Effect of heavy-ion on frequency selectivity of semiconducting polymer/electrolyte heterojunction

Abstract

Heavy ion Nd³⁺ is introduced into the electrolyte layer to study frequency selectivity of a semiconducting polymer/electrolyte double-layer cell. This cell exhibits long-term depression under low-frequency stimulations and potentiation under high-frequency stimulations by positive triangular pulses, suggesting a conventional learning protocol, i.e., spike-rate-dependent plasticity. The frequency selectivity depends significantly on the input shape due to large ionic size and mass. The input threshold of the frequency selectivity is around the voltage inducing a negative differential resistance (V_NDR) influenced by the loading rate. The typical value of V_NDR is 0.3 V for a loading rate of 100 V s⁻¹, but V_NDR disappears when the loading rate exceeds 1000 V s⁻¹. Besides, the frequency selectivity has not been observed under rectangular pulse input. Moreover, the possibility of bidirectional signal transfer has been tested simply by anti-connecting two individual cells. Our study suggests the possibility to realize signal pruning and synthetizing by changing ionic types.

---

Introduction

Neural network computing and synaptic learning simulations have attracted great attention recently.^1–7 Simple memristive cells have been found to emulate synaptic functions and may require less space and energy than silicon-based neuromorphic circuits.^3–5,7–14 Conventional synaptic plasticity and learning protocols have been achieved, such as spike-rate-dependent plasticity (SRDP) and spike-time-dependent plasticity.^15–17 However, it has been suggested that in biology, the synaptic weight is actually regulated by secondary state-variable items such as the postsynaptic Ca²⁺ concentration, but not directly regulated by carefully engineered spike shapes or overlaps.^18,19 A recent study confirmed that the dynamic evolution of internal state variables in an oxide-based memristive system could exhibit Ca²⁺ dynamics and encode timing information for synaptic modulation.^20,21 Therefore, finding a new type of memory media with an internal ion flux might be an effective way to approximate real synaptic plasticity, which becomes a direction of common efforts.

Our previous work has proved the great potential of semiconducting polymer/electrolyte cells in synaptic simulation and brain-like computing.^9,22,23 Some elementary synapse-like characteristics such as frequency selectivity were achieved by modulating ion migration between a semiconducting layer (poly(3-hexyl)thiophene (P3HT) or poly[2-methoxy-5-(2-ethylhexyloxy)-1,4-phenylenevinylene] (MEH-PPV)) and polyethylene oxide (PEO)–Li⁺ electrolyte. As we have known, real neural systems mostly consist of ions with small atomic mass, such as Na⁺ and Ca²⁺,²⁴ which is one reason we started with Li⁺. We adopted a measurement protocol in which a train of spikes is fired and excitatory postsynaptic current (EPSC) or inhibitory postsynaptic current (IPSC) is detected usually,^7,17 which is rarely used in conventional electronics but commonly in neuroscience. Thus, the semiconducting polymer/electrolyte system owes working model approximating biological synapse, upon which the signal transferring, memory and learning depends on ionic kinetics at the interface. Moreover, for an artificial neural network circuitry, there should be many more choices of electrolyte ions that could exhibit more types of plasticity and be suitable for various types of electronic functions. Naturally, we consider in this study to examine the Nd³⁺ ion, which has a large-scale molecular and atomic size as well as multiple valence states. Nd³⁺ salts have been proved to be soluble in a PEO matrix,²⁵ so they could become a suitable candidate for an electrolyte substance. In addition, P3HT was replaced here by poly[2-methoxy-5-(2-ethylhexyloxy)-1,4-phenylenevinylene] (MEH-PPV) as the semiconducting layer exhibiting better doping properties with electrolytes. The mixture of MEH-PPV and the PEO–ion electrolyte has been applied in light-emitting electrochemical cell systems.^26,27 Long term memory (LTM) has been obtained for the cell composed of PEO + Li⁺/MEH-PPV double layer.²³

This study found that large ionic mass and triplet valence influenced current–voltage (I–V) properties significantly. The Nd-doped cells presented SRDP learning protocol as well as Li-doped cells. However, the ionic kinetics was sensitive to loading rate greatly. Negative differential resistance (NDR) appeared but varied with loading rate. Frequency selectivity was also selective to shape of input pulse. Our study suggests signal handling by modulating ionic type.

