Opportunities and dilemmas of in vitro nano neural electrodes

Developing electrophysiological platforms to capture electrical activities of neurons and exert modulatory stimuli lays the foundation for many neuroscience-related disciplines, including the neuron–machine interface, neuroprosthesis, and mapping of brain circuitry. Intrinsically more advantageous than genetic and chemical neuronal probes, electrical interfaces directly target the fundamental driving force—transmembrane currents—behind the complicated and diverse neuronal signals, allowing for the discovery of neural computational mechanisms of the most accurate extent. Furthermore, establishing electrical access to neurons is so far the most promising solution to integrate large-scale, high-speed modern electronics with neurons that are highly dynamic and adaptive. Over the evolution of electrode-based electrophysiologies, there has long been a trade-off in terms of precision, invasiveness, and parallel access due to limitations in fabrication techniques and insufficient understanding of membrane–electrode interactions. On the one hand, intracellular platforms based on patch clamps and sharp electrodes suffer from acute cellular damage, fluid diffusion, and labor-intensive micromanipulation, with little room for parallel recordings. On the other hand, conventional extracellular microelectrode arrays cannot detect from subcellular compartments or capture subthreshold membrane potentials because of the large electrode size and poor seal resistance, making it impossible to depict a comprehensive picture of a neuron's electrical activities. Recently, the application of nanotechnology on neuronal electrophysiology has brought about a promising solution to mitigate these conflicts on a single chip. In particular, three dimensional nanostructures of 10–100 nm in diameter are naturally fit to achieve the purpose of precise and localized interrogations. Engineering them into vertical nanoprobes bound on planar substrates resulted in excellent membrane–electrode seals and high-density electrode distribution. There is no doubt that 3D vertical nanoelectrodes have achieved a fundamental milestone in terms of high precision, low invasiveness, and parallel recording at the neuron–electrode interface, albeit with there being substantial engineering issues that remain before the potential of nano neural interfaces can be fully exploited. Within this framework, we review the qualitative breakthroughs and opportunities brought about by 3D vertical nanoelectrodes, and discuss the major limitations of current electrode designs with respect to rational and seamless cell-on-chip systems.


Introduction
Electrogenic neurons, as the control units of most biological living beings, have great potential in advancing life technologies and articial intelligence. In the central nervous system, neuronal networks are able to learn adaptively from environmental inputs, form cognitions, and carry memory storage. In the peripheral nervous system, neurons can sense a diversity of mechanical, chemical, and thermal stimuli, while delivering accurate controls through neuromuscular junctions for both long-range, high-strength and short-range, delicate motions. Moreover, the highly efficient transformation from chemical energy to ionic gradients allows a neuron to generate and transmit electric signals. Therefore, it has long been a major pursuit in neuroscience, bioengineering, and electrical engineering to develop seamless neural interfaces for probing, understanding, and modulating neural activities. And, despite advances in engineering large-scale electrophysiological approaches for in vivo applications, [1][2][3][4][5][6][7] establishing neuronal interfaces at the cellular and subcellular levels in vitro is still imperative to answer the fundamental questions of how to achieve high-delity and reliable cell-to-chip communications.
At the implementation level, constructing a bridge between electronic devices and neurons requires the electrodes not only to have appropriate electrical properties for signal detection and/or current injection, but also to be able to adapt to the dynamic and fragile nature of cells. Conventional tools of intracellular micropipettes (e.g. patch clamps) and extracellular microelectrode arrays (MEAs) have intrinsically suffered from the trade-off between invasiveness and precision. Patch clamping oen leads to severe cell damage within hours of electrode insertion, and is difficult to expand to multisite, parallel recordings. While planar MEAs impose no invasiveness to cells, the large dimension of electrode size (oen close to or bigger than neuronal soma) and poor electrode-membrane seal make it difficult to reveal subcellular and subthreshold neuronal activities. Looking at nature, the nanoscale ion channels are the driving forces behind the great diversity of neuronal dynamics, which provide biomimetic inspiration to overcome this trade-off by shrinking the electrode dimensions to the nanoscale. Although a variety of nanostructures, such as planar nanowires and suspended nanoparticles, [8][9][10][11][12][13] have been applied to neuronal recording and stimulation, a revolutionary breakthrough towards an organized and high-resolution neural interface has been enabled by 3D nanofabrication techniques. Vertical nanoelectrode arrays signicantly reduce the projection area on the planar substrate to which they are bound without compromising the interfacial area they share with the local cell membrane, allowing for the realization of high-density fabrication and parallel recording from different compartments of single neurons. More importantly, their topography of 3D protrusions provides intimate electrode-membrane seals, drastically improving the reliability and delity of recording. In this way, nanoelectrode arrays not only induce sufficient invasiveness to ensure high-quality signals, but also cause minimal cellular damage even aer electrical or optical poration. Therefore, these advantages that have arisen from scaling down to the nanoscale have provided new possibilities for solid-state electronics to "talk" to electrogenic neurons.
