materials Article Multi-Terminal Transistor-Like Devices Based on Strongly Correlated Metallic Oxides for Neuromorphic Applications Alejandro Fernández-Rodríguez 1 , Jordi Alcalà 1 , Jordi Suñe 2, *, Narcis Mestres 1 and Anna Palau 1, * 1 Institut de Ciència de Materials de Barcelona, ICMAB-CSIC, Campus UAB, 08193 Bellaterra, Barcelona, Spain; afernandez3@icmab.es (A.F.-R.); jalcala@icmab.es (J.A.); narcis@icmab.es (N.M.) 2 Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, 08193 Bellaterra, Barcelona, Spain * Correspondence: jordi.sune@uab.cat (J.S.); palau@icmab.es (A.P.)  Received: 14 November 2019; Accepted: 3 January 2020; Published: 8 January 2020  Abstract: Memristive devices are attracting a great attention for memory, logic, neural networks, and sensing applications due to their simple structure, high density integration, low-power consumption, and fast operation. In particular, multi-terminal structures controlled by active gates, able to process and manipulate information in parallel, would certainly provide novel concepts for neuromorphic systems. In this way, transistor-based synaptic devices may be designed, where the synaptic weight in the postsynaptic membrane is encoded in a source-drain channel and modified by presynaptic terminals (gates). In this work, we show the potential of reversible field-induced metal-insulator transition (MIT) in strongly correlated metallic oxides for the design of robust and flexible multi-terminal memristive transistor-like devices. We have studied different structures patterned on YBa2 Cu3 O7−δ films, which are able to display gate modulable non-volatile volume MIT, driven by field-induced oxygen diffusion within the system. The key advantage of these materials is the possibility to homogeneously tune the oxygen diffusion not only in a confined filament or interface, as observed in widely explored binary and complex oxides, but also in the whole material volume. Another important advantage of correlated oxides with respect to devices based on conducting filaments is the significant reduction of cycle-to-cycle and device-to-device variations. In this work, we show several device configurations in which the lateral conduction between a drain-source channel (synaptic weight) is effectively controlled by active gate-tunable volume resistance changes, thus providing the basis for the design of robust and flexible transistor-based artificial synapses. Keywords: strongly correlated oxides; resistive switching; neuromorphic computing; transistor-like devices 1. Introduction Digital computers can process a large amount of data with high precision and speed. However, compared to the brain, the computer still cannot approach a comparable performance considering cognitive functions such as perception, recognition, and memory. Neuromorphic computing, operating with a parallel architecture connecting low-power computing elements (neurons) with multiple adaptive memory elements (synapses), appears as a very attractive alternative to von-Neuman based algorithms in future cognitive computers [1,2]. The advantages of using analogue with very large-scale integration include: Inherent parallelism, as well as reducing the chip area and power consumption in comparison with digital implementations [3]. Design of computational systems mimicking the way that brain works, with intrinsically massive parallel information processing, is completely unfeasible Materials 2020, 13, 281; doi:10.3390/ma13020281 www.mdpi.com/journal/materials Materials 2020, 13, 281 2 of 11 by using the existing hardware which is based on conventional digital logic. Although stable learning has been achieved with digital logic for low-precision applications using binary weights [4,5], the development of novel functional materials, and individual device components able to resemble the properties of neurons and synapses, are mandatory to bring a revolutionary technological leap toward the implementation of a fully neuromorphic computer. Resistive-switching devices, modeled as memristors, have become a leading candidate to mimic basic functionalities of biological components in a neural network, while providing clear advantages in energy and scalability [3,6]. The resistive switching effect consists of a non-volatile reversible switch between different resistance states, induced by an electric field [7]. Strongly correlated metal oxides showing metal–Mott insulating transitions (MIT) appear as particularly interesting materials for future neuromorphic device architectures, because they show large resistance variations, induced by small carrier concentration modulations, driven by an electric field, allotted to obtain multilevel analogue states [8–10]. The ability to continuously tune the electrical resistance, as well as to induce both volatile and non-volatile transitions, put them in a unique position to mimic neurons and synapses on a device level [11,12]. In particular, to achieve useful synaptic plasticity, a multistate behavior should be changed in an analog continuous fashion with long retention time so that the device resistance continuously depends on the electrical history. Spike time dependent plasticity (STDP) has been successfully demonstrated in different memristive devices based on two-terminal (2T) metal-insulator-metal passive circuit elements [13,14]. However, in biological systems, signal transmission and synapse learning are both generally regarded to occur concurrently in synapse-connected neuron pairs. Current 2T artificial synaptic devices operate by separating the signal transmission and self-learning processes in time. In this context, three-terminal (3T) synaptic devices, being able to realize both functions simultaneously, offer a promising solution for efficient synapse simulation [11,15,16]. Another reason why research on multiterminal devices is relevant is the possibility of having several gates which can obtain signals from different sources simultaneously, and they can therefore experience spatiotemporal effects, which 2T devices cannot [15]. Lately, new multiterminal devices that are able to mimic important aspects of biological sensing functions have been developed. In this sense, through the utilization of a simple organic electrochemical transistor based device with multiple gates, a sensing system has been demonstrated that is analogous to the orientation selectivity from the thalamus (the center part of the brain) to the visual cortex, which governs the vision process in the brain [17]. In a second example, a series of split-gate molybdenum sulfide transistors were implemented to mimic the coincidence nerve network in the owl’s brain [18]. Given the huge number of neurons and synaptic connections in the human brain, multi-terminal memristors are also needed to perform complex functions as heterosynaptic plasticity [19–22]. We have recently demonstrated stable volume field-induced resistive switching in structures based