Experimental details

MEH-PPV, PEO (MW = 100 000), neodymium trifluoromethanesulfonate [Nd(CF₃SO₃)₃], and cyclohexanone were purchased from Sigma-Aldrich Co. Ltd. and used as received. Commercially purchased silicon substrates with a previously deposited 100 nm Pt film were cut into 10 mm × 10 mm squares and used as bottom electrodes (BEs). MEH-PPV was dissolved in cyclohexanone to form an 8 mg mL⁻¹ solution and then stirred on a magnetic hot plate at 38 °C for 24 h. PEO–Nd³⁺ electrolyte solution was prepared by dissolving PEO in deionized water at a concentration of 2 wt% and adding Nd(CF₃SO₃)₃ with a matching principle of EO : Nd³⁺ = 16 : 1/3. During fabrication of the MEH-PPV/PEO–Nd³⁺ double-layer media, all the steps were performed in a glove box filled with nitrogen. MEH-PPV solution (3.5 μL) was first spin-coated on the BEs at a speed of 2000 rpm for 70 s and then dried on a hot plate at 50 °C for 6 h. Next, 3.5 μL of PEO–Nd³⁺ electrolyte solution was drop-cast on the prepared MEH-PPV film and dried at 85 °C for 20 min. Pt top electrodes (TEs) with a thickness of 70 nm and diameter of 300 μm were deposited using electron-beam evaporation through a shadow mask at an evaporation rate of 0.5–0.8 Å s⁻¹. A schematic of the MEH-PPV/PEO–Nd³⁺ double-layer media is shown in Fig. 1(a).

Fig. 1 Structure and component characterization of Pt/MEH-PPV/PEO–Nd(CF₃SO₃)₃/Pt device. (a) Schematic of the device structure. (b) Raman spectra of MEH-PPV, PEO–Nd³⁺, and their double layers. (c) XRD patterns of MEH-PPV, PEO–Nd³⁺, and their double layers. (d) Optical image of PEO–Nd³⁺ crystal on MEH-PPV. (e) and (f) Cross-sectional images of MEH-PPV/PEO–Nd(CF₃SO₃)₃ double-layer device obtained using (e) SEM and (f) electron back-scattering.

All the electrical measurements, including DC sweeps and spike stimulations, were conducted using an Agilent B1530A arbitrary function generator at room temperature. During the measurements, the BEs were grounded unless otherwise noted. The testing details have been described in ref. 23. Raman spectra were obtained using an HR-800 Raman system, in which a 633 nm He–Ne laser was used as the excitation source. X-ray images were obtained by a DMAX 2500V X-ray diffraction (XRD) instrument. Scanning electrical microscopy (SEM) and back-scattered electron images were acquired using an LEO-1530 scanning electron microscope.

Results and discussions

The device structure of Pt/MEH-PPV/PEO–Nd(CF₃SO₃)₃/Pt is shown in Fig. 1(a). To avoid reactions between the electrodes and the organic layers, all electrical measurements were made using Pt electrodes. The electrical properties should be determined only by ionic migration in MEH-PPV, the PEO–Nd³⁺ layer, and the MEH-PPV/PEO–Nd³⁺ junction. The microstructures of the cell were first examined and are presented in Fig. 1(b)–(f). Fig. 1(b) shows Raman spectra of MEH-PPV, PEO–Nd³⁺, and the MEH-PPV/PEO–Nd³⁺ double layer obtained using a 633 nm He–Ne laser as a light source. Because the stimulated wavelength of PEO–Nd³⁺ was much smaller than that of MEH-PPV under the same excitation light source, we amplified the spectra of PEO–Nd³⁺ tenfold for better visibility, whereas the images of MEH-PPV and the MEH-PPV/PEO–Nd³⁺ double layer are shown at their original size. The spectra of both MEH-PPV and the MEH-PPV/PEO–Nd³⁺ double layer exhibited Raman band frequencies at 1112 cm⁻¹ (C–H in-plane bending liberation), 1282 cm⁻¹ and 1308 cm⁻¹ (C=C stretching band of benzene ring), 1583 cm⁻¹ (C–C stretching band of benzene ring), and 1624 cm⁻¹ (C=C stretching band of vinyl groups).^28,29 No significant difference was observed in either the shape or relative intensity between MEH-PPV and the MEH-PPV/PEO–Nd³⁺ double layer, indicating that the chemical structure and gap energy of MEH-PPV was unchanged after it was covered by the PEO–Nd³⁺ layer. That is, no discernable doping occurred in the initial state.

