as Na + and K + . Thus, the most important application of nanopores is that they are used as building blocks for artificial neurons by regulating ion-specific effects. [5][6][7][8] Ions diffuse in a very "aggregated" style inside the nanochannels, which are the basis of the matter, information, and energy exchange in the biological system especially for nervous system. [9,10] Using nanopores as a biomimetic model for simulating the ion channels in the neurons will assist in investigating the ion-specific effect in confined space from a neurophysiological perspective. Meanwhile, developing computational neuroscience to investigate the ion behaviors in the neurons is also of great significance. [11,12] Herein, we will introduce the impact of ion-specific effects inside the nanochannels and its application to develop artificial neurons and neural network. It is noteworthy that the "aggregated" environment of ions inside the nanochannels will tremendously affect the ion-specific behaviors due to the strong electrostatic interaction between ions and ion-channel interaction.

MECHANISM OF NANOPORE SENSORS
Signal transduction depends on the rapid changes in the membrane potential of cells. Due to this selective distribution of fine-tuned functional elements, the breakdown of ion channels in neurons can lead to various nervous system diseases. [13,14] In addition, ion channels are usually sites where drugs, poisons, or toxins act. Therefore, ion channels play a key role in cellular physiology and pathophysiology, particularly for the nervous system. [15] In the mid-1990s, a new analytical method-nanopore sensing technique-began to develop. [16] Nowadays, it has become a very important detection method. [17] Nanopores can be used for the applications of DNA sequencing, single-molecule chemistry, ion effect study, etc. [18][19][20][21] F I G U R E 1 Schematic depiction of the ion aggregation in the confined nanochannels F I G U R E 2 Structure and signal of biological and solid-state nanopores. Reproduced with permission. [25] Copyright 2020, American Chemical Society The mechanism of the nanopore sensor is by monitoring the change of the current signal using a patch clamp when the electrolyte flows through the nanopore. It can evaluate the ion or molecule information and behaviors. [22] At present, nanopore sensors are mainly divided into two categories. One is the family of biological protein nanopores, including αhemolysin, aerolysin, and Mycobacterium smegmatisporin A(MspA), etc. [23] The other type is solid-state nanopores made of artificial materials, including silicon, silicon nitride, graphene, and organic polymer films, etc. [24] (Figure 2). Both have been well applied for studying ion-specific effects in confined nanochannels and have their own advantages and limitation. [25] 3

ION-SPECIFIC EFFECTS IN CONFINED NANOCHANNELS AND NANOPORES
Ion channels are vital to the signal transduction pathways, which have three important characteristics: (1) ion transportability; (2) ion-specific recognition; (3) they switch according to certain signals. Channels guide ion diffusion across cell membranes at extremely fast speeds, thereby providing massive ionic currents: as many as 100 million ions per second. [26] 3.1 K + /Na + distribution inside and outside the cell Although ions diffuse across the ion channels at a very high rate, they are surprisingly selective: each ion channel allows only one or a few types of ions to pass through. For example, the resting membrane potential in neurons depends mainly on channels that are selectively permeable to K + . [27] Generally, the permeability of these channels to K + is 100 times higher than the permeability to Na + . However, during the action potential, the permeability of ion channels to Na + is 10-20 times higher than the permeability to K + .
Bernstein first proposed that the unbalanced distribution of K + inside and outside the cell and the cell membrane is mainly permeable to K + in a resting state may be the basis for the cell to maintain a polarized state of internally negative and externally positive. [28] It is known that the intracellular K + concentration of all normal biological cells exceeds the extracellular K + by a lot, while the extracellular Na + concentration is much higher than the intracellular Na + concentration, which is the result of the Na + pump activity. In this case, K + must have a tendency to diffuse out of the membrane, while Na + has a tendency to diffuse into the membrane. Assuming that the membrane is only permeable to K + in the resting state, only then K + can move out of the membrane. At this time, because the negatively charged proteins inside the membrane cannot move out of the cell, as K + moves out, the inside of the membrane becomes negative and the outside becomes more positive. This outward diffusion of K + cannot be carried out indefinitely because the externally positive and inner negative electric field force caused by K + moving out of the membrane will hinder the continued outward movement of K + , and the more K + moves out, the greater the obstacle. Therefore, when the potential energy difference of K + concentration on both sides of the membrane that promotes K + to move out is equal to the potential energy difference of K + that prevents K + from moving out, there will be no net K + movement across the membrane. The potential difference between the inside and the outside of the membrane that has been formed is also stabilized at a certain value that will not increase. This stable potential difference is called the K + equilibrium potential. Bernstein used this theory to explain the generation mechanism of cell transmembrane resting potential.
According to the fact that not only the negative potential disappears in the membrane during excitation, but also a positive potential appears, Hodgkin believed that the emergence of the ascending branch of the action potential could be as simple as Bernstein. [29] Because this can only make the original negative potential in the membrane rise back to zero at most. Based on this, they hypothesized that when the membrane was stimulated, there might be a sudden increase in the permeability of the membrane to Na + , exceeding the permeability of K + . Due to the high Na + outside the cell, and the negative potential in the membrane at rest also attracts the inflow of Na + , the Na + inflows rapidly, resulting in the rapid disappearance of the negative potential in the membrane. Moreover, due to the high concentration potential of Na + outside the membrane, Na + can continue to move inward when the negative potential in the membrane decreases to zero potential, until the positive potential formed by the inwardmoving Na + in the membrane is enough to prevent the net diffusion of Na + . However, the duration of the action potential is extremely short; the recovery of the intramembrane potential to the resting state soon followed; that is, repolarization occurred, resulting in a rapid descending branch of the spike curve. As later experiments proved, the appearance of this descending branch is due to the disappearance F I G U R E 3 Scheme of ion transport measurements setup and conditions of nanopores. (A) Experimental setup of ion transport measurements through biological nanopores. Reproduced with permission. [44] Copyright, Royal Chemical Society. (B) Schematic illustrating the gated ionic transmembrane transport through a Si conical nanopore. (C) Optical image of the measurement setup. (D) Schematic of the circuit used to measure the switching behavior of the conical nanopore device. Reproduced with permission. [45] Copyright 2012, American Chemical Society of Na + permeability and the accompanying increase of K + permeability.
Therefore, one of the keys to the versatility of neuronal signals is to activate different types of ion channels, each of which has high selectivity for specific ions. The selectivity of ion channels is based on special hydration effect, electrostatic effect, and molecular sieve size effect. [30]

