Electrolyte-gated transistors for synaptic electronics,

neuromorphic computing, and adaptable biointerfacing

Citation for published version (APA):
Ling, H., Koutsouras, D. A., Kazemzadeh, S., van de Burgt, Y. B., Yan, F., & Gkoupidenis, P. (2020). Electrolyte-
gated transistors for synaptic electronics, neuromorphic computing, and adaptable biointerfacing. Applied
Physics Reviews, 7(1), [011307]. https://doi.org/10.1063/1.5122249

Document license:
TAVERNE

DOI:
10.1063/1.5122249

Document status and date:

Published: 01/03/2020

Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:
[email protected]
providing details and we will investigate your claim.

Download date: 01. Apr. 2023

Electrolyte-gated transistors for synaptic
electronics, neuromorphic computing, and
adaptable biointerfacing
Cite as: Appl. Phys. Rev. 7, 011307 (2020); https://doi.org/10.1063/1.5122249
Submitted: 30 July 2019 . Accepted: 05 December 2019 . Published Online: 16 January 2020

Haifeng Ling , Dimitrios A. Koutsouras, Setareh Kazemzadeh , Yoeri van de Burgt , Feng Yan ,
and Paschalis Gkoupidenis

COLLECTIONS

Paper published as part of the special topic on Brain Inspired Electronics

Note: This paper is part of the special collection on Brain Inspired Electronics.

This paper was selected as an Editor’s Pick

ARTICLES YOU MAY BE INTERESTED IN

Brain-inspired computing with memristors: Challenges in devices, circuits, and systems

Applied Physics Reviews 7, 011308 (2020); https://doi.org/10.1063/1.5124027

The building blocks of a brain-inspired computer

Applied Physics Reviews 7, 011305 (2020); https://doi.org/10.1063/1.5129306

Ferroic tunnel junctions and their application in neuromorphic networks

Applied Physics Reviews 7, 011304 (2020); https://doi.org/10.1063/1.5120565

Appl. Phys. Rev. 7, 011307 (2020); https://doi.org/10.1063/1.5122249 7, 011307

© 2020 Author(s).
 Applied Physics Reviews REVIEW scitation.org/journal/are

Electrolyte-gated transistors for synaptic

electronics, neuromorphic computing, and
adaptable biointerfacing
Cite as: Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249
Submitted: 30 July 2019 . Accepted: 5 December 2019 .
Published Online: 16 January 2020

Haifeng Ling,1,2,3 Dimitrios A. Koutsouras,3 Setareh Kazemzadeh,4 Yoeri van de Burgt,4 Feng Yan,2,a)
and Paschalis Gkoupidenis3,a)

AFFILIATIONS
1
Institute of Advanced Materials (IAM), Nanjing University of Posts & Telecommunications, Nanjing 210023, China
2
Department of Applied Physics, Hong Kong Polytechnic University, Hong Kong, China
3
Department of Molecular Electronics, Max Planck Institute for Polymer Research, Mainz 55128, Germany
4
Microsystems, Institute for Complex Molecular Systems, Eindhoven University of Technology, Eindhoven 5612AJ, The Netherlands

Note: This paper is part of the special collection on Brain Inspired Electronics.
a)
Authors to whom correspondence should be addressed: [email protected] and [email protected]

ABSTRACT
Functional emulation of biological synapses using electronic devices is regarded as the first step toward neuromorphic engineering and
artificial neural networks (ANNs). Electrolyte-gated transistors (EGTs) are mixed ionic–electronic conductivity devices capable of efficient
gate-channel capacitance coupling, biocompatibility, and flexible architectures. Electrolyte gating offers significant advantages for the real-
ization of neuromorphic devices/architectures, including ultralow-voltage operation and the ability to form parallel-interconnected net-
works with minimal hardwired connectivity. In this review, the most recent developments in EGT-based electronics are introduced with
their synaptic behaviors and detailed mechanisms, including short-/long-term plasticity, global regulation phenomena, lateral coupling
between device terminals, and spatiotemporal correlated functions. Analog memory phenomena allow for the implementation of
perceptron-based ANNs. Due to their mixed-conductivity phenomena, neuromorphic circuits based on EGTs allow for facile interfacing
with biological environments. We also discuss the future challenges in implementing low power, high speed, and reliable neuromorphic
computing for large-scale ANNs with these neuromorphic devices. The advancement of neuromorphic devices that rely on EGTs high-
lights the importance of this field for neuromorphic computing and for novel healthcare technologies in the form of adaptable or trainable
biointerfacing.
Published under license by AIP Publishing. https://doi.org/10.1063/1.5122249

TABLE OF CONTENTS I. INTRODUCTION

I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Present digital computation relies on the von Neumann architec-
II. ION TRANSPORT IN NEUROMORPHICS . . . . . . . . . . 4 ture as implemented with complementary metal-oxide-semiconductor
III. NEUROMORPHIC DEVICES AND FUNCTIONS . . . . 6 (CMOS) technology. Due to current limitations in semiconductor
A. Synaptic plasticity and global phenomena . . . . . . . 6 manufacturing technology, modern computer systems typically employ
B. Biofriendly materials and mechanically flexible hierarchical storage and memory-processor discrete approaches to
electrolytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 achieve an optimal balance between computational performance and
C. Spatially correlated functions . . . . . . . . . . . . . . . . . . 10 computational cost. With this approach, parallel processing is challeng-
D. Perceptron-based artificial neural network. . . . . . . 13 ing, an issue commonly referred to as the von Neumann bottleneck.1
E. Neuromorphic sensing and biointerfacing . . . . . . . 15 A neural system, such as the mammalian brain, is an efficient
1. Synaptic coupling of living neurons . . . . . . . . . 15 information-processing system, which contains about 1011 neurons to
2. Artificial sensory systems . . . . . . . . . . . . . . . . . . 15 form 1015 synaptic connections.2 Through this specific structure, our
IV. SUMMARY AND PERSPECTIVES . . . . . . . . . . . . . . . . . 18 brain is massively parallel and offers distributed computation, which

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-1

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

combines processing and memory together with a very low power purpose of the brain-related research projects is to enable a better
consumption of 20 W. Neuromorphic devices, which aim to emulate understanding of the principles of information processing in the brain,
functions of biological neurons and synapses, may provide new build- with an ultimate goal to understand malfunctions and find new ways
ing blocks for the post-Moore law era. In this scenario, neuromorphic to cure brain diseases.
computing is regarded as a promising computing paradigm for future Neurons and synapses in the brain are the smallest unit of learn-
artificial intelligence (AI), big data analysis, Internet of Things (IoT), ing and memory. As shown in Fig. 1(a), a synapse is the nanogap
etc.3,4 Since the European Union and the United States took the lead (3.5 nm) that connects two neurons, which can receive and process
in launching their Human Brain Project (HBP) and Brain Research massive presynaptic inputs to determine the postsynaptic outputs.5
through Advancing Innovative Neurotechnologies (BRAIN) in 2013, The synaptic weight (W), i.e., the connection strength between neu-
more and more countries and research institutes have been accelerat- rons, depends on the concentrations of ionic species (e.g., Ca2þ, Naþ,
ing investments in brain-inspired technology worldwide. Another and Kþ) upon presynaptic action potentials (APs), which modulate

FIG. 1. Schematic illustration of (a) a biological synapse and (b) the EGT-based artificial synapse. (c) Ionic circuits used to model EGTs. (d) The circuits for a voltage-
controlled crossbar array. Synaptic EGTs that encode the synaptic weights are present at each cross point. Diagrams indicated in the dashed frame are two typical device
structures of EGTs. (e) Schematics of a single-layer perceptron (SLP)-based network. Artificial neural networks connect an input layer to an output layer using hidden layers
(synaptic weights, W). The synaptic weight (Wm,n) between each input (Xm) node and output (Yn) node can be modulated to train the network to perform the desired operation.

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-2

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

the release of neurotransmitters. The synaptic weight permits the cou- only energy efficient but has great potential in a mixed-signal neural
pling between neurons, conveying electrical or chemical signals and network.21 Although the appropriate candidates for artificial synapses
influencing the spiking behavior of neighboring neurons, giving thus and neurons is still under debate, tunable memory devices including
rise to the neuronal development. Changes in synaptic strength are two-terminal memristors and three-terminal neuromorphic transis-
known as synaptic plasticity. In neurology, synaptic plasticity is activ- tors are promising.3,22 The operation principles of these devices are
ity dependent at either or both sides of the synapse.6 Generally, synap- based on coupled ionic-electronic features.23 The original definition
tic plasticity can be classified into short-term plasticity (STP) and for memristor (“memory resistor”) was predicted from symmetry
long-term plasticity (LTP), respectively. STP corresponds to the tran- arguments by Chua in 1971.24 In 1976, Chua et al. generalized the
sient modification of synaptic strength after stimulation, which lasts memristor concept to a much broader class of nonlinear dynamical
for tens of milliseconds to a few minutes, while LTP is a persistent systems they called memristive systems.25 In 2008, Strukov et al.
modification of synaptic strength, which can last from hours to years. correlated the resistive switching devices with memristors, in which
In biological systems, STP is required for short-term memory (STM) memresistance was experimentally observed in Pt/TiOx/Pt memo-
and allows synapses to perform critical computational functions in ries.26,27 The resistance of the devices depends on the history of cur-
neural circuits such as transmission, encoding, and filtering of neuro- rent that has flowed through it under an external bias voltage. Note
nal signals.7 However, LTP is obviously needed for storing the proc- that the roots of resistive switching phenomena date back to the
essed information, i.e., the long-term memory (LTM), and is thought 1960s,28,29 while resistive switching devices were proposed as a new
to underpin learning and memory.8 STP can be converted to LTP after generation of nonvolatile memories even before the experimental evi-
sufficient training or persistent neuronal activities. A putative neuronal dence of memristors.30 The pinched hysteresis loop has been identified
mechanism of learning and memory is Hebbian synaptic plasticity, as the fingerprint of memristive systems,31,32 thereby unifying a broad
i.e., the synaptic strength between the pre- and postsynaptic neurons class of two-terminal nonvolatile memories as memristors. These devi-
depends on the time-correlation of their activity.9,10 Two classic para- ces include resistive switching RAM (RRAM),33–35 magnetic RAM
digms for the induction of Hebbian plasticity is spike-timing-depen- (MRAM),36 and phase-change memory (PCM).37 Over the last
dent plasticity (STDP) and spike-rate-dependent plasticity (SRDP).11 decade, memristors were intended for applications in data storage,
STDP is the refinement of the Hebbian rule, where the synaptic modi- logic circuits, and neuromorphic computing. In particular, for neuro-
fication relies on the relative timing of activity between the pre- and morphic applications, abrupt switching that is present in conventional
postsynaptic neurons.12 STDP is considered to be the main learning binary memristors (i.e., digital-type memories) is unfavorable, and
and memory mechanism of the brain and also forms the basis for analog memory phenomena are desirable for the training of neural
autonomous, unsupervised learning in neuromorphic computing.13 networks with high accuracy/speed.38,39 The gradual switching can be
SRDP is another widely observed learning rule, which reflects the engineered by redox reactions (e.g., electrochemical metallization, and
influence of the activity frequency of presynaptic inputs on synaptic valence change),40 the current-induced Joule heating effect,41 the spin-
modification.14 transfer torque effect,42 and even ferroelectric polarization.43 For the
Hardware implementation of synaptic functionalities using redox-based resistive switching memristors that are coupled by ionic
microelectronic devices is regarded as the footstone for neuromorphic and electronic transport dynamics, aspects such as electrolyte materi-
engineering. Multiple and nonvolatile conductance levels of electronic als,44,45 electrode activity,46 filament nucleation,47,48 and ambient49
devices are needed to mimic the basic functions of synaptic plasticity. have been extensively studied to achieve controllable switching charac-
Presently, there are two possible ways of synapse realization: the tradi- teristics. Nowadays, memristors have attracted great interest for single
tional silicon hardware and the emerging memory devices. The main synaptic units with high scalability (<2 nm), 3D integration capability,
advantage of the former is the matureness of silicon and its full inte- and fast switching speed (nanosecond).50 Large area arrays of mem-
gration with the standard CMOS technology. Mead et al. coined the ristors could be easily integrated in a crossbar architecture to perform
term “neuromorphic” in the 1980s and proposed the concept of a sin- vector-matrix multiplication directly utilizing Ohm’s law and
gle transistor learning synapse in 1995.15,16 The silicon approach relies Kirchhoff’s law.21 Complex computational tasks and artificial neural
on charge-based mechanisms as in conventional flash memory and networks (ANN) have been implemented with such post-CMOS
random access memory (RAM), such as static and dynamic RAM device arrays.51–53
(SRAM and DRAM). Several MOS transistors are usually needed to On the other hand, neuromorphic transistors could provide an
build a silicon neuron.17 With the development of semiconductor alternative platform for synaptic electronics because of their structural
manufacturing technologies over the past two decades, nonbiomimetic nature, with physically separated input and output terminals.54–56
CMOS chips have been the advanced electronic implementation of Hysteresis in the transfer curves of transistors with gradually changed
neuron circuitry. Brain-like circuits are now commercially available, conductance presents a history-dependent memory behavior. Hence,
by integrating billions of transistors on a square centimeter.18–20 the synaptic weight (i.e., channel conductance, G) can be precisely
However, the need for high speed and design complexity for CMOS controlled in a tightly coupled fashion between the control terminal
architectures complicate the path to achieve the interconnectivity, (gate electrode) and the transduction terminals (source–drain electro-
information density, and energy efficiency of the brain. Unlike the “0” des). Meanwhile, the training or “write” operation is on the gate that is
and “1”-based digital transistors that make up modern computer spatially separated from the signal transmission or the “read” process
chips, alternative analog devices with nonlinear transmission charac- on the channel. This endows synaptic circuits with concurrent actuali-
teristics and slow ion motions are closer to the biophysical properties zation of inference and learning, hence facilitating the implementation
of neurons and synapses. In addition, analog computing without using of more complex neuromorphic functions.57 Besides, the write current
analog-to-digital or digital-to-analog (ADC/DAC) conversion is not (i.e., gate current) could be much lower than the read current (drain