In this focused review, we will discuss the fundamental advantages and issues of in vitro nano neural electrodes towards the goal of accurate and rational cell-machine interfaces, primarily focusing on vertical nanoelectrode arrays that are most promising for large-scale, parallel neuronal interfaces. Specically, we start with the neuronal computation process enabled by nanoscale ion channels and neuronal projections, as well as a brief overview on the electrical interface between nanoelectrodes and neurons. Next, we highlight the opportunities provided by nanoelectrodes regarding resolution, signal quality, intracellular access, and fabrication exibility. And nally, we discuss the challenges of current in vitro nanoelectrode platforms from the system level of mapping of neuron electrical dynamics and constructing functional cell-on-chip systems.

Ionic operation and neuronal computation
Despite the complexity and plasticity of large-scale neural networks, neuronal operations are still governed by the laws of thermodynamics, in which the interactions between chemical and electric potentials are the sole factors that control ionic transport. Various species of cations, such as Na + , K + , and Ca 2+ , serve as charge carriers for a neuron to generate transmembrane currents. Assisted by active ion transporters, 14 a neuron can maintain constant gradients for each ionic species across its membrane (Table 1). At rest, for each permeable ionic species, its chemical concentration difference across the intracellular and extracellular spaces gives rise to an electric potential difference governed by the Nernst equation (eqn (1)), which tends to counterbalance the transmembrane concentration gradient. The result is an equilibrium where its chemical and electrical potentials are balanced: This Nernst potential is the driving force behind charge carrier transports across the neuronal membrane, and selective ion channels serve as gates to recruit certain ion species for an action potential and to regulate the temporal dynamics of ionic currents. For example, neuronal membrane at rest has very low permeability (non-zero) to Na + , thus the cytoplasm and extracellular uid are nearly isolated systems for Na + . Although extracellular Na + is both high in chemical potential and electric potential, the equilibrium potential of Na + (dened by eqn (1)) has little effect on the resting membrane potential due to such separation. The activation of Na + channels during the early stage of an action potential, once it occurs, brings two spaces into a single system where Na + diffuses from the extracellular uid to the cytoplasm, tending to reach its equilibrium electrochemical potential. Such an alternative electricity is a direct result of the neuron's contact-separate strategy in that it uses selective ion channels to switch the transports of its charge carriers at different stages of an action potential.
At the cellular level, a single neuron is an independent computational unit and the generation of action potentials is just the result of its computation process. The post-synaptic potentials (PSPs) from distal dendrites, either excitatory or inhibitory, are integrated at the neuronal soma that determines if an action potential will be red or not (Fig. 1a). 15 Unlike the digital all-or-none feature of action potentials, these synaptic inputs are analog signals with different amplitudes and durations (Fig. 1b). 16 More importantly, because of neurite migration and synaptic plasticity, PSPs are both spatial and temporal variants. Taking into account the non-uniform distributions of ion channels, these unique characteristics of neurons will certainly require an electrophysiological platform capable of high-resolution, long-term, and seamless recording/ stimulation, if we want to govern the detailed neuronal dynamics.