on strongly correlated metallic perovskite oxides (La1−x Srx MnO3 (LSMO) and YBa2 Cu3 O7-δ (YBCO)), modulated through oxygen diffusion [23,24]. Optimally doped LSMO and YBCO materials are metallic in its initial state and they evolve into the insulating state by decreasing the oxygen content [25,26]. Multiple resistance states, needed for synaptic applications, can be achieved by tuning the oxygen doping with the applied voltage. In order to elucidate this behavior, Figure 1 displays different curves obtained by measuring the R(T) evolution of a switched contact in a YBCO film, after applying different voltage pulses. A clear MIT transition from the optimally doped metallic state to an underdoped insulating state is observed. Materials Materials2020, 13,13, 2020, 281281 3 3ofof 1111 Materials 2020, 13, 281 3 of 11 Figure Figure1. Resistance versus temperature obtained for aaYBCO film after applying aaseries of voltage Figure1.1.Resistance Resistanceversus versustemperature temperatureobtained obtainedfor for aYBCO YBCOfilm filmafter afterapplying applying aseries seriesofofvoltage voltage pulses to a silver contact of 100 µm. pulses pulsestotoa asilver silvercontact 100m. contactofof100 m. The key advantage of these systems is that the resistive switching is based on the MIT, which being The Thekeykey advantage advantage of these thesesystems is that thatthetheresistive switching isisbased on thetheMIT, MIT,which an intrinsic property of the of material, systems causes aishomogeneous resistive change switching of resistancebasedin aon gate modulable which being an intrinsic being allowing an intrinsic property of the material, causes a homogeneous change of resistance in a gate volume, the property of the material, design of flexible causes transistor-like a homogeneous devices (memristors) change [23]. Itofis resistance worth noting in athat gate modulable modulable volume, volume, allowing allowing the thedesign design ofofflexible flexible transistor-like transistor-like devices devices (memristors) (memristors) [23]. [23].ItItisis the volume resistance modulation, observed in metallic perovskites, offers enhanced robustness, in worth worth noting notingthat thatthe thevolume volumeresistance resistance modulation, modulation, observed ininmetallic perovskites, offers terms of cycle-to-cycle and device-to-device variations, when observed compared with metallic perovskites, that induced offers in strongly enhanced enhanced robustness, robustness, in terms in terms ofofcycle-to-cycle cycle-to-cycle and and device-to-device device-to-device variations, variations, when when compared compared with with correlated oxides that are insulating in the pristine state, where the switching phenomena is strongly that induced that induced ininstrongly strongly correlated correlated oxides oxides that that are areinsulating insulating ininthe pristine the pristine state, state, where where the theswitching switching localized at the contact interface or in confined filaments [7,27]. phenomena phenomena isisstrongly strongly localized localized atatthe thecontact contactinterface interfaceorin orinYBCO inconfined confined filaments [7,27]. Here, we report on the study of the oxygen diffusion basedfilaments [7,27]. memristor multi-terminal Here, Here, wewe report report on on the thestudy study of the of theoxygen oxygen diffusion diffusion in inYBCO YBCO based based multi-terminal multi-terminal memristor memristor devices in which the oxygen redistribution, and thus the conductance of a drain-source channel, may devices devicesininwhich which the theoxygen oxygen redistribution, redistribution, and and thus thus the theconductance conductance ofofa drain-source a drain-source channel, channel, may may be tuned by using various gates. A sketch of the oxygen diffusion mechanism occurring below the bebetuned by using tuned by using various various gates. A sketch gates.isAshown of the sketchinofFigure oxygen the oxygen diffusion mechanism occurring diffusion mechanism occurring below the below the gate, emulating a synaptic process, 2. gate, gate,emulating emulatinga asynaptic synapticprocess, process,isisshown shownininFigure Figure2.2. Figure 2. Schematic representation of a transistor-like device emulating a biological synapse. Blue Figure and red2.in Figure 2.Schematic Schematicrepresentation the bottomrepresentationofofa atransistor-like pictures depict device transistor-like optimally dopeddeviceemulating emulating and a abiological biological under-doped YBCO,synapse. synapse.Blue Blueand and respectively. red in Accordingthe bottom red in thetobottompictures pictures the oxygen depict optimally depict the doping, optimally doped schemesdoped and under-doped and under-doped represent YBCO, (a) high, (b)YBCO, respectively. respectively. intermediate, According and According (c) low totothe theoxygen oxygen source-drain doping, doping,the conductance. theschemes schemesrepresent represent(a)(a)high, high,(b) (b)intermediate, intermediate,and and(c)(c)low lowsource-drain source-drain conductance. conductance. The conductance between a source-drain channel (post-synaptic membrane) is controlled through The Theconductance modulatory gate terminals conductance between a asource-drain (pre-synaptic between inputs). Bychannel source-drain the (post-synaptic application channel membrane) of a gate voltage, (post-synaptic oxygen membrane) isisvacancies controlled controlled through arethrough modulatory redistributed gate withingate modulatory terminals the terminals (pre-synaptic YBCO channel, inputs). locally changing (pre-synaptic By the itsthe inputs). By application doping of a level, thereby application gate voltage, of a gatetheir oxygen resistance voltage, oxygen vacancies vacanciesare (conductance).areredistributed Synaptic within plasticity redistributed the theYBCO YBCOchannel, characteristics within may be channel, locally changing obtained locally itsitsdoping dopinglevel, with intermediate changing thereby synaptic level, their weight thereby their resistance states resistance(conductance). achieved by tuning the (conductance). Synaptic plasticity amountplasticity Synaptic characteristics of oxygencharacteristics may vacancies created. be may be obtained Multiple obtained with intermediate pre-synaptic input with intermediate synaptic terminals weight synaptichave been weight states achieved emulated states by bytuning by using achieved the multiple tuning the amount amountofofoxygen intermediate gates oxygen vacancies between vacancies thecreated. created.Multiple drain-source pre- channel. Multiple pre- Materials 2020, 13, 281 4 of 11 synaptic Materials input 2020, terminals have been emulated by using multiple intermediate gates between the drain- 13, 281 4 of 11 source channel. 2. 