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-3

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

current) so that this three-terminal configuration results in improved architectural flexibility), a unique opportunity arises to explore these
state retention and energy efficiency.58 From the mechanism perspec- devices as candidates for neuromorphic building blocks with new and
tive, neuromorphic transistors are desirable for the decoupling of STP unconventional form factors [Fig. 1(e)].22,68,78
and LTP through the use of different functional regions and physical In this review, the recent advances in EGT-based synaptic elec-
mechanisms,59–61 thus being a naturally suitable medium in which tronics, from a single device to neural networks, are comprehensively
STP and LTP can be induced concomitantly and expressed indepen- summarized and discussed. A brief theoretical background of static
dently.62 From other performance and functionality viewpoints, neu- and transient characteristics for EGTs is introduced in Sec. II. In Sec.
romorphic transistors offer advantages of multimodal control (e.g., III, recent developments of synaptic responses and artificial perception
electrical, optical, mechanical stimuli, and physicochemical responses neurons are reviewed based on the electrolyte gating of metal oxides,
such as gas and ion/molecular sensing63–66), thereby allowing the organic materials, and 2D materials. Finally, challenges and perspec-
implementation of artificial synapses that mimic biological sensory tives for future research of synaptic electronics and neuromorphic sys-
(afferent) and motor (efferent) neurons.67,68 tems based on electrolyte-gated transistors are discussed in Sec. IV.
Among all kinds of neuromorphic transistors, electrolyte-gated
II. ION TRANSPORT IN NEUROMORPHICS
transistors (EGTs), in which the semiconducting channel is in contact
with a gate electrode via an electrolyte [Fig. 1(b)], have shown to be In biological synapses, signals are carried via the exchange of var-
very promising for the implementation of artificial synapses. The elec- ious ionic or molecular species. The action potential opens the
trolyte is an ionic conductor but an electronic insulator dielectric.69 voltage-gated calcium channels of the presynaptic membrane, leading
An electrolyte could be either in a liquid or a solid state, having ions to the secretion of neurotransmitter vesicles at the synaptic cleft.79
(anions and cations) displaced in opposite charges at the electrolyte/ Neurotransmitters are defused through the synaptic cleft and bind to
electrode interfaces in response to an electric field. Typically, two receptors of the postsynaptic membrane to activate ion channels. The
major categories of EGTs have been employed for synaptic electronics influx of ions at the postsynaptic membrane alters the polarization
depending on the permeability of the semiconductor channel to the state of the neuron, and action potentials are fired if a depolarization
ions of the electrolyte. Specifically, the impermeable one is the threshold is exceeded.80 Therefore, neuromorphic devices that involve
electrolyte-gated field-effect transistor (EG-FET) with the channel cur- signals of ionic or molecular nature are desirable due to the their bio-
rent modulated by the gate voltage via a capacitive field-effect mecha- logical relevance during device operation.
nism at the channel/electrolyte interface. Due to the ultrathin There are different approaches to implement coupled ionic and
“electrical double layers” (EDLs, 1.0 nm) formed at the gate/electro- electronic conductivities in artificial synapses to emulate the biological
lyte and electrolyte/semiconductor interfaces, EG-FET is often known synaptic functionality. The concept of using electrolytes in contact
as electrical double layer transistors (EDLTs). The other type is called with semiconducting materials is not a new one. It dates back in the
an electrochemical transistor (ECT), which is based on electrochemical ‘50s when researchers at Bell labs were experimenting with germanium
doping/dedoping processes upon the bulk injection of ionic species to electrodes interfacing the aqueous solution of potassium hydroxide,
the redox-active channel material. Previously, EGTs have been widely potassium chloride, and hydrochloric acid.81 These experiments
used in chemical and biological sensing, which was enabled by their showed that electrolytes can be efficiently used to modulate the semi-
high transconductance and compatibility with aqueous solutions and conductor surface potential, validating the idea that they can also be
biological systems.70–72 The interest in EGTs dramatically increased in employed for transistor gating. The motivation behind that was to
recent years after being associated with the concept of neuromorphic take advantage of the large capacitance electrolytes can deliver, which
transistors. Compared to dielectric thickness-dependent field-effect allows for extremely low operation voltages. In this part, as a case
transistors (FETs, nF/cm2), EGTs employ a high parallel plate capaci- study, we mainly focus on organic electrolyte-gated transistors, and
tance (1–10 lF/cm2) and/or volumetric capacitance (370 F/cm3 or specifically, we are going to present their basic operation mechanism
500 lF/cm2 equivalent capacitance/unit area) to realize the high and provide the main equations that govern their steady state and
coupling efficiency of the gate to the channel.73 This feature endows transient response. To achieve this, we categorize electrolyte-gated
synaptic EGTs with the ability to alter conductance at ultralow vol- transistors in two major classes.82 The first one includes transistors the
tages (millivolt), leading to an attractive alternative in energy- semiconducting film of which is impermeable to the electrolyte ions
efficient neuromorphic circuits.74 Moreover, coupled and tunable ionic [Fig. 2(a)], while in the second class, these ions can penetrate the semi-
and electronic conductances make EGTs valuable for electronic synap- conductor, changing its redox state [Fig. 2(b)].
ses [Fig. 1(c)] since their operation mode approaches the biological In the impermeable mode of operation, an applied gate voltage
counterparts when compared to other technologies. For these reasons, forces ions to migrate to and pile up at the gate/electrolyte and electro-
EGTs can be exploited for the realization of electronic prostheses to lyte/semiconductor interfaces. These ions screen charges at the gate
directly interface with living neurons that have signals of low ampli- and accumulate (or deplete) carriers in the semiconducting film. As a
tude.75 Another benefit of the synaptic EGTs is their structural flexibil- result, EDLs are formed at both interfaces, which can be considered as
ity in constructing the synaptic networks. A shared electrolyte would capacitors with a Debye screening length k at the nanometric scale. In
be an interesting feature to provide a global control mechanism (one- this case, the thickness of the dielectric is reduced to an interface level,
input to multioutputs).76 On the other hand, an electrolyte is particu- resulting in a high parallel plate capacitance. The specific capacitance
larly suitable for the lateral-gated transistor configuration [Fig. 1(d)], of these nanocapacitors can be estimated to 10 lF/cm2, a value signifi-
enabling multi-inputs to one-output.77 In this regard and given the cantly larger than the typical 0.1 lF/cm2 achieved by solid-state dielec-
specific characteristics of EGTs (e.g., ion/electron interaction, high spe- tric capacitors.82,83 It is also worth noticing that for these devices, the
cific capacitance, physiological environmental compatibility, and applied gate potential drops predominantly at the formed double layer

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-4

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

interface in this case may or may not be formed depending on the gate
material employed (polarizable or nonpolarizable gate electrodes).
Transistors that work in this configuration are called electrochemical
transistors (ECTs), and the redox is a reversible process taking advan-
tage of the entire volume of the conducting film and delivering large
volumetric capacitance values.86 As a result, small changes in gate
biases result in large modulations in drain current, which is the reason
for ECTs to operate as efficient switches and powerful amplifiers.
Organic electrochemical transistors (OECTs) often employ a conduct-
ing polymer that is electrochemically active and ion permeable. With
the right choice of the channel material, the device can operate either
in the accumulation or the depletion mode [Fig. 2(c)].87 Specifically,
for the latter mode, [poly(3,4-ethylenedioxythiophene) doped with
poly(styrene sulfonate) (PEDOT:PSS)], an archetypical conductive p-
type polymer blend consisting of hole conductive PEDOT oligomers
polymerizing in the ion conductive PSS template, has been widely
exploit in biosensing and neuromorphic applications. PEDOT:PSS can
undergo an oxidation/reduction reaction and switch between the con-
ducting (oxidized) PEDOTþ and the insulating (neutral) PEDOTo
electrochemical states. The PSS chain is hydrophilic and ion permeable.
Applying a positive bias at the gate leads to the injection of cations into
the polymer blend [Fig. 2(d)]. Therefore, the positive bias at the gate
leads to the diffusion of cations into the PSS structure and the compen-
sation of the sulfonic acid groups of PSS. At the same time, PEDOT
FIG. 2. Typical architecture of EGTs. Accumulation-mode operation of an EGT for charge carriers (polarons and bipolarons) hop between PEDOT
(a) undoped ion-impermeable and (b) permeable semiconductors. Depletion-mode regions through p-p stacking. This mechanism thus provides both
operation for conducting polymers (e.g., PEDOT:PSS) (c) without and (d) with posi- ionic and electrical conductivities in the channel. The charge neutrality
tive gate voltage.
of the PEDOT:PSS layer implies a reduction in the number of holes in
the polymer, which is similar to electrochemically dedoping. These
capacitors, while the voltage remains constant in the bulk of the (electri-
excess holes are extracted at the drain electrode, and since the drain
cally neutral) electrolyte. The above fact implies that the operation of an
current is proportional to the quantity of mobile holes in the channel,
impermeable EDL transistor can be considered an extreme case of a FET.
it probes the doping state of the organic polymer.73
Therefore, for the steady state regime, the channel current is
The steady state current in a device like that is given by a formula
given by an equation similar to the one derived for FETs84 that bares similarities to (1)
 
W V2  
IDS ¼ lC 0 ðVG  VTh ÞVD  D ; (1) Wd  V2
L 2 IDS ¼ lC ðVTh  VG ÞVD  D ; (4)
L 2
where W and L are the channel width and length, l is the charge car-
rier mobility, C0 is the capacitance of the dielectric per unit area, VTh is where W, L, and d are the channel width, length, and thickness,
respectively, l is the charge carrier mobility, C is the capacitance per
the threshold voltage, VD is the source drain bias, and VG is the voltage
unit volume of the channel, VTh is the threshold voltage, VD is the
applied at the gate electrode.
source drain bias, and VG is the voltage applied at the gate electrode.
Regarding the transient analysis, the current can be expressed as
The figure of merit of the conducting polymer can be defined by the
the sum of the initial and the charging currents.85 In this approach,
product of the charge-carrier mobility and volumetric capacitance
the Ward-Dutton model was adopted in an electrolyte-gated FET in
(lC).86 Equation (4) for ECTs is similar to Eq. (1) for EG-FETs, with
order for the terminal charges and capacitances to be calculated and
the difference that the product of channel thickness (d) and volumetric
consequently the charging currents to be obtained. Finally, the tran- capacitance (dC) replaces C0 . This variation defines the difference
sient drain and source currents are given as follows: between the two devices.
dQD ðt Þ Finally, when it comes to the transient response, a quasistatic
iD ðt Þ ¼ ID0 ðt Þ þ ; (2) approximation ignores the spatial voltage and charge density varia-
dt
tions and averages the ionic current and charge density. As a result,
dQs ðt Þ
iS ðt Þ ¼ IS0 ðt Þ  ; (3) the transient drain current is simplified to88
dt    
se t
where ID0 ðt Þ and IS0 ðt Þ are the initial currents and QD ðt Þ and QS ðt Þ I ðt; VG Þ ¼ Iss ðVG Þ þ DIss 1  f i exp  ; (5)
si si
the charges at the drain and source, respectively.
For the permeable mode, the electrolyte ions can penetrate the where Iss (VG) is the steady-state source-drain current at a gate voltage
semiconductor film, thereby modulating its conductivity, a process VG and DIss is the difference between the current for gate voltage VG
called electrochemical reaction. The EDL at the gate/electrolyte minus the current voltage for gate voltage VG ¼ 0 and f a proportionality

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-5

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

constant to account for the spatial nonuniformity of the dedoping The electrostatic coupling effect has been widely used to mimic the
process. STP functions in dense semiconductor-based EDLTs.123 Chen et al.
Various semiconducting materials have been actively examined in reported on a carbon nanotube (CNT) channel synapse with a
synaptic EGTs. Note that the ionic dynamics in EGTs are complicated, hydrogen-doped poly(ethylene glycol) (PEG) electrolyte.105,124 When
and ionic gating modes could be modulated by varying the stimulation the presynaptic spike was low and short (5 V, 1 ms), the dynamic change
conditions on the gate terminal, such as the width, frequency, and of postsynaptic current (PSC, channel current) was induced by the
polarity of the gate pulses, which serve as the driving force of ion hydrogen ion accumulation at the electrolyte/channel interface, which in
migrations toward the rigid or soft channel materials. The most com- turn modified the electron concentration in the CNT channel.83 Such an
mon classes of channel materials consist of rigid metal-oxide semicon- interfacial EDL electrostatic modulation process was reversible when the
ductors [indium-zinc-oxide (IZO),89,90 indium gallium zinc oxide gate voltage was removed, representing the temporal analog phenome-
(IGZO),91–93 and indium-strontium-zinc-oxide (ISZO)94], binary non in biological synapses. The CNT synaptic device showed dynamic
metal-oxides [zinc oxide (ZnO)95], perovskite oxides [tungsten oxide signal processing and learning functions with an extremely low energy
WO3 (Ref. 61)], 2D materials [graphene,96–98 MoS2,59,99,100 a- of 7.5 pJ/spike. In contrast to the two-terminal memristive devices in
MoO3,101,102 WSe2,103 etc.], 1D materials [carbon nanotubes which the energy consumption is mainly contributed by the write opera-
(CNTs),104,105 InP nanowires,106 and polymer nanowires107,108], and tion, the read energy of synaptic EGT is comparable or larger than the
soft organic semiconductors [PEDOT:PSS,58,109 poly(3-hexylthio- write energy because the gate leakage current is generally much lower
phene) (P3HT),110,111 polyaniline (PANI),112,113 2,7-dioctyl[1]benzo- than the channel current, and thus, the energy efficiency is read lim-
thieno[3,2-b][1]benzothiophene (C8-BTBT),114 pentacene,115 etc.]. ited.101,125 The overall energy consumption scales with the channel
The large free volumes in the conjugated polymer (CP) bulk can also area.74 Channel materials with inherently low conductivity are more
transport ions, leading to an extremely high transconductance dictated suitable to construct energy-efficient synaptic EGTs.58
by the volumetric capacitance (C).86 Amorphous oxide The STP functions based on the electrochemical doping concept
semiconductor-based synaptic EGTs have a combination advantages of were initially demonstrated by Gkoupidenis et al. in PEDOT:PSS
high mobility, transparency, and inherent persistent photoconductivity based organic electrochemical transistors (OECTs) gated with an
(PPC) by optimizing the metallic composition ratios.94 These inorganic aqueous KCl electrolyte (0.1 M).109 In this depletion-mode transistor,
semiconductors could be deposited either by solution-processing or synaptic functions were reproduced by applying positive presynaptic
magnetron sputtering with uniform properties in large scale produc- voltage pulses (VPre) at the gate electrode (with amplitude VP, width
tion. They can act as both an electrode and a channel,116 like the way tP, period TP, and time interval between the pulses Dt ¼ TP  tP).
of conducting polymer PEDOT:PSS.117 Cations (Kþ) were injected from the electrolyte to the soft
PEDOT:PSS blend, resulting in the compensation of the sulfonic acid
III. NEUROMORPHIC DEVICES AND FUNCTIONS groups of PSS and the charge neutrality of the PEDOT:PSS layer. This
The biologically inspired neuromorphic systems are expected to electrochemically dedoping lost mobile holes in PEDOT:PSS and
be capable of dealing with complex and intelligent tasks, where neuro- thereby induced an inhibitory postsynaptic current (IPSC). After
morphic functionalities need to be implemented at the single device or removing the pulse, the injected cations relax to the electrolyte, and
circuit level. Below, the basic principles of neuromorphic devices and the PEDOT:PSS layer was reversibly doped to its initial high-
functions are presented. conductance state, showing a typical short-term depression behavior.
The PPD behavior was mimicked by applying a pair of pulses at the
gate electrode [Fig. 3(a)]. The depression percentage decreased with
A. Synaptic plasticity and global phenomena
the increasing time interval Dt. High frequency presynaptic stimuli
Synaptic plasticity, i.e., the ability of synapses to modulate the were heavily suppressed, enabling the depressive PEDOT:PSS OECT
coupling between the pre- and the postsynaptic neurons, leads to the to serve as a low-pass filter that impedes supernumerary bursts of pre-
dynamic development of the neuronal network. For synaptic plasticity, synaptic pulses. Frequency-dependent high- or low-pass filtering could
there are two forms of synaptic responses, namely, potentiation and be realized in a single device depending on the patterns of presynaptic
depression, which are linked to the strengthening and weakening of activity.126,127 For example, Ling et al. reported an identical spike-
synaptic transmission, respectively.118 For STP, short-term potentia- polarity method to mimic the concomitance of excitatory and inhibi-
tion (STPo) and short-term depression (STD) in synapses act as high/ tory short-term plasticities in a PEDOT:PSS OECT.126 Owing to the
low-pass filtering that plays a significant role in the information proc- distinctive volumetric capacitance (C) and rapid electrochemical dop-
essing of auditory and visual system.119 STP in the form of amplitude ing process (0.4 ms) of OECTs, the dynamical reconfiguration
encoding is also believed to enhance the bandwidth of neurons when- between the excitatory and inhibitory responses with 10 mV stimulus
ever saturation in their rate code is reached.120 For LTP, it is now clear resolution and multilevel synaptic strength was realized by controlling
that long-term potentiation (LTPo) and long-term depression (LTD) the doping degree of PEDOT:PSS, without preforming operations or
are used for multiple brain functions in addition to learning and mem- introducing additional modulation terminals. High frequency input
ory, as it contributes to the neuronal development.121,122 Fundamental signals (50 Hz) could induce a strong suppression/potentiation effect
synaptic plasticity behaviors have been successfully mimicked in to filter out the corresponding low/high frequency input signal, and
electrolyte-gated synaptic transistors, such as excitatory/inhibitory thus, both low-pass/high-pass filtering functions can be implanted in a
postsynaptic current (EPSC/IPSC), paired pulse facilitation/depression single synaptic device by modulating the synapse operation modes.
(PPF/PPD), and high/low-pass filtering effects, and spike-timing- An enormous effort has been made in understanding the mecha-
dependent plasticity (STDP). nism that underlies learning and memory.128 Learning may be

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-6

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 3. (a) PPD effect in PEDOT:PSS-based synaptic OECTs.109 Reproduced with permission from Gkoupidenis et al., Adv. Mater. 27, 7176 (2015). Copyright 2015 Wiley-
VCH. (b) STM to LTM transition as a function of the pulse training sequence.132 Reproduced with permission from Gkoupidenis et al., Appl. Phys. Lett. 107, 263302 (2015).
Copyright 2015 AIP Publishing LLC. (c) STM to LTM transition as a function of pulse duration (Vp ¼ 2.5 V).102 Reproduced with permission from Yang et al., Adv. Mater. 29,
1700906 (2017). Copyright 2017 Wiley-VCH. (d) Schematic of the ion gel–gated P3HT core-sheath nanowire synaptic transistor. (e) LTP and LTD triggered by 60 negative and
60 positive pulses.108 Reproduced with permission from Xu et al., Sci. Adv. 2, e1501326 (2016). Copyright 2016 AAAS. (f) Zoomed-in HRTEM image and schematic illustra-
tions of Liþ adsorption on the surface (left, STP) and Liþ intercalations into the van der Waals gaps (right, LTP).103 Reproduced with permission from Zhu et al., Adv. Mater.
30, 1800195 (2018). Copyright 2018 Wiley-VCH. (g) Asymmetric STDP function implemented in the WO3-based synaptic EGTs.61 The inset shows the schematic representa-
tion of the synaptic EGTs. Reproduced with permission from Yang et al., Adv. Mater. 30, 1801548 (2018). Copyright 2018 Wiley-VCH. (h) Schematic of the configuration of the
array of PEDOT:PSS-based synaptic OECTs and principle of the global input. (i) Spatial maps show that the global input forces a global restriction on every output.76
Reproduced with permission from Paschalis et al., Nat. Commun. 8, 15448 (2017). Copyright 2017 Springer Nature Publishing.