To explore the crystallization condition of the PEO–Nd³⁺ electrolyte, the XRD patterns were then measured and are shown in Fig. 1(c). It has been reported that PEO–LiCF₃SO₃ consists of crystalline PEO, crystalline polymer compound P(EO)₃:LiCF₃SO₃, and an amorphous phase, among which the amorphous part is the primary contributor to the ionic conductivity.^30–34 Here the examined range was 15° to 30°, where MEH-PPV does not make any contribution. We found that both pure PEO–Nd³⁺ and PEO–Nd³⁺ on the MEH-PPV layer contained domains of crystalline PEO, as indicated by XRD peaks at 19.2° and 23.3°.^30,33 However, the pure PEO–Nd(CF₃SO₃)₃ exhibits other diffraction peaks at 17.4°, 19°, 19.6°, and 22.2°, which are thought to represent the crystalline polymer compound, whereas on the MEH-PPV layer, PEO–Nd³⁺ showed only crystalline PEO and an amorphous component. Thus, it could be concluded that in the MEH-PPV/PEO–Nd³⁺ double layer, most Nd³⁺ ions existed only in the amorphous part but did not form crystal compounds with the PEO matrix.

Earlier studies of solid polymer electrolytes such as PEO–LiCF₃SO₃ and PEO–MgCl₂ have already presented optical images of those PEO matrix electrolytes. Typical spherulitic morphology was clearly observed under an optical microscope (Fig. 1(d)), although the ionic size of Nd³⁺ increased greatly, demonstrating electrolyte crystallization with a chain-folding microstructure.³⁴ Compared with Li⁺ and Mg²⁺ doping,^30,35 the grain size of PEO–Nd³⁺ was apparently smaller at a similar (ethylene oxide) EO : ion ratio, with a diameter of only about 100 μm on average, indicating much better nucleation in the PEO–Nd³⁺ system. The microstructure was then examined in more detail by SEM and back-scattered electron imaging. Fig. 1(e) and (f) show cross-sectional images of the same part of our device obtained using SEM and back-scattered electrons, respectively. In theory, back-scattered electron images are bright in regions where the atoms have a large atomic number and dark in regions where they have a small atomic number. Thus, the relatively bright region in the upper and middle parts of Fig. 1(e) are attributed to Nd³⁺ and Pt, respectively. Ordered PEO fibers on the nanoscale were clearly visible (Fig. 1(e)), building a chain fold region in which metal ions were distributed homogenously without segregation (upper bright area in Fig. 1(f)). Ion conduction in electrolytes is reportedly confined within the chain fold region and guided by the crystalline lamellae.^34,36 Thus, this result, combined with the XRD result in Fig. 1(c), allows us to conclude that crystalline PEO has an orderly orientation on the MEH-PPV layer, and Nd³⁺ ions migrated through the amorphous region within the chain fold area. In this way, ion channels analogous to the bio-membrane in synapses were realized, improving the confinement of ion migration between the semiconducting polymer and electrolyte layers, as illustrated schematically in Fig. 2(a).

Fig. 2 Schematic of microstructure in an MEH-PPV/PEO–Nd(CF₃SO₃)₃ cell in (a) its pristine state without any bias and (b) its modified state with positive bias.

The electrical characteristics of the Pt/MEH-PPV/PEO–Nd³⁺/Pt device were first examined by direct current (DC) sweeping at a rate of 100 V s⁻¹, as shown in Fig. 3(a) and (b). Although both exhibit hysteresis loops and residual current, our device behaved much differently at positive and negative bias. The negative differential resistance (NDR) effect was observed only under positive sweeping but not under negative sweeping, indicating a single-directional dynamic doping process in the MEH-PPV/PEO–Nd³⁺ junction resulting from the different ionic mobility of Nd³⁺ and CF₃SO₃⁻ at the interface; a schematic illustration is shown in Fig. 2(b). V_NDR ranged from 0.18 to 0.25 V when the sweeping rate was 100 V s⁻¹ and shifted slightly upward as the circulation increased. This shift in V_NDR resulted from the different residual current at the end of the previous sweeping cycle when the voltage was swept back to 0 V (Fig. 3(a)), indicating that the electrical state was modified by the processing history and that the ion flux could modulate the internal state variables of our cells; these features hold great potential for learning and memory simulation.

Fig. 3 Direct current–voltage (I–V) properties of Pt/MEH-PPV/PEO–Nd(CF₃SO₃)₃/Pt device. I–V curves measured at a sweeping rate of 100 V s⁻¹ in sweep cycles of (a) 0–2 V, (b) 0 to −2 V. Fifth I–V curve out of ten 0–2 V sweeping cycles at scanning rates of (c) 10, 20, 50, and 100 V s⁻¹ and (d) 200, 500, and 1000 V s⁻¹. (e) Time at which NDR effect appeared and (f) NDR position (V_NDR) for various sweeping rates.