Nanopore channels and ion transport inside nanopores
Before solving the small amount of current flowing through a single ion channel in biological membranes, the channel function of artificial lipid bilayer membranes has been studied. In the early 1960s, Mueller and Rudin developed a technique to form a functional lipid bilayer by applying a small drop of phospholipid to a hole in a non-conductive compartment separating two salt solutions. [31] Although the lipid membrane is highly resistant to ions, when certain peptide antibodies are added to the salt solution, the ion transport across the lipid membrane is significantly increased. Early studies on nanopore channels, such as α-hemolysin nanopore and gramicidin A, are especially helpful in providing information. [16,32] These short current pulses reflect all or all of the opening and blockage of a single ion nanopore channel.
The measurement of ion transport inside nanopores varies from biological to solid-state nanopores ( Figure 3) and the setup also depends on different materials' size, shape, charge, wettability, recognition, and other factors related to the interaction between ions and nanopores. The measurement inside the biological nanopore is mostly based on electrophysiological setup and the sensitivity has been improved over the years. The solid-state nanopores have more choices in materials and fabrication methods. Their measurement methods also depend largely on the physicochemical properties of the nanopore materials. Karnik did pioneering research in the ionic transport of ultrasmall twodimensional graphene nanopores. [33] Radenovic and Feng reported specific and unique ionic behaviors inside two-F I G U R E 4 Schematic of the mechanisms of K + ion selective and nonselective channel permeation. Reproduced with permission. [46] Copyright 2018, Nature Publishing Group dimensional solid-state nanopores including ionic Coulomb blockade (subnanometer measurement), nonlinear electrohydrodynamic effect, and Ionic conductance oscillations. [34][35][36] Bocquet further explored the physical mechanism of ionic Coulomb blockade inside nanochannels through theoretical modeling. [37] Kondrat studied how to speed up ion transport using supercapacitor nanopores and ionic liquid as the model system without compromising capacitive characteristics. [38] Grey achieved first direct observation of ion transport dynamics in the nanopores of supercapacitor electrodes using in situ diffusion NMR spectroscopy. [39] Some nice reviews have introduced and summarized the ion transport measurements setup and conditions of nanopores in very detail recently. [7,[40][41][42][43]