described as the mechanism by which new information about the synaptic OECTs by modulating the electronic structure of channel
world is acquired and memory as the mechanism by which that materials.58,130,131 Gkoupidenis et al. used a poly(tetrahydrofuran)
knowledge is retained. At the cellular level, it has been revealed that (PTHF)-based PEDOT derivative (PEDOT:PTHF) as the OECT chan-
long-term potentiation consists of distinct phases involving different nel gated with an aqueous KCl electrolyte (0.1 M).132 When a high
molecular mechanisms.129 The storage of long-term memory is associ- reduction potential was applied, the polymer structure PEDOT:PTHF
ated with gene expression, de novo protein synthesis, and formation of underwent a structural collapse and an opposite polarity oxidation
new synaptic connections.8 At the device level, ion penetration into potential was required for the reversal of this conformational
the channel is the identifying characteristic of OECTs, resulting in an change.133 As a result, the PEDOT:PTHF OECTs exhibited a nonvola-
operation mechanism that is distinctly different from EDLTs. Given tile phenomenon and could operate either in the STM or LTM regime,
the inherent reversibility of ion injection/extraction in/from the loose depending on the number of training pulses [Fig. 3(b)]. The amplitude
bulk of the channel volume, the implementation of long-term memory of PSCs persisted a permanent increase in synaptic efficacy after the
properties in conjugated polymer-based OECTs is still challenging. repetitive stimulation, resulting in long-term phenomena. The coexis-
Generally, the nonvolatile conductance tuning was obtained in tence of a STP function upon an LTM state is similar to biological

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-7

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

memory and supports merged processing and storage capabilities in a effectively manipulated by tailoring the diffusion dynamics by varying
single device level. Gerasimov et al. reported an evolvable synaptic thicknesses and structures of the vdW materials, producing LTP and
OECT by electropolymerizing a self-doped conjugated monomer LTD functionalities with enhanced linearity, symmetry, and reproduc-
sodium 4-(2-(2,5-bis(2,3-dihydrothieno[3,4-b][1,4]dioxin-5-yl)mthio- ibility. Ge et al. reported concomitant STP and LTP behaviors using
phen-3-yl)ethoxy)butane-1-sulfonate (ETE-S) as the channel and 0.01 the insulator–metal transition mechanism in an n-type tungsten oxide
M NaCl as the electrolyte, which exhibited STM and LTM functionali- (WO3) synaptic transistor, which was gated with an ionic liquid of
ties.60 The p-type PETE-S could operate in the hybrid accumulation– N,N-diethyl-N-(2-methoxyethyl)-N-methylammonium bis-(trifluoro-
depletion mode, enabling the mimicking of short-term potentiation methylsulphonyl)-imide [DEME]þ[TFSI].61 When a positive VG
(STP) and depression (STD) when applying a small negative/positive (0.6 V, 70 ms) was applied, DEMEþ ions accumulated at the ionic liq-
VG below 0.3 V with short pulse duration (100 ms). However, the uid (IL)/channel interfaces to form EDLs and induced additional elec-
LTPo and LTD properties were attained by electropolymerization and trons in the channel. Once VG was removed, cations and anions in the
electrochemical overoxidation of the channel material under 30 VG ILs relaxed and mixed together in short time, leading to volatile phe-
spikes of 0.5 V and 2 V (1s), respectively. Classical conditioning nomena. The results of X-ray photoelectron spectroscopy (XPS),
was also demonstrated by connecting a preformed PETE-S resistor Raman spectroscopy, transmission spectroscopy, and secondary ion
and an evolvable OECT. mass spectrometry (SIMS) clarified that when VG (1.8 V, 210 ms) was
The STM to LTM transition has been extensively studied in higher than the threshold value of the hydrolysis reaction (VT), the
EDLTs and can be accomplished by modifying the ionic gating effects trace water molecules contained in the ILs could dissociate into pro-
from ideal electrostatic coupling (surface accumulation and without tons (Hþ) and hydroxyls, both of which had smaller ion sizes and rela-
interfacial electrochemical processes) to electrochemical reactions (ion tively high chemical reactivity.102 The protons then penetrated into the
intercalation) through higher VP and/or longer tP gate presynaptic WO3 channel, resulting in the formation of a stable metallic HxWO3
spikes [Fig. 3(c)].61,102,103,134 Xu et al. reported organic core-sheath phase and the nonvolatile resistance state. The synaptic potentiation
nanowire artificial synapses that were gated with an ion gel composed and depression could be reproduced by consecutive positive (1.8 V,
of a poly(styrene-block-methyl methacrylate-block-styrene) (PS- 210 ms) and negative spikes (1 V, 210 ms), respectively. A typical
PMMA-PS) triblock copolymer and 1-ethyl-3-methylimidazolium asymmetric STDP was obtained in this WO3 synaptic transistor,
bis(trifluoromethyl sulfonyl) imide [EMIM]þ[TFSI] ionic liquid [Fig. which was induced by temporal correlations of pre- and postsynaptic
3(d)].108 The p-type poly(3-hexylthiophene) (P3HT) core was printed spikes [Fig. 3(g)].
and wrapped with a polyethylene oxide (PEO) sheath. The nanowire Various mechanisms have been adopted to implement the non-
printing technique could precisely define the channel length down to volatility and modulate the functional regions from STP to LTP, such
300 nm, ensuring that the synaptic device has a low energy consump- as electrostatic coupling, electrochemical doping/dedoping, charge
tion down to 1.2 fJ per spike. When the synaptic transistor was stim- trapping/detrapping, ferroelectric polarization, and phase transforma-
ulated by 30 presynaptic spikes (1 V or 1 V, 50 ms), the anions/ tion (electropolymerization, insulator-to-metal, semiconductor-to-
cations that accumulating near the nanowire would penetrate into the metal, and amorphous-to-crystalline).60,61,98,136 Note that most of the
P3HT core through the PEO sheath and gradually increased/decreased reported LTP is based on large structural transformations in channels,
the charge carrier density in the channel. Due to the PEO sheath, the resulting in high write noises and even irreversible electrochemical
spontaneous release of the trapped ions in the P3HT nanowire was doping.134 If operated in their subthreshold regime (e.g., with the use
slow, and this restricted mobility induced LTP or LTD. As shown in of low amplitude and/or narrow width input pulses), the nonvolatile
Fig. 3(e), the organic nanowire-based synaptic transistor operated in transistors can practically only show STM dynamics without affecting
an analog fashion with gradual but nonlinear conductance changes. the LTM. However, nonvolatile neuromorphic devices with write-
However, a large range of linearly and symmetrically programmable erase ability and structural stability are highly desired. For example,
conductance states are required to avoid complex programming Fuller et al. reported an all solid-state, nonvolatile electrochemical
schemes in large arrays and to facilitate near-ideal training and infer- transistor. In this device, a negative gate voltage induced the intercala-
ence accuracy in neural network simulations.131,135 In this regard, tion of Liþ dopants into the channel of Li1xCoO2.137 The channel
Yang et al. reported a polymer electrolyte PEO:LiClO4 gated synaptic then undergoes an insulator-to-metal transition with nearly six orders
transistor based on 2D van der Waals (vdW) layered WSe2 that could of magnitude increase in electronic conductivity. A positive gate volt-
biorealistically emulate both STP and LTP.103 As shown in Fig. 3(f), age could reintercalate Liþ and return the channel to its initial conduc-
the high-resolution transmission electron microscopy (HRTEM) char- tivity. This process was highly reversible and without large structural
acterization supported that when a gate pulse (5 V, 50 ms) was applied, transformations. In a device based on the conductive polymer
the Liþ ions were adsorbed to the surface of the 2D materials. Ions dif- PEDOT:PSS, a similar LTP behavior was demonstrated.58 In this
fused back into the ion gel after the removal of the gate bias, leading to example, the device resembles an electrochemical battery, where ionic
the STP property. In contrast, when the Liþ ions got intercalated into charges (protons) can move through the electrolyte that separates the
the vdW material after the application of 100 gate pulses (5 V, 50 ms), two organic electrodes, while electronic charges move through the
they could not be liberated spontaneously after removing the gate bias, electronic circuit. This results in enhanced state stability and analog
and hence, backward diffusion of the ions was retarded, leading to the tuning. To further enhance the state and cycle stability, smart material
LTP behavior. Because the STP property was dominated by surface design and the definition of the number of separable states have to be
adsorption, the short-term synaptic response was related to the effec- taken into account.130
tive area of the vdW channel, whereas it was insensitive to the layer Hebbian plasticity alone through potentiation and depression is
number of the vdW material. The long-term synaptic activity could be expected to be insufficient to explain activity-dependent development

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-8

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

because it tends to destabilize the activity of neural circuits (i.e., higher biopolymers are attracting growing interest for solid-state electrolytes
temporal correlation between pre- and postsynaptic signals leads to a of green and flexible synaptic transistors since the transport of protons
higher coupling between neurons, with no upper limit).138 is found in many natural phenomena such as muscle contraction, taste
Homeostatic plasticity is another important function that is believed receptor cells, and mammalian brain.162,163 Typical biopolymer electro-
to stabilize the neuronal activity that was induced by the positive- lytes that have been used include polysaccharides (chitosan,164–167
feedback nature of Hebbian plasticity.139 Homeostatic plasticity mech- sodium alginate,168 pectin,169 and starch170), cellulose (lignin114), and
anisms include global changes in synaptic strengths and changes in proteins (albumin171). Shi et al. used chicken albumen as the electrolyte
neuronal excitability. Due to the fact that biological neural networks film in n-type IZO-based synaptic EGTs.171 According to the leakage
are immersed in a common electrolyte environment, global factors of and electrochemical impedance spectroscopy (EIS) measurements [Fig.
this environment are forcing specific normalization functions that reg- 4(a)], the hydrated albumen film was an electron insulating but ionic
ulate the overall network behavior.139 Gating OECTs with electrolytes conducting electrolyte with a maximum leakage current density of
offers a straightforward way to emulate homeostatic regulation, a phe- 5  107 A/cm2 at 0.02 MV/cm and a high proton conductivity (r)
nomenon that is not easily accessible with other solid-state technolo- of 3.7  104 S/cm. The ionic conductivity was mainly attributed to
gies.76,140,141 Gkoupidenis et al. demonstrated the global control of a the migration of protons induced by the amino acid interaction with
whole array of PEDOT:PSS-based OECTs that were immersed in a 0.1 H2O. Protons could move along the hydrogen-bond network following
M NaCl electrolyte [Fig. 3(h)]. The weights of these individual artificial the Grotthuss mechanism. Thus, it was necessary to maintain a certain
synapses could be modulated globally by the voltage applied on the water content for the fabrication and characterization of the biopoly-
electrolyte and by its ion concentration, in the way that was analogous mer electrolyte-based synaptic transistors. As shown in Fig. 4(b), the
to homeostasis. The synchronization of I/O transmission could be specific capacitance increased with the decreasing frequency and
reproduced in this system [Fig. 3(i)]. Additionally, the electrolyte reached the maximum value (>1.0 lF/cm2) at 1.0 Hz due to the forma-
established soft connections between individual devices (grids) without tion of huge interface EDL capacitance. Note that at a higher voltage
hard connectivity (i.e., physical wiring). It was also shown that the out- level of >2.0 V and a frequency of <10 Hz [Fig. 4(c)], the mobile pro-
put (Oi) of these grids could be synchronized by a global oscillatory tons penetrated across the albumen/IZO interface and reacted with
input despite the fact that individual local inputs (Ii) were stochastic IZO, resulting in surface hydrogenation and an increase in nonvolatile
and independent.141 The synchronization effect was more pronounced IZO conductance. For the biodegradable electronics, Guo et al.
at its amplitude extrema and high frequency of the global oscillatory reported a starch-based electrolyte for indium-tin-oxide (ITO) synaptic
input. This global temporal coupling resembled well the phase locking EGTs.170 Starch is an environmentally friendly and naturally abundant
of neurons to brain oscillations, leading to a functional type of connec- polymeric carbohydrate. During the gelatinization process, water mole-
tivity found in brain oscillations (i.e., binding through synchrony). cules are bound to the starch chain network. The proton conductivity
(r) of the 14 lm thick starch film was estimated to be 2.6  103
B. Biofriendly materials and mechanically flexible S/cm. The proton related EDL capacitance was 1.6 lF/cm2 at 1 Hz.
electrolytes Both short-term and long-term synaptic plasticities were mimicked
In recent years, flexible, wearable, and implantable electronics are under voltage spikes of 0.5 V (10 ms) and 4 V (600 ms), respectively.
attracting increasing interest for healthcare and biomedical applica- Due to the water solubility of the starch film, the devices could easily be
tions.142 In addition, the use of biocompatible and biodegradable dissolved in DI water after soaking for 60 s [Fig. 4(d)].
materials in electronic devices can be an important trend in the devel- In addition to the nontoxicity and biodegradability, biopolymer-
opment of green electronics.143 Compared with metal-oxide semicon- based electrolyte films have the advantage of being active flexible sub-
ductors, biofriendly conjugated polymers (CPs) as channel materials strates in forming or conforming to complicated textures and
of EGTs have received growing attention in flexible neuromorphic shapes.172 Liu et al. reported a freestanding chitosan membrane to act
platforms for their advantages such as chemical tunability, low as both the electrolyte and the flexible substrate in the IZO EGTs.164
cost, mechanical flexibility, and compatibility with printing and roll- The 100 lm chitosan membrane was able to form transparent
to-roll processes.144 Conjugated polymers can not only transport ions pinhole-free conformal coatings with a surface strain (e) of 0.5% at
but also transport holes (p-type; popular examples include 10 mm bending radius. Since the chitosan membrane is a three-
PEDOT:PSS,145,146 PANI,113 P3HT,147,148 and poly(2-(3,30 -bis(2-(2- dimensional proton conductor, when a positive presynaptic spike was
(2-methoxyethoxy)ethoxy)ethoxy)-[2,20 -bithiophen]-5-yl)thieno [3,2- applied on the in-plane gate, protons were driven laterally and accu-
b]thiophene) [p(g2T-TT)]149,150) or electrons [n-type, such as mulated at the chitosan/IZO channel interface [Fig. 4(e)]. The drain
p(gNDI-gT2),151 P-90,152 and poly(benzimidazobenzophenanthroline) current was increased due to the accumulated electrons by the proton/
(BBL)153]. The simulation of synaptic functions in organic EGTs has electron electrostatic coupling effect. Figure 4(f) shows the paired pulse
extended the functionality into E-skins, artificial afferent nerve, etc., facilitation (PPF) behavior as a function of the interval between two
which highly benefits from the mixed conductivity and biocompatibil- successive presynaptic spikes (1.5 V and 50 ms) for the freestanding
ity features of conjugated polymers.144,154,155 synaptic transistor. The PPF ratio decreased gradually with increasing
Biocompatible and biodegradable electrolyte materials are also Dt, and a maximum PPF value of 222% was obtained at Dt ¼ 10 ms.
highly recommended for green electronics since the electrolyte is the The distance between the laterally coupled gate electrodes and the
integral part of an EGT. Electrolyte materials consist of ionic liquids channel layer was shown to play a key role in the operation frequency.
(ILs),61,156 ion-gels,92,157,158 polyelectrolytes,159 polymer electro- For example, the PPF value decreased much slower with a larger gate-
lytes,100,103,160 aqueous salts [e.g., NaCl, KCl, and phosphate buffered to-channel distance due to the fact that longer diffusion time was
saline (PBS) solution], and even water.161 Among them, proton-rich needed for the protons to relax back to the balanced state in the case