Nanoelectrode-neuron interface
The recording/stimulation principle of 3D vertical nanoelectrodes is the same as planar MEAs. However, the small dimensions and vertical protrusions of nanoelectrodes can bring about signicant improvements to some critical parameters. The equivalent circuit of a nanoelectrode-neuron interface is shown in Fig. 2, and a comprehensive discussion on recording mechanisms is covered in ref. 17. The nanoelectrodeelectrolyte interface is an electric double layer (EDL) of capacitive nature during recording. Electrical activities of local membrane, in the form of transmembrane ionic current, cause charge redistribution in the EDL. The electrical potential variation caused by such a charge redistribution is recorded by the amplifying circuit. Because cell membrane spontaneously wraps around the nano-protrusions (like endocytosis), the resulting tight adhesion can greatly reduce the ionic cle between the electrode and lipid membrane, supressing the current leakage through the seal resistance, R seal . Moreover, the membrane curvature is spontaneously formed without forced insertions, thus imposing minimum invasiveness and damage to the neuron. Since the junctional area is also at nanoscale, the signal source from the neuronal membrane is highly localized, allowing for the recovery of spatial distribution of the ion channels and membrane properties.

Qualitative breakthroughs by nano probes
Even for animals with the seemingly simplest behaviors, there might be a rather complicated nervous system functioning in the background. In contrast to the rapid development in solidstate electronics, in which billions of transistors are integrated per chip, it took decades for scientists to resolve the complete connectivity of C. elegan's nervous system, an invertebrate worm with only 280 neurons. 18 As mentioned before, a single neuron is a complex computational device that receives inputs from McCulloch-Pitts neuron performs a weighted sum of its synaptic inputs (each input i is multiplied by a synaptic weight w i ), and then a threshold operation. Each incoming presynaptic signal produces a PSP at the postsynaptic terminal, which spreads passively to the cell body (spatial summation). The cell body will perform temporal summation of all of the PSPs from different synapses. If the resulting average PSP at the soma exceeds the potential threshold, an action potential will be fired. This figure has been adapted from ref. 15 with permission from Elsevier. (b) Examples of an excitatory post-synaptic potential (EPSP) and an inhibitive postsynaptic potential (IPSP). The reversal potential (dotted line) of an EPSP is more positive than the action potential threshold, increasing the probability of triggering an action potential. For IPSPs, the reversal potential is more negative than the threshold, producing an inhibitive effect on action potential generation. This figure has been adapted from ref. 16 with permission from Sinauer Associates. This journal is © The Royal Society of Chemistry 2020 RSC Adv., 2020, 10, 187-200 | 189 multiple pre-synaptic terminals followed by neurocomputation to produce action potentials of various frequencies. Although the properties of ion channels have been extensively characterized using patch-clamp electrophysiology, 19,20 their nonuniform spatial distribution and time variance make it challenging to decode the computation process. The situation becomes more frustrating for neural networks, where the variance of synapse location and time-dependent synaptic strength are involved. Therefore, to sketch the complete picture of nervous system operations, it is imperative, at the fundamental level, to rst develop in vitro sensors with high resolution and long-term robustness to reveal the electrical activities of single neurons.

Precise subcellular interrogation
Fundamentally more advanced than microelectrodes, nanoelectrodes are not only a matter of size-scaling, but a breakthrough regarding both the quantity and quality of information extracted from cells. MEAs, developed for the purpose of large scale and parallel neural recordings, are intrinsically limited by their large dimensions and defects at the neuronal interface. First of all, with nanoscale ion channels and submicron neurites and synapses, it is difficult for MEAs to pinpoint specic cellular compartments and extract localized information from plasma membrane, as the electrode size of 5-30 mm can only characterize the averaged electrical activities of the attached membrane. As a result, subcellular information reecting the operation of nanoscale ion channels and submicron neurites and synapses is oen attenuated or buried in the signals recorded from larger areas. Second, planar microelectrodes suffer from relatively weak electrode-cell coupling that originates from a 70-100 nm cle between the cell membrane and solid-state probes. 24 Such a cle, lled with highly conductive electrolyte, contributes to most of the current leakage of the entire recording system. 25 Although the topographical improvement using mushroom protrusions has greatly enhanced the membrane adhesion for cardiomyocytes and Aplysia neurons, 26 the inevitable cell membrane deformation of micrometer size occupies a considerable amount of membrane area, making it difficult to simultaneously extract information from multiple sites of a small mammalian neuron.