Ion selectivity inside the nanochannels
Most cells can conduct local cell-to-cell signal transduction, but only neurons and muscle cells can conduct long-distance fast signal transduction. In neurons, the transportation of ions through the ion channel is passive and does not require the consumption of metabolic energy. The direction and final balance of this flux are not determined by itself, but by the static electricity and the driving force of diffusion through the membrane. Ion channels select the type of ions they allow to pass through the membrane, allowing cations or anions to permeate. As shown in Figure 4, the mechanism of ion transport through the membrane can be divided into two categories: selective permeation and non-selective permeation. [46] However, most cation-selective channels can only transport a single type of ion. Most types of anion-selective channels also have a high degree of discrimination; they only conduct one type of physiological ion, Cl − . In the past, much research has indicated nanopores can distinguish ions of opposite charges such as Na + and Cl − . However, developing a nanopore that can differentiate the ions of the same charge like Na + and K + remains challenging. Supramolecular chemistry is very promising in enhancing ion-selectivity of nanopores. [47] Siwy reported a biomimetic K + -selective nanopores decorated with crown ether. [48] In this nanopore, the K + -selective channel shows that the current of K + is about 80 times that of Na + , which F I G U R E 5 Designing K + -selective solid-state nanopores. Reproduced with permission. [48] Copyright 2019, American Association for the Advancement of Science is significantly higher than that of any other artificial system ( Figure 5). The benefit of utilizing solid-state nanopores is enhancing the design control, which is more stable than the biological nanopore channel.
Significant K + selectivity over other monovalent cations is not available in most nanopore systems. Solid-state nanopores modified with crown ethers exhibit excellent K + selectivity both theoretically and experimentally. However, demonstrating the intrinsic K + selectivity of individual synthetic nanopores remains challenging, as it is difficult to unambiguously measure ionic currents from individual stably formed macromolecular nanopores. Feng designed a supramolecular nanopore based on pillararene macrocycles that exhibit outstanding selectivity for K + over Na + . [49] Ionic current measurements show that a single nanopore is stably inserted into the lipid bilayer membrane with remarkable cation-selective transport, up to 21-fold selectivity for K + /Na + . Direct chemical synthesis provides a promising route to design a new class of synthetic nanopores with tailored transport functions within their atomically defined chemical structures.
In addition, another work from Siwy revealed the difference in ion selectivity inside the nanopore between divalent cation Ca 2+ and monovalent cation K + . Ca 2+ plays important roles in many physiological processes and there are many specific Ca 2+ channels, but the concentrations of Ca 2+ are much lower than those of K + and Na + ( Figure 6). [50] They prepared the mixture solutions of CaCl 2 and KCl to probe the Ca 2+ selectivity and the Ca 2+ molar fraction varying from 0 to 1. The high selectivity of Ca 2+ in biological Ca 2+ channels largely depends on the presence of Ca 2+ binding sites on the channel walls. Inspired by the properties of Ca 2+ -selective channels, they developed Si 3 N 4 nanopores which are negatively charged by silanization. These nanopores exhibited similar Ca 2+ -selectivity to biologcal channels. Nanopores smaller than 20 nm in diameter are blocked by Ca 2+ , making the ionic current in the mixture of KCl and CaCl 2 as well as in CaCl 2 even ten times smaller than that in KCl solution.
Golovchenko prepared graphene nanopore with outstanding K + /Cl − selectivity [51] and Xia developed only function-elements at outer-surface strategy to improve the ion selectivity inside the nanochannels without change in resistance. [52] Yu reported the controllable modification of polyimidazolium cationic brushes on the inner surface of F I G U R E 6 Schematic illustration of Ca 2+ induced charge inversion in carboxylated nanopores. Left: in low magnitudes of Ca 2+ mole fraction, the surface charge on the pore walls remains negative, and the concentration of positive ions is enhanced in the pore. Right: when the concentration of Ca 2+ exceeds a threshold value, the effective surface charge density, σ s , switches sign from negative to positive due to the effect of charge inversion. Voltage applied across a nanopore induces an asymmetric distribution of Ca 2+ ions that leads to an asymmetric σ s . Due to concentration polarization, the entrance at the bottom, in contact with a positively biased electrode, is negatively charged; the entrance at the top, in contact with a negatively biased electrode, is positively charged. Note that the accumulated Ca 2+ at the top entrance becomes correlated at the Stern layer; the effective positive surface charge leads to the enhanced concentration of Cl − in the diffuse layer at this entrance. Reproduced with permission. [50] Copyright 2020, American Chemical Society glass micro-nanotubes by means of surface-initiated atom transfer radical polymerization, and studied the specific effect of monovalent anions on rectification. [53] It is found that unlike kosmotropes (such as Cl − ), chaotropes like ClO 4 − are driven by hydrophobic interactions and are more easily adsorbed on the surface of polyimidazolium cations, and over-adsorption occurs at high salt concentrations, thus making the significant charge reversal. The ion transport here exhibits a concentration-dependent rectification inversion. Meanwhile, the corresponding rectification ratios of monovalent anions were sorted, and it was found that the sequence was consistent with the series of Hofmeister anions. This study provides a unique perspective based on the anion selectivity for the design of nanopore-based ionic devices and sensors.
Bezrukov systematically studied the ion selectivity of a series of biological nanopores and explored the cationselective channel regulated by anions. [54] The ion selectivity of biological nanopores is mainly owing to the interaction between ions and residues of the protein and the electrostatic interactions. Thus, the ion selectivity could be potentially improved by modulating the surface charge of and substituting the amino acid residue of the pore-forming protein. Such a strategy has also been applied to enhance the ion selectivity of solid-state nanopores.

Ion current rectification and other ion-specific effects in nanochannels
Unraveling the ion transportation process in nanopore channels is of great significance to studying the ion-specific effect in confined space. Very interestingly, Chen and Li found that ion mobility in nanopores is greatly reduced at high concentrations ( Figure 7) due to enhanced pairing and collisions between partially dehydrated ions. [55] This is contrary to F I G U R E 7 The mobility of Na + in nanopores as a function of the NaCl concentration in the reservoirs. Reproduced with permission. [55] Copyright 2019, American Chemical Society F I G U R E 8 Schematics of the fundamental processes pertaining to the ionic current rectification in nanopores. (A) Ion fluxes and the buildup of concentration enrichment-depletion in the positively biased nanopore with negative surface charge. (B) Effective length (L eff ), measuring the length of high electric field region marked in red in a neutral nanopore as a reference is the sum of the upper part, L eff _ u , and lower part, L eff _ l . (C,D) L eff in a nanopore with a negative surface charge at positive and negative bias for cases A and B, respectively. In case A, the border between region I and region II lies inside the nanopore. The black dashed line marks where L eff ends without surface charge as in (B), while the red dash line marks the actual L eff . In case B, this border is significantly distanced from the high electric field region (marked as the yellow line in corresponding figures) and lies far above the nanopore thickness. Reproduced with permission. [59] Copyright 2020, American Chemical Society the common expectation and provides essential insights into designing novel artificial neural devices.
Moreover, ion current rectification (ICR) is a very important ion-specific effect inside the nanopores (Figure 8). [56] Efficient ion rectification channels are the structural basis for organisms to regulate various physiological functions. For example, K + ion channels on the surface of biological membranes can achieve rectification and transmission of ions and protons in opposite directions to ensure electrolyte and pH balance on both sides of the membrane and maintain each normal functioning of life activities. Ion rectification channels also have broad application prospects in the fields of clean energy conversion and storage. Apel first proved that nanopore geometry has a great impact on ICR. [57] Siwy conducted experiments with single-conical nanopores in polymer films to investigate the dependence of ion rectification on the transport ion type of the same charge, and recorded current-voltage curves for three electrolytes, LiCl, NaCl, and KCl. [58] The continuous model based on the Poisson-Nernster-Planck equation cannot explain the experimental observations. Molecular dynamics simulations revealed differential binding between Li + , Na + , and K + ions and carboxyl groups on the pore walls, resulting in changes in the effective surface charge of the nanopore.
The emergence of ionic rectification as a specific current response within nanopores typically requires pore size greater than 500 nm. With the shortening of the pore size, the ion rectification in the pore will gradually weaken until it disappears. Through multi-field coupled simulation studies, Qiu found that the charged inner surface can generate strong ion concentration polarization at the small end of the cone hole, which in turn causes reverse ion rectification. [59] The ionic electric double layer existing near the small end of charged outer surface of the cone hole can cause the enrichment and dissipation of ions in the hole under the electric field, thereby producing forward ion rectification. The reason for the weak ion rectification phenomenon in the ultrashort nanopores (length 100 nm) is revealed, and the controllable adjustment of the rectification ratio from 2 to 170 is achieved by setting different nanopore conditions. Mao and Yu developed polyimidazolium brush (PimB)-modified micropipets for study of Hofmeister effect-induced ICR and also achieved microscale ICR observation with the same polyelectrolyte brush. [55,60] Szleifer and Huang designed an artificial nanopore wrapped by an asymmetric polyzwitterion brush as a model system to explore asymmetric transport in confinement space. [61] They developed a non-equilibrium theory to explain the complex charge-regulating effects of weak polymeric zwitterions and resolved the coupling between the polymer conformation and electric field. On this basis, they provide a systematic theory for stimuli-responsive behavior and transport properties within the nanopore, taking into account all details. The model shows that by incorporating pH-sensitive gradients in the polymer coating of nanopores, various charge and structure patterns can be obtained, allowing multiple functionalities in the designed system in a pH-responsive manner.