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-9

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 4. (a) A typical Nyquist plot of the albumen film. The inset shows the diagram of the schematic diagram of the albumen-gated synaptic EGT. (b) Specific capacitance of
the albumen film as a function of frequency under different AC potentials. The inset shows the schematic diagram of the IZO/albumen interface hydrogenation process under
high bias potential. (c) STM to LTM transition as a function of pulse amplitudes (tp ¼ 1 s).171 Reproduced with permission from Wu et al., Sci. Rep. 6, 23578 (2016). Copyright
2016 Springer Nature Publishing. (d) Dissolution processes of the starch electrolyte gated ITO synaptic transistors on a glass substrate in de-ionized water at room tempera-
ture.170 Reproduced with permission from Guo et al., Org. Electron. 61, 312 (2018). Copyright 2018 Elsevier. (e) Schematic diagram of the chitosan-gated IZO freestanding
synaptic transistor. (f) PPF index as a function of presynaptic spike interval (Dt) between two successive presynaptic spikes (1.5 V, 50 ms). The inset shows the optical image
of the freestanding synaptic transistors.164 Reproduced with permission from Liu et al., Adv. Mater. 27, 5599 (2015). Copyright 2015 Wiley-VCH. (g) Spatiotemporally correlated
EPSC as a function of Dtpre2-pre1 between the two presynaptic spikes (0.5 V and 1 V, 20 ms).165 Reproduced with permission from Wu et al., J. Mater. Chem. C 2, 6249 (2014).
Copyright 2014 Royal Society of Chemistry. (h) Schematic of WCN-gated C8-BTBT freestanding synaptic transistors. (i) AFM image of WCNs. (j) A high-pass filtering behavior
in the C8-BTBT synaptic transistor.114 Reproduced with permission from Dai et al., ACS Appl. Mater. Interfaces 10, 39983 (2018). Copyright 2018 American Chemical Society.

of a larger migration distance.173 Due to the lateral coupling property, 7.3  103 S/cm and a high laterally coupled effective capacitance of
the freestanding solid-state electrolyte has been proven suitable for 18.65 nF/cm2 at 30 Hz. The spatial summation function was repro-
coplanar-gated transistors to emulate temporal summation func- duced in the C8-BTBT synaptic transistors with dual coplanar-gates,
tions.165 As can be seen in Fig. 4(g), the obtained EPSC was asymmet- which showed a sublinear integration. The lateral effective capacitance
rical with respect to the time interval between the temporally and the dendritic integration effect were tuned by controlling the dis-
correlated spikes from two separate gate inputs. Huang et al. reported tance of the in-plane gates. A high-pass filtering effect was also simu-
the use of wood-derived cellulose nanopapers (WCNs) as a freestand- lated in Fig. 4(j), and the increase in the presynaptic spike frequency
ing electrolyte in C8-BTBT-based synaptic transistors [Fig. 4(h)].114 would induce a strong EPSC amplitude increase.
The as-fabricated 30 lm WCNs were composed of densely packed
nanosized fibers with around 1 nm roughness [Fig. 4(i)]. This smooth
active substrate exhibited high transparency (close to 90%) in the C. Spatially correlated functions
visible-light region. Freestanding WCNs were rich in hydrogen bonds In the brain, information processing is spatiotemporal. The abil-
and thus able to absorb moisture from the air and thus introduce pro- ity of the human brain to convert input information from multipath-
tons into WCNs. The active film exhibited an ionic conductivity of ways into specific output patterns is important for learning, memory,

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-10

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

executing event-driven behaviors, and enabling parallel computa- obtained at the orientation of 0 , ten presynaptic spikes induced by
tions.174 Dendritic integration plays an important role in information ten edges were triggered successively. Orientation selectivity was
transformation including the addition of nonsimultaneous unitary shown to be dependent on the ionic conductance of the aqueous elec-
events (temporal summation) and addition of unitary events occurring trolyte. Besides, the alcohol solution gated synaptic transistor demon-
simultaneously in separate regions of the dendrite (spatial summa- strated the highest degree of orientation selectivity since the delivery of
tion).175 An electrolyte shows frequency sensitivity and bulk ion conduc- hydronium and hydroxyl ions was inhibited (facilitated) by alcohol
tivity, enabling the laterally gated transistor configuration in which both molecules (salt ions). An individual multigate neuromorphic transistor
the gate electrodes and a source-drain channel are patterned concur- has also been used for object detection. Wan et al. designed a visual
rently in one plane, decreasing fabrication complexity. Spike pulses system for emulating the Lobula Giant Movement Detector (LGMD)
applied on multi-in-plane gates are in analogy to parallel synaptic inputs neuron.177 The system was constructed by a 20  20 photoreceptor
from various dendritic positions. Both the gate-to-channel distance (den- array connecting to the multiple in-plane gate arrays of the IZO neu-
drite to postsynapse) and gate electrode area contribute to the change in romorphic transistor in a one-to-one correspondence [Fig. 5(f)].
the synaptic weight.77,100 Thus, electrolyte-gated neuromorphic transis- Proton conducting graphene oxide (GO) electrolytes were fabricated
tors with lateral multigates would be helpful for realizing spatially corre- on graphene/poly(ethyleneterepthalate) (PET) substrates. Both of the
lated functions and advance the capability of neuromorphic square object and the photoreceptor array were parallel to the xoy
performance on single synaptic devices. Interesting spatiotemporal infor- plane with the center coordinates of (0, 0, zo) and (0, 0, za), respec-
mation processing functions such as spatial co-ordination and visual ori- tively. The gray dashed box denotes the image of the object in the pho-
entation recognition,77,100,111,127,161,176 visual detection,177 sound location toreceptor array, and the edge movement thus could be detected by
functionality,178 spiking logic response,99,110,164 spatial summation,114,179 comparing the difference between the successive images. The excit-
and classical conditioning180 have been demonstrated. atory stimuli (0.5 V, 1.0 ms) were triggered and then sent to the corre-
Orientation selectivity is a broadly investigated phenomenon in sponding presynaptic terminals (gates) of the neuromorphic transistor
the primary visual cortex.181,182 Gkoupidenis et al. demonstrated this once the object edge was detected by using a photoreceptor. Three
function in a PEDOT:PSS-based OECT array with 3  3 coplanar Au object approaching modes (toward, away, and parallel) to the photode-
gate electrodes [Fig. 5(a)].77 An enhanced current response I0 (i.e., tector array were then distinguished according to the EPSC dynamics.
PSC amplitude) was obvious for the gates that are closer to the drain The computing functions in the neural network are based on
electrode, which was attributed to the decrease in the electrolyte resis- synaptic integrations. The spatial summation has been leveraged to
tance (RE) for smaller gate-drain electrode distances (d). Similar to the reproduce Boolean logic operations. As shown in Fig. 5(g), two types
current response I0, inhomogeneity toward the drain electrode was of logic “AND” and “YESV2” were constructed, respectively, by chang-
also evident in the relaxation time (tR) mapping. The closest to the ing the applied gate voltage drop across the channel that was con-
drain, gate electrode (x ¼ 1, y ¼ 1) resulted in the lowest tR value. This trolled by the area of gate electrodes. The input 10 ms voltages of 0
spatial inhomogeneity in PSC could thus be used for implementing and 1.0 V were defined as 0 and 1, respectively. The integrated EPSC
the orientation selectivity function. As shown in Fig. 5(b), the variable amplitude was defined as the output, and the threshold value was set
orientation could be created by superimposing eight gate pulses at gate to 750 nA. For the same electrode area, the EPSC amplitude was larger
electrode 0 and another gate electrode x (x¼ 1  6), simultaneously.111 than the threshold line only when input signals were “11,” which indi-
The spatial orientations of the input pulse were defined by the angle cated the AND logic. When the area of G2 was 2.5-fold larger than
range from 0 to 180 . Figure 5(c) shows the polar diagram of the that of G1, the capacitance of the G2-electrolyte interface would be
EPSCs for different spatial orientations of the input pulse for this ion- larger to induce the efficient EDL gating. As long as input V2 is 1, the
gel gated P3HT synaptic transistor. The average peak EPSC was gradu- output EPSC amplitude was larger than the threshold value, which
ally increased with the orientation angle changed from 0 to 50.2 or indicated the YESV2 logic. Besides, the “OR” logic, synaptic weight reg-
from 180 to 129.8 , which was similar to the orientation tuning curve ulation, and coincidence detection could be achieved by introducing a
in the primary visual cortex. The maximum tuning response (2.22 lA) modulatory terminal (Gm), and thus, the network functionalities can
was obtained at the orientation angle of 129.8 , for gate electrode 5 been enriched.110,164 The spatial summation effect was also demon-
that was the closest to the drain electrode. Neurons in the primary strated with two spatial isolated presynaptic inputs in GO-coupled
visual cortex respond preferentially to edges with a particular orienta- synaptic transistors [Fig. 5(h)].179 When the input spikes were first
tion. Inspired by this, a neuromorphic system with a combination of triggered individually and then simultaneously on two in-plane gates
photodetectors and a multigated synaptic transistors was further (G1 and G2), the three EPSCs can be obtained. The red and blue curves
developed to mimic the edge recognition function.161,176,179 As shown were the EPSCs (A1, A2) triggered by single presynaptic spikes (V1
in Fig. 5(d),161 a square panel with five pairs of black-white grating and V2) ranged from 0.2 to 1.4 V on G1 and G2, respectively. The
patterns was moved along the y axis in the yoz plane. The orientation expected sum (SE, green dashed curves) was defined as the arithmetic
angle (h) was defined as the angle between the z axis and the grating sum of two individual EPSC responses (A1 þ A2), and the measured
orientation. The coordinate of the photodetector is (x, 0, 0), and the sum (SM, black curves) was the EPSC stimulated by the two simulta-
coordinate of the panel center was (0, y, 0). Each time the edge of the neously triggered spikes. Summation would be expected to range from
grating pattern moved across the coordinate origin, it was detected by sublinear to superlinear depending on the stimulus intensity of the
the photodetector, and the processing circuit would provide a voltage individual excitatory postsynaptic potentials (EPSP). Figure 5(i) shows
pulse (0.5 V, 10 ms) to the presynaptic terminal of the aqueous solu- that such spatial summation was nearly linear for low spike voltages. A
tion gated n-type indium–gallium–zinc oxide (IGZO) synaptic transis- linear summation model has been postulated to facilitate coincidence
tor. Figure 5(e) shows that when the maximum value of EPSC was detection by cortical neurons.183 Strong superlinear for intermediate

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-11

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 5. (a) Spatial mapping the resulting amplitude I0 of the IPSC. Each (x, y) gate electrode was pulsed separately (0.3 V, 50 ms).77 Reproduced with permission from
Gkoupidenis et al., Sci. Rep. 6, 27007 (2016). Copyright 2016 Springer Nature Publishing. (b) Schematic illustrations of the ion gel-gated P3HT synaptic transistor with a multi-
in-plane-gate structure. (c) Polar diagram of the EPSCs for different spatial orientations of the input pulse.111 Reproduced with permission from Qian et al., Appl. Phys. Lett.
110, 083302 (2017). Copyright 2017 AIP Publishing LLC. (d) A schematic diagram showing the measurements for the orientation tuning experiment. (e) The normalized gain
of EPSC responses plotted as a function of the orientation angle for the solution-gated IGZO synaptic transistor.161 Reproduced with permission from Wan et al., ACS Appl.
Mater. Interfaces 8, 9762 (2016). Copyright 2016 American Chemical Society. (f) The EPSC output of the visual system recorded in response to an object when it moves
toward the photoreceptor array. (g) Input–output characteristics of the AND and YESV2 logics from the two combinations of presynaptic input terminals.177 Reproduced with per-
mission from Wan et al., Adv. Mater. 28, 5878 (2016). Copyright 2016 Wiley-VCH. (h) The schematic diagram of the spatial summation with two spatial isolated synapses. (i)
The measured sum (SM) plotted as a function of expected sum (SE).179 Reproduced with permission from Wan et al., Adv. Mater. 28, 3557 (2016). Copyright 2016 Wiley-VCH.
(j) Schematic of highly interconnected P3HT neural devices with an ion-gel membrane as both an electrolyte and a substrate. (k) Simulation of Pavlov’s learning.180
Reproduced with permission from Fu et al., ACS Appl. Mater. Interfaces 10, 26443 (2018). Copyright 2018 American Chemical Society. (l) The proposed multigated architec-
ture of analogous artificial MoS2 synapses. (m) Controlled facilitation and depression of synaptic weights with the electroiono-photoactive multigated architecture.59
Reproduced with permission from John et al., Adv. Mater. 30, 1800220 (2018). Copyright 2018 Wiley-VCH.

spike voltages and sublinear for high spike voltages were necessary for the electrolyte-gated neuromorphic system. For example, classical con-
the exponential function and logarithmic function, respectively.184 ditioning according to the Pavlovian associative learning rule has been
The nature and geometry of the gate electrode can be regarded as implemented in a multiterminal P3HT synaptic transistor.180 The
a key factor for designing the initial synaptic weight distribution in topology of multi-inputs to one-output was mapped with lateral gates

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-12

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

through solid-state ion-gel consisting of poly(vinylidenefluoride- (SLP) comprises an input layer and an output layer and can perform
cohexafluoropropylene) (PVDF-co-HFP) and 1-ethyl-3-methylimida- binary classification and thus solve linearly separable problems. The
zolium bis-(trifluoromethylsulfonyl)imide ([EMI]þ[TFSA]) ionic liq- capacity of ANNs can be further enhanced by introducing additional
uid [Fig. 5(j)]. Based on protonic electrochemical doping/dedoping intermediate layers (i.e., hidden layers) to construct multilayer percep-
processes at P3HT/ion gel interfaces, four types of STDP learning trons (MLPs). Thus, MLP networks are capable of solving nonlinear
behaviors of Hebbian STDP, anti-Hebbian STDP, symmetrical STDP, separable problems for multiclass classification.192 The performance of
and visual STDP were successfully mimicked on a single neuromor- hardware-based ANNs has been simulated with some specific learning
phic transistor. The classical conditioning was demonstrated as shown algorithms, mainly including the backpropagation (BP) method in
in Fig. 5(k), and four 3 V pulse signals (tp ¼ 50 ms and Dt ¼ 50 ms) which the synaptic weights are iteratively adjusted, while error back
were applied to the G2 electrode to simulate the unconditioned stimu- propagates from the output to input.137,193,194 In these simulations, the
lus of “bell ringing,” which induced a smaller change in the synaptic channel conductance change (DG) was used as the weight update for
weight value (DWpeak) compared to the defined threshold (45 lA, executing the learning algorithm.195 Since the weights represent the
green dotted line), corresponding to the absence of salivary response. synaptic strength, these processes are commonly referred to as potenti-
When four 5 V pulse signals (tp ¼ 50 ms and Dt ¼ 50 ms) were ation and depression in synaptic devices, inspired from the biological
applied to the G1 electrode to simulate the conditioned stimulus of terminology.196 In terms of computing, cognitive tasks like pattern rec-
“sight of food,” a DWpeak value larger than the threshold was triggered, ognition and classification demand synaptic devices with a wide
which corresponds to the salivary response. Then, four 3 and 5 V dynamic linear conductance range131 and distinguishable readout
input signals were applied simultaneously to the G1 and G2 electrodes states.130 The symmetric weight update along with low write noise and
to simulate the training process. After training, a salivary response low switching voltages and currents determines the efficiency of cross-
could be produced even with four 3 V pulse signals (bell ringing) bar neuromorphic computational kernels and significantly improves
due to the nonvolatile feature. Therefore, an effective link between the the classification accuracy of backpropagation schemes.137,189 Owing
input signals applied to the G1 and G2 electrodes was established, to the decoupled input and output electrodes, as well as the symmetry
demonstrating that dynamic processes of memorizing were incorpo- between the material at the gate and the channel, synaptic EGTs can
rated into the gate matrix simulation. achieve gradual nonvolatile conductance changes in a range, which
The coexistence of multiple forms of synaptic plasticity would shows nearly linear and symmetric weight update.137 In addition, the
increase the processing capability and memory storage capacity of large specific capacitance enables low energy consumption, making
neuromorphic transistors.126 With nonplanar and multigated architec- EGTs as important building blocks favorable for large-scale, energy-
tures, neuromorphic EGTs could operate independently either in a efficient neuromorphic computing networks.
pristine ionotronic mode, an electronic mode, or a photoactive Van de-Burgt et al. reported the simulations of the neural net-
mode.59,185,186 This configuration enables a higher order of plasticity work using a PEDOT:PSS-based electrochemical neuromorphic
that emulates the effects of neuromodulators such as dopamine or organic device (ENODe) gated with a proton conductor Nafion
noradrenaline, also called heterosynaptic plasticity or three-factor [Fig. 6(a)].58 As shown in Fig. 6(b), the ENODes showed a battery-like
learning.187,188 More importantly, they can operate synergistically in a operation by decoupling the read and write operations with an exter-
dual-gated additive/subtractive mode, allowing the programming of nal switch. This results in enhanced state stability and analog tuning.
weight changes with fine precision, with the net weight change defined Hence, a low energy switching (<10 pJ for 103 lm2 devices) and long
by the overall capacitive coupling across the semiconducting chan- retention times (25 h with a 0.04% standard deviation in conductance)
nel.189 Mathews et al. reported the synergistic gating effect of the elec- could be achieved simultaneously. In an 1 V range, long-term poten-
tro-iono-photo in MoS2 EGTs [Fig. 5(i)].59 This 2D channel-based tiation and depression displaying 500 discrete and nonvolatile conduc-
EGT addressed different levels of charge-trapping probabilities to tance states were obtained by applying 500 6 1.5 mV (1 s) presynaptic
finely tune the synaptic weights: for example, electron trapping- pulses [Fig. 6(c)]. According to the statistical distribution of 15 000
detrapping at the SiO2/MoS2 interface though the back-gate (Si) was experimentally measured conductance levels, the numerical weights
responsible for the electronic-mode, ion migration–relaxation kinetics showed small nonlinearity (NL) and low noise (<1%), meeting the
at the ionic liquid/MoS2 interface although the top-gate (Beryllium requirements for high ANN training accuracy when executing
copper probe) accounted for the ionotronic-mode, and persistent pho- updates. A three-layer network (a single-layer perceptron with one
toconductivity though an additional light gate was used in the photo- hidden layer) based ANN was simulated with the backpropagation
active-mode.94 The three modes were combined to modulate Hebbian method. The simulated ENODe-based learning circuit achieved a high
STDP plasticity with metaplasticity and homeostatic regulation [Fig. classification accuracy between 93% and 97% for three types of image
5(m)]. Classical conditioning was emulated using simultaneous paired recognition: an 8  8 pixel image version of handwritten digits;
stimulation of unconditioned optical and conditioned voltage pulses, Modified National Institute of Standards and Technology (MNIST), a
resulting after the training process in associative learning. 28  28 pixel version of handwritten digits; and a Sandia file classifica-
tion dataset. Moreover, plastic ENODes fabricated on PET substrates
D. Perceptron-based artificial neural network enable the potential integration of neuromorphic functionality in flexi-
Perceptron is a model of feedforward neural networks that com- ble electronic systems.
prise densely interconnected adaptive processing subunits (artificial Shang et al. reported the simulation of handwritten digit recogni-
neurons).190 Perceptron-based artificial neural networks (ANNs) are tion using a 2D layered a-phase molybdenum oxide (a-MoO3)-based
intended to emulate the brain’s ability to recognize and distinguish the synaptic EGT array gated by a PEO:LiClO4 electrolyte [Fig. 6(d)].101
difference between objects or meanings.191 A single-layer perceptron Voltage pulses with higher amplitudes (>62 V) or long duration