The development of nanoelectrodes, especially vertical nanoprotrusions, allows probes to integrate with cell membrane at a much ner scale. This intrinsic feature, on the one hand makes it possible for nanoelectrodes to accommodate for the nonuniform distributions of ion channels on the neuronal membrane and to record potentials generated from local ionic currents, which may reveal detailed information on neural computation dynamics. 27 Nanopillars, with a diameter of about 100 nm, can interface with only a small portion of the cell membrane, allowing for multiple detection at different sites from a cell. Moreover, nanowire eld-effect transistors (FETs) have reached the 10 nm scale, approaching the dimensions of single ion channels 28 (though this device has not achieved extracellular recordings from single ion channels). Early studies oen fabricated multiple nanopillars on the same conductive pad, resulting in a loss of individuality, as they are electrically interconnected (Fig. 3a). However, the parallel probing advantage of nanoelectrodes has been realized by the breakthrough in high-density and independent nanowire arrays with a site-to-site spacing of only 750 nm, 23 where each nanoprobe, as an independent unit, can acquire high-delity electrical signals from its local interface with the neuronal membrane (Fig. 3b). Such submicron-scale information from different sites (Fig. 3c-e), combined together, will be valuable for mapping the electrical properties of cell membrane and to understand the spatial and temporal mechanism of how neuronal signals are generated and transmitted.
It should be noted that the nanoscale resolution discussed above is limited to extracellular recordings because the potential change in the EDL relies on the transmembrane current. Once electrodes gain intracellular access, the recorded signals will reect intracellular potentials that are affected by the space constant of cellular compartments.

Enhancement of seal resistance
The spontaneous membrane engulfment around nanoelectrodes provides an excellent seal for recording subthreshold membrane potentials. Various studies have characterized the plasma membrane deformation on vertical nanoprotrusions using uorescence and electron microscopic techniques (Fig. 4a). Although the value of cle thickness has shown substantial variance among devices, from less than 5 to 18 nm, 29,30 there is still no doubt that their membrane attachment is much tighter than that of planar microelectrodes (i.e., a 70-100 nm cle), giving rise to a seal resistance of 80 29 to 500 MU, 22 which is a more than two orders of magnitude enhancement. Besides the morphologically tight adhesion, the membrane curvature caused by vertical nano-protrusions may also induce the aggregation of ion channels, increasing the local ionic current. 31,32 Furthermore, treated using hydrophobic bands, nanopillars can even enhance their seal resistance to the GU range, which is comparable to that of patch-clamp seals (Fig. 4b). 33,34 Despite the increased electrode impedance caused by the reduced surface area, the improvement in the membrane-electrode seal clearly overweighs the sacrice of electrode impedance, which has been proved through both theoretical study 25 and electrophysiological tests. [21][22][23]35,36 More importantly, membrane engulfment was also observed at large neurites that are difficult to access by patch-clamps and microelectrode arrays (Fig. 4c), 29 providing the promise for the parallel monitoring of dendrites, axons, and pre-and post-synaptic terminals, as these compartments play more signicant roles than the soma regarding neural network plasticity.

Intracellular access through membrane poration
Membrane poration can be conveniently applied to nanoelectrodes to directly record from the cytoplasm with higher accuracy and less invasiveness comparing to patch clamps and MEAs. By penetrating the cell membrane either spontaneously or articially, intracellular signals will pass through a resistive interface to nanoelectrodes, rather than being "ltered" by the membrane capacitance (Fig. 2b). Also, due to the rather large dimensions of pore openings comparing to leaky K + channels, the access resistance is signicantly lowered, yielding a high signal-to-noise ratio (Fig. 5d3).
Unlike the invasive sharp electrodes that oen lead to cell death aer impalement, vertical nanostructures impose little damage on the lipid membrane. The sub-100 nm dimensions and high aspect ratio of the nanostructures induce endocytosis, allowing them to fuse into cells (Fig. 5a). 38 With biomimetic surface modications 38,39 or hydrophobic coatings, 34 the uid leakage between the cytoplasm and extracellular medium can also be reduced signicantly, making nanoscale penetration suitable for repeatable intracellular access over long-term recordings.
