Pre-synaptic DC bias controls the plasticity and dynamics of three-terminal neuromorphic electrolyte-gated organic transistors

The recent effort towards the development of efficient neuromorphic components and devices, aimed at coupling high computational efficiency with low power consumption by mutating processing and learning paradigms for neural systems, led to identify a number of candidate materials and device architectures [1]. Organic (semi-)conductive materials possess highly desirable figures of merit for devising neuromorphic organic electronic systems. Neuromorphic properties, such as spike time dependent plasticity [2], control on global connectivity [3], multiterminal operations [4] and time-dependent pattern classification [5] were demonstrated with organic device architectures. Organic neuromorphic devices are often termed organic artificial synapses in analogy with neural synapses.

The neuromorphic response of organic electronic devices arises by the dynamics following the out-of-equilibrium displacement (by means of an input signal) of the active material (the organic (semi-)conductor) initially at equilibrium with the operational electrolyte [6]. Since the equilibrium involves all the species dissolved in the solution, the artificial synapses will respond differently to different compositions of the solution. In the presence of strong interactions between some solute species and the material, the dynamics of recovery the equilibrium condition, once the input signal is removed, can be particularly slow. Artificial synapse architectures effectively sense those moieties in solution that give rise to strong, albeit not necessarily specific, interactions. Indeed, administration of square voltage pulses at a fixed frequency (as input signal) results for the species with strong interactions in the kinetic unbalance between the processes of adsorption/retention and desorption/expulsion in/out the organic active material [7, 8]. Each input voltage pulse drives the organic (semi) conductor to a 'metastable' state with a different conductance level with respect to the initial one (the baseline). The response to the sequence of input voltage pulses is a ladder of current spikes of different intensities, either decreasing (depressive short-term plasticity—STP—response) or increasing (facilitating STP response). Once in the final metastable state, the conductance/current relaxes to the initial baseline value with a time scale characteristic of the composition of the solution, of the baseline and of the features (amplitude, frequency and number of spikes) of the input signal. This means that the initial and final conductance states of artificial synapses in response to a defined input signal are connected through a precise pathway. One can also regard the relaxation time from the final metastable state to the initial state as the result of training the device by means of the input signal, because the relaxation is characteristic of the two states connected by the input signal. If the relaxation time is long, then the artificial synapse has kept memory of the input signal, and in case of very long time scales we term this long-term plasticity—LTP.

It is clear that the behavior of artificial synapse arises from the interplay between active material conductance and ion density in close proximity to the material itself, resulting in the newly coniated definition of such materials as organic mixed ionic-electronic conductors—OMIECs [9].

If this is true in the steady-state, it becomes even more critical when dealing with transient phenomena, in which the dramatic difference between ion dynamics and charge transport dynamics gives rise to a complex out-of-equilibrium behavior, which is regarded as the origin of the neuromorphic response.

The key role of the Gibbs free energy in determining kinetics and magnitude of doping/dedoping processes in OMIECs, and, in particular, in the most widespread poly(3,4-ethylenedioxythiophene):polystyrene sulfonate—PEDOT:PSS—has been demonstrated and exploited in the development of binding-energy-specific sensors [6]. Lately, it was unambiguously demonstrated how OMIEC conductance is a cumulative observable which provides a comprehensive insight on the thermodynamic and kinetic condition of the electrolyte/OMIEC system [10]. Interestingly, such phenomena can be observed also in non-OMIEC-based architectures, provided that they integrate ionic density in proximity with a (semi-)conductive channel in their working principle (e.g. using poly-electrolytes as dielectric layers) [11, 12].

Among organic electronic device architectures, the one which majorly benefits from the above discussed features and integrates them in its functional principle is the electrolyte-gated organic transistor—EGOT [13]. In EGOTs, a semi-conductive film bridges two contacts—termed source (S) and drain (D)—and it is directly exposed to an electrolyte, whose potential is set by a third contact—termed gate (G)—directly bathing in the electrolyte. In such architectures, small gate voltage—V_GS—variations result in significant channel current—I_DS—variations, making EGOTs ideal platforms for transduction [14–16] and sensing applications [17–19]. A quantitative figure of merit for such applications of EGOT technology is the transconductance, g_m = dI_DS /dV_GS.