3.5
The application of ion-selective nanopore sensor toward neuroscience To apply the outstanding ion selectivity of nanopore sensor to neuroscience, Ling, Hyeon, and Chen recently reported pioneering research using silica nanopore sensor with great Na + /K + selectivity for the non-invasive dynamic monitoring of brain K + ions. [62] They used an ultra-thin K + permeable membrane to wrap the mesoporous silica nanoparticles containing K + indicator to form a K + nanosensor with high sensitivity and specificity, thus realizing the non-invasive detection of K + in the brain of active mice. This sensor can specifically recognize K + and block the entry of other ions through ion channels on the ultra-thin permeable membrane on its surface, so it has high selectivity. At the same time, due to the existence of the ultra-thin permeable membrane, the interference signals of other ions are eliminated, and the sensitivity of the system to the K + signal recognition has also been improved (Figure 9).
The researchers co-incubated hippocampal neurons cultured in vitro with artificial cerebrospinal fluid-contained nanosensors and used drugs to induce the epileptic activity of neurons simultaneously. Then, they monitored the nanosensor's response to K + signals in an in vitro model. They found that in a complex simulated environment with Na + F I G U R E 9 Atom-level design and performance of the K + nanosensors. Reproduced with permission. [62] Copyright 2020, Nature Publishing Group interference, the sensitivity of the nanoparticles wrapped with an ultra-thin permeable membrane to signals of K + was much higher than that of the nanoparticles without the permeable membrane. By stimulating the neurons of the mouse brain with a short current pulse, the K + signal changes in the brain slices can also be well recognized by the nanoparticles wrapped in the ultra-thin permeable membrane ( Figure 10).
Next, the researchers injected the K + nanosensor intracranially into the CA3 area of the hippocampus that was extremely excited during epileptic seizures and induced epilepsy in the mice. By recording the fluorescent signal generated by the nanosensor in the active mouse and comparing it with the EEG signal, the sensitivity and applicability of the nanosensor to the K + signal recognition in the brain of the active mouse are confirmed. In addition, by injecting the nanosensor into different regions of the mouse brain (the hippocampus, amygCadala, and cerebral cortex) and releasing stimulus signals from the hippocampus, researchers can dynamically observe the changes in K + fluorescence signal pattern in different regions consisting of EEG ( Figure 11).
The biological system uses ion selectivity to store energy in the form of chemical potential on the cell membrane. This energy can then be applied to power processes such as signal transduction in the neurons. Studying the mechanism of ion-specific effects in neurons and nanopore sensors will significantly improve the understanding and treatment of neural activities and central nervous system diseases. Materials used for fabricating nanopore sensors such as graphene has been widely used for neural activity detection. [63] Also, developing synthetic nanopores with excellent ion selectivity and other ion-specific effects through various strategies is still a very important task in the future. [64][65][66][67][68][69][70]