Spontaneous membrane penetration enabled by vertical nanoprotrusions was observed and widely applied in delivering biomolecules into the cytoplasm (Fig. 5b). Without external forces such as centrifuging or manual penetration, the high aspect-ratio of vertical nanowires can induce penetration of the cell membrane by gravity or adhesive force driven internalization aer cell plating, 43,44 which has been reported for various cell lines, primary neurons, broblasts, and immune cells. [45][46][47][48][49][50] Theoretically, for nanoelectrode recordings, spontaneous intracellular access has the advantage of a stable electrodeneuron interface in which internalized electrodes can maintain their cytoplasm access for a long period of time. 40 Moreover, excluding external forces will reduce the probability of cell damage. However, the performance of this approach has not yet met the requirements of parallel intracellular electrophysiology. The bottleneck lies in the unclear penetration mechanism and low penetration rate of the nanowires. Although surface modi-cation with cell adhesion molecules (CAMs) or cell penetrating pipettes (CPPs) have enhanced the internalization efficiency to 15%, 44,51 it is still not affordable for neuronal recordings, since a nanoelectrode array has much fewer probes than relatively simple drug delivery platforms. In addition, a systematic study using nanopillars, nanocones, and sharp nanopillars has revealed that vertical nanostructures spontaneously penetrate the cellular membrane to form a steady intracellular coupling only in rare cases, and suggested that most spontaneous penetrations might occur only during the initial hours of cell plating with membrane resealing aerwards. 53 Therefore, it is important to address the signicant issues in terms of both mechanism and application before spontaneous poration can be exploited further for nano neural electrodes.
On the other hand, articially induced membrane poration through electrical or optical approaches has been well applied to intracellular recordings. In electroporation, cell membrane can be ruptured under a strong electric eld applied at the nanoelectrode tip, allowing for signicant improvement in the signal amplitude and intracellular-like shapes. The mechanism of this process was investigated by computer simulations. Specically, water defects, caused by the interaction of water dipoles with the electric eld gradient at the water/lipid interface, are signicantly promoted by the external electric eld. As a result, water can penetrate the lipid bilayer from both sides, leading to pore formation (Fig. 5c). 41 However, since electroporation has to be conducted by the same electrodes as used for the recordings, the switching between pulse injection and recording modes will inevitably result in a blind recording period caused by overcharging. Additionally, the number, size, 21,54 and distribution of membrane pores from electroporation are difficult to predict and characterize. 55 To address these issues, the combination of vertical nanoelectrodes and plasmonic optoporation was developed for opening transient pores at the tips of nanopillars (Fig. 5d1). 37 Upon irradiation by laser pulses, the Au surface of the nanoelectrode generates highly energetic electrons (hot electrons) into the water conduction band, which induce a chain reaction of photon absorption and the release of more hot electrons. Eventually, water molecules are accelerated by hot electron impact, producing a mechanical shock wave that ruptures the cell membrane (Fig. 5d2). 42 So far, the electrode material has been limited to Au due to its low energy threshold (3.7-2.2 eV) for hot electron excitation 42,56 at the solid-water interface. Comparing to electroporation, this approach also has the advantage of localized pore formation, in addition to continuous recording. Because of its essence of mechanical nanoshockwaves, plasmonic optoporation is likely to produce single pores that are highly localized and easier to modulate. Considering that the lipid membrane reseals aer poration, a single and localized pore could make the recorded signals better in terms of delity and consistency than a group of sparsely distributed pores. This might also be the reason for the prolonged period of intracellular access of up to 80 min, 37 as it takes more time for the cell membrane to reseal a large pore than many small pores.
For both electroporation and optoporation, the recording window is relatively short (10-80 min, depending on the poration method and parameters). However, nanoelectrode penetration is a highly local process that has little impact on the adjacent membrane as well as seal resistance. In fact, the cell membrane spontaneously reseals aer the recording window. Thus, the poration could be repeatedly performed (once every several hours) over a long-term cell culture. Although this repeated sampling is not a continuous recording, considering that the neuronal long-term plasticity (learning and memory) is a much slower process, repeated intracellular access with nanoelectrodes still holds great promise for studying long-term neural network dynamics.