We then changed the sweeping rate to explore ion migration at the heterojunction in more detail. Seven loading rates between 10 and 1000 V s⁻¹ were applied to the same cell, and the results for the fifth cycle out of 10 are presented in Fig. 3(c) and (d). We found that as the rate increased, the loop area increased. Because the junction at the MEH-PPV/PEO–Nd³⁺ interface had a capacitive effect, it was reasonable that a larger sweeping rate could generate a lower capacitive reactance as well as a higher corresponding current. Fig. 3(c) also shows that the NDR position (V_NDR) moved to a higher voltage range as the sweeping rate increased (V_NDR ≈ 0.1 V at 10 V s⁻¹, and V_NDR ≈ 0.25 V at 100 V s⁻¹). Our previous work suggested that NDR phenomena would be generated when the internal electric field E_i became larger than the external field E_x at the polymer/electrolyte interface.^9,22 For a single device structure, E_x changed only with the applied voltage, whereas E_i was related to the amount of ion accumulation at the MEH-PPV/electrolyte interface, which depended greatly on the timing and ion mobility. The external and internal fields are denoted as E_x(v) and E_i(μ, t). At a faster sweeping rate, both E_x(v) and E_i(μ, t) would increase more rapidly than the previous values. The total time of the doping process, as well as the time at which NDR appears, would then decrease (Fig. 3(e)). Further, E_i(μ, t) would increase more slowly than E_x(v) because of the time needed for ion accumulation. Therefore, when the sweeping rate was increased, the accumulated ions at the original V_NDR would decrease. A higher voltage was needed for E_i(μ, t) to reach E_x(v) so as to generate the discharge process as well as the NDR phenomena, as shown in Fig. 3(f).

Furthermore, the NDR effect disappeared when the sweeping rate was too fast (i.e., 1000 V s⁻¹, blue line in Fig. 3(d)), and only an inflection point at around 0.7 V could be observed. In addition, the DC curve was not as smooth as those under lower sweeping rates. The extreme case is to load a rectangular pulse with a direct rising edge. The pulse response would be composed of a sharp current rise and subsequent relaxation. No NDR could be observed in this case. The loading rate caused the NDR to disappear earlier than in the case with Li⁺ doping, in which an evident NDR effect appeared at a loading rate of 1000 V s⁻¹. Considering that the Nd³⁺ ion is larger than Li⁺, it was reasonable that it was much more difficult for Nd³⁺ than for Li⁺ to pass through the semiconducting polymer/electrolyte interface. At a rate of 1000 V s⁻¹, the external electric field E_x increased so rapidly that the internal field E_i could not catch up to it. In other words, the response speed of Nd³⁺ ions would become much slower, leading to a flattened potential of the loop line as well as eliminating the NDR effect. Thus, there is a limited sweeping rate for obtaining effective dynamic doping and the NDR effect in the MEH-PPV/PEO–Nd³⁺ heterojunction. The rate of increase of the external electric field would affect Nd³⁺ migration as well as its NDR phenomenon much more than that of other ions with a smaller size and atomic mass. This might lead to signal selectivity of the loading rate depending on the ionic type. The loading rate could become an important factor for encoding timing information.

The frequency selectivity was observed in a P3HT/PEO–Li⁺ double-layer device that exhibited similar NDR phenomena under positive bias.^9,22 This is also an interesting issue for the new system of MEH-PPV/PEO–Nd³⁺ in this work. Two types of stimulations, triangular pulses and rectangular pulses, were used to measure its pulse response. The triangular stimulation increased at a rate of 100 V s⁻¹, and the width of the rectangular pulses was 5 ms. The last discharging peak value in a train of 40 spikes was used to calculate the synaptic weight. This could be regarded as the excitatory postsynaptic current (EPSC) or inhibitory postsynaptic current (IPSC) related to the release probability of neurotransmitters in bio-synapse.^15,17 Some works used the integration value of the EPSC or IPSC to calculate the weight modification and obtained a weight variation that was consistent with that calculated from the peak values of the EPSC. Here we used only the peak values of the EPSC to calculate the weight modification for the following reasons. First, the shape of the synaptic current recorded from biological neurons is typically not regular and is difficult to convert into an integral formula.¹⁷ The noisy disturbed the decayed current after a relaxation period. This problem would also appeared when using the current after a relaxation period as EPSC.³⁷ Second, the synaptic current sometimes contains several peaks resulting from action potential superposition; each peak has a unique biological significance, and they cannot be simply integrated. Thus, if a simple neuromorphic circuit includes two or more types of ions, it would be difficult to use the integration value to distinguish the characterized signal relative to the specific ion. At last, using the peak values of the EPSC would be more efficient and save energy.^15,17

