Artificial neuromorphic electronics that mimic the workingprinciple of neural synapses implement a uniquecomputing paradigm emphasizing cognitive computing

capability1–6. Synapse-motivated device networks make highpower-efficiency and fast parallel processing feasible due toinherent architectural characteristics7–12. For instance, a simplesignal—transmission action between neurons through a synapseonly consumes a millionth of the energy required to execute theequivalent action in a traditional von Neumann computingsystem13.

Non-volatile memory and history-dependent analogue-likestates are two elemental characteristics for synapse-simulatingdevices14–16. Binary transition metal oxide-based two-terminalmetal–insulator–metal memory resistance structures have beenexplored as the building blocks of neuromorphic systems9,12,17–19.Self-learning ability of the synapse is also a crucial attribute forcognitive computing20. Spike-timing-dependent plasticity (STDP),in which the change of synapse weight (w) is a function (f1) of thetime difference between postneuron (tpost) and preneuron (tpre)signals (w! f1(tpost" tpre)), is one common self-learning processin human brains (Fig. 1a)12. Physically, the increase of synapseweight is manifested by augmentation of the quantity of neuro-transmitters and dendritic receptors21. In biological systems,signal transmission and synapse learning are both generallyregarded to occur concurrently in synapse-connected neuronpairs22. Current two-terminal metal–insulator–metal artificialsynapses operate by separating the signal transmission and self-learning processes in time8,9. Three-terminal synaptic devices,

being able to realize both functions simultaneously, therefore offera promising solution for efficient synapse simulation23–28.

Herein, we demonstrate a three-terminal rare-earth nickelate(RNiO3, R! rare-earth element) thin-film transistor that mimicsa biological synapse. By implementing such synaptic nickelatedevice, we successfully realize the first concurrent operation ofsignal transmission and STDP learning in correlated oxides,which provides a new opportunity and strategy to exploreneuromorphic-correlated oxide electronics including program-mable fluidic circuits.

ResultsDevice architecture and operation principle. As correlatedoxides with sharp thermally driven insulator–metal transition,nickelates are of interest in areas spanning from physics toelectronics29,30. Figure 1b illustrates the schematic and operationof a synaptic transistor, in which the source and drain are ana-logues of the preneuron and postneuron terminals, respectively.Perovskite SmNiO3 (SNO) is utilized as the channel material andis engineered to have properties of non-volatile memory, analo-gue states and learning function triggered by gate pulses. Themetallic phase resistance of SNO is very sensitive to the stoi-chiometry31. Oxygen vacancies often lead to the destabilization ofNi3# , which is manifested by the increase of nickelates’ resistancein the metallic regime. Previously, the resistance modulation ofbulk nickelates was achieved by adjusting the degree of oxygendeficiency during synthesis31–33. Recently, suppression orenhancement of the metallic phase of the nickelates wasrealized by introducing either tensile or compressive strain,which is believed to be able to regulate the Ni3# /Ni2# ratio34–36.The resistance modification from both methods is permanent andirreversible. Here, by modulating the stoichiometry of SNO byionic liquid (IL) gating, the conductivity of the SNO is expected tobe regulated in an in situ manner due to the stabilization anddestabilization of Ni3# , which has been reported to have afundamental role in the metal–insulator transition (MIT)mechanism of SNO29. This is fundamentally different fromelectrostatic modulation, in which the resistance tuning is limitedby the extremely short screening length of correlated oxides andthe resistance recovers relatively quickly upon the removal of gatebias37–40. In the synaptic device, the conductance (s) of SNOemulates synapse weight and it is modified by gate pulses, whichare received from a multiplexer and are a function (f3) (ref. 26) ofthe time difference between drain (tdrain) (preneuron) and source(tsource) (postneuron) pulses. In total, s! f2(tdrain" tsource).

Figure 1c illustrates the proposed conductance modulationmechanism. For synapse processes, the transmission andreception of ions are essential for non-volatile and analoguebehaviours. We here simulate the synaptic process by the creationand annihilation of oxygen vacancies in the SNO channelby electrochemical reactions through the IL–SNO interface.Two principal chemical reactions occurring in this process are(1) and (2):

OxO $ V2#

O # 2e" # 12O2 $1%

Ni3# # e" $ Ni2# $2%The overall defect formation reaction within SNO can thereforebe written as:

2Ni3# #OxO $ 2Ni2# #V2#

O # 12O2 $3%

The first reaction is commonly observed in oxides when anexternal bias is applied41–45. When oxygen leaves the SNO lattice,the Ni3# is destabilized and transforms to Ni2# , as shown in