THE IMPACT OF STUDYING ION BEHAVIORS WITH NANOPORES ON NEUROSCIENCE AND NEURAL NETWORK
The most important neuroscience application of studying the ion behaviors in nanopore channels is the development of artificial neurons with memristive effects based on neural network models. [71][72][73] The artificial neurons will enable the emergence of brain-inspired computing in the near future to realize real artificial intelligence. [74] Bocquet developed an artificial neuron based on the signal transduction of nanopore ion channels. [75] The artificial neurons consist of extremely thin graphene slits containing a monolayer of water molecules ( Figure 12).
Under the action of an electric field, ions from this layer of water can assemble into elongated clusters and develop a property known as the memristor effect. Beniaguev's research shows a deep neural network requires 5 to 8 layers of interconnected artificial neurons to represent the complexity of a single biological neuron. [76] Many studies have designed artificial neural networks with increasing scale and layers F I G U R E 1 0 Imaging of K + release in cultured neurons. Reproduced with permission. [62] Copyright 2020, Nature Publishing Group F I G U R E 1 1 Multipoint [K + ] o measurements in freely moving mice. Reproduced with permission. [62] Copyright 2020, Nature Publishing Group F I G U R E 1 2 Building an artificial Hodgkin-Huxley neuron from 2D ionic memristors. (A) Electronic representation of the Hodgkin-Huxley model. I Na , R Na , and V Na ; I K , R K , and V K ; and I L , R L , and V L are the current, resistance, and Nernst potential of Na + , K + , and other species, respectively. (B) Prototype ionic machine implemented in BD simulations, exhibiting primitiveneuronal behavior. Two slits with different ionic concentrations are simulated over long time scales (t ∼ 1 ms). Each slit is connected to a pair of reservoirs, imposing a given Nernst potential on the slit, as in the original computation by Hodgkin and Huxley. (C) Spontaneous voltage spikes emitted by the prototype ionic machine and qualitative explanation for the observed spiking effect. The right panel shows different phases of a single spike, corresponding to the red rectangle on the left. Black line, quiescent neuron; red line, spike initiation; blue line, discharge. Reproduced with permission. [75] Copyright 2019, American Association for the Advancement of Science by simulating the information processing mechanism of the human brain.
The biological neural network is a complex dynamic system in which a large number of various types of neurons are connected to each other through various synapses and interactions to realize the complex transmission and processing of various types of neural signals. In recent years, in order to study the complex transmission process of various neural signals explore and master various functional mechanisms of the biological nervous system, computational neuroscientists have established a large number of mathematical models for biological neural networks and developed a large number of system simulation software platforms through computer software and hardware technology, such as GENESIS, NEURON, NEST, etc. [77][78][79] The dynamics of biological neurons and synapses are very complicated. The amount of calculation in the simulation of the physical neural network system is huge. For example, applying neurotransmitter-receptor ion channel kinetics to a full connection with a neuron number N = 10 4 , if the biological neural network is built on the complex synapse model, each synapse has 10 3 receptors. If the traditional clock synchronization algorithm is used to simulate the neural network, within each time step, N(10 4 ) neuronal membrane potential calculations and about N 2 (10 8 ) synaptic current calculations are included, and each synaptic current involves the calculation of the receptor concentration including the calculation of 10 3 receptor binding and debinding transformation process. This means that the calculation of every input current of a synapse involves a large number of complicated ion fluxes. The calculation workload of the entire system simulation will be tremendous. Thus, under the single-processor framework, simulation of the biological neural network will become very difficult. Therefore, When the amount of neurons and synapses in the neural network is large, the simulation of biological neural networks requires a high-efficient parallel algorithm. [80] Thus, connecting the biological and artificial neural networks is of great significance for both understanding neuroscience mysteries and developing artificial neural intelligent systems ( Figure 13). Hodgkin-Huxley model was proposed by Alan Hodgkin and Andrew Huxley in 1952. [82] They were awarded the Nobel Prize in Medicine in 1963 for pioneering contributions to explaining the generation and spread of action potentials. The Hodgkin-Huxley model accurately depicts the biological characteristics of the membrane voltage, which can be in good agreement with the results of the electrophysiological experiment of biological neurons. However, the computational complexity is high, and it is difficult to realize the real-time simulation of large-scale neural networks (Figures 14 and 15).
When all transmembrane conductance has fast dynamics, the Hodgkin-Huxley model description of membrane potential and voltage-gated conductance dynamics can be simplified to a one-dimensional system. However, all physical, chemical, and biological systems near the saddle-node bifurcation have specific general characteristics that do not depend on the details of the system. Therefore, all nervous systems close to this branch have the same neural computation properties. In addition, there is a rest attractor with a slow transition in the real biological system. Therefore, the neuron model constructed comprehensively based on the characteristics of fast and slow dynamics can richly express various complex firing patterns. Chay model is a typical one-dimensional model simplified by the H-H model and integrates the fast and slow subsystems. [83] In 2001, Izhikevich proposed an RF model. In this model, the meta-modeling was preliminarily divided into two types: F I G U R E 1 3 Overview of the biological and artificial neural networks. Reproduced with permission. [81] Copyright 2020, Nature Publishing Group

F I G U R E 1 5
Ion-specific effects inside the nanochannels toward artificial neuroscience integrator and resonator. [84] In 2003, Izhikevich reported another neural model. This model is superior to the RF model. It can show all the known neural behaviors of cortical neurons and has a simple structure. [85] It only requires 13 floatingpoint calculations to complete a 1 ms simulation. So even if you use a laptop, you can simulate a network composed of thousands of neurons. The Izhikevich model also has a rich physiological meaning, and each parameter has a corresponding meaning. As Izhikevich himself said, since then, the contradiction between the intuitiveness of the neuron model and the computational efficiency did not exist. By modifying the parameters, the Izhikevich model can simulate more than 20 neuron behaviors. For the ionic effects of cortical and thalamic neurons, the Izhikevich model has extremely high accuracy (See Table 1).
The TrueNorth model is the neural model used by the brain-like computing chip launched by IBM in 2013. [86,87] The realization of the TureNorth neuron provides a new idea for the circuit realization of the neural model, combining the conductance-dependent model and the non-conductivitydependent model simultaneously. The TrueNorth model balances simulation capabilities and physical implementation costs. While supporting rich ionic conductance behavior simulations, it also has the advantages of low energy consumption, high efficiency, and a small footprint. The model can be implemented using 1272 logic gates (924 implementation models Operations, 348 are used to form a random number generator).
Biological neural networks have always been the focus of artificial intelligence simulation. Through the exploration of its structure and ionic conductance behavior, people can build high-efficiency, low-power neural networks, which can effectively promote the progress of related fields. Due to the diversification of neural network functions and structures, researchers are increasing the demands for neurons. Neuron models must approximate the specific ion effects of biological neurons as much as possible so that people can build more extensive neural networks. According to the requirements of circuits and algorithms, the neuron model must be as simple as possible to achieve excellent functions, largescale integration, and production. The development of neuron model is tightly related to the development of neuroscience. With the deeper exploration of biological neuroscience and the advancement of neural network research, the mathematical and physical aspects of neuron models are constantly improving and enriching. The development process of the neuron model runs through the entire history of the development of neural networks. Over the past century, it has experienced three stages of "ionic conductance-dependent model"-"non-conductance-dependent model"-"combination of ionic conductance-dependent model and non-conductancedependent model". [88]