The synaptic weight (W = I_f/I_{f=1 Hz}) at various stimulation frequencies is presented in Fig. 4(a). For triangular pulses, the synaptic weight exhibited depression (W < 100%) under low-frequency stimulation (1–30 Hz) and potentiation (W > 100%) under high-frequency stimulation (30–140 Hz); that is, it exhibited selectivity of the stimulation frequency (red line in Fig. 4(a)). A threshold θ_f at around 30–40 Hz, at which the synaptic weight switched from depression to potentiation, was observed. This result could resemble the CLO learning model proposed by Cooper, Liberman, and Oja,^16,38,39 which solved the unlimited potentiation problem of the traditional Hebbian learning rule by inducing a depression range and a threshold θ_f before potentiation.^16,39,40

Fig. 4 Weight modification with pulse frequencies and corresponding LTM. (a) Weight modification for various frequencies under stimulation consisting of 0.3 V rectangular or 0.3 V triangular pulses. (b) Weight modification obtained 60 min after 0.3 V triangular stimulation. (c) Weight modification for various pulse frequencies under triangular stimulation of different amplitudes from 0.07 to 0.6 V.

However, this frequency selectivity did not occur under rectangular pulses, under which the synaptic weight remained larger than 100% in the entire frequency range (black line in Fig. 4(a)). Considering the disappearance of the NDR effect when the sweeping rate exceeded 1000 V s⁻¹ (Fig. 3(d)), the different responses to various stimulation shapes were reasonable. The voltage obviously increased much more rapidly than 1000 V s⁻¹ in the rectangular pulses, which weakened the dynamic doping and de-doping process at the MEH-PPV/PEO–Nd³⁺ heterojunction. This identification of stimulation rates and the corresponding different responses to spike shapes could possibly be useful in coding timing information. If each type of media ion could correspond to a particular operating pulse shape, it might be possible to realize parallel computing with a huge calculation capacity.

LTM was further tested using 0.3 V triangular pulses. Frequency sequences of 1 Hz (baseline frequency) → n Hz (n > 1, stimulation) → 1 Hz → 1 Hz → 1 Hz → 1 Hz were applied on our device, in which the number of 1 Hz spikes after n Hz stimulation provided a reading function. As shown in Fig. 4(b), both depression (i.e., at 10 Hz) and potentiation (i.e., at 100 Hz) could last for more than 50 min, suggesting an excited state generated by stimulation that could be maintained for a relatively long period before relaxing to its pristine state. This LTM effect was attributed to the introduction of the heavy ion Nd³⁺, which was quite easily entangled with MEH-PPV molecules during doping and then remained for some time before relaxing because of its large ionic size. In the biological synapse system, LTM is crucial to realizing learning and memory. Previously reported frequency selectivity without long-term depression or long-term potentiation could be regarded as an important signalling method but is not related to the synaptic learning function.^9,41 Thus, our results here showing LTM in frequency selectivity represent a great breakthrough for mimicking the learning protocol, i.e., the conventional SRDP.^17,42

Subsequently, we found a strong relationship between the NDR effect and frequency selectivity, in which the latter appeared only when the stimulation voltage was slightly larger than the voltage of the NDR effect if the loading rate was fixed (V_NDR = 0.18–0.25 V, loading rate = 100 V s⁻¹ in Fig. 3(a)). As shown in Fig. 4(c), the synaptic weight switched from depression to potentiation with increasing frequency in the stimulation range from 0.2 to 0.5 V, but monotonous potentiation was observed only when the stimulation value was 0.07 or 0.6 V. We have proposed that the frequency selectivity was induced by timing modulation of the ion doping, de-doping, and re-doping at the semiconducting polymer/electrolyte interface, which was also the reason that frequency selectivity could occur only under a positive bias but not under a negative one.⁹ It could be understood that when the stimulation amplitude smaller than the value of V_NDR was applied (such as 0.07 V in Fig. 4(c)). The smaller driving voltage let the ions move slowly and superficially into the semiconducting layer and then resulted slow and weak de-doping process. Thus, discharging peaks could respond only to the superposition effect of ion doping into the MEH-PPV layer, causing continuous potentiation of the synaptic weight. When the stimulation amplitude was much higher than the value of V_NDR for a fixed loading rate, the re-doping process might be sufficient, so the current values at the pulse end were comparable. Therefore, no significant weight variation appeared, and the pulse response resembled that of an ion-doped electrolyte. However, this explanation of Fig. 4(c) requires further study in depth, i.e., regarding the upper limit of the pulse amplitude (θ_v), because we have obtained frequency selectivity in the P3HT/PEO–Li⁺ system using a voltage amplitude much higher than the value of V_NDR. Determining the value of θ_v is very significant because it might correspond to the depolarization potential in the action potential in the biological neuron synapse, which contains trains of action potential with a fixed loading rate and a certain amplitude range.¹⁵