Fabrication of nanoelectrodes
Most substrate-bound vertical nanoelectrodes with diameters of 100-200 nm can be either fabricated top-down from bulk materials or bottom-up through crystal growth or deposition. So far, fabrication techniques have been developed for only doped semiconductors (Si) and noble metals (Pt, Au) due to their compatibility with photo-and e-beam lithography as well as chemical and physical etching. The top-down process has been widely used for metal-coated semiconductor nanopillars and nanotubes, in which high-resolution lithography was rst applied to dene the diameters of electrodes on photoresists, followed by reactive ion etching to remove uncovered material and leave out vertical nanostructures of high aspect-ratios. Additionally, since the top-down approach is a wellestablished process in dealing with semiconductors, it promotes innovative improvements to achieve various electrode geometries, smaller diameters, and higher densities. For example, the diameter of Si nanowire electrodes was reduced from 600 to 150 nm by thermal oxide thinning, based on that Si can be thermally oxidized to SiO 2 , which can then be selectively etched by HF (Fig. 6a). 22 Fabrication of hollow nanotubes exploited the tone-inversion of photoresists by secondary electrons during focused ion-beam (FIB) milling (Fig. 6b), resulting in a cylinder wall that could not be removed by developers. 52 Plus, the density of independent nanoelectrodes was greatly enhanced by thermally bonding a Si wafer onto patterned metal leads before e-beam lithography and plasma etching (Fig. 6c). 23 Noble metals of Pt and Au have excelled in microelectrodes because of their inertness in biological uids and excellent conductivity, rendering robust neural interfaces for long-term sensing. For nanoelectrodes, however, metals are highly resistant to plasma etching, and it would be expensive and unreliable to use conventional li-off processes for high-aspect-ratio protrusions that are 1-3 mm in height. Therefore, the bottom-up approach by selective deposition has been mostly adopted for metal nanoelectrodes. Specically, nanoscale holes can be milled on a layer of photoresist by e-beam lithography, followed by electrochemical deposition or FIB-assisted deposition to create the protrusion structure (Fig. 7a). 21 In comparison to semiconductor-based processes, bottom-up deposition involves much fewer steps that greatly reduce fabrication failure. Moreover, it is insensitive to substrates so that transparent glass and quartz can conveniently serve as substrates for cell observation under inverted microscopes. However, because of the limited maneuverability of noble metals at the nanoscale, bottom-up fabricated electrodes are oen pillar-shaped without sharpened tips (like Si electrodes), which may more or less affect the cell membrane-electrode interface due to their restrained bending ability under the pulling force from the cell membrane and the larger curvature at the electrode tips. 57 Another factor that makes semiconductor electrodes more popular is their compatibility with CMOS (complementary metal-oxide-semiconductor) technologies that integrate on-chip ampliers in the vicinity of electrode probes, signicantly reducing the amplitude loss by eliminating the stray capacitance. 58 Besides electrode-based nanoprobes, bottom-up crystal growth through vapor-liquid-solid (VLS) deposition is also the most critical process in the fabrication of nanowire eld-effect transistors (NWFETs). 38,[59][60][61] Composed of three terminals: source (S), drain (D), and gate (G), an FET's charge carrier density in the S-D channel is modulated by the voltage applied on the gate, thus resulting in a conductance controlled by the gate. During neural recordings, the alternating electrical potential from a neuron serves as the gate voltage that modulates the S-D current (Fig. 7b). The original neuronal potentials can then be recovered by tting this current into the known transconductance characteristics of the FET. Despite their unique structure, the working principles of NWFETs are not essentially different from those of vertical nanoelectrodes, as both devices operate upon capacitive charge redistribution at the solid-electrolyte interface. Instead of transmitting signals to ampliers through metallic wirings, FETs function as amplifying units integrated directly with EDL capacitance, thus eliminating the stray capacitance and wiring resistance that compromise the signal-to-noise ratio. 58

Dilemmas related to seamless neuralchip interface
In situ characterization of the electrode-neuron seal Neuronal signals acquired by nanoelectrodes are strongly affected by the seal resistance, R seal , due to high electrode impedance. More importantly, it is imperative to measure the R seal for each electrode in order to recover a neuron's actual electrical activities. Indeed, the tight adhesion between the cell membrane and vertical nanostructures yields detectable intracellular-like signals aer membrane poration. However, interfacing with different neuronal compartments of different dimensions, each electrode will possess a unique R seal , causing substantial variance across electrodes. Moreover, neuronal migration may also contribute to the time-variation of R seal . Apparently, if this critical parameter is not available or well characterized, the quantitative information of the recorded data will be lost.