Synapse

Preneuron Postneuron

Three-terminal synapse

Preneuron Postneuron

Source Drain

SNO

Gate

Gate electrode

Source Ionic liquid Drain

O2,ads + e –

1

2

f3NeurotransmittersDendritic receptors

w = f1(tpost – tpre)! = f2(tdrain – tsource)

O •–2,ads

Oxo Ni3+ + e – Ni2+V2+ + 2e – +o O2

SmNiO3

Figure 1 | Three-terminal nickelate synaptic transistor device. (a) In aneural synapse, the synapse weight (w), manifested by the quantity ofneurotransmitters and dendritic receptors, is a function (f1) of the timedifference between preneuron and postneuron spikes (tpost" tpre). (b)Three-terminal SmNiO3 (SNO) transistor gated by ionic liquid with itschannel conductance (s) tuned by the time difference (tdrain" tsource)between source and drain spikes. Function f3 is applied to simulate the timedifference between source and drain, which is manifested by the gate bias.(c) Proposed resistance modulation mechanism, in which oxidation andreduction of Ni species, through the creation/annihilation of oxygenvacancies in SNO channel by external electric field, is designed to enableSNO conductance switching.

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms3676

2 NATURE COMMUNICATIONS | 4:2676 |DOI: 10.1038/ncomms3676 | www.nature.com/naturecommunications

& 2013 Macmillan Publishers Limited. All rights reserved.

2 ISRN Nanotechnology

Moreover, the complex view on BS requires further stages ofprecise processing in order to decode the received or controlinformation. There are di!erent kinds of transducers/sensorsfor picking up NI: room-temperature and superconducting,external, and implantable. Development of such devicesis increasing the penetration into bioprocess while simul-taneously simplifying the exploitation of the measuringsystems in order to bring them closer to the wide rangeof applications. For this reason, the magnetometer with aroom-temperature pickup coil (PC) for detecting signals,which can clearly be detected in higher frequency range, wasdeveloped in order to simplify the SQUID system. The PC isset outside the cryostat and is connected to the input coil ofthe SQUID [9] or a channel of superconducting field-e!ecttransistor (SuFET) [10]. On the other hand, implantable-into-nerve fiber transducers are evolving from the ordinarySi-chip microelectronics devices [11] into superconductingand nanodevices [12, 13].

The recent achievements in nanoelectronics can beregarded as a further step in the progress of BS transduction.They give us the possibility to create the most advancedand universal device on the basis of known microsystems.Such a sensor/transducer is suitable for picking up BS—NI, electrically active (ionized) molecules, and the base-pair recognition event in DNA sequences—and transformingit into recognizable information in the form of electricvoltage, or a concentration of organic or chemical substances.Moreover, this process can be executed in reverse. Substancesand/or voltages influence BSs, thereby controling or creatingthem (BS) [14]. Steady and rapid progress in the roboticsfield requires ever quicker and better human-machine inter-action and the development of a new generation of interfacesfor intelligent systems. Such advances give rise to markedlyincreased biophysical research on the one hand and the needfor new bioelectronic devices on the other. Transductionand measurement of BS are key elements of MIs design.There are two means involved in signal transduction: (1)biochemical—by hormones and enzymes; (2) biophysical—by nerve impulses (ionic currents). Let us consider thebiophysical ones as useful for the said interfaces design above.There are two values—voltage and electric current—whichcharacterize the pathway of transduction [15].

Calculations of PC arrays were performed with theprimary sensor flux transformer sites distributed uniformlyon a spherical sensor shell, extending from the vertex toa maximum angle max [16]. The radial magnetometersand gradiometers occupy one site each, there are twoorthogonal planar gradiometers at each site and there arethree orthogonal magnetometers at each site for vectormagnetometers. Coverage can be achieved by designing somekind of density control mechanism, that is, scheduling thesensors to work alternatively to minimize the power wastagedue to the overlap of active nodes’ sensing areas. The sensingarea of a node is a disk of a given radius (sensing range). Thesensing energy consumption is proportional to the area ofsensing disks or the power consumption per unit [17].

There are two broad ways of brain-computer interface(BCI): invasive and noninvasive. The invasive technique cancapture intracortical action potentials of neurons and thus,

provides high signal strength spatiotemporally, for example,prediction of movement trajectory. In noninvasive tech-nique, EEG and MEG have emerged as viable options; bothof them have time resolutions in milliseconds. Any activityin brain is accompanied by change in ionic concentrations inneuron leading to polarization and depolarization. Such anelectrical activity is measured by EEG, while MEG measuresthe magnetic field associated with these currents. Electric andmagnetic fields are oriented perpendicular to each other [18].