SRDP learning protocol had been realized previously in the MEH-PPV/PEO–Li cells and a random channel mode had been proposed for the mechanism.²³ In brief, the Li-doped PEO is fibrous with a core of amorphous conductive phase. Li ions in such nano- conductive channel migrate to the MEH-PPV layer under the external electric field. When low frequency stimulation (LFS) is loaded, part of ions in the active channels will be trapped in the MEH-PPV layer. That results in weak ionic recoiling at the end of a pulse and then smaller IPSC and weight depression. When high frequency stimulation (HFS) is loaded, the intensive stimulation will activate the ions in dormant channels and the ions in MEH-PPV corresponding to these channels. The result is that the ionic recoiling increase to induce high EPSC and weight potentiation. We think that the random channel model is suitable for the Nd-doped cells but the mechanism is a little modified due to the two apparent experimental differences. First, the Raman spectra demonstrated that the Li ions were doped into the MEH-PPV layer for the pristine cell, which was not found for the Nd-doped pristine cell. This is the reason that there was not evident NDR effect appearing in this Li-doped cell. Second, long term frequency selectivity could be observed under the rectangular stimulations for this Li-doped cell.²³

Thus, the main reasons for LTM of frequency selectivity of the Nd-doped cells are easily understood. The large size of Nd ions and high valence (+3) led them to relax slowly because the coordination number is large. The depression mechanism of the Nd-doped cells is the same as that the Li-doped cells in LFS. However, though there are not the pristine ions in the MEH-PPV layer could be activated for the Nd-doped cells, the activated channels in the PEO layer could compensate more recoiling ions to result in potentiation weight under HFS, which is relatively small. When the loading rate was very high, i.e., using rectangular pulse, the Nd ions are difficult to be doped into the MEH-PPV layer in significant depth. The valence of Nd might be changed under high loading rate because Nd has several external electrons. In this situation, the Nd ions were not trapped into the semiconducting layer under either LFS or HFS so that weight depression disappeared. That is the reason we did not observe frequency selectivity of the Nd-doped cells.

It is widely known that in the human neural system, signals are generated by an axon, pass through a synapse, and are accepted by a dendrite. As mentioned above, our double-layer cells exhibited frequency selectivity as well as SRDP only under a positive triangular bias (Fig. 4(a) and (b)), which agreed well with the unidirectional procedure in bio-synapse. However, when a nerve fiber was removed from an organism and then operated, neural signals would transmit bidirectionally inside the stimulated nerve fiber, which meant that reactions to both positive and negative stimulation could be realized.¹⁷ The Hebbian learning model also requires an information flow generating an anterograde input signal flow to realize synaptic modification.¹⁷ Thus, in artificial neuromorphic circuits, bidirectional selective functionality would be necessary to meet different types of calculation demands, for the purpose of which the basic components need to be combined.

The above results let us know that our devices have rectifying-like property, and that the ions movements through interface induced the state modification. However, we did not found corresponding symbol to describe our device. Recent studies have used memristor to mimic synaptic plasticity, which demonstrates the variation of resistance.^3–5 The memristor need to combine other component, such as capacitor, to simulate intrinsic ionic kinetics of synapse, e.g., action potential.¹ The electrical properties of our device demonstrated that both resistance and capacity were changed. Thus, we introduced a symbol similar to that of memristor in Fig. 5 to describe the hetero-junction of our device.⁴³ The label ‘X’ indicates that impedance will be modified, which includes resistance, capacity and induction. The bold short bar indicates the rectifying direction because the cation driven to the semiconducting polymer layer induced inverse electrical field and increased resistance, which was also the reason for the NDR effect.

Fig. 5 (a) Schematic of the anti-connecting device structure. The individual cell was indicated by a inset symbol, of which X indicates impedance and short bold bar indicates the rectifying direction. I–V curves measured at a sweeping rate of 100 V s⁻¹ in sweep cycles of (b) 0–2 V and (c) 0 to −2 V. (d) Weight modification for various pulse frequencies under stimulation by 0.5 V (red) and −0.5 V (black) triangular pulses.