A variety of methods have been developed to topographically characterize the membrane-surface gaps, which have been comprehensively covered in ref. 62. In general, imaging of the highest resolution can be achieved using TEM, where the cellon-chip devices are thin-sliced before observation (Fig. 4a). 29,30 Alternatively, FIB and SEM (FIB-SEM) can target the biointerface at any desired locations through FIB milling (Fig. 8), 63,64 which greatly improves the technical exibility over that of TEM. However, electron microscopy is a destructive process that can only be conducted at the nal stage, as the cells have to be killed during sample preparation. On the other hand, live cell imagings, such as through uorescent confocal microscopy 65,66 and curvature-sensing of proteins, [67][68][69][70] can qualitatively reveal the membrane engulfment around the nanopillars but cannot provide quantitative information on the seal resistance.
Despite the lack of techniques for in situ characterization, the membrane-surface seal can still be estimated from electrical models. However, the values of R seal were signicantly different among the different devices, from tens of MU to GU. Typically, these estimations were obtained by tting the recorded waveforms into the electrical model of the electrode-neuron interface. 23,35 For example, in an equivalent circuit model, the electrode impedance, stray capacitance, and amplier input impedance can be either measured or estimated with reasonable accuracy. Given the signals generated by the target cell, the R seal value can be estimated by sweeping until the simulation results t well with the recorded data. However, in order to calculate R seal , the neuronal signal sources, which are supposed to be revealed by nanoelectrodes, are assumptively predened. This approach might work well when recording action potentials with known waveforms, but when it comes to subthreshold signals from neurites, it will no longer be valid, as there are no predened signals available to calibrate R seal . Therefore, for nanoelectrodes to work independently as robust neural interfaces, it is important to develop in situ measurement approaches.

Bidirectional recording and modulation
High-delity bidirectional communication is critical for accurate and intelligent cell-on-chip systems and brain-machine interfaces, as neural probes should not only detect neural activities but also exert control over neural networks. For nanoelectrodes based on a solid-electrolyte interface, the capacitive nature of charge transport makes the recording quality more vulnerable to poor seal resistance. For a typical nanoelectrode with a 1 pF surface capacitance, its impedance is about 150 MU at 1 kHz, and will increase signicantly for slower changing signals. Since such an impedance is comparable or higher than the seal resistance (100-500 MU), both the amplitude and shape of recorded signals will be affected substantially by the electrode-membrane seal. Moreover, low-frequency eld potentials may have important links to the brain's perception, motion, and memory, 71,72 but will be severely attenuated by the high impedance of nanoelectrodes. Additionally, considering the missing approach to accurately measure seal resistances in situ, the delity of acquired data has to be put under question.
On the other hand, this issue becomes more problematic for neural modulation purposes. The small surface area of nanoelectrodes results in an extremely low charge delivery capacity, which might be able to inject miniature post-synaptic currents (mPSCs, 10-20 pA amplitude, 10 ms duration 73 ) but not simulate presynaptic spikes to evoke excitatory post-synaptic currents (eEPSCs, 0.1-0.2 nA amplitude, more than 50 ms duration 74 ). Although larger current injection can be realized by breaking into the faradaic regime, 22 the electrochemical reactions will inevitably cause electrode degradation and even water electrolysis for polarizable electrodes such as Au or Pt. Additionally, the slow recovery/discharging process aer stimulation makes it difficult to immediately switch an electrode from stimulation to recording mode.
Direct ionic access with a resistive nature has the potential solution to these intrinsic constraints of solid-electrolyte  interfaces. For example, conventional electrophysiological technology using uid-lled micropipettes is capable of both signal acquisition and current injection at the same time. Despite major drawbacks in terms of destructive membrane penetration, uid diffusion, and limited parallel access, the direct ionic interface established between the uid-lled micropipettes and neuronal cytoplasm allows them to yield reliable, high-delity, and consistent results, due to their low access resistance (about 25 MU for sharp electrodes with a 10 nm tip). More importantly, combined with non-polarizable Ag/AgCl electrodes and a Wheatstone Bridge circuit, a micropipette electrode in current-clamp mode is able to implement simultaneous stimulation and recording. 75 Such a concept was partially realized using nanotube intracellular probes made from chemically grown Si nanowires (Fig. 7b). 28,59 With FETs integrated at the bottom of nanotubes, this bioelectronic probe provides high-resolution intracellular mapping of electrogenic cells. Yet the use of FETs as an interface also precluded the compatibility for current injection. Surprisingly, the breakthrough in nanoscale ionic neural interface occurred mainly in the eld of drug delivery, where vertical nanostraws (hollow nanotubes) were used to impale cell membrane, allowing for the diffusion of molecules into the cytoplasm (Fig. 5b). 40,[76][77][78][79][80] Specically designed for drug delivery, the nanostraws in these devices are not individually addressable. In addition, the unclear mechanism of spontaneous membrane penetration gives rise to unstable and low percentage of cytoplasm access. Despite their current shortcomings, the success of nanostraw uid platforms and nanotube FETs still demonstrated that the ionic nano-neural electrophysiology is fundamentally and technically feasible, providing great promise for future improvements in engineering.