Application of organic-, chemical-, and carbon-nanotubes- (CNT-) based FETs for design of thesuperconducting transducers (SuFETTrs) of BS intodi!erent quantities (electrical and biochemical) is theproposed variant of interfacing [19]. The placement of thedevices can be carried out both in vivo and in vitro withthe possibility of forming the controlling BS from the saidquantities. The range of picked up BS varies from 0.6 nA to10 µA with frequencies from 20 to 2000 Hz.

A further step should be the synthesis of the said twomethods in order to develop the internal (implantable)nano-bio-interface arrays. This means wrapping of molec-ular nanowired PC around the axons of a nerve fibre orsynapses of neurons in order to obtain the natural biosignalsfrom the nervous system and brain. This leads to sensingaccess across a vast range of spatial and temporal scales,including the ability to read neural signals from a selectsubset of single neural cells in vivo. Moreover, this processcan be executed in reverse for introducing the artificialcontrol signals with the local neural code into the single cellelectrical activity.

2. Biosignals and Nanoelements for TheirTransduction

As an electrical signal, the biosignal has two components:electrical potential or voltage and ionic or electronic cur-rents. The first component is su"ciently developed anddoes not require penetration into the substances of biosignalpropagation. The marketable progress in transducing of thesecond component began when the necessary instrumenta-tion for measurement of micro- and nanodimensions hadbeen created [14].

Short platinum nanowires (NWs) already have beenused in submicroscopic sensors and other applications. Amethod of making long (cm) Pt NW of a few nanometersin diameter from electrospinning was described [20]. Thosewires could be woven into the first self-supporting webs ofpure platinum. Double-gated silicon NW structures (DG-SiNW), where the position and/or type of the charge couldbe tuned within the NW by electric field, have been studied[21]. Self-assembled molecular nanowires were found tobe composed of a single crystal, allowing good electricaltransport with low resistivity [22].

An interesting structure is that of helical CNT ornanocoils for PCs. Nanocoils o!er unique electronic prop-erties that straight CNT do not have. The plasticity ofCNT will be relevant to their use in nanoscale devices [23].Carbon nanocoils (CNCs), as a new type of promising

4 ISRN Nanotechnology

Myelin

Gate

Axons

NI

Ionic currents

Channel’s ibio

Vout/in

SuFET

Figure 2: Transducing of the nerve impulses and introducing of therelevant artificial signals [15].

B2a

B1a

B0a

B2b

B1b

B0b

Figure 3: PC for “two-dimensional” gradiometer that detects bothaxial-second-order gradient and planar-first-order gradient of MF[37]. Copyright 2007 The Japan Society of Applied Physics.

biosensors of the relevant physical and chemical quantities[19, 39].

The proposal to measure the biosignal values of di!erentorigins with advanced nanosensors of EM quantities isjustified when allowing for superconducting abilities of thedevices. They are composed in full-scale arrays. The saidarrays can be both implantable into ionic channels of anorganism and sheathed on the sources of the EM emanation.Nanowired head sensors function both in passive mode forpicking up the biosignals and with additional excitation of adefined biomedium through the same head (in reverse) [40–42].

3.1. The Arrangements of a NanoFET-Based Delivery andTransducing. The advances in nanotechnology are opening

HP

(a)

(b)

(c)

(d)

(e)

Figure 4: Comparison of the discrete geometry of three self-assembly models. (a) Unfolding an HP chain. (b)–(d) Severalrepresentations of unfolding a cube [38]. Copyright 2011 NationalAcademy of Sciences, USA.

the way to achieving direct electrical contact of nano-electronic structures with electrically and electrochemi-cally active subcellular structures, including ion channels,receptors, and transmembrane proteins. The method ofcombining the bioelectric nature of the nerve impulses NIand synaptic currents between neighboring neurons withbody-temperature PC and zero-resistance-CNT-based inputof the SuFET device in order to obtain most advantageousbiosensor/transducer was recently advanced [43]. On theother hand, neuroelectronic systems for two-way interfacingof the neuronal and the electronic components by capacitivecontacts and by FETs with an open gate were developed.A nanoSuFET with a high-temperature superconductingchannel is introduced into the nerve fibre or brain tissue fortransducing their signals in both directions.

Such a sensor/actuator is suitable for picking up BS—NIand electrically active (ionized) molecules—and transform-ing it into recognizable information in the form of electric