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-13

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 6. (a) Schematic of the PEDOT:PSS-based flexible all solid-state neuromorphic device. (b) Schematic explaining the decoupling of the read and write operations. (c) LTP
and LTD displaying 500 discrete states.58,214 The inset shows a zoom-in image showing the individual states. Reproduced with permission from van de-Burgt et al., Nat. Mater.
16, 414 (2017). Copyright 2017 Springer Nature Publishing. (d) Schematics of a three layer (one hidden layer) neural network. (e) Schematics of a synaptic weight layer com-
posed of voltage programmed Li-ion synaptic transistor crossbar array and access devices. (f) The recognition accuracy evolution with training epochs for the 28  28 pixel
handwritten digit image. The inset shows the asymmetric ratio (AR) between the LTP and LTD and the temporal variation are calculated over 50 cycles.101 Reproduced with
permission from Yang et al., Adv. Funct. Mater. 28, 1804170 (2018). Copyright 2018 Wiley-VCH. (g) The IFG array is mapped to a three-layer neural network used to classify
XOR logic. (h) G modulation read current <10 nA while maintaining a high signal-to-noise ratio during nearly linear and symmetric programming. (i) XOR classification func-
tion.74 Reproduced with permission from Fuller et al., Science 364, 570 (2019). Copyright 2019 AAAS.

times (10 ms) induced the reversible intercalation of Liþ dopants into 10 lm. The calculated asymmetric ratio (AR) value of the synaptic
the a-MoO3 lattice, forming molybdenum bronze (LixMoO3) and transistor was as small as 0.31 6 0.12 [the inset in Fig. 6(f)]. The cycle-
leading to a nonvolatile DG. By alternatively applying 50 identical to-cycle variation, i.e., write noise, was <12% in 15 a-MoO3 EGT
pulses (62.5 V, 10 ms) with 10 s interval, bidirectional analog switch- nanosheets with the channel thickness of 16.8–28.0 nm, indicating
ing was obtained, where the near-linear channel conductance was set good device-to-device uniformity. Based on the experimentally mea-
to numerous states between 42 and 75 nS in both the LTP and LTD sured LTPo and LTD characteristics, a three-layer network was used
processes. This low conductance was promising for the application of to perform supervised learning with BP methods. Figure 6(e) schemat-
neuromorphic networks composed of large-scale device arrays. The ically shows a crossbar array. The simulated recognition accuracy
minimum energy consumption value for a nonvolatile DG was 1.8 pJ approached 87.3% for a large image (28  28 pixels) of handwritten
for a single pulse (2.5 V, 10 ms) event with the channel length of digits after 40 training epochs [Fig. 6(f)]. Park et al. reported a

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-14

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

photonic synaptic EGT with a mixed weight updating mechanism for sensitive for interfacing biological substances and signals.198,199 Inspired
the simulation of high-speed and low-power optic-neural networks.92 by the fact that nervous systems can sense, process, memorize, and clas-
GO nanosheets modified with long alkyl chains (alkylated GO) were sify various external stimuli, synaptic EGTs have been developed to
embedded as the charge-trapping sites between the ion-gel blocking mimic stimulisensitive artificial synapses and smart sensorimotor nerve-
gate electrolyte and the IGZO semiconducting channel. With the assis- tronics.64,200 Synaptic EGTs could enable potential applications in
tance of a light pulse, a larger conductance change ratio (Gmax/Gmin) adaptable biointerfacing, neuroinspired actuation, with long-term appli-
without the degradation of the nonlinearity property was obtained and cations such as local diagnosis and treatment through closed-loop con-
the recognition rate of MNIST training patterns was improved from trol of biological environment, and local control of a neural activity.67
49% to 62% with 100 weight states. To further improve the recognition
accuracy, more conductance states and a higher ratio of Gmax/Gmin 1. Synaptic coupling of living neurons
were required through adjusting the gate pulse duration and/or ampli-
tude but without sacrificing the linearity. A promising solution is to In the field of bioelectronics, EGTs have been interfaced with
improve the ion capacity of the channel materials and the conductance neuronal cells for biochemical signal recording and transduction of
sensitivity to doping. The nonlinear potential drop between the series bioelectrical signals from cells and tissues.86,201 The amplitude of the
resistance also affects the linearity.58 intracellular action potential (AP) is in the range of few tens of milli-
Emelyanov et al. demonstrated experimentally a double hidden lay- volt, and it is about 2 orders of magnitude larger than the amplitude of
ered ANN to solve a nonlinearly separable task (implementation of an the extracellular potential (few hundreds lV).202 Owing to the large
XOR logic gate).197 In their synaptic EGTs, HCl p-doped PANI was used specific capacitance/transconductance, EGTs have a unique advantage
as the channel materials and interfaced with a PEO:LiClO4 solid electro- of being able to sense ultralow spike amplitudes, which are in the range
lyte. In the simulation, every decision boundary was performed by one of intracellular APs. Functional coupling of living neurons through
neuron in the two hidden layers, and the points inside the triangle-like artificial synapses is the primary requirement for their implementation
class corresponded to class 1, while the others corresponded to class 0. as prosthetic devices or in building hybrid networks.203,204 Vuillaume
The double hidden layers allowed the perceptron to classify not only et al. reported a pentacene/AuNP-based synapse transistor interfaced
“black” (logic 0) and “white” (1) classes but also “gray” ones (some range with neurons [Fig. 7(a)].115 The NaCl-gated pentacene/AuNP EGTs
of signal amplitude between logics 0 and 1). The possible position of deci- showed short-term facilitating and depressing behavior when chang-
sion boundaries between classes could be well defined after the learning ing the input spike frequency at a low spike voltage of 50 mV. Human
procedure. The ability of artificial neurons to classify the “exclusive or” neuroblastoma stem cells (SH-SY5Y) were then adhered, grown, and
(XOR) logic function is a basic demonstration of the nonlinear separable differentiated into neurons on top of the 15 nm pentacene channel
task. Fuller et al. experimentally implemented this function into a 3  3 [Fig. 7(b)]. The STP response was monitored before and after cell
ionic floating-gate memory array (IFG).74 The memory array was based growth. When cells differentiated into neurons for day 6 [Fig. 7(c)],
on a PEDOT:PSS EGT connected to a selector device of two-terminal there was a detectable amplitude change in the STP response.
volatile conductive-bridge memory (CBM). This CBM permitted the A synapse is a biological structure, which connects two neurons
individual addressing of redox transistors in a crossbar configuration (at enabling specific and unidirectional information flow (excitation or
the circuit level, not the physical crossbar configuration). Determined by inhibition) from one neuron to another. In this regard, Erokhin et al.
the competition between write gate voltage (VW) and the ON threshold demonstrated this functional interface between two living neurons in
Vth of the CBM, electron injection (extraction) through the CBM into the rat brain slices via a PANI-based synaptic EGT [Fig. 7(d)].75 Initially,
top PEDOT:PSS gate results in the reversible electrochemical oxidation APs evoked by suprathreshold depolarizing current injection in either
(reduction) of the bottom PEDOT:PSS channel,58 thereby increasing neuron failed to evoke any response in another cell of the pair, indicat-
(decreasing) G. The redox transistors operated similar to a flash memory ing that this pair of cells was not connected by natural synapses in either
but with more than an order of magnitude lower voltage operation and direction. The two neurons were then connected through an electronic
weight readout <10 nA while maintaining a high signal-to-noise ratio. In circuit with a PANI synaptic EGT to play the role of a synapse analog.
addition, the devices showed finely spaced conductance levels with near- As the resistance of the PANI channel reduced upon depolarization, the
ideal analog behavior [Fig. 6(g)]. Moreover, the downscaled devices could consecutive depolarizing steps and the APs in “presynaptic” cell 1
enable a remarkable >1-MHz write-read frequency (<1 ls) for the induced a gradual increase in voltage responses from the PANI EGT
polymer-based redox transistors. For the demonstration of the XOR logic [Fig. 7(e), plot 3] and in “postsynaptic” cell 2 (plot 4). When the depola-
function, an input example X ¼ [1, 0] [blue in Fig. 6(h)] was fed to the rizing response in cell 2 reached the AP threshold (40 mV) at sweep
first layer of the network (orange), whereas the output YT ¼ [1, 1], where #113 (plot 4), cell 2 started to reliably fire APs. Moreover, the PANI
T denotes the matrix transpose, was sent to the last layer (green). The synapses efficiently supported synchronized delta-oscillations in this
final output of the network correctly classified X ¼ [1, 0] as ZT ¼ [1] two-neuron network. As a result, the unidirectional, activity-dependent
according to the XOR truth table. The network was used to execute ana- coupling of living neurons through an organic synaptic EGT was real-
log dot products during inference in a 3  3 prototype array with 100% ized. For further perspective of implantable prosthetic synapses, one
accuracy [Fig. 6(i)]. Such highly efficient neuromorphic computers could must fulfill requirements such as size scaling, biocompatibility, flexibil-
extend ANN learning to new low-power platforms. ity, and stretchability of the artificial synapse.

E. Neuromorphic sensing and biointerfacing 2. Artificial sensory systems

EGTs offer high and tunable transconductance to electrostatic Neuromorphic EGTs are physical computing nodes that can
potential changes and also to electrochemical phenomena, making them directly interface with analog signals from sensors, enabling the

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-15

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 7. (a) Optical bright field images of SH-SY5Y cells grown on interdigitated Au electrodes coated by a pentacene/AuNP thin-film. (b) Immuno-fluorescence images of SH-
SY5Y differentiation on the synaptic EGT at day 6. (c) STP recorded as a function of electrical spike frequency (after cell differentiation).115 Reproduced with permission from
Desbief et al., Org. Electron. 38, 21 (2016). Copyright 2016 Elsevier. (d) Infrared differential interference contrast microphotograph of a P7 rat brain slice with visually identified
L5/6 neocortical cells (cell 1,2) recorded simultaneously. (e) Activity-dependent coupling of two cortical neurons by PANI-based synaptic EGT.75 Reproduced with permission
from Juzekaeva et al., Adv. Mater. Technol. 4, 1800350 (2019). Copyright 2019 Wiley-VCH.

efficient construction of artificial sensory systems with integrated sens- tactile-perception system (DOT-TPE) with integrated sensing and
ing/actuation and signal processing functionalities. Since skin is the neuron-like information-processing functionalities of pressure stimuli
largest organ of the human body, which contains a variety of sensors, [Fig. 8(a)].205 The proposed DOT-TPE consisted of both a pressure-
it offers the ability to perceive and interact with the surrounding envi- sensing organic field-effect transistor (OFET) and a signal-processing
ronment.154,200 Zang et al. reported a dual-organic-transistor-based OFET using poly(diketopyrrolopyrrole-terthiophene) (PDPP3T) as

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-16

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

FIG. 8. (a) Equivalent electrical circuit for the DOT-TPS. (b) The relative changes in current in the pressure sensing device and the Ipost responses of the synaptic transistor
under different pressures.205 The inset shows the schematic illustrations of the tactile speed-perception functionality of the DOT-TPE array. Reproduced with permission from
Zang et al., Adv. Mater. 29, 1606088 (2017). Copyright 2017 Wiley-VCH. (c) Schematic illustrations of the hybrid reflex arc made of an artificial afferent nerve and a biological
efferent nerve. (d) Braille reading functionality of the artificial afferent nerve system.206 Reproduced with permission from Kim et al., Science 360, 998 (2018). Copyright 2018
AAAS. (e) Photograph of a 5  6 artificial visual-perception system array. Scale bar, 5 mm. (f) Color recognition functionality of the ferroelectric/electrochemical modulated artifi-
cial visual-perception system (Left: NIR, Right: green).136 Top is the signal mapping immediately recorded after light exposure, and bottom is the remnant signals after 1800 s.
Reproduced with permission from Wang et al., Adv. Mater. 30, 1803961 (2018). Copyright 2018 Wiley-VCH. (g) Schematic illustrations of the organic optoelectronic synapse
and neuromuscular electronic system.107 Reproduced with permission from Lee et al., Sci. Adv. 4, eaat7387 (2018). Copyright 2018 AAAS.

the channel layer. Chitosan was selected as the electrolyte to establish with a larger Victor-Purpura distance (DVP) of 300. Furthermore,
the proton–electron coupling at the semiconductor/dielectric interface. the artificial afferent nerve was connected to the biological efferent
The suspended-gate OFET exhibited a high sensitivity of over nerves of a detached cockroach leg to complete a hybrid monosynaptic
50 kPa1, which was sufficient to mimic the tactile sensing properties reflex arc. The flow of information from multiple pressure (amplitude
of human skin (8–192 kPa1). The pressure sensor acted as the signal and frequency) triggered the isometric contraction actuation of the tib-
transduction element for converting the external pressure into presyn- ial extensor muscle accordingly, indicating the successful emulation of
aptic current pulses. Due to the instantaneous change in the potential a biological reflex arc. Recently, Chen et al. developed a self-powered
drop across the sensing device in the DOT-TPE, the triggered presyn- ion gel-gated PDVT-10 synaptic transistor that mimicked the tactile
aptic signals were processed simultaneously by monitoring the output synapse.207 This synaptic transistor was actuated by a triboelectric
current Ipost of the synaptic EGT [Fig. 8(b)]. Hence, the cognitive tac- nanogenerator (TENG) that functioned as the tactile sensor to realize
tile information was collected, and the increase in the strength and the self-powered tactile synapse. The self-powered synaptic system was
duration of the sustained pressure caused an increase in Ipost. By taking able to realize AND and OR logic functions and Pavlovian condition-
advantage of the short-term synaptic facilitation function of DOT- ing. The biomimetic functionality of tactile perception systems, com-
TPE, 3  3 pixel arrays were built to mimic the tactile perception of bined with their promising features of flexibility and large-area
dynamic mechanical contact. Tactile perception with comprehensive fabrication, represents a step forward for pressure-sensory synapses
information on pressure strength and frequency was demonstrated towards novel E-skins for neuromorphic sensing and actuation.
[inset in Fig. 8(b)]. Kim et al. reported a bioinspired flexible organic Photostimulated synapses using light as the input signals not
artificial afferent nerve made of a resistive pressure sensor, an organic only combine visual perception, information processing, and memory
ring oscillator, and an ion gel-gated conjugated polymer-based synap- together but also provide advantages such as higher bandwidth,
tic transistor [Fig. 8(c)].206 The artificial afferent nerve could collect robustness, and potential for parallelism, which is suitable for simulat-
pressure signals (1–80 kPa) and converted them into action potentials ing retinal neurons.208–210 The transduction of the color and intensity
(0–100 Hz) by using ring oscillators. The function of braille character of the incident light into neural signals is a main process for visual per-
reading was implanted in this artificial afferent nerve by connecting ception. Guo et al. reported a ferroelectric/electrochemical modulated
three synaptic transistors with two ring oscillators and a 2  3 array of organic synaptic EGT for an ultraflexible, artificial visual-perception
pressure sensors [Fig. 8(d)]. The synaptic transistor integration could system [Fig. 8(e)].136 A p-type copolymer poly(isoindigo-co-bithio-
significantly improve the discrimination among the braille characters phene) [P(IID-BT)] was used as the channel material, and bilayered