2. Materials Materials and and Methods Methods The Thegeometry geometryofofYBCO YBCOtransistor-like transistor-like devices devicesconsist consistof aofdrain-source a drain-source channel withwith channel different gate different terminals to modulate the channel conductance. Optimally gate terminals to modulate the channel conductance. Optimally doped, epitaxial doped, epitaxial YBa Cu 2 YBa O (YBCO) Cu3O7−δ (YBCO) 3 27−δ thin films, with thickness thin films, of 100ofnm, with thickness 100were nm, grown were grownby pulsed laser deposition by pulsed (PLD) on laser deposition (001)-LaAlO (PLD) 3 single3 on (001)-LaAlO crystal substrates. The parameters used in this process were previously single crystal substrates. The parameters used in this process were previously optimized for our optimized for our purposes. The substrate purposes. Thewas heatedwas substrate up to T = 800–810 heated ◦ up to T =C,800–810 with an°C, O2 with partial anpressure O2 partial of pressure 0.3 mbar of during 0.3 mbarthe deposition 2 during theand a fixed target-substrate deposition distance of 52.5 and a fixed target-substrate mm. of distance A 52.5 highmm.fluence laserfluence A high (∼2 J/cm ) working laser (∼2 J/cmat2) aworking frequency at a frequency of 5 Hz was used. During the cooling ramp, we 2increase the P(Oin of 5 Hz was used. During the cooling ramp, we increase the P(O ) in the chamber order 2) in the to obtain well oxygenated samples. The thickness of the film is mainly chamber in order to obtain well oxygenated samples. The thickness of the film is mainly determined determined by the number of pulses. For these by the number ofsamples, pulses. For 2600 pulses these were applied samples, 2600 pulsesobtaining a thickness were applied of 100 anm. obtaining Topography thickness of 100 shows a high-quality flat surface in all cases, the root-mean-square (rms) nm. Topography shows a high-quality flat surface in all cases, the root-mean-square (rms) value of value of surface roughness is foundroughness surface to be below 1 nm. to is found The bestructural below 1 nm. features of YBCOfeatures The structural films have been studied of YBCO films have by been theta-2theta studied X-ray diffraction (Siemens Diffractometer D5000, Siemens AG, Munich, by theta-2theta X-ray diffraction (Siemens Diffractometer D5000, Siemens AG, Munich, Germany). Germany). The epitaxial nature of the films The epitaxial naturewas of evidenced the films was by the detection evidenced byofthe only (001) peaks detection along of only with (001) the corresponding peaks along with the (001) peaks originating corresponding (001) peaksfrom originating the (001)-LAO from substrates in the theta-2theta the (001)-LAO substrates in X-ray the diffraction theta-2theta spectra. X-ray Photolithography and wet etching were used to pattern channels with different diffraction spectra. Photolithography and wet etching were used to pattern channels with different widths, w = 5–100 µm. 2 silver contacts, spaced widths, w = 5–100 m. After the patterning, multiple 50 nm thick, 100 × 100 m silver contacts, spaced After the patterning, multiple 50 nm thick, 100 × 100 µm 2 different distances apart, differentd = distances 100–300 µm, apart,were 100–300 m, d =deposited by sputtering were depositedand lift-off. by sputtering and lift-off. Electrical measurements were performed at Electrical measurements were performed at room temperature room temperature with awith Keithley 2450 source-meter a Keithley 2450 source- at ambient meter atpressure ambientand temperature. pressure Voltage pulses and temperature. of 4 s,pulses Voltage were applied between of 4 s, were two top applied gates (top–top between two top configuration), located at different positions of the channel, while measuring gates (top–top configuration), located at different positions of the channel, while measuring the current–voltage (I–V), the and associated resistance–voltage (R–V) characteristics, in a two-point current–voltage (I–V), and associated resistance–voltage (R–V) characteristics, in a two-point configuration. The variation of the drain-source configuration. The conductance variation of the through the channel, drain-source obtainedthrough conductance after applying the gate the channel, pulses, after obtained was evaluated by measuring the resistance at intermediate segments applying the gate pulses, was evaluated by measuring the resistance at intermediate of the channel, N, (R N = N segments of4,the , 1, 2, 3, 5) in a standard channel, N, (Rfour-point N, N = 1, 2, method, 3, 4, 5) using two external in a standard electrodes four-point to inject method, usingthetwo current and intermediate external electrodes to contacts inject the to current measureand the voltage. intermediateIn this way, wetoavoid contacts measurethe contribution the voltage.ofInthe thiscontact way, resistance we avoidand the thus we obtain the bulk resistance change. Figure 3a shows a schematic representation contribution of the contact resistance and thus we obtain the bulk resistance change. Figure 3a shows of the proposed device and Figure a schematic 3b an optical representation microscopy of the proposedimage deviceofandseveral Figuredevices 3b anwith different optical channelimage microscopy widths of (w = several devices with different channel widths (w = 100, 50, 20, 10, and 5 m). 100, 50, 20, 10, and 5 µm). Figure 3. (a) Schematic representation of a transistor-like device with a source-drain channel and Figure 3. (a) Schematic representation of a transistor-like device with a source-drain channel and multiple tunable gates. The channel conductance is evaluated by measuring intermediate resistances multiple tunable gates. The channel conductance is evaluated by measuring intermediate resistances (RN ); (b) optical microscope image of several devices patterned with different channel widths from 5 to (RN); (b) optical microscope image of several devices patterned with different channel widths from 5 100 µm. to 100 m. 3. Results and Discussion 3.1. Switching Characteristics between Two Gates Figure 4 shows repeated I–V scans (Figure 4a), and the associated R–V curves (Figure 4b), obtained for a device with a channel of w = 50 µm, by applying positive and negative voltage pulses within two gate electrodes placed d = 200 µm apart. A complementary switching behavior was reproduced, since the two gates, see opposite voltage polarities in opposite directions [28,29]. Thus, for a given polarity, 3.1. Switching Characteristics between Two Gates Figure 4 shows repeated I–V scans (Figure 4a), and the associated R–V curves (Figure 4b), obtained for a device with a channel of w = 50 m, by applying positive and negative voltage pulses within Materials two gate 2020, 13, 281electrodes placed d = 200 m apart. A complementary switching behavior5 was of 11 reproduced, since the two gates, see opposite voltage polarities in opposite directions [28,29]. Thus, for a given polarity, one electrode undergoes a set process (incorporating oxygen and thus switching one electrode from undergoesstate a high resistance a set (HRS) processto(incorporating oxygen a low