We connected the BEs of two Pt/MEH-PPV/PEO–Nd(CF₃SO₃)₃/Pt devices and tested through two TEs, as shown in the schematic in Fig. 5(a). The two devices were equivalent to each other, so a new symmetrical connecting device was built, and either side could be grounded. Fig. 5(b) and (c) show the electrical characteristics of the connected device obtained after DC sweeping; the NDR effect could be observed regardless of whether the bias was positive or negative. Because the NDR effect appeared at approximately ±0.5 V, we further tested its response to spike frequencies separately using ±0.5 V triangular waves. Fig. 5(d) shows that frequency-dependent selectivity could be obtained under both positive and negative bias, though there was a slight difference in the threshold at which the synaptic weight switched from depression to potentiation, possibly as a result of inhomogeneous thickness or salt distribution in the PEO layer. Therefore, it could be proved that by using a simple electrical series, the unidirectional frequency selectivity of the Pt/MEH-PPV/PEO–Nd(CF₃SO₃)₃/Pt device could be expanded to two directions, showing great potential for future neuromorphic circuits.

Combining the previous studies, we think that it is possible to realize signal pruning and synthetizing by changing ionic types.⁴⁴ For example, connecting a Li-doped cell with a Nd-doped cell in parallel could modulate the band width of either depression or potentiation, or recognize loading rate. Anti-collecting Li & Nd doped cells would result in asymmetric signal transferring between the upstream and downstream neurons, which leads to mutual signal feedback. Moreover, the ionic concentration in the cells are able to be modified to modulate signal strength and direction. These are our considerations in the following studies that we suppose to be helpful exploring learning rules and prompting their applications.

Conclusions

A semiconducting polymer/electrolyte cell using the heavy ion Nd³⁺ to provide an ionic medium was fabricated. Crystalline PEO was found to have an orderly orientation on the MEH-PPV polymer layer, and Nd³⁺ ions were thought to migrate through the amorphous region within the chain fold area. The NDR effect was observed under positive DC sweeping. The position of V_NDR was found to change with the sweeping rate, and the NDR effect disappeared when the rate exceeded 1000 V s⁻¹. This cell exhibited frequency selectivity under a positive triangular bias, responding with depression under low-frequency stimulation and with potentiation under high-frequency stimulation. Both the depression and potentiation effects could last for about an hour, so LTM as well as the conventional SRDP learning model were realized. However, under a positive rectangular bias, the synaptic weight remained larger than 100% in all frequency ranges and exhibited no selectivity, indicating recognition of the external sweeping rate. Frequency selectivity was related to NDR phenomena because it existed only when the stimulation value was close to V_NDR. By simply connecting the BEs of two individual Pt/MEH-PPV/PEO–Nd³⁺/Pt devices, the original single-directional frequency selectivity was successfully expanded to both the positive and negative directions, which may be useful for future artificial neuromorphic circuits.

Acknowledgements

This work was supported by National Natural Science foundation of China (Grant No. 51231004 and 51371103) and National Hi-tech (R&D) project of China (Grant Nos. 2014AA032901 and 2012AA03A706) and Brain Inspired Computing Research, Tsinghua University (20141080934).