PERSPECTIVE
The ion transport process is of great significance to neuroscience, especially for neural signal transduction. The human brain is an extremely efficient neural device that can TA B L E 1 Ion specific effects from various nanopores and their limitations System Ion specific effects Limitations Ref.

Si 3 N 4 nanopore with a cylindrical hole Ionic memcapacitive effects in nanopores
To observe the additional memory effect, one needs both high and low frequencies due to the polarization of the ionic solutions [5] Nanochannel-based interfacial memristor consisting of an ionic liquid and a KCl solution Tunable nanochannel conductance KCl concentration can affect the magnitude of conductance tuning [10] Ion-channel switch biosensor The conductance of a population of molecular ion channels is switched by the recognition event [28] Graphene nanopores Subcontinuum transport probe Intrinsic pore defects are not uniform and not all in the subnanometer range [35] Sub-nm MoS 2 pore Ionic Coulomb blockade effect Cannot use voltage gate to observe Coulomb oscillations, but using alternative pH gating leads to addition of excess negative charges [36] Graphene nanopore Ion channel mechanosensitivity Difficulties of precision nanofabrication and integration of mechanical force [37] Sub-nm MoS 2 pore Ionic Coulomb blockade effect Current classical molecular dynamics does not fully account for possible chemical reactions in the angstrom-scale pores [38] Novolac-derived porous carbon particles The potential difference with a linear sweep can avoid pore clogging caused by counterions Computing non-linear voltage sweep U(t) requires the knowledge of the in-pore counter-ion density, which is not straightforward to measure [40] Atomistic simulations of K + channels' selectivity filter (SF) Rapid permeation of K + and ion selectivity; the direct knock-on of completely desolvated ions in the channel's selectivity filter Although recorded K + and Na + density profiles do not strictly reflect equilibrium free energy, they are used to approximately quantify differences in binding affinity for K + and Na + in the SF and kinetic barriers [48] Silicon-nitride solid-state nanopores modified with 18-Crown-6 or ssDNA Ionic selectivity Ion selectivity depends on the presence of the modified crown ether or DNA [50] Synthetic macrocycle nanopore Ion selectivity EPM nanopores exhibit high potassium selectivity, but is still lower than the highest reported selectivity of natural potassium channel KcsA The molecular simulation has not been parameterized so it cannot fully reproduce the interaction between the ions and the residues of the nanopores [51] Electrostatically charged graphene nanopores created with freestanding graphene membranes Ion selectivity The GHK model does not encapsulate the vital electrostatic effects of surface charge. In comparison, the PNP model is more complete but also inconvenient because it does not offer analytical solutions [53] Nano-channel system with only function elements at outer surface and an anodic aluminum oxide membrane Ion transport regulation [54] Poly-imidazolium brush (PimB)-modified nanopipettes Hofmeister-effect induced ICR The mechanism of Hofmeister effect inside the nanopores is still controversial [55] Gramicidin A channels Ion selectivity Not all parameters concerning the mechanisms of grA channel sensitivity are considered such as the regulation of the dwell time, etc. [56] Silicon-nitride nanopore chip sandwiched by two poly PMMA reservoirs Bulk ion mobility More salt solutions with various ion types (charge and size) need to be further explored [57] Artificial asymmetric nanopore and nanopore membranes with tip geometries Ion current rectification (ICR) Irregular behavior in the membrane due to irregular pore shapes The diameter of the fractured pore mouths cannot be determined because the lips were strained and destroyed in the fracturing process [59] Conically-shaped nanopores in polymer film with opening diameters 3-25 nm, measured in KCl, NaCl, and LiCl

ICR
The continuum modeling based on PNP equations cannot capture properties of ionic current carried by different monovalent cations [60] COMSOL Multiphysics 3D simulations of ultrashort conical nanopores ICR [61] (Continues)

System
Ion specific effects Limitations Ref.