resistance and state thus switching (LRS)) fromother while in the a high resistance one a reset state (HRS) to a low resistance state (LRS)) while in the other one a reset transition transition is produced (losing oxygen and switching from an LRS to an HRS). The set (Vset) and is produced (losing reset oxygen (V and switching from an LRS to an HRS). The set (Vset ) and reset (Vreset ) voltages, associated reset) voltages, associated to each gate electrode, are depicted in Figure 4b. In general, for both to each gate polarities, Vsetelectrode, occurs atare depicted lower voltagein Figure values 4b. thanInVgeneral, for both polarities, Vset occurs at lower reset, due to a fast motion of oxygen in the low voltage values than conductivity regions [23]. V , due to a fast motion of oxygen in reset In a device with symmetrical gates, the the low conductivity resistance valuesregions [23]. obtained at In thea device HRS with and LRS symmetrical for a givengates, thepulse voltage resistance values are the sameobtained at the HRS thus providing and LRS forloop, a symmetrical a given voltage as the one pulse are the shown in Figure 4b. same thus providing a symmetrical loop, as the one shown in Figure 4b. Figure 4. Typical (a) I–V and (b) R–V, obtained by using a two-point measurement, for a YBCO Figure 4. Typical (a) I–V and (b) R–V, obtained by using a two-point measurement, for a YBCO transistor-like device with a channel width of w = 50 µm, obtained by applying several voltage pulses transistor-like device with a channel width of w = 50 m, obtained by applying several voltage pulses through two identical gates separated at a distance of d = 200 µm, in a top–top electrode configuration. through two identical gates separated at a distance of d = 200 m, in a top–top electrode configuration. The evolution of Vset and Vreset have been investigated by evaluating different I–V curves obtained The evolution for devices of Vset and with different Vreset have channel beenapplying widths investigated the by evaluating minimum different voltage pulse I–V curves able obtained to reversibly for devices with different channel widths applying the minimum voltage pulse able switch the gates at different distances (see Figure 5a). The values of Vset and Vreset increase linearly to reversibly switch with the the gates atdistance, electrode differentaccording distances to (see Figure 5a). a constant The valueswith dependence of Vthe set and Vreset increase linearly with set and reset electric field (Eset the electrode distance, according to a constant dependence with and Ereset , respectively) which are of the order of, Eset ~ (1–1.5) × 104 V/m, the set and reset Ereset electric×field ~ (1.5–2) (Eset 104 V/m. and E reset, respectively) which are of the order of, Eset  (1–1.5) × 104 V/m, Ereset  (1.5–2) × 104 V/m. Bias Bias voltages of V ~ 2 V are needed to switch gates placed 10 µm apart and lower values, favorable for voltages practical of V  2 V areare applications, needed to switch expected gates the by reducing placed 10 m device apart andFigure dimensions. lower5bvalues, showsfavorable for the HRS and practical applications, are expected by reducing the device dimensions. Figure 5b LRS resistances obtained for devices of different widths. The HRS have been read at V ~ 0 V, whereasshows the HRS and LRS resistances for the LRS we obtained considered forthe devices minimum of different widths. value of Theresistance. the loop HRS have been Both read values at are V0indicated V, whereasin for the LRS we considered Figure 4b by dashed arrows. the minimum value of the loop resistance. Both values are indicated in Figure 4bclearly It is by dashed arrows. observed that the resistance values for the HRS and LRS increase with decreasing the channel width, with a behavior that is essentially linear, which is completely consistent with a volume resistive switching process [23]. Deviations of the linear dependence at low channel widths may be attributed to fabrication factors or border effects. In this way, by changing the gate area or unbalancing the applied positive/negative voltage pulse, one can modulate the weight of resistance variation in each gate. Figure 6a shows a typical example of a device with different gate areas in which the HRS and LRS of each gate exhibit different resistance changes, thus producing an asymmetrical R–V curve. Figure 6b shows an example of an asymmetrical R–V curve obtained in a device with equal gates but applying asymmetrical voltage pulses. In this case, the maximum applied negative voltage is lower than Vreset and thus the contact that should switch at the HRS at this polarity does not change. Both situations will be better described in the following. Materials 2020, 13, 281 6 of 11 Materials 2020, 13, 281 6 of 11 Figure 5. (a) Set (closed symbols) and reset (open symbols) voltages as a function of the gate distance obtained for devices of different widths; (b) evolution of the HRS and LRS resistance values with the device width. Dashed lines correspond to a linear dependence of R with the channel width, solid lines are guides to the eye. All values have been obtained by using a two-point configuration. It is clearly observed that the resistance values for the HRS and LRS increase with decreasing the channel width, with a behavior that is essentially linear, which is completely consistent with a volume resistive switching process [23]. Deviations of the linear dependence at low channel widths may be attributed to fabrication factors or border effects. In this way, by changing the gate area or unbalancing the applied positive/negative voltage pulse, one can modulate the weight of resistance variation in each gate. Figure 6a shows a typical example of a device with different gate areas in whichFigure the HRS 5.Figure and (a) Set 5. (a) LRS (closed Set of symbols) (closed each and symbols) gate exhibit reset and reset (opendifferent (open symbols) resistance voltages symbols) voltages changes, as aasfunction a function of theof thus producing thedistance gate gate distance an asymmetrical R–V obtainedobtained for curve. devices ofFigure for devicesdifferent 6bwidths; of different shows widths; an exampleof (b)evolution (b) evolution ofof thean theHRSasymmetrical HRS and LRSLRS and R–V resistance curve values resistance with obtained values with the in a the devicedevice with width. equal gates device width. Dashed but Dashed applying lines lines asymmetrical correspond correspond toto a alinear voltage pulses. lineardependence dependence ofof Inthethis R with R with case, channel the thewidth, width, channel maximum solid lines applied solid lines are guides to the eye. All values have been obtained by using a two-point configuration. negative voltage is lower than V and thus the contact that should are guides to the eye. All values have been obtained by using a two-point configuration. reset switch at the HRS at this polarity does not change. Both situations