Focusing on different aspects of their working principle, EGOTs are referred to as electrolyte-gated organic field-effect transistors—EGOFETs [20], organic electro-chemical transistors—OECTs [21] or ion-gated transistors—IGT [22]. The key aspect of all these architectures is the role of the gate electrode in setting the organic semiconductor conductance. As discussed above, this implies a direct role of the gate electrode on the Gibbs free-energy of the system and, subsequently, on its response kinetics when it is operated as an artificial synapse. According to this and to some previous observations in similar architectures [23], the hypothesis addressed in this work is that a modulation of the DC offset at the gate electrode—usually termed pre-synaptic terminal in neuromorphic architectures—should imply not only a modulation of steady-state observables such as channel (trans-)conductance but also a difference in transient aspects, such as the temporal evolution of the 'post-synaptic' channel current.

In this work, the hypothesis is tested against standard neuromorphic characterization protocols and rationalized in terms of displaced charge. We highlight the key role of transconductance as an estimator for amplitude, resolution, time evolution and 'direction' of xeromorphic phenomena in artificial synapses based on EGOT architectures. This work provides a comprehensive discussion on the use of pre-synaptic DC offset at the gate electrode, which governs trans conductance, as a prompt tuning tool towards the response of such devices, achieving fine modulation of transient STP and LTP effects in the post-synaptic drain-source current, reversibly switching from depressing to potentiating behavior, without the need of changing active material [24, 25], of permanently altering the device layout [26] or of rearranging the connection layout [23].

PEDOT:PSS-based EGOTs were fabricated as described in the Experimental section and their steady-state I–V response has been characterized. Results are summarized in figure 1. Figure 1(a) schematically shows the chosen EGOT architecture with the common-source/common-ground connection layout. Figure 1(b) reports a representative transfer characteristic (forward and backward gate sweep), I_DS vs V_GS, and the corresponding transconductance (estimated as first derivative from the forward scan branch), g_m vs V_GS. The plot also evidences a negligible leakage I_GS current, a modest clockwise hysteresis and a maximum g_m ≈ 12 mS at V_GS ≈ −0.75 V. Output characteristics (figure 1(c)) show ohmic behavior at V_DS ⩽ 0 with gate dependent saturation plateaus, as expected for p-type depletion mode EGOTs. These EGOTs are in line with the state-of-the-art PEDOT:PSS-based devices and exhibit the well-known dependence of I_DS on V_GS and on V_DS, as evidenced by the 3D plot in figure 1(d).

Zoom In Zoom Out Reset image size

The role of DC pre-synaptic bias in governing the neuromorphic response of EGOTs was investigated by characterizing the short-term plasticity (STP) response of the post-synaptic current—I_DS—to square voltage pulses superimposed to the DC offset at the pre-synaptic terminal (i.e. the gate electrode). In figure 2(a) the schematic drawing of the device layout with the connections evidences the input voltage generator connected to the gate, and the amperometer to the drain electrode. A representative STP characterization is shown in figure 2(b): upon a train of voltage square pulses, either positive or negative, termed programming phase, the response of neuromorphic EGOTs is either depressed or potentiated. These responses are due to the progressive accumulation (in the case of depression) or depletion (in the case of potentiation) of cations in close proximity of the PEDOT:PSS channel upon repeated pulsing. Upon each pulse, the device responds with an I_DS current spike that partially relaxes before the next voltage pulse is administered. The result is the mean I_DS current (and its conductance) drifting towards a novel, dynamical steady state different from the initial equilibrium state. When the programming phase is over, the channel current/conductance slowly relaxes back to its initial DC value.