References

1. M. D. Pickett, G. Medeiros-Ribeiro and R. S. Williams, Nat. Mater., 2013, 12, 114–117.
2. J. J. Yang, D. B. Strukov and D. R. Stewart, Nat. Nanotechnol., 2013, 8, 13–24.
3. T. Ohno, T. Hasegawa, T. Tsuruoka, K. Terabe, J. K. Gimzewski and M. Aono, Nat. Mater., 2011, 10, 591–595.
4. S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumder and W. Lu, Nano Lett., 2010, 10, 1297–1301.
5. Z. Q. Wang, H. Y. Xu, X. H. Li, H. Yu, Y. C. Liu and X. J. Zhu, Adv. Funct. Mater., 2012, 22, 2759–2765.
6. C. Zamarreño-Ramos, L. A. Camuñas-Mesa, J. A. Pérez-Carrasco, T. Timothée Masquelier, T. Serrano-Gotarredona and B. Linares-Barranco, Front. Neurosci., 2011, 5, 26.
7. G. Rachmuth, H. Z. Shouval, M. F. Bear and C. S. Poon, Proc. Natl. Acad. Sci. U. S. A., 2011, 108, E1266–E1274.
8. S. Z. Li, F. Zeng, C. Chen, H. Liu, G. S. Tang, S. Gao, C. Song, Y. S. Lin, F. Pan and D. Guo, J. Mater. Chem. C, 2013, 1, 5292–5298.
9. F. Zeng, S. H. Lu, S. Z. Li, X. J. Li and F. Pan, PLoS One, 2014, 9, e108316.
10. W. He, K. J. Huang, N. Ning, K. Ramanathan, G. Q. Li, Y. Jiang, J. Y. Sze, L. P. Shi, R. Zhao and J. Pei, Sci. Rep., 2014, 4, 4755.
11. Y. Li, Y. P. Zhong, L. Xu, J. J. Zhang, X. H. Xu, H. J. Sun and X. S. Miao, Sci. Rep., 2013, 3, 1619.
12. Y. Li, L. Xu, Y. P. Zhong, Y. X. Zhou, S. J. Zhong, Y. Z. Hu, L. O. Chua and X. S. Miao, Advanced Electronic Materials, 2015, 1, 1500125.
13. D. S. Jeong, I. Kim, M. Ziegler and H. Kohlstedt, RSC Adv., 2013, 3, 3169–3183.
14. F. Zeng, S. Z. Li, J. Yang, F. Pan and D. Guo, RSC Adv., 2014, 4, 14822.
15. S. M. Dudek and M. F. Bear, Proc. Natl. Acad. Sci. U. S. A., 1992, 89, 4363–4367.
16. L. N. Cooper and M. F. Bear, Nat. Rev. Neurosci., 2012, 13, 798–810.
17. G. Q. Bi and M. M. Poo, J. Neurosci., 1998, 15, 10464–10472.
18. H. Z. Shouval, M. E. Bear and L. N. Cooper, Proc. Natl. Acad. Sci. U. S. A., 2002, 99, 10831–10836.
19. M. Graupner and N. Brunela, Proc. Natl. Acad. Sci. U. S. A., 2012, 109, 3991–3996.
20. S. Kim, C. Du, P. Sheridan, W. Ma, S. H. Choi and D. W. Lu, Nano Lett., 2015, 15, 2203–2211.
21. C. Du, W. Ma, T. Chang, P. Sheridan and D. W. Lu, Adv. Funct. Mater., 2015, 25, 4290–4299.
22. S. H. Lu, F. Zeng, W. S. Dong, A. Liu, X. J. Li and J. T. Luo, Nano-Micro Lett., 2015, 7, 121–126.
23. W. S. Dong, F. Zeng, S. H. Lu, A. Liu, X. J. Li and F. Pan, Nanoscale, 2015, 7, 16880–16889.
24. M. E. Bear, B. W. Connors and M. A. Paradiso, Neuroscience: Exploring the Brain, High Education Press, Beijing, 2007.
25. A. Aabloo and J. O. Thomas, Electrochim. Acta, 1998, 43, 1361–1364.
26. Q. B. Pei, G. Yu, C. Zhang, Y. Yang and A. J. Heeger, Science, 1995, 269, 1086–1088.
27. J. Gao, Y. F. Li, G. Yu and A. J. Heeger, J. Appl. Phys., 1999, 86, 4594.
28. T.-T. Cuong, M. R. Phillips and N. Thien-Phap, J. Lumin., 2008, 128, 2031–2034.
29. P. Damlin, C. Kvarnstrom and A. Ivaska, Electrochim. Acta, 1999, 44, 4087–4097.
30. C. D. Robitaille and D. Fauteux, J. Electrochem. Soc., 1986, 133, 315–325.
31. A. Vallee, S. Besner and J. Prudhomme, Electrochim. Acta, 1992, 37, 1579–1583.
32. C. Berthier, W. Gorecki, M. Minier, M. B. Armand, J. M. Chabagno and P. Rigaud, Solid State Ionics, 1983, 11, 91–95.
33. C. P. Rhodes and R. Frech, Macromolecules, 2001, 34, 2660–2666.
34. W. S. Young, P. J. Brigandi and I. T. H. Epps, Macromolecules, 2008, 41, 6276–6279.
35. Y. Yang and H. Huo, J. Polym. Sci., Part B: Polym. Phys., 2013, 51, 1162–1174.
36. S. K. Fullerton-Shirey and J. K. Maranas, Macromolecules, 2009, 42, 2142–2156.
37. Q. X. Lai, L. Zhang, Z. Y. Li, W. F. Stickle, R. S. Williams and Y. Chen, Adv. Mater., 2010, 22, 2448–2453.
38. M. F. Bear, L. N. Cooper and F. F. Ebner, Science, 1987, 237, 42–48.
39. L. N. Cooper, F. Liberman and E. Oja, Biological Cybernetics, 1979, 33, 9–28.
40. Y. Fregnac and M. Imbert, J. Physiol., 1978, 278, 27M24.
41. Z. Rotman, P. Y. Deng and V. A. Klyachko, J. Neurosci., 2011, 31, 14800–14809.
42. N. Intrator and L. N. Cooper, Neural Networks, 1992, 5, 3–17.
43. A. Adamatzky and L. Chua, Memristor Networks, Springer, Switzerland, 1st edn, 2014.
44. L. F. Abbott and W. G. Regehr, Nature, 2004, 431, 796.

---