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-17

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

poly(vinylidenefluoride-cotrifluoroethylene) [P(VDF-TrFE)] and First, a more profound insight should be obtained into the ion
poly[(1-vinylpyrrolidone)-co-(2-ethyldimethylammonioethyl methac- transport/trapping related device physics to ensure stable and reliable
rylate ethyl sulfate)] [P(VP-EDMAEMAES)] acted as the polyelectro- neuromorphic functionalities. The most fundamental synaptic behaviors
lytes. Besides the fundamental STP/LTP operations by electrochemical of STP and LTP are selectively induced by controlling the degree of
doping, the large EDL capacitance allowed P(VDF-TrFE)’s polariza- device volatility through hybrid mechanisms. The LTP behavior suffers
tion switching, to achieve an extra ferroelectric LTP function with reli- in many cases from challenges in large structural transformations of
able nonvolatility. This synaptic EGT was integrated with an organic channel materials and/or parasitic oxidation reactions, leading to perfor-
light-sensitive electronic component to construct an artificial visual- mance degradation. Novel channel materials and gating mechanisms are
perception system. This light sensory system displayed an incident highly desired to maintain electronic structures for the implementation
light intensity and frequency dependent electrochemical LTP to ferro- of LTP functions. Next to the simulated synaptic behavior, additional
electric LTP transition. A 5  6 pixel array was demonstrated to allow metrics should be systematically reported to evaluate the reconfiguration
for color recognition by converting photons from different wave- ability, device stability and (cycling) lifetime, and endurance.
lengths with identical intensity (10.80 mW/cm2) into volatile (850 nm, An increase in the write-read speed and a decrease in energy con-
64 Hz) and nonvolatile (550 nm, 64 Hz) synaptic signals, respectively sumption are needed to be able to build neuromorphic networks. The
[Fig. 8(f)]. The wavelength-recognition functionality characterized by operation speed of EGTs is mainly limited to kilohertz (millisecond),
signal’s degree of nonvolatility was attributed to the higher-energy with the exception of a few examples, by the ion drift and diffusion in
green light-triggered extra ferroelectric LTP compared to the electro- the gate electrolyte and/or the channel. Selecting gate electrolytes and
chemical only LTP with near infrared (NIR) irradiation. Lee et al. channel materials with high ion mobility to shorten the ion migration
reported an optoelectronic sensorimotor synapse based on a stretch- time are currently suggested. The speed could also be increased by scaling
able organic nanowire synaptic transistor (s-ONWST), integrated with down device dimensions, especially by reducing the channel and electro-
a photodetector and a neuromuscular system [Fig. 8(g)].107 The ion lyte thickness to shorten the ion diffusion distance. When the device is
gel-gated fused thiophene diketopyrrolopyrrole (FT4-DPP)-based syn- scaled down to sub-100 nm scale, it is possible to increase the switching
aptic EGT was built onto a 100% prestrained styrene ethylene butylene speed to megahertz (microsecond), which can also decrease the power
styrene (SEBS) rubbery substrate to introduce the stretchability. Each consumption to the subfemtojoule level for a single spike event.
visible light pulse induced an output presynaptic voltage of 1.1 V Furthermore, internal ion-gated electrochemical transistors, with ions
from the organic photodetector to drive the s-ONWST. Spike duration that are preloaded on the channel, have shown faster operation speeds,
but their applicability in the neuromorphic device configuration should
and number-dependent plasticity were similar in the stretchable
be further investigated.213 Vertical configurations of transistors can also
s-ONWST at 100% strain to those in the device at 0% strain. The
further reduce ion diffusion time, compared with a planar geometry.
optoelectronic sensorimotor system was then used for optical wireless
Moreover, the high-density integrated circuits (ICs) of synaptic
communication. The International Morse code was coded in the form
EGTs suffer from device-to-device uniformity (e.g., device noise, non-
of patterns of visible light to trigger the postsynaptic potentiation of
linearity, and asymmetry) and environmental dependence (e.g.,
the s-ONWST, and every letter of the English alphabet was decoded
humidity > 50%). In addition to the optimization of thin film prepara-
according to the sum of EPSC amplitude peak values. Photoelectric
tion techniques, cumulative probability and related statistics of chan-
neuromorphic devices are promising for the development of next-
nel conductance change (DG) extracted from a large amount of
generation human–machine interface applications, soft robotics,
experimental measurements should be considered to properly simulate
neurorobotics, and electronic prostheses.211,212 device nonidealities. Another issue is the requirement of a liquid elec-
trolyte and humidity ambience in some synaptic devices. From a tech-
IV. SUMMARY AND PERSPECTIVES
nological point of view, liquid gates are difficult to scale below 100 nm
Overall, electrolyte-gated transistors are promising candidates for dimensions or to integrate with the existing solid-state circuits. From
synaptic electronics and neuromorphic computing owing to their an application point of view, solid electrolyte-based synaptic EGTs
mixed ionic/electronic transduction, physiological environmental show that these platforms can be downscaled for space-efficient elec-
compatibility, mechanical/architectural flexibility, and in particular tronic devices, while studies performed with a liquid electrolyte show
low energy consumption. Basic synaptic functions have been success- their potential applications for interfacing with biology, such as
fully mimicked. Unique functions arising from the inherent properties human/machine interaction through neuromorphic sensing/actuation,
of electrolyte-gated devices are also demonstrated, including global personalized healthcare, and artificial nerves and organs.
regulation phenomena, coplanar coupling, and integration with free- Neural network training algorithms suitable for such device net-
standing active substrates and with living neurons, among other exam- works and supporting (external) hardware circuits should be further
ples. A variety of complex tasks, such as dendritic integration, classic developed. The single-layer perceptron (SLP) model is the simplest
conditioning, pattern recognition, and sensory function, are realized in kind of neural network and has been widely employed to implement
EGT-based arrays and circuits. basic training and learning. Nevertheless, more complex neural net-
Currently, emulating the synapse behavior is still the main focus works are required to solve demanding tasks. Toward this direction,
in this emerging field. Some challenges (reconfigurability, stability, reli- there are only scarce attempts of hardware multilayer perceptrons
ability, speed, power consumption, and supporting circuitry such as with EGTs. Apart from feedforward networks, recurrent networks
selector devices) are being faced for the realization of large-scale with EGTs are yet to be explored.
ANNs using EGTs. The current scientific challenges and perspectives In the long-term, materials science and neuroscience should go
can be summarized as follows: hand in hand in order to define the computational primitives of the

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-18

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

21
brain, which are necessary for processing, and up to which level of Q. Xia and J. J. Yang, Nat. Mater. 18(4), 309–323 (2019).
22
complexity these primitives have to be emulated at the device level. Y. van de Burgt, A. Melianas, S. T. Keene, G. Malliaras, and A. Salleo, Nat.
Electron. 1(7), 386–397 (2018).
This bidirectional interaction between the two disciplines should be 23
Z. Y. Wang, L. Y. Wang, M. Nagai, L. H. Xie, M. D. Yi, and W. Huang, Adv.
extended beyond the functional device implementation of the synapses Electron. Mater. 3(7), 1600510 (2017).
and neurons. 24
L. Chua, IEEE Trans. Circuit Theory 18(5), 507–519 (1971).
25
L. O. Chua and K. S. Mo, Proc. IEEE 64(2), 209–223 (1976).
26
ACKNOWLEDGMENTS D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, Nature
453(7191), 80–83 (2008).
27
F. Yan wishes to thank the Research Grants Council (RGC) of J. J. Yang, M. D. Pickett, X. Li, D. A. A. Ohlberg, D. R. Stewart, and R. S.
Hong Kong (Project No. C5015-15G) and the Hong Kong Polytechnic Williams, Nat. Nanotechnol. 3, 429 (2008).
28
J. G. Simmons, R. R. Verderber, and N. F. Mott, Proc. R. Soc. London, Ser. A
University (Project Nos. 1-ZVGH, G-YBJ0, and 1-ZVK1). H. F. Ling 301(1464), 77–102 (1967).
wishes to thank the Alexander von Humboldt Foundation for 29
T. W. Hickmott, J. Appl. Phys. 33(9), 2669–2682 (1962).
financial support via Humboldt Research Fellowships for Postdoctoral 30
R. Waser and M. Aono, Nat. Mater. 6(11), 833–840 (2007).
31
researchers. H. F. Ling also acknowledges the support of the National S. P. Adhikari, M. P. Sah, H. Kim, and L. O. Chua, IEEE Trans. Circuits Syst. I
Natural Science Foundation of China (No. 61905121), the Natural 60(11), 3008–3021 (2013).
32
Science Foundation of Jiangsu Province, China (No. BK20190734), I. Valov, E. Linn, S. Tappertzhofen, S. Schmelzer, J. van den Hurk, F. Lentz,
and R. Waser, Nat. Commun. 4(1), 1771 (2013).
and Nanjing University of Posts and Telecommunications Start-up 33
T. Ohno, T. Hasegawa, T. Tsuruoka, K. Terabe, J. K. Gimzewski, and M.
Fund (No. NY219157). Y. van de Burgt acknowledges funding from Aono, Nat. Mater. 10(8), 591–595 (2011).
European Union’s Horizon 2020 Research and Innovation 34
Z. Wang, S. Joshi, S. E. Savel’ev, H. Jiang, R. Midya, P. Lin, M. Hu, N. Ge, J. P.
Programme (Grant Agreement No. 802615). Strachan, Z. Li, Q. Wu, M. Barnell, G.-L. Li, H. L. Xin, R. S. Williams, Q. Xia,
and J. J. Yang, Nat. Mater. 16, 101 (2017).
35
Y. Yang, P. Gao, L. Li, X. Pan, S. Tappertzhofen, S. Choi, R. Waser, I. Valov,
REFERENCES and W. D. Lu, Nat. Commun. 5(1), 4232 (2014).
1 36
M. A. Zidan, J. P. Strachan, and W. D. Lu, Nat. Electron. 1(1), 22–29 A. Sengupta and K. Roy, Appl. Phys. Rev. 4(4), 041105 (2017).
37
(2018). X. Zhu, D. Li, X. Liang, and W. D. Lu, Nat. Mater. 18(2), 141–148 (2019).
2 38
K. Duygu, Y. Shimeng, and H. S. P. Wong, Nanotechnology 24(38), 382001 X. Sun, P. Wang, K. Ni, S. Datta, and S. Yu, paper presented at the 2018 IEEE
(2013). International Electron Devices Meeting (IEDM), 2018.
3 39
D. Ielmini and H. S. P. Wong, Nat. Electron. 1(6), 333–343 (2018). B. C. Jang, S. Kim, S. Y. Yang, J. Park, J.-H. Cha, J. Oh, J. Choi, S. G. Im, V. P.
4
C. D. James, J. B. Aimone, N. E. Miner, C. M. Vineyard, F. H. Rothganger, K. Dravid, and S.-Y. Choi, Nano Lett. 19(2), 839–849 (2019).
40
D. Carlson, S. A. Mulder, T. J. Draelos, A. Faust, M. J. Marinella, J. H. Naegle, G. Milano, S. Porro, I. Valov, and C. Ricciardi, Adv. Electron. Mater. 5(9),
and S. J. Plimpton, Biol. Inspired Cognit. Archit. 19, 49–64 (2017). 1800909 (2019).
5 41
C. D. Harvey and K. Svoboda, Nature 450, 1195 (2007). D. Kuzum, R. G. D. Jeyasingh, B. Lee, and H. S. P. Wong, Nano Lett. 12(5),
6
L. F. Abbott and S. B. Nelson, Nat. Neurosci. 3(11), 1178–1183 (2000). 2179–2186 (2012).
7 42
L. F. Abbott and W. G. Regehr, Nature 431(7010), 796–803 (2004). A. F. Vincent, J. Larroque, N. Locatelli, N. B. Romdhane, O. Bichler, C.
8
M. A. Lynch, Physiol. Rev. 84(1), 87–136 (2004). Gamrat, W. S. Zhao, J. Klein, S. Galdin-Retailleau, and D. Querlioz, IEEE
9
D. Hebb, The Organization of Behavior (Wiley, New York, 1949). Trans. Biomed. Circuits Syst. 9(2), 166–174 (2015).
10 43
D. O. Hebb and D. Donderi, Textbook of Psychology (Psychology Revivals) S. Boyn, J. Grollier, G. Lecerf, B. Xu, N. Locatelli, S. Fusil, S. Girod, C.
(Psychology Press, 2013). Carretero, K. Garcia, S. Xavier, J. Tomas, L. Bellaiche, M. Bibes, A.
11
M. Bear, L. Cooper, and F. Ebner, Science 237(4810), 42–48 (1987). Barthelemy, S. Sa€ıghi, and V. Garcia, Nat. Commun. 8(1), 14736
12
N. Caporale and Y. Dan, Annu. Rev. Neurosci. 31(1), 25–46 (2008). (2017).
13 44
T. Masquelier, R. Guyonneau, and S. J. Thorpe, Neural Comput. 21(5), A. Wedig, M. Luebben, D.-Y. Cho, M. Moors, K. Skaja, V. Rana, T.
1259–1276 (2009). Hasegawa, K. K. Adepalli, B. Yildiz, R. Waser, and I. Valov, Nat.
14
G. Rachmuth, H. Z. Shouval, M. F. Bear, and C.-S. Poon, Proc. Natl. Acad. Sci. Nanotechnol. 11, 67 (2016).
45
108(49), E1266–E1274 (2011). H. F. Ling, K. M. Tan, Q. Y. Fang, X. S. Xu, H. Chen, W. W. Li, Y. F. Liu, L.
15
C. Diorio, P. Hasler, A. Minch, and C. A. Mead, IEEE Trans. Electron Devices Y. Wang, M. D. Yi, R. Huang, Y. Qian, L. H. Xie, and W. Huang, Adv.
43(11), 1972–1980 (1996). Electron. Mater. 3(8), 1600416 (2017).
16 46
P. E. Hasler, C. Diorio, B. A. Minch, and C. Mead, paper presented at the W. Zhongrui, R. Mingyi, M. Rivu, J. Saumil, J. Hao, L. Peng, S. Wenhao, A.
Advances in Neural Information Processing Systems, 1995. Shiva, Z. Ye, L. Can, W. Huaqiang, X. Qiangfei, and Y. J. Joshua, Adv. Funct.
17
G. Indiveri, B. Linares-Barranco, T. Hamilton, A. van Schaik, R. Etienne- Mater. 28(6), 1704862 (2018).
47
Cummings, T. Delbruck, S.-C. Liu, P. Dudek, P. H€afliger, S. Renaud, J. Schemmel, I. Valov, I. Sapezanskaia, A. Nayak, T. Tsuruoka, T. Bredow, T. Hasegawa, G.
G. Cauwenberghs, J. Arthur, K. Hynna, F. Folowosele, S. Sa€ıghi, T. Serrano- Staikov, M. Aono, and R. Waser, Nat. Mater. 11(6), 530–535 (2012).
48
Gotarredona, J. Wijekoon, Y. Wang, and K. Boahen, Front. Neurosci. 5, 73 (2011). H. Ling, M. Yi, M. Nagai, L. Xie, L. Wang, B. Hu, and W. Huang, Adv. Mater.
18
P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. 29(35), 1701333 (2017).
49
Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo, I. Vo, S. T. Tsuruoka, K. Terabe, T. Hasegawa, I. Valov, R. Waser, and M. Aono, Adv.
K. Esser, R. Appuswamy, B. Taba, A. Amir, M. D. Flickner, W. P. Risk, R. Funct. Mater. 22(1), 70–77 (2012).
50
Manohar, and D. S. Modha, Science 345(6197), 668–673 (2014). M. Lanza, H. S. P. Wong, E. Pop, D. Ielmini, D. Strukov, B. C. Regan, L.
19
F. Akopyan, J. Sawada, A. Cassidy, R. Alvarez-Icaza, J. Arthur, P. Merolla, N. Larcher, M. A. Villena, J. J. Yang, L. Goux, A. Belmonte, Y. Yang, F. M.
Imam, Y. Nakamura, P. Datta, G. Nam, B. Taba, M. Beakes, B. Brezzo, J. B. Puglisi, J. Kang, B. Magyari-K€ ope, E. Yalon, A. Kenyon, M. Buckwell, A.
Kuang, R. Manohar, W. P. Risk, B. Jackson, and D. S. Modha, IEEE Trans. Mehonic, A. Shluger, H. Li, T.-H. Hou, B. Hudec, D. Akinwande, R. Ge, S.
Comput. Aided Des. 34(10), 1537–1557 (2015). Ambrogio, J. B. Roldan, E. Miranda, J. Su~ ne, K. L. Pey, X. Wu, N. Raghavan,
20
M. Davies, N. Srinivasa, T. Lin, G. Chinya, Y. Cao, S. H. Choday, G. Dimou, E. Wu, W. D. Lu, G. Navarro, W. Zhang, H. Wu, R. Li, A. Holleitner, U.
P. Joshi, N. Imam, S. Jain, Y. Liao, C. Lin, A. Lines, R. Liu, D. Mathaikutty, S. Wurstbauer, M. C. Lemme, M. Liu, S. Long, Q. Liu, H. Lv, A. Padovani, P.
McCoy, A. Paul, J. Tse, G. Venkataramanan, Y. Weng, A. Wild, Y. Yang, and Pavan, I. Valov, X. Jing, T. Han, K. Zhu, S. Chen, F. Hui, and Y. Shi, Adv.
H. Wang, IEEE Micro 38(1), 82–99 (2018). Electron. Mater. 5(1), 1800143 (2019).