It is clearly observed that willthebe better described resistance in the values for the HRSfollowing. and LRS increase with decreasing the channel width, with a behavior that is essentially linear, which is completely consistent with a volume resistive switching process [23]. Deviations of the linear dependence at low channel widths may be attributed to fabrication factors or border effects. In this way, by changing the gate area or unbalancing the applied positive/negative voltage pulse, one can modulate the weight of resistance variation in each gate. Figure 6a shows a typical example of a device with different gate areas in which the HRS and LRS of each gate exhibit different resistance changes, thus producing an asymmetrical R–V curve. Figure 6b shows an example of an asymmetrical R–V curve obtained in a device with equal gates but applying asymmetrical voltage pulses. In this case, the maximum applied negative voltage is lower than Vreset and thus the contact that should switch at the HRS at this polarity does not change. Both situations will be better described in the following. Figure 6. Typical I–V characteristics, obtained by using a two-point measurement, for YBCO Figure 6. Typical I–V characteristics, obtained by using a two-point measurement, for YBCO transistor-like devices by applying (a) symmetrical voltage pulses using two gates with different transistor-like devices by applying (a) symmetrical voltage pulses using two gates with different switching performance; (b) asymmetrical voltages pulses to switch just one gate. switching performance; (b) asymmetrical voltages pulses to switch just one gate. Next, we will demonstrate that the reversible gate resistance modulation, which occurs through a Next, we will demonstrate that the reversible gate resistance modulation, which occurs through field-induced MIT driven by oxygen diffusion, is not just occurring below the gates, but also effectively a field-induced MIT driven by oxygen diffusion, is not just occurring below the gates, but also modifies the volume resistance of the drain-source channel. effectively modifies the volume resistance of the drain-source channel. Figure 6. Typical I–V characteristics, obtained by using a two-point measurement, for YBCO 3.2. Conductance Modulation in a Drain-Source Channel transistor-like devices by applying (a) symmetrical voltage pulses using two gates with different The switching performance; conductance (b) asymmetrical modulation voltages pulses of the device to switch channels hasjust one gate. been evaluated by measuring the relative variation of volume resistance at different segments of it (in a four-point configuration), after Next, we will demonstrate that the reversible gate resistance modulation, which occurs through applyinga several positive field-induced MITand negative driven voltage by oxygen pulsesis between diffusion, two intermediate not just occurring below the gates, in aalso gates, but two-point configuration. Figure effectively 7a shows modifies a schematic the volume resistance representation ofchannel. of the drain-source the active gates (in yellow) and voltage probe positions considered for a device with a track of w = 10 µm and gate distance d = 100 µm. Figure 7b shows the R–V curves measured through the yellow gates by applying a maximum voltage of 12 V. A complementary switch of the two contact gates is clearly evidenced with a rather symmetrical R–V hysteresis curve. We plot in Figure 7c the percentage resistance change, measured at different segments of the channel, ∆RN , calculated by using Equation (1). ∆RN = 100 × [RN (t) − RN (i)]/RN (i) (1) probe positions considered for a device with a track of w = 10 m and gate distance d = 100 m. Figure 7b shows the R–V curves measured through the yellow gates by applying a maximum voltage of 12 V. A complementary switch of the two contact gates is clearly evidenced with a rather symmetrical R–V hysteresis curve. We plot in Figure 7c the percentage resistance change, measured at different Materials segments of13, 2020, the281channel, RN, calculated by using Equation (1). 7 of 11 RN = 100 × [RN(t) − RN(i)]/RN(i) (1) were RN (i) and RN (t) are the resistance values measured at the N segment at the initial state and after were RN(i) and RN(t) are the resistance values measured at the N segment at the initial state and after several voltage pulses, respectively, using a four-point configuration. A clear correlation between several voltage pulses, respectively, using a four-point configuration. A clear correlation between ∆RN and the applied pulses is observed in Figure 7c. Large resistance variations (∆RN ~ 15–35%) ∆RN and the applied pulses is observed in Figure 7c. Large resistance variations (∆RN  15–35%) are are obtained at the segments close to the gates (N = 2, 3, 4), whereas the resistance does not change obtained at the segments close to the gates (N = 2, 3, 4), whereas the resistance does not change on on those segments located further away from the gates (N =1, 5). It is worth pointing out that the those segments located further away from the gates (N =1, 5). It is worth pointing out that the variation of resistance at different regions of the channel compensate each other, providing a nearly variation of resistance at different regions of the channel compensate each other, providing a nearly constant resistance when measured through the whole channel (RTOTAL ). This is in agreement with a constant resistance when measured through the whole channel (RTOTAL). This is in agreement with a redistribution of oxygen vacancies within the channel, with no external oxygen exchange, as modeled redistribution of oxygen vacancies within the channel, with no external oxygen exchange, as modeled in [23]. in [23]. Figure 7. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel of Figure 7. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel w = 10 µm, by applying positive (top) and negative (bottom) voltage pulses between the two yellow of w = 10 m, by applying positive (top) and negative (bottom) voltage pulses between the two yellow gates separated at d = 100 µm. Red and blue colors represent HRS and LRS, respectively; (b) R–V gates separated at d = 100 m. Red and blue colors represent HRS and LRS, respectively; (b) R–V characteristics, obtained by applying voltage pulses between the gates in two-point configuration; characteristics, obtained by applying voltage pulses between the gates in two-point configuration; (c) (c) percentage resistance change, measured at different segments of the device, using a four-point percentage resistance change, measured at different segments of the device, using a four-point configuration, after a series of gate voltage pulses. The initial resistance of all segments was RN ~ configuration, after a series of gate voltage pulses. The initial resistance of all segments was RN  1500– 1500–2000 µm. 2000 m. The major effect of field induced oxygen diffusion