Rational alignment with neurites and synapses
Despite the extraordinary signal acquisition performance of 3D nanoelectrodes, these devices should not be limited to the vision of capturing intracellular-like signals or gaining larger amplitudes, and a lot of engineering needs to be applied to fully exploit their potential for precise and parallel neural interfacing. If we treat a neuron as a system, its dendrites, axons, and synapses are where the ionic current inputs, outputs, and intercellular signal passage take place. In particular, it is the temporal and spatial diversity of chemical synapses that give rise to the plasticity of neural networks. Both excitatory and inhibitory synapses can exist on the same neuron, and the synaptic strength can be inuenced by Hebbian plasticity 81 or heterosynaptic modulation. 82 These critical but small neuronal features at the 1-2 mm dimensions, although having been qualitatively studied using biological approaches, 83 still need to be quantied by electrophysiology in order to establish interfaces with modern electronic chips. Li et al. have demonstrated that carbon ber nanometric electrodes (CFNEs) can accurately access individual synapses and monitor the dynamics of neurotransmitter uxes 84 (Fig. 9), which suggests that nanoelectrodes hold great promise in hacking the neurites and synapses. However, CFNEs are mounted on glass micropipettes operated through a conventional patch-clamp setup, eliminating their potential for large-scale parallel recording. For vertical nanoelectrode arrays, however, the question still remains as to how to precisely access to the critical neuronal compartments while maintaining comprehensive coverage of the entire neuron. Therefore, engineering efforts in the areas of cell manipulation, neurite guidance, and nanoelectrode alignment are to be expected in the future to functionally integrate nanoelectrode arrays with neuronal circuits.

Conclusions
The advancement of nano neural electrodes is a joint effort involving nanofabrication, neuroscience, electronic engineering, and biophysics. Behind this development lies the rationale of matching to the nanoscale neuronal compartments (i.e., local ion channels, synapses, and neurites) that are critical to information processing. Ideally, neural interfaces should be able to selectively monitor the electrical activities of neurons without inducing acute damage or substantially altering the natural status of the neurons. For 3D vertical nanoelectrodes, scaling down the electrode sizes to less than 200 nm has caused minimal membrane deformation (relative to the total area), while enabling precise electrical neuronal interfaces at the subcellular level. Arranging the nanoprotrusions into individually addressable arrays further opens up the possibility of parallel, large-scale neuronal recording without labor-intensive manipulations.
Besides the rationale of accurate probing, organized vertical nanoelectrode arrays are more prominent over other nanoscale interfaces (such as dispersed nanoparticles and nanowires, 85 and lying-down nanowires 86 ), because the unique topography of substrate-bound vertical nanostructures addresses two signicant barriers towards robust neuron-electronic integration: membrane-electrode seal and reliable intracellular access. The spontaneous membrane engulfment greatly reduces the current leakage through the seal, ensuring a good signal-to-noise ratio. And, membrane poration via electrical or optical approaches can be realized with low power injection, high repeatability, and membrane recovery. Overcoming these fundamental issues has made vertical nanoelectrodes a promising platform for neuronchip communications.
Nevertheless, several key challenges still need to be addressed to push this technology further towards functional cell-on-chip systems. As mentioned in the remaining dilemmas, the prerequisite for understanding and decoding neuronal computation from recorded data is awareness of the recording conditions. The lack of approaches for characterizing the seal resistance in situ will inevitably make the recorded signals less trustworthy. On the other hand, the high electrode impedance from the EDL capacitance has put an intrinsic limitation on both recording and stimulation, which might be resolved by switching to direct ionic interfaces. Moreover, regarding device functionality, aligning nanoelectrodes with neurites and synapses needs to be addressed in the future to fully exploit the nanoscale advantages.

Conflicts of interest
There are no conicts to declare.