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-19

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

51 84
Z. Wang, C. Li, W. Song, M. Rao, D. Belkin, Y. Li, P. Yan, H. Jiang, P. Lin, M. C. R. Newman, C. D. Frisbie, D. A. da Silva Filho, J.-L. Bredas, P. C. Ewbank,
Hu, J. P. Strachan, N. Ge, M. Barnell, Q. Wu, A. G. Barto, Q. Qiu, R. S. and K. R. Mann, Chem. Mater. 16(23), 4436–4451 (2004).
85
Williams, Q. Xia, and J. J. Yang, Nat. Electron. 2(3), 115–124 (2019). D. Tu, L. Kergoat, X. Crispin, M. Berggren, and R. Forchheimer, “Transient
52
P. M. Sheridan, F. Cai, C. Du, W. Ma, Z. Zhang, and W. D. Lu, Nat. analysis of electrolyte-gated organic field-effect transistors,” Proc. SPIE 8478,
Nanotechnol. 12, 784 (2017). 84780L (2012).
53 86
Z. Wang, S. Joshi, S. Savel’ev, W. Song, R. Midya, Y. Li, M. Rao, P. Yan, S. J. Rivnay, P. Leleux, M. Ferro, M. Sessolo, A. Williamson, D. A. Koutsouras,
Asapu, Y. Zhuo, H. Jiang, P. Lin, C. Li, J. H. Yoon, N. K. Upadhyay, J. Zhang, D. Khodagholy, M. Ramuz, X. Strakosas, R. M. Owens, C. Benar, J.-M.
M. Hu, J. P. Strachan, M. Barnell, Q. Wu, H. Wu, R. S. Williams, Q. Xia, and Badier, C. Bernard, and G. G. Malliaras, Sci. Adv. 1(4), e1400251 (2015).
87
J. J. Yang, Nat. Electron. 1(2), 137–145 (2018). M. Moser, J. F. Ponder, Jr., A. Wadsworth, A. Giovannitti, and I. McCulloch,
54
Y. Yu, Q. Ma, H. Ling, W. Li, R. Ju, L. Bian, N. Shi, Y. Qian, M. Yi, L. Xie, and Adv. Funct. Mater. 29(21), 1807033 (2019).
88
W. Huang, Adv. Funct. Mater. 29, 1904602 (2019). D. A. Bernards and G. G. Malliaras, Adv. Funct. Mater. 17(17), 3538–3544 (2007).
55 89
S. Dai, Y. Zhao, Y. Wang, J. Zhang, L. Fang, S. Jin, Y. Shao, and J. Huang, J. Zhou, Y. Liu, Y. Shi, and Q. Wan, IEEE Electron Device Lett. 35(2),
Adv. Funct. Mater. 29, 1903700 (2019). 280–282 (2014).
56 90
H. Han, H. Yu, H. Wei, J. Gong, and W. Xu, Small 15(32), 1900695 (2019). L. Q. Zhu, C. J. Wan, L. Q. Guo, Y. Shi, and Q. Wan, Nat. Commun. 5, 3158
57
C. D. Danesh, C. M. Shaffer, D. Nathan, R. Shenoy, A. Tudor, M. Tadayon, Y. (2014).
91
Lin, and Y. Chen, Adv. Mater. 31, 1808032 (2019). H. K. Li, T. P. Chen, P. Liu, S. G. Hu, Y. Liu, Q. Zhang, and P. S. Lee, J. Appl.
58
Y. van de Burgt, E. Lubberman, E. J. Fuller, S. T. Keene, G. C. Faria, S. Phys. 119(24), 244505 (2016).
92
Agarwal, M. J. Marinella, A. Alec Talin, and A. Salleo, Nat. Mater. 16, 414 J. Sun, S. Oh, Y. Choi, S. Seo, M. J. Oh, M. Lee, W. B. Lee, P. J. Yoo, J. H. Cho,
(2017). and J. H. Park, Adv. Funct. Mater. 28(47), 1804397 (2018).
59 93
J. R. Abraham, L. Fucai, C. N. Anh, K. M. R. Z. Chao, F. Qundong, B. Q. Wu, J. Wang, J. Cao, C. Lu, G. Yang, X. Shi, X. Chuai, Y. Gong, Y. Su, Y.
Arindam, L. Zheng, and M. Nripan, Adv. Mater. 30(25), 1800220 (2018). Zhao, N. Lu, D. Geng, H. Wang, L. Li, and M. Liu, Adv. Electron. Mater.
60
J. Y. Gerasimov, R. Gabrielsson, R. Forchheimer, E. Stavrinidou, D. T. Simon, 4(12), 1800556 (2018).
94
M. Berggren, and S. Fabiano, Adv. Sci. 6(7), 1801339 (2019). M. Lee, W. Lee, S. Choi, J. W. Jo, J. Kim, S. K. Park, and Y. H. Kim, Adv.
61
J.-T. Yang, C. Ge, J.-Y. Du, H.-Y. Huang, M. He, C. Wang, H.-B. Lu, G.-Z. Mater. 29(28), 1700951 (2017).
95
Yang, and K.-J. Jin, Adv. Mater. 30(34), 1801548 (2018). R. A. John, J. Ko, M. R. Kulkarni, N. Tiwari, N. A. Chien, N. G. Ing, W. L.
62
Z. Zhang, T. Li, Y. Wu, Y. Jia, C. Tan, X. Xu, G. Wang, J. Lv, W. Zhang, Y. Leong, and N. Mathews, Small 13(32), 1701193 (2017).
96
He, J. Pei, C. Ma, G. Li, H. Xu, L. Shi, H. Peng, and H. Li, Adv. Mater. 31(3), H. Tian, W. Mi, X.-F. Wang, H. Zhao, Q.-Y. Xie, C. Li, Y.-X. Li, Y. Yang, and
1805769 (2019). T.-L. Ren, Nano Lett. 15(12), 8013–8019 (2015).
63 97
Z. Song, Y. Tong, X. Zhao, H. Ren, Q. Tang, and Y. Liu, Mater. Horiz. 6(4), M. T. Sharbati, Y. Du, J. Torres, N. D. Ardolino, M. Yun, and F. Xiong, Adv.
717–726 (2019). Mater. 30(36), 1802353 (2018).
64 98
D.-G. Seo, Y. Lee, G.-T. Go, M. Pei, S. Jung, Y. H. Jeong, W. Lee, H.-L. Park, Y. Yao, X. Huang, S. Peng, D. Zhang, J. Shi, G. Yu, Q. Liu, and Z. Jin, Adv.
S.-W. Kim, H. Yang, C. Yang, and T.-W. Lee, Nano Energy 65, 104035 Electron. Mater. 5(5), 1800887 (2019).
99
(2019). J. Jie, G. Junjie, W. Xiang, Y. Yi, X. Haipeng, N. Dongmei, Y. Junliang, H. Jun,
65
N. Liu, Y. Liu, J. Hu, Y. He, X. Zhang, and Q. Wan, Appl. Surf. Sci. 481, G. Yongli, and W. Qing, Small 13(29), 1700933 (2017).
100
1412–1417 (2019). D. Xie, J. Jiang, W. Hu, Y. He, J. Yang, J. He, Y. Gao, and Q. Wan, ACS Appl.
66
M. Giordani, M. Berto, M. Di Lauro, C. A. Bortolotti, M. Zoli, and F. Mater. Interfaces 10(31), 25943–25948 (2018).
101
Biscarini, ACS Sens. 2(12), 1756–1760 (2017). C.-S. Yang, D.-S. Shang, N. Liu, E. J. Fuller, S. Agrawal, A. A. Talin, Y.-Q. Li,
67
E. R. W. van Doremaele, P. Gkoupidenis, and Y. van de Burgt, J. Mater. B.-G. Shen, and Y. Sun, Adv. Funct. Mater. 28(42), 1804170 (2018).
102
Chem. C 7, 12754–12760 (2019). C. S. Yang, D. S. Shang, N. Liu, G. Shi, X. Shen, R. C. Yu, Y. Q. Li, and Y. Sun,
68
Y. Lee and T.-W. Lee, Acc. Chem. Res. 52(4), 964–974 (2019). Adv. Mater. 29(27), 1700906 (2017).
69 103
J. Li, W. Tang, Q. Wang, W. Sun, Q. Zhang, X. Guo, X. Wang, and F. Yan, J. Zhu, Y. Yang, R. Jia, Z. Liang, W. Zhu, Z. U. Rehman, L. Bao, X. Zhang, Y.
Mater. Sci. Eng., R 127, 1–36 (2018). Cai, L. Song, and R. Huang, Adv. Mater. 30(21), 1800195 (2018).
70 104
P. Lin and F. Yan, Adv. Mater. 24(1), 34–51 (2012). S. Kim, J. Yoon, H.-D. Kim, and S.-J. Choi, ACS Appl. Mater. Interfaces 7(45),
71
N. Wang, A. Yang, Y. Fu, Y. Li, and F. Yan, Acc. Chem. Res. 52(2), 277–287 25479–25486 (2015).
105
(2019). K. Kim, C. L. Chen, Q. Truong, A. M. Shen, and Y. Chen, Adv. Mater. 25(12),
72
S. M. Won, E. Song, J. Zhao, J. Li, J. Rivnay, and J. A. Rogers, Adv. Mater. 1693–1698 (2013).
106
30(30), 1800534 (2018). D. Sarkar, J. Tao, W. Wang, Q. Lin, M. Yeung, C. Ren, and R. Kapadia, ACS
73
J. Rivnay, S. Inal, A. Salleo, R. M. Owens, M. Berggren, and G. G. Malliaras, Nano 12(2), 1656–1663 (2018).
107
Nat. Rev. Mater. 3, 17086 (2018). Y. Lee, J. Y. Oh, W. Xu, O. Kim, T. R. Kim, J. Kang, Y. Kim, D. Son, J. B.-H.
74
E. J. Fuller, S. T. Keene, A. Melianas, Z. Wang, S. Agarwal, Y. Li, Y. Tuchman, Tok, M. J. Park, Z. Bao, and T.-W. Lee, Sci. Adv. 4(11), eaat7387 (2018).
108
C. D. James, M. J. Marinella, J. J. Yang, A. Salleo, and A. A. Talin, Science W. Xu, S. Y. Min, H. Hwang, and T. W. Lee, Sci. Adv. 2(6), e1501326 (2016).
109
364(6440), 570–574 (2019). P. Gkoupidenis, N. Schaefer, B. Garlan, and G. G. Malliaras, Adv. Mater.
75
E. Juzekaeva, A. Nasretdinov, S. Battistoni, T. Berzina, S. Iannotta, R. 27(44), 7176–7180 (2015).
110
Khazipov, V. Erokhin, and M. Mukhtarov, Adv. Mater. Technol. 4(1), C. Qian, J. Sun, L-a Kong, G. Gou, J. Yang, J. He, Y. Gao, and Q. Wan, ACS
1800350 (2019). Appl. Mater. Interfaces 8(39), 26169–26175 (2016).
76 111
P. Gkoupidenis, D. A. Koutsouras, and G. G. Malliaras, Nat. Commun. 8, C. Qian, L.-A. Kong, J. Yang, Y. Gao, and J. Sun, Appl. Phys. Lett. 110(8),
15448 (2017). 083302 (2017).
77 112
P. Gkoupidenis, D. A. Koutsouras, T. Lonjaret, J. A. Fairfield, and G. G. D. A. Lapkin, A. V. Emelyanov, V. A. Demin, T. S. Berzina, and V. V. Erokhin,
Malliaras, Sci. Rep. 6, 27007 (2016). Microelectron. Eng. 185–186, 43–47 (2018).
78 113
S. Jia, F. Ying, and W. Qing, J. Phys. D: Appl. Phys. 51(31), 314004 (2018). S. Battistoni, V. Erokhin, and S. Iannotta, Org. Electron. 65, 434–438 (2019).
79 114
P. Greengard, Science 294(5544), 1024–1030 (2001). S. Dai, Y. Wang, J. Zhang, Y. Zhao, F. Xiao, D. Liu, T. Wang, and J. Huang,
80
A. E. Pereda, Nat. Rev. Neurosci. 15, 250 (2014). ACS Appl. Mater. Interfaces 10(46), 39983–39991 (2018).
81 115
W. H. Brattain and C. G. B. Garrett, Bell Syst. Tech. J. 34(1), 129–176 (1955). S. Desbief, M. di Lauro, S. Casalini, D. Guerin, S. Tortorella, M. Barbalinardo,
82
S. H. Kim, K. Hong, W. Xie, K. H. Lee, S. Zhang, T. P. Lodge, and C. D. A. Kyndiah, M. Murgia, T. Cramer, F. Biscarini, and D. Vuillaume, Org.
Frisbie, Adv. Mater. 25(13), 1822–1846 (2013). Electron. 38, 21–28 (2016).
83 116
Y. He, Y. Yang, S. Nie, R. Liu, and Q. Wan, J. Mater. Chem. C 6(20), Y. H. Liu, L. Qiang Zhu, Y. Shi, and Q. Wan, Appl. Phys. Lett. 104(13), 133504
5336–5352 (2018). (2014).