is a drift of oxygen vacancies going from the The major effect of field induced oxygen diffusion is a drift of oxygen vacancies going from the negatively charged gate to the positive one. Thus, an accumulation of oxygen vacancies confined below negatively charged gate to the positive one. Thus, an accumulation of oxygen vacancies confined the right gate occurs for positive voltage pulses (top Figure 7a), that is detected with a switching of this below the right gate occurs for positive voltage pulses (top Figure 7a), that is detected with a gate to the HRS in the two-point configuration measurement (Figure 7b). The complementary effect is switching of this gate to the HRS in the two-point configuration measurement (Figure 7b). The obtained for negative voltage pulses with an accumulation of oxygen vacancies below the left gate complementary effect is obtained for negative voltage pulses with an accumulation of oxygen (bottom Figure 7a) thus inducing a switching of this gate to the HRS. We have depicted this localized gate effect (from now on referred as gate switching) by coloring blue and red regions (not at scale) below the gates for LRS and HRS, respectively. The gate switching cycles produce a non-trivial reversible oxygen redistribution within the channel, that have been evaluated by measuring the resistance at different segments in a four-point configuration. The measured resistance at a given segment of the channel is directly correlated with its local oxygen concentration, being higher by increasing the amount of oxygen vacancies [25,30]. Thus, assuming a homogenous switch and considering that the resistivity of the HRS is much higher than that of the LRS, the amount of switched channel volume can be directly correlated from the resistance variation as schematically represented in Figure 7a. When one of the gate electrodes is switched to the LRS, there is a region nearby the gate that loses oxygen, cycles produce a non-trivial reversible oxygen redistribution within the channel, that have been evaluated by measuring the resistance at different segments in a four-point configuration. The measured resistance at a given segment of the channel is directly correlated with its local oxygen concentration, being higher by increasing the amount of oxygen vacancies [25,30]. Thus, assuming a Materials 2020, 13,switch homogenous 281 and considering that the resistivity of the HRS is much higher than that of 8 of 11 the LRS, the amount of switched channel volume can be directly correlated from the resistance variation as i.e.,schematically segment N = represented in Figure 2 for the left gate 7a. When (top Figure onesegments 7a) and of the gate N= electrodes 3 and N =is4switched for the rightto the LRS, (bottom there is a region nearby the gate that loses oxygen, i.e., segment N = 2 for Figure 7a). Although the device presents a rather symmetrical R–V curve (Figure 7b), the slightly the left gate (top Figure 7a) and segments different N = 3 and N produces gate performance = 4 for thean right (bottom in asymmetry Figure 7a). Although the oxygen the device redistribution. Thatpresents a rather is, the right gate symmetrical R–V curve (Figure 7b), the slightly different gate performance produces (with a higher HRS value) is able to inject oxygen vacancies in a wider lateral distance than the left one. an asymmetry in It isthe oxygen worth redistribution. noting that the oxygenThatvacancy is, the right gate (with a higherproduces accumulation/decrease HRS value) is able a large to inject effect in the oxygen oxygen vacancies in a wider lateral distance than the left one. It is worth noting redistribution within the channel that induces resistance changes not only in the segments localized that the oxygen vacancy accumulation/decrease between the gates but also produces a large effect in the concomitant in the ones, oxygen redistribution envisaging the large oxygen within the channel mobility occurring that in induces resistance these devices. changes not only in the segments localized between the gates but also in the concomitant ones, modulation Conductivity envisaging the alonglargetheoxygen mobility track can be tunedoccurring by furtherin these devices. the weight of the unbalancing resistance switch in each gate. Figure 8 shows conductance measurements performedthe Conductivity modulation along the track can be tuned by further unbalancing in weight a deviceofwith the resistance w = 100 µm switch and aingate eachdistance gate. Figure d = 5008 shows µm, conductance considering the measurements configuration performed in a device schematically shown within w = 100 m and a gate distance d = 500 m, considering the configuration Figure 8a. In this case the R–V hysteresis curves measured are clearly asymmetric, with very different schematically shown in Figure 8a. In thisstates high-resistance case the R–V hysteresis achieved in the two curves measured gates (Figure are8b).clearly asymmetric, The resistance with very variation at different different high-resistance states achieved in the two gates (Figure 8b). The resistance segments of the channel for this device is shown in Figure 8c. In this case, the oxygen distribution is variation at different segments of the channel with highly inhomogeneous, for this deviceresistance a larger is shownchange in Figure 8c. In this (~200%) case, the oxygen concentrated close todistribution the right gate is highly inhomogeneous, (segment with a larger resistance N = 4), and a complementary smootherchange (200%) resistance concentrated variation, close atolarge that occupies the right gate region of (segment N = 4), and a complementary smoother resistance variation, that the channel, nearby the left gate (segments N = 1, 2, and 3). As in the previous case, the gate that occupies a large region of the channel, undergoes nearby thetoleft a transition gate insulating a more (segmentsstate N = (left 1, 2, gate and in3).this As case) in theisprevious the one thatcase,is the ablegate that to inject undergoes oxygen vacanciesa transition to a more to further lateralinsulating distances.state (left gate in this case) is the one that is able to inject oxygen vacancies to further lateral distances. Figure 8. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel of Figure 8. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel w = 100 µm, by applying positive (top) and negative (bottom) voltage pulses between the two yellow of w = 100 m, by applying positive (top) and negative (bottom) voltage pulses between the two gates separated at d = 500 µm. Red and blue colors represent HRS and LRS, respectively; (b) R–V yellow gates separated characteristics, obtainedatby 500 m. Red d =applying andpulses voltage blue colors represent between HRSinand the gates LRS, respectively; two-point (b) configuration; R–Vpercentage (c) characteristics, obtained resistance change,bymeasured applyingat different voltage pulses segments between the gates of the device, usingina four-point two-point configuration; configuration, after a series of gate voltage pulses. The initial resistance of all segments wasusing (c) percentage resistance change, measured at different segments of the device, R ~ N 150–200 µm. The maximum applied voltage in the I–V curves may also be used as a tuning knob to control the oxygen redistribution (and thus the conductance) through the channel. We show in Figure 9 an extreme case, for a device with w = 30 µm and d = 500 µm, in which we just switch one of the two gates, maintaining the other at the LRS. To do so, we apply a negative voltage pulse higher than Vreset to switch the left contact to the HRS (Figure 9b). This contact is then switched back by to the LRS by was RN  150–200 m. The maximum applied voltage in the I–V curves may also be used as a tuning knob to control the oxygen redistribution (and thus the conductance) through the channel. We show in Figure 9 an Materials extreme2020, case, 281 a device with w = 30 m and d = 500 m, in which we just switch one of the 13, for 9 of two11 gates, maintaining the other at the LRS. To do so, we apply a negative voltage pulse higher than Vreset to switch the left contact to the HRS (Figure 9b). This contact is then switched back by to the LRS by applying a positive voltage higher than Vset but lower than that needed to switch the right contact to applying a positive voltage higher than Vset but lower than that needed to switch the right contact to the HRS (Vreset ). the HRS (Vreset). Figure 9. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel of Figure 9. (a) Schematic representation of the oxygen redistribution in a YBCO device with a channel w = 30 µm, by applying positive (top) and negative (bottom) voltage pulses between the two yellow of w = 30 m, by applying positive (top) and negative (bottom) voltage pulses between the two yellow gates separated at d = 500 µm. Red and blue colors represent HRS and LRS, respectively; (b) R–V gates separated at d = 500 m. Red and blue colors represent HRS and LRS, respectively; (b) R–V characteristics, obtained by applying voltage pulses between the gates in two-point configuration. characteristics, obtained by applying voltage pulses between the gates in two-point configuration. Maximum applied positive voltage pulse has been kept below Vreset in order to maintain the right Maximum applied positive voltage pulse has been kept below Vreset in order to maintain the right contact at the LRS; (c) percentage resistance change, measured at different segments of the device, using contact at the LRS; (c) percentage resistance change, measured at different segments of the device, a four-point configuration, after a series of gate voltage pulses. The initial resistance of all segments using a four-point configuration, after a series of gate voltage pulses. The initial resistance of all was RN ~ 400–500 µm. segments was RN  400–500 m. The resistance variation through the channel is shown in Figure 9c and the associated oxygen The resistance redistribution variation through is schematically depicted the inchannel Figure 9a.is shown In this in Figure case, 9c and the we observe thatassociated by switching oxygen the redistribution left gate to theisLRS, schematically depictedare oxygen vacancies in injected Figure 9a. in In thethis case, wenearby segments observe (Nthat= 1,by2).switching Subsequentthe left gate to the LRS, oxygen vacancies are injected in the segments nearby voltage cycles produce a reversible motion of oxygen vacancies form N = 1, 2 to N = 3. Note that the (N = 1, 2). Subsequent voltage cycles conductance is produce a reversible not affected motionnear in the sections of oxygen the right vacancies gate, N =form 4, 5, N = 1, 2istokept which N = at 3. the Note that the LRS. conductance Resistanceis experiments not affected in the sections directly confirmnearthatthe theright gate, N =between conductance 4, 5, which is kept at the a drain-source LRS. can channel Resistance experiments directly confirm that the conductance between a drain-source be effectively modulated by using different active gates and that the oxygen redistribution in it strongly channel can be effectively depends modulated on the switching by using different performance of each active gate, andgates theand that the applied oxygen voltage redistribution pulse. in it These results strongly depends on the switching performance of each gate, and the applied provide a proof-of-concept of resistance modulation in multi-gate memristor structures based on the voltage pulse. These results provide strongly a proof-of-concept correlated materials showing of resistance modulation the MIT. However, theintop–top multi-gate memristor application structures of voltage based strongly on thenot limits strongly only thecorrelated materials showing device performance but also the theMIT. However, complete the top–top application characterization of synaptic of voltage functions. strongly limits not only the device performance but also the complete characterization For practical applications, top–bottom configurations should be used to increase the on-off ratio, to of synaptic functions. allow For programming reliable practical applications, top–bottom intermediate states configurations and to design should be used the required to increase devices the on-off characteristics ratio, to (linear allow reliable conductance change,programming symmetric setintermediate states and and reset, retention, etc.).toThedesign proposedthe devices requiredprovide devicesa characteristics (linear conductance change, symmetric set and reset, retention, wide design space based on material engineering (oxide doping, oxygen scavenging layers, etc.) and etc.). The proposed devices provide geometrical a wide engineering design (oxide space separation thickness, based on between materialgates,engineering number of (oxide gates, doping, etc.) which oxygen may help to achieve the desired device characteristics for synapses and neurons. 4. Conclusions Our work shows the potential of multi-terminal memristive structures, based on strongly correlated YBCO metallic oxide, as a promising approach for the design of neuromorphic devices, exploiting the Materials 2020, 13, 281 10 of 11 tuneability of field-induced oxygen doping. Results demonstrate that multiple gates can be used to change the conductance (local oxygen doping) between a source-drain channel, thus emulating the synaptic weight. The movement and redistribution of oxygen vacancies within the channel, and thus its conductance, may be controlled by the device geometry, gate dimensions and position, and bias voltage. A large design flexibility can be obtained by changing the switching performance of different gates, thus offering the possibility to locally adjust the conductance response as required to implement neuromorphic functionalities. Author Contributions: A.P., N.M., and A.F.-R. conceived and designed the experiments. A.F.