Poly-Imidazolium brush (PimB)-Modified micropipets
Micrometer scale ion current rectification (MICR) The distribution of ions at the pore opening is nonuniform so the total amount of ions was used to predict the occurrence of ICR as opposed to the cross-section (used for nanopores) [62] Multifunctional nanopore using polyampholyte brush with composition gradient Ion selectivity Study is limited to a system of fixed parameters such as pore dimension, polymer grafting density, and compositional gradient. It is expected that the system performance can be further optimized by tuning the system parameters [63] Highly sensitive and selective K + silica nanosensor Sensitive and specific monitoring and capturing of potassium ions Further development of tissue-penetrating NIR-based K + nanosensors would allow the precise detection of the epileptic foci in the whole-brain imaging [64] perform many very complex tasks while consuming very little energy. Achieving this depends on the basic unit of the human brain-the neuron. Neurons are unique in that they are structured like nanopore ion channels. Ion channels can open and close depending on the stimulus they receive and generate different membrane potentials-resting and action potentials. The resulting flow of ions generates electrical currents responsible for firing action potentials, the signals that allow neurons to communicate with each other. Nanopores are an advantageous tool based on biomimetic simulation of ion channels, which are very useful for studying the behavior of ions in confined space. At present, the understanding of ion channel proteins from a structural biology perspective is still improving, so the biophysical simulation of ion behavior using nanopores is crucial. The current nanopores are mainly divided into biological nanopores based on pore-forming proteins and solid-state nanopores based on different porous/pore-shaped materials. These solid-state nanopores can be further divided into inorganic materials, supramolecular materials, organic polymer polyelectrolyte materials, etc. We have previously mentioned that biological nanopores have atomic-scale precision, which facilitates the study of the specific behavior of ions and their interactions with nanopores. The stability, sensitivity, and integrability of solid-state nanopores can be used to better understand the specific behavior of confined ions and their advantages in the development of biomimetic artificial neurons from the perspective of nanofluidics. The ion-specific behaviors of nanopores mainly include ion selectivity, ICR, ion gating and transport regulation, ion interaction with nanopore(residues) (equivalent to Hofmeister effect in some cases), ionic Coulomb blockade effect, etc. Furthermore, due to the need for developing artificial neurons, studying ionic memcapacitive effect is also of researchers' particular interest. Improvements in nanopore measurement setup to study the ion behaviors have never stopped, and Long recently reported single-molecule frequency fingerprint for ion interaction networks in a confined nanopore. [89] This makes a major contribution to improving the sensitivity of detecting the very minute ion current signals in nanopores. The optimization of nanopore materials is also on the way, and the current exploration of many ionspecific effects is still limited by materials. The development of novel functional nanoscale porous materials is very helpful for improving the selectivity of ions and developing artificial neurons. Combining advanced optoelectronic detection and characterization equipment and theoretical simulation methods is also helpful for exploring ion-specific effects in nanopores. It is worth mentioning here that it is still quite challenging to improve the ion selectivity of nanopores.
In the future, researchers will focus on optimizing the sensitivity of the instrument, modifying the physicochemical properties of current nanopore materials, or designing new inorganic, organic and supramolecular materials. Crown ethers with excellent ion selectivity are good examples, and coordination chemistry can inspire the development of ion-selective nanopores a lot.
The current artificial neural network-based artificial intelligence can already complete similar tasks of the human brain, but it consumes a huge amount of energy. Therefore, developing the artificial nervous system based on artificial neurons that is as energy-efficient as the human brain is a major direction that needs to be explored deep in the future. The structural biology understanding of ion channels in detail still has a long way to go. However, the biological function study of ion channels and the exploration of the electrophysiological mechanism of neural signal transduction based on membrane potential are relatively mature. The ion channels of the cell membrane are mainly based on several major ions K + , Na + , Ca 2+ , Cl − , H + in the body. The formation of the membrane potential mainly depends on the potential difference caused by the transmembrane transport of K + and Na + .
Here, we need to spend some energy on introducing the natural evolution background of the Na + -K + pump as the evolution process and learning behavior can be a useful guide and reference for the development of artificial neuroscience. The cell membrane potential of animals and electrical signaling in the nervous system mainly depend on the presence of Na + -K + pumps. Most plants rely more on proton pumps. In addition, Cl − and Ca 2+ are widely present in animals and plants to participate in the formation and change cell membrane potential. From an evolutionary perspective, the selection of these ions is highly dependent on the living environment of animals and plants. Sea is the birthplace of biological evolution, so the main ions of seawater, K + , Na + , Cl − , Ca 2+ , become the main ions involved in the internal environment of organisms. Coupled with its charge properties, it has gradually evolved into a part of biologically unique electrical signal transduction of nervous system. Plants do not have a nervous system. For terrestrial plants, after entering freshwater, the pressure to maintain the balance of Na + and K + became much smaller, while the pressure to obtain various other ions increased, and plants developed more proton pumps.
The Na + -K + is the basic unit of electrical signal transduction in the nervous system, and it is also the basis for the informative properties of the nervous system. The action potential formed by the potential difference caused by the change of ion concentration is rapidly conducted in the myelin sheath of the neuron in the form of saltatory conduction. This information transmission method is extremely efficient and fast, which consumes very little energy and delivers a huge amount of information.