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-20

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

117 151
X. Fan, B. Xu, N. Wang, J. Wang, S. Liu, H. Wang, and F. Yan, Adv. Electron. A. Giovannitti, C. B. Nielsen, D.-T. Sbircea, S. Inal, M. Donahue, M. R. Niazi,
Mater. 3(5), 1600471 (2017). D. A. Hanifi, A. Amassian, G. G. Malliaras, J. Rivnay, and I. McCulloch, Nat.
118
R. S. Zucker and W. G. Regehr, Annu. Rev. Physiol. 64(1), 355–405 (2002). Commun. 7, 13066 (2016).
119 152
E. S. Fortune and G. J. Rose, J. Neurosci. 20(18), 7122–7130 (2000). A. M. Pappa, D. Ohayon, A. Giovannitti, I. P. Maria, A. Savva, I. Uguz, J.
120
B. James, L. Darnet, J. Moya-Dıaz, S.-H. Seibel, and L. Lagnado, Nat. Rivnay, I. McCulloch, R. M. Owens, and S. Inal, Sci. Adv. 4(6), eaat0911
Neurosci. 22(7), 1140–1147 (2019). (2018).
121 153
J. A. Kauer and R. C. Malenka, Nat. Rev. Neurosci. 8, 844 (2007). H. Sun, M. Vagin, S. Wang, X. Crispin, R. Forchheimer, M. Berggren, and S.
122
R. A. Nicoll, Neuron 93(2), 281–290 (2017). Fabiano, Adv. Mater. 30(9), 1704916 (2018).
123 154
Q. Lai, L. Zhang, Z. Li, W. F. Stickle, R. S. Williams, and Y. Chen, Adv. Mater. B. C.-K. Tee, A. Chortos, A. Berndt, A. K. Nguyen, A. Tom, A. McGuire, Z. C.
22(22), 2448–2453 (2010). Lin, K. Tien, W.-G. Bae, H. Wang, P. Mei, H.-H. Chou, B. Cui, K. Deisseroth,
124
C. L. Chen, K. Kim, Q. Truong, A. Shen, Z. Li, and Y. Chen, Nanotechnology T. N. Ng, and Z. Bao, Science 350(6258), 313–316 (2015).
155
23(27), 275202 (2012). L. Shenghua, F. Ying, L. Guijun, L. Li, L. H. Ka-wai, C. Xianfeng, and Y. Feng,
125
J. Zhou, N. Liu, L. Zhu, Y. Shi, and Q. Wan, IEEE Electron Device Lett. 36(2), Adv. Mater. 29(35), 1701733 (2017).
156
198–200 (2015). C. Ge, C. X. Liu, Q. L. Zhou, Q. H. Zhang, J. Y. Du, J. K. Li, C. Wang, L. Gu,
126
H. Ling, N. Wang, A. Yang, Y. Liu, J. Song, and F. Yan, Adv. Mater. Technol. G. Z. Yang, and K. J. Jin, Adv. Mater. 31(19), 1900379 (2019).
157
4(9), 1900471 (2019). K. H. Lee, M. S. Kang, S. Zhang, Y. Gu, T. P. Lodge, and C. D. Frisbie, Adv.
127
L.-A. Kong, J. Sun, C. Qian, C. Wang, J. Yang, and Y. Gao, Org. Electron. 44, Mater. 24(32), 4457–4462 (2012).
158
25–31 (2017). S. K. Hwang, T. J. Park, K. L. Kim, S. M. Cho, B. J. Jeong, and C. Park, ACS
128
T. V. P. Bliss, G. L. Collingridge, R. G. M. Morris, and G. M. Morris Richard, Appl. Mater. Interfaces 6(22), 20179–20187 (2014).
159
Philosoph. Trans. R. Soc. London, Ser. B 358(1432), 643–647 (2003). A.-M. Pappa, S. Inal, K. Roy, Y. Zhang, C. Pitsalidis, A. Hama, J. Pas, G. G.
129
R. C. Malenka and M. F. Bear, Neuron 44(1), 5–21 (2004). Malliaras, and R. M. Owens, ACS Appl. Mater. Interfaces 9(12), 10427–10434
130
S. T. Keene, A. Melianas, Y. v d Burgt, and A. Salleo, Adv. Electron. Mater. (2017).
160
5(2), 1800686 (2019). C. Zou, J. Sun, G. Gou, L.-A. Kong, C. Qian, G. Dai, J. Yang, and G.-H. Guo,
131
S. T. Keene, A. Melianas, E. J. Fuller, Y. V. D. Burgt, A. A. Talin, and A. Salleo, Appl. Phys. A 123(9), 597 (2017).
161
J. Phys. D: Appl. Phys. 51(22), 224002 (2018). C. J. Wan, Y. H. Liu, L. Q. Zhu, P. Feng, Y. Shi, and Q. Wan, ACS Appl.
132
P. Gkoupidenis, N. Schaefer, X. Strakosas, J. A. Fairfield, and G. G. Malliaras, Mater. Interfaces 8(15), 9762–9768 (2016).
162
Appl. Phys. Lett. 107(26), 263302 (2015). F. Yu, L. Q. Zhu, W. T. Gao, Y. M. Fu, H. Xiao, J. Tao, and J. M. Zhou, ACS
133
B. Winther-Jensen, B. Kolodziejczyk, and O. Winther-Jensen, APL Mater. Appl. Mater. Interfaces 10, 16881–16886 (2018).
163
3(1), 014903 (2015). S. D. Yambem, J. Timm, M. Weiss, A. K. Pandey, and R. Marschall, Adv.
134
L. A. Kong, J. Sun, C. Qian, Y. Fu, J. X. Wang, J. L. Yang, and Y. L. Gao, Org. Electron. Mater. 3(12), 1700316 (2017).
164
Electron. 47, 126–132 (2017). Y. H. Liu, L. Q. Zhu, P. Feng, Y. Shi, and Q. Wan, Adv. Mater. 27(37),
135
D. Lee, J. Park, K. Moon, J. Jang, S. Park, M. Chu, J. Kim, J. Noh, M. Jeon, B. 5599–5604 (2015).
165
H. Lee, B. Lee, B. Lee, and H. Hwang, paper presented at the 2015 IEEE G. Wu, J. Zhang, X. Wan, Y. Yang, and S. Jiang, J. Mater. Chem. C 2(31),
International Electron Devices Meeting (IEDM), 2015. 6249–6255 (2014).
136 166
H. Wang, Q. Zhao, Z. Ni, Q. Li, H. Liu, Y. Yang, L. Wang, Y. Ran, Y. Guo, W. A. Cifarelli, A. Parisini, T. Berzina, and S. Iannotta, Microelectron. Eng. 193,
Hu, and Y. Liu, Adv. Mater. 30(46), 1803961 (2018). 65–70 (2018).
137 167
E. J. Fuller, F. E. Gabaly, F. Leonard, S. Agarwal, S. J. Plimpton, R. B. Jacobs- F. Yu, L. Q. Zhu, H. Xiao, W. T. Gao, and Y. B. Guo, Adv. Funct. Mater.
Gedrim, C. D. James, M. J. Marinella, and A. A. Talin, Adv. Mater. 29(4), 28(44), 1804025 (2018).
168
1604310 (2017). Y. Liu, X. Wan, L. Q. Zhu, Y. Shi, and Q. Wan, IEEE Electron Device Lett.
138
G. G. Turrigiano and S. B. Nelson, Curr. Opin. Neurobiol. 10(3), 358–364 35(12), 1257–1259 (2014).
169
(2000). A. Cifarelli, A. Parisini, S. Iannotta, and T. Berzina, Microelectron. Eng.
139
G. G. Turrigiano and S. B. Nelson, Nat. Rev. Neurosci. 5(2), 97–107 (2004). 185–186, 55–60 (2018).
140 170
D. A. Koutsouras, G. G. Malliaras, and P. Gkoupidenis, MRS Commun. 8(2), L. Q. Guo, J. Tao, L. Q. Zhu, H. Xiao, W. T. Gao, F. Yu, and Y. M. Fu, Org.
493–497 (2018). Electron. 61, 312–317 (2018).
141 171
D. A. Koutsouras, T. Prodromakis, G. G. Malliaras, P. W. M. Blom, and P. G. Wu, P. Feng, X. Wan, L. Zhu, Y. Shi, and Q. Wan, Sci. Rep. 6, 23578 (2016).
172
Gkoupidenis, Adv. Intell. Syst. 1(1), 1900013 (2019). J. Zhang, J. Dai, L. Zhu, C. Chen, and Q. Wan, IEEE Electron Device Lett.
142
H. F. Ling, S. H. Liu, Z. J. Zheng, and F. Yan, Small Methods 2(10), 1800070 35(8), 838–840 (2014).
173
(2018). X. Wan, P. Feng, G. D. Wu, Y. Shi, and Q. Wan, IEEE Electron Device Lett.
143
S.-W. Hwang, H. Tao, D.-H. Kim, H. Cheng, J.-K. Song, E. Rill, M. A. 36(2), 204–206 (2015).
174
Brenckle, B. Panilaitis, S. M. Won, Y.-S. Kim, Y. M. Song, K. J. Yu, A. G. J. Stuart and N. Spruston, Nat. Neurosci. 18, 1713 (2015).
175
Ameen, R. Li, Y. Su, M. Yang, D. L. Kaplan, M. R. Zakin, M. J. Slepian, Y. N. Spruston, Nat. Rev. Neurosci. 9, 206 (2008).
176
Huang, F. G. Omenetto, and J. A. Rogers, Science 337(6102), 1640–1644 P. Gkoupidenis, S. Rezaei-Mazinani, C. M. Proctor, E. Ismailova, and G. G.
(2012). Malliaras, AIP Adv. 6(11), 111307 (2016).
144 177
S. Inal, J. Rivnay, A.-O. Suiu, G. G. Malliaras, and I. McCulloch, Acc. Chem. C. J. Wan, Y. H. Liu, P. Feng, W. Wang, L. Q. Zhu, Z. P. Liu, Y. Shi, and Q.
Res. 51(6), 1368–1376 (2018). Wan, Adv. Mater. 28(28), 5878–5885 (2016).
145 178
Y. Anneng, L. Yuanzhe, Y. Chenxiao, F. Ying, W. Naixiang, L. Li, and Y. Feng, Y. He, S. Nie, R. Liu, S. Jiang, Y. Shi, and Q. Wan, Adv. Mater. 31(21),
Adv. Mater. 30(23), 1800051 (2018). 1900903 (2019).
146 179
J. Rivnay, P. Leleux, M. Sessolo, D. Khodagholy, T. Herve, M. Fiocchi, and G. C. J. Wan, L. Q. Zhu, Y. H. Liu, P. Feng, Z. P. Liu, H. L. Cao, P. Xiao, Y. Shi,
G. Malliaras, Adv. Mater. 25(48), 7010–7014 (2013). and Q. Wan, Adv. Mater. 28(18), 3557–3563 (2016).
147 180
S. Inal, G. G. Malliaras, and J. Rivnay, Nat. Commun. 8(1), 1767 (2017). Y. Fu, L.-A. Kong, Y. Chen, J. Wang, C. Qian, Y. Yuan, J. Sun, Y. Gao, and Q.
148
A. Laiho, L. Herlogsson, R. Forchheimer, X. Crispin, and M. Berggren, Proc. Wan, ACS Appl. Mater. Interfaces 10(31), 26443–26450 (2018).
181
Natl. Acad. Sci. 108(37), 15069–15073 (2011). R. Shapley, M. Hawken, and D. L. Ringach, Neuron 38(5), 689–699 (2003).
149 182
A. Giovannitti, D.-T. Sbircea, S. Inal, C. B. Nielsen, E. Bandiello, D. A. Hanifi, N. J. Priebe and D. Ferster, Neuron 75(2), 194–208 (2012).
183
M. Sessolo, G. G. Malliaras, I. McCulloch, and J. Rivnay, Proc. Natl. Acad. Sci. J. S. Nettleton and W. J. Spain, J. Neurophysiol 83(6), 3310–3322 (2000).
184
113(43), 12017–12022 (2016). A. Polsky, B. W. Mel, and J. Schiller, Nat. Neurosci. 7(6), 621–627 (2004).
150 185
C. Cendra, A. Giovannitti, A. Savva, V. Venkatraman, I. McCulloch, A. Salleo, Y. Chen, W. Qiu, X. Wang, W. Liu, J. Wang, G. Dai, Y. Yuan, Y. Gao, and J.
S. Inal, and J. Rivnay, Adv. Funct. Mater. 29(5), 1807034 (2019). Sun, Nano Energy 62, 393–400 (2019).

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-21

Published under license by AIP Publishing
 Applied Physics Reviews REVIEW scitation.org/journal/are

186 202
Y. Yang, Y. L. He, S. Nie, Y. Shi, and Q. Wan, IEEE Electron Device Lett. T. Cramer, B. Chelli, M. Murgia, M. Barbalinardo, E. Bystrenova, D. M.
39(6), 897–900 (2018). de Leeuw, and F. Biscarini, Phys. Chem. Chem. Phys. 15(11), 3897–3905
187
M. Chistiakova, N. M. Bannon, M. Bazhenov, and M. Volgushev, (2013).
203
Neuroscientist 20(5), 483–498 (2014). G. Tarabella, P. D’Angelo, A. Cifarelli, A. Dimonte, A. Romeo, T. Berzina, V.
188
N. Fremaux and W. Gerstner, Front. Neural Circuits 9, 85 (2016). Erokhin, and S. Iannotta, Chem. Sci. 6(5), 2859–2868 (2015).
189 204
R. A. John, N. Tiwari, C. Yaoyi, Ankit, N. Tiwari, M. Kulkarni, A. Nirmal, A. A. Romeo, A. Dimonte, G. Tarabella, P. D’Angelo, V. Erokhin, and S.
C. Nguyen, A. Basu, and N. Mathews, ACS Nano 12(11), 11263–11273 (2018). Iannotta, APL Mater. 3(1), 014909 (2015).
190 205
G. Daniel, Principles of Artificial Neural Networks (World Scientific, 2013). Y. Zang, H. Shen, D. Huang, C. A. Di, and D. Zhu, Adv. Mater. 29(18),
191
B. Widrow and M. A. Lehr, Proc. IEEE 78(9), 1415–1442 (1990). 1606088 (2017).
192 206
I. A. Basheer and M. Hajmeer, J. Microbiol. Methods 43(1), 3–31 (2000). Y. Kim, A. Chortos, W. Xu, Y. Liu, J. Y. Oh, D. Son, J. Kang, A. M. Foudeh,
193
D. E. Rumelhart, G. E. Hinton, and R. J. Williams, in Cognitive Modeling, C. Zhu, Y. Lee, S. Niu, J. Liu, R. Pfattner, Z. Bao, and T.-W. Lee, Science
edited by T. A. Polk and C. M. Seifert (MIT Press, 2002), Vol. 5, p. 1. 360(6392), 998–1003 (2018).
194 207
R. J. Erb, Pharm. Res. 10(2), 165–170 (1993). Y. Liu, J. Zhong, E. Li, H. Yang, X. Wang, D. Lai, H. Chen, and T. Guo, Nano
195
Y. Xin, Proc. IEEE 87(9), 1423–1447 (1999). Energy 60, 377–384 (2019).
196 208
S. Battistoni, V. Erokhin, and S. Iannotta, J. Phys. D: Appl. Phys. 51(28), S. Wang, C. Chen, Z. Yu, Y. He, X. Chen, Q. Wan, Y. Shi, D. W. Zhang, H.
284002 (2018). Zhou, X. Wang, and P. Zhou, Adv. Mater. 31(3), 1806227 (2019).
197 209
A. V. Emelyanov, D. A. Lapkin, V. A. Demin, V. V. Erokhin, S. Battistoni, G. Y. Wang, Z. Lv, J. Chen, Z. Wang, Y. Zhou, L. Zhou, X. Chen, and S.-T. Han,
Baldi, A. Dimonte, A. N. Korovin, S. Iannotta, P. K. Kashkarov, and M. V. Adv. Mater. 30(38), 1802883 (2018).
210
Kovalchuk, AIP Adv. 6(11), 111301 (2016). Q. Shuchao, W. Fengqiu, L. Yujie, W. Qing, W. Xinran, X. Yongbing, S. Yi,
198
D. Khodagholy, J. Rivnay, M. Sessolo, M. Gurfinkel, P. Leleux, L. H. Jimison, W. Xiaomu, and Z. Rong, 2D Mater. 4(3), 035022 (2017).
211
E. Stavrinidou, T. Herve, S. Sanaur, R. M. Owens, and G. G. Malliaras, Nat. S. Seo, S.-H. Jo, S. Kim, J. Shim, S. Oh, J.-H. Kim, K. Heo, J.-W. Choi, C. Choi,
Commun. 4, 2133 (2013). S. Oh, D. Kuzum, H. S. P. Wong, and J.-H. Park, Nat. Commun. 9(1), 5106
199
N. Wang, Y. Liu, Y. Fu, and F. Yan, ACS Appl. Mater. Interfaces 10(31), (2018).
212
25834–25840 (2018). H. Tian, X. Wang, F. Wu, Y. Yang, and T. Ren, paper presented at the 2018
200
C. Wan, G. Chen, Y. Fu, M. Wang, N. Matsuhisa, S. Pan, L. Pan, H. Yang, Q. IEEE International Electron Devices Meeting (IEDM), 2018.
213
Wan, L. Zhu, and X. Chen, Adv. Mater. 30(30), 1801291 (2018). G. D. Spyropoulos, J. N. Gelinas, and D. Khodagholy, Sci. Adv. 5(2), eaau7378
201
E. Stavrinidou, R. Gabrielsson, E. Gomez, X. Crispin, O. Nilsson, D. T. Simon, (2019).
214
and M. Berggren, Sci. Adv. 1(10), e1501136 (2015). J. J. Yang and Q. Xia, Nat. Mater. 16, 396 (2017).

Appl. Phys. Rev. 7, 011307 (2020); doi: 10.1063/1.5122249 7, 011307-22

Published under license by AIP Publishing