-R., J.A., and A.P. performed sample fabrication and the experiments. A.F.-R analyzed the results. J.S., A.P., and N.M. wrote the manuscript. All authors contributed to results interpretation and reviewed the manuscript. All authors have read and agreed to the published version of the manuscript. Funding: This research was founded by the Spanish Ministry of Economy and Competitiveness through the “Severo Ochoa” Programme for Centres of Excellence in R&D (SEV-2015-0496), SUPERSWITCH project (FUNMAT-FIP-2017), COACHSUPENERGY (MAT2014-51778-C2-1-R), and SuMaTe (RTI2018-095853-B-C21) projects, co-financed by the European Regional Development Fund, MCIU/AEI/FEDER, UE. We also thank support from the European Union for NanoSC Cost Action NANOCOHYBRI (CA 16218) and from the Catalan Government with 2017-SGR-1519. A.F.-R thanks the Spanish Ministry of Economy for the FPI Spanish grant (BES-2016-077310). J.S. was supported by the Spanish Ministerio de Ciencia, Innovación y Universidades and the ECSEL EU Joint Undertaking through the WAKeMeUP 783176 Project. Conflicts of Interest: The authors declare no conflict of interest. References 1. Solomon, P.M. Analog neuromorphic computing using programmable resistor arrays. Solid-State Electron. 2019, 155, 82–92. [CrossRef] 2. Gokmen, T.; Vlasov, Y. Acceleration of deep neural network training with resistive cross-point devices: Design considerations. Front. Neurosci. 2016, 10, 333. [CrossRef] 3. Burr, G.W.; Shelby, R.M.; Sebastian, A.; Kim, S.; Sidler, S.; Virwani, K.; Ishii, M.; Narayanan, P.; Fumarola, A.; Sanches, L.L.; et al. Neuromorphic computing using non-volatile memory. Adv. Phys. X 2017, 2, 89–124. [CrossRef] 4. Suri, M.; Querlioz, D.; Bichler, O.; Palma, G.; Vianello, E.; Vuillaume, D.; Gamrat, C.; De Salvo, B.; Suri, M.; Querlioz, D.; et al. Bio-inspired stochastic computing using binary CBRAM synapses. IEEE Trans. Electron Devices 2013, 60, 2402–2409. [CrossRef] 5. Yu, S.; Gao, B.; Fang, Z.; Yu, H.; Kang, J.; Wong, H.-S.P. Stochastic learning in oxide binary synaptic device for neuromorphic computing. Front. Neurosci. 2013, 7, 186. [CrossRef] 6. Zidan, M.A.; Strachan, J.P.; Lu, W.D. The future of electronics based on memristive systems. Nat. Electron. 2018, 1, 22–29. [CrossRef] 7. Sawa, A. Resistive switching in transition metal oxides. Mater. Today 2008, 11, 28–36. [CrossRef] 8. Janod, E.; Tranchant, J.; Corraze, B.; Querré, M.; Stoliar, P.; Rozenberg, M.; Cren, T.; Roditchev, D.; Phuoc, V.T.; Besland, M.P.; et al. Resistive switching in Mott insulators and correlated systems. Adv. Funct. Mater. 2015, 25, 6287–6305. [CrossRef] 9. Bagdzevicius, S.; Maas, K.; Boudard, M.; Burriel, M. Interface-type resistive switching in perovskite materials. J. Electroceramics 2017, 39, 157–184. [CrossRef] 10. Bi, C.; Meng, X.; Almasi, H.; Rosales, M.; Wang, W. Metal based nonovolatile field-effect transistors. Adv. Funct. Mater. 2016, 26, 3490–3495. [CrossRef] 11. Zhou, B.Y.; Ramanathan, S. Mott memory and neuromorphic devices. Proc. IEEE 2015, 103, 1289–1310. [CrossRef] 12. Stoliar, P.; Tranchant, J.; Corraze, B.; Janod, E.; Besland, M.P.; Tesler, F.; Rozenberg, M.; Cario, L. A leaky-integrate-and-fire neuron analog realized with a Mott insulator. Adv. Funct. Mater. 2017, 27, 1604740. [CrossRef] 13. Kuzum, D.; Yu, S.; Wong, H.S. Synaptic electronics: Materials, devices and applications. Nanotechnology 2013, 24, 382001. [CrossRef] [PubMed] 14. Jo, S.H.; Chang, T.; Ebong, I.; Bhadviya, B.B.; Mazumder, P.; Lu, W. Nanoscale memristor device as synapse in neuromorphic systems. Nano Lett. 2010, 10, 1297–1301. [CrossRef] Materials 2020, 13, 281 11 of 11 15. Han, H.; Yu, H.; Wei, H.; Gong, J.; Xu, W. Recent progress in three-terminal artificial synapses: From device to system. Small 2019, 15, 1900695. [CrossRef] 16. Yang, C.S.; Shang, D.S.; Liu, N.; Fuller, E.J.; Agrawal, S.; Talin, A.A.; Li, Y.Q.; Shen, B.G.; Sun, Y. All-solid-state synaptic transistor with ultralow conductance for neuromorphic computing. Adv. Funct. Mater. 2018, 28, 1804170. [CrossRef] 17. Gkoupidenis, P.; Koutsouras, D.A.; Lonjaret, T.; Fairfield, J.A.; Malliaras, G.G. Orientation selectivity in a multi-gated organic electrochemical transistor. Sci. Rep. 2016, 6, 27007. [CrossRef] 18. Das, S.; Dodda, A.; Das, S. A biomimetic 2D transistor for audiomorphic computing. Nat. Commun. 2019, 10, 1–10. [CrossRef] 19. Sangwan, V.K.; Lee, H.-S.; Bergeron, H.; Balla, I.; Beck, M.E.; Chen, K.-S.; Hersam, M.C. Multi-terminal memtransistors from polycrystalline monolayer molybdenum disulfide. Nature 2018, 554, 500–504. [CrossRef] 20. Stoddart, A. Electronic devices: Making multi-terminal memtransistors. Nat. Rev. Mater. 2018, 3, 18014. [CrossRef] 21. Gou, G.; Sun, J.; Qian, C.; He, Y.; Kong, L.A.; Fu, Y.; Dai, G.; Yang, J.; Gao, Y. Artificial synapses based on biopolymer electrolyte-coupled SnO2 nanowire transistors. J. Mater. Chem. C 2016, 4, 11110–11117. [CrossRef] 22. Zhu, L.Q.; Wan, C.J.; Guo, L.Q.; Shi, Y.; Wan, Q. Artificial synapse network on inorganic proton conductor for neuromorphic systems. Nat. Commun. 2014, 5, 3158. [CrossRef] [PubMed] 23. Palau, A.; Fernandez-rodriguez, A.; Gonzalez-rosillo, J.C.; Granados, X.; Coll, M.; Bozzo, B.; Ortega-hernandez, R.; Sun, J.; Obradors, X.; Puig, T. Electrochemical tuning of metal insulator transition and nonvolatile resistive switching in superconducting films. ACS Appl. Mater. Interfaces 2018, 10, 30522–30531. [CrossRef] [PubMed] 24. Gonzalez-Rosillo, J.C.; Ortega-Hernandez, R.; Arndt, B.; Coll, M.; Dittmann, R.; Obradors, X.; Palau, A.; Suñe, J.; Puig, T. Engineering oxygen migration for homogeneous volume resistive switching in 3-terminal devices. Adv. Electron. Mater. 2019, 5, 1800629. [CrossRef] 25. Wuyts, B.; Moshchalkov, V.V.; Bruynseraede, Y. Resistivity and Hall effect of metallic oxygen-deficient YBa2 Cu3 Ox films in the normal state. Phys. Rev. B 1996, 53, 9418. [CrossRef] 26. Sakai, J.; Ito, N.; Imai, S. Oxygen content of La1−x Srx MnO3−y thin films and its relation to electric-magnetic properties. J. Appl. Phys. 2006, 99, 08Q318. [CrossRef] 27. Waser, R.; Dittmann, R.; Staikov, C.; Szot, K. Redox-based resistive switching memories nanoionic mechanisms, prospects, and challenges. Adv. Mater. 2009, 21, 2632–2663. [CrossRef] 28. Linn, E.; Rosezin, R.; Kügeler, C.; Waser, R. Complementary resistive switches for passive nanocrossbar memories. Nat. Mater. 2010, 9, 403–406. [CrossRef] 29. Nardi, F.; Balatti, S.; Larentis, S.; Gilmer, D.C.; Ielmini, D. Complementary switching in oxide-based bipolar resistive-switching random memory. IEEE Trans. Electron Devices 2012, 60, 70–77. [CrossRef] 30. Lee, Y.S.; Segawa, K.; Li, Z.Q.; Padilla, W.J.; Dumm, M.; Dordevic, S.V.; Homes, C.C.; Ando, Y.; Basov, D.N. Electrodynamics of the nodal metal state in weakly doped high-Tc cuprates. Phys. Rev. B-Condens. Matter Mater. Phys. 2005, 72, 054529. [CrossRef] © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).