In addition, as anions, inhibitory neurotransmitters can increase the permeability of the posterior membrane to Cl − , make the potential difference between the inside and outside of the membrane larger, and form an inhibitory postsynaptic potential. Protons also participate in the formation of membrane potential, but in animals, neutral cations with high concentration like K + and Na + are preferred due to protons transport will cause the change in pH which is an unfavorable factor. It is noteworthy and worth mentioning here that, in the living body, why does the conduction of neural electrical signals not choose electrons? It is common sense that conductors and many semiconductor materials count on electrons to conduct electricity. In living organisms, mitochondria and chloroplasts, which are highly demanded for redox reactions, also have abundant electron transport chains on their membrane structures. But why does the generation of cell membrane potential depend on ions rather than electrons? The answer is obvious, the cost of redox reactions is very high. As an unstable species with strong reducibility, high-energy electrons are transported through a series of proteins on the membrane, and their energy is gradually released through the oxidation and reduction of membrane proteins. In mitochondria, for example, oxygen is converted into water. The whole process is very complicated and costly. The high reliance on electron donors and acceptors also makes electron transfer difficult to achieve in a pervasive chemical environment around cell membranes. Therefore, ions are the best choice for membrane potential generation and neural electrical signaling. But it is undeniable that the ion pumps that assist the ion transport also consume a lot of energy. The cells in our body are pumping out the Na + inside the cells and pumping in the K + outside the cells all the time, and the energy consumed accounts for about 20% of the total energy consumed by the cells. Among them, in order to pump out the Na + entering the cells in the neural activity, the energy consumed by the nerve cells even accounts for 60% of the total energy consumption of the nerve cells! But even so, it still has unparalleled advantages for nerve signals to form action potential conduction through the potential difference generated by ion concentration.
The potential model formed by ions at both ends of the cell membrane allows the cell membrane to be regarded as a leaky capacitor. There are a large number of ion channels on the cell membrane, and the opening of a single ion channel is "all-ornone", and the permeability of the membrane to ions depends on the number of ion channels in the open state. There is currently no unified explanation for the specific mechanism by which voltage affects ion channels. A popular claim is that voltage-gated ion channels have "potential-sensing domains" that can be displaced by the force of an electric field to open or close the channel. To summarize, action potentials have the following characteristics:(1) The generation of action potentials requires stimulation above the threshold stimulation; (2) conduction of action potentials is not attenuated; (3) action potentials are discrete.
A typical neuron can receive thousands of messages at a time through its dendrites and soma. When the soma is sufficiently aroused, its own information is passed on to the axon, which transmits the information to the synaptosome via action potentials. This neurotransmitter-containing vesicle ruptures, releasing the neurotransmitter into the synaptic cleft. When a neurotransmitter molecule of the right shape arrives at the postsynaptic membrane, it stays on the receptor and stimulates the "next" neuron. Excess neurotransmitters are recycled into "last" neurons through a reuptake process. So neurons work by the close cooperation of electrical and chemical signals. So far, researchers have had great difficulty simulating these high-speed signals, even with microfluidics. Now researchers have developed a number of ion pumps (flows, clusters) based on slits in ultrathin materials. Using theoretical and computational tools, the researchers have shown how to assemble these to simulate the physical mechanism of action potential firing and thus the transmission of information. In addition, the ion pumps could potentially be used in combination with chemotherapy in today's deep brain stimulation therapies, such as treating central nervous system disorders by mimicking signals that the body does not normally generate.
In addition to the electrophysiological model, some scholars have also proposed the theory of mechanical waves for membrane potential. At present, this theory is still controversial. However, it has to be said that the cellular electrophysiological model of neural signaling must take into account the mechanical properties, which will greatly promote the nanofluidics design of artificial neurons and the intelligent development of artificial neural networks.
Simulating artificial neurons and artificial neural networks from the perspective of artificial intelligence is the main task in future. Real-time replication of large-scale neural networks via computational process is absolutely non-trivial. The emergence of supercomputers provides high efficiency and flexibility for processing large amounts of information at the cost of energy consumption. Therefore, developing artificial neurons that are as energy-efficient as those in the human brain holds great promise. In addition to efficiency and flexibility, real-time simulation is a desirable feature of neural networks. This is possible with existing neuromimetic architectures, which simulate complex models of neuronal electrical impulses in real time with high efficiency and flexibility. At present, the neural network simulation system developed by people has very powerful functions, which can process a huge amount of neurons, synapses, and event simulations in a very short time with very low energy consumption.
In addition, by simulating the complex mechanism of the human brain and animal nervous system to process information and the diversity of neurons, it is also very important to design artificial neural networks with increasing scale, layers, and types. From the current level of artificial intelligence, it is still extremely challenging to develop neural networks that accurately simulate biological neurons. Beniaguev's research shows that the computational complexity of deep learning is crushed by biological neurons. [76] The types of human neurons are numerous and complex, and the number of neurons in the human brain is as high as 86 billion. However, by training artificial deep neural networks to simulate the computation of biological neurons, they found that deep neural networks require 5 to 8 layers of interconnected artificial neurons to compare to the complexity of a single biological neuron. Deep learning algorithms build deep neural networks by processing large amounts of data in hidden layers of interconnected nodes. The development of deep neural networks was inspired by real neural networks in the human brain, whose nodes mimic real neurons. Computational neuroscientists use input-output functions to model the relationship between the input received by the dendrites of biological neurons and the signals emitted by the neurons. Then, they kept increasing the number of layers of neurons in the deep neural network until they reached a millisecond-level with 99% accuracy. 5-8 layers of neurons mean that the computational complexity of about 1000 artificial neurons is equivalent to 1 biological neuron. Though people have tried many various optimization methods, they are still not able to achieve simplification. More unfortunately, neuroscientists currently cannot record the full input-output function of real neurons, so there may be more information that cannot be captured. In other words, real neurons may be more complex, and more artificial neurons are needed to achieve greater computational complexity, which will consume more time and data for the algorithm to learn. From the above perspectives to guide the direction of future research: from basic ion-specific effects in nanochannels, to cell membrane potential models of neural signaling, from biological ion channels, to biomimetic nanopores, from nanofluidics and biophysical properties to artificial intelligence and big data analysis, from biological neurons to artificial neurons to artificial neural networks, it is believed that these promising research achievements from different fields and perspectives will eventually be integrated, resulting in more efficient and powerful artificial intelligence based on artificial neurons and artificial neural networks.

A C K N O W L E D G M E N T S
This work was financially supported by the Natural Science Foundation of China (22102029) and the Natural Science Foundation of Fujian Province (2021J01158).

C O N F L I C T O F I N T E R E S T
The authors declare no conflict of interest.

D ATA AVA I L A B I L I T Y S TAT E M E N T
The data supporting the findings of this study are available within the article.