Zoom In Zoom Out Reset image size

Figure 2(c) shows the I_DS response to two subsequent (first depressive, then potentiating) STP programming cycles. The current baseline moves towards lower conductance steady states for V_GS increasing from −0.7 V to 0.7 V. This modulation of the baseline conductance value by the pre-synaptic DC bias is indeed expected from EGOT I–V characteristics (figure 1). Surprisingly, instead, the transient current response to the pre-synaptic DC bias exhibits much larger relative variation as the V_GS becomes more positive. This can be evidenced by expressing the STP response in terms of percentage current variation with respect to the baseline.

$$\begin{align}{I_{\text{DS}},\operatorname{var}} = 100 \times \frac{{I_{\text{DS}}}(t) - {I_{\text{DS}}}(0)}{{I_{{\text{DS}}}}(0)}.\end{align}$$

Figure 2(d) shows that more positive V_GS values (lower PEDOT:PSS baseline conductance) yield greater I_DS,var for both programming and relaxation phases. Importantly, these effects of pre-synaptic DC bias are conserved also upon repeated reversible switching between depressive and potentiating STP programming phases (figure 2(e)).

Moreover, as evidenced in figure 2(f), the relaxation kinetics is also slower. This is inferred by setting an arbitrary reading time, t_read (vertical dashed line in figure 2(b)), after the programming phase end, and measuring the relative increase/decrease of the current from the minimum/maximum of the last programming pulse. Figure 2(f) shows the linear correlation between the DC V_GS bias and the residual I_DS,var at t_read, for both potentiation and depression. The combined effects of modulated relative amplitude and relaxation time scale open unprecedented possibilities for the use of such devices in neuromorphic circuitry. Indeed, by setting the DC pre-synaptic bias, it is possible to decide if a certain programming phase results in the storage of information and also the relative magnitude of the stored information. An additional critical parameter which is directly controlled by the DC V_GS bias is the number of discrete accessible conductance states during programming (figure 2(g)). In detail, by setting a discretization threshold of ΔI_DS,var = 2% to discriminate between two states, it is evident that the higher the I_DS,var (i.e. the more positive the V_GS), the larger the number of conductance states. Figure 2(g) shows a zoom on the I_DS,var maxima during a depressive STP programming cycle, V_GS values spanning from −0.7 V to 0.7 V. Conductance states are identified by horizontal lines at multiples of 2% I_DS,var, accessible states (i.e. states in which maxima actually fall) are highlighted in green and blue and counted by dots on the right axis, prohibited states are highlighted in red. As shown in figure 2(h), also the number of accessible states scales linearly with V_GS and goes from 3 to 10 while moving from negative V_GS values to positive ones.

A useful tool to provide an a priori estimate of the number of accessible conductance states, n, for a given discretization threshold, d_th, which shall be set according to the resolution of the readout strategy in the desired neuromorphic application, can be given in terms of STP_span, defined as the absolute value of the difference between I_DS,var in response to the first programming pulse and I_DS,var in response to the last one. In particular,

$$\begin{equation*}n = \operatorname{int} \left( {{\raise0.7ex\hbox{${{\text{ST}}{{\text{P}}_{{\text{span}}}}}$} \!\mathord{\left/ \right.} \!\lower0.7ex\hbox{${{d_{th}}}$}}} \right) + 1.\end{equation*}$$

As a consequence, in the proposed neuromorphic EGOTs, the linear relation between STP_span and V_GS is mirrored by a linear n vs V_GS dependence in the considered V_GS range (figure 2(i)).

Such profound effect of DC gate voltage on neuromorphic phenomena can be rationalized in terms of exchanged charge upon programming. As a consequence, it is necessary to extensively characterize the channel response upon the application of square pulses (figures 3(a) and (b)). A single square voltage pulse—either positive or negative (top panels)—is applied at the pre-synaptic electrode while measuring the elicited current in the post-synaptic channel between the source and drain electrode (center panels). The measured post-synaptic I_DS is then integrated in time to obtain chronocoulograms (Q vs t), here reported in terms of charge variation (ΔQ) with respect to the charge value before pulse administration (bottom panels). This protocol is repeated for various DC V_GS,Offset values, ranging from −0.7 V to 0.7 V, resulting in the trends depicted in figures 3(c) and (d).

Zoom In Zoom Out Reset image size

Concerning the single negative pulse, figure 3(c), it is possible to observe how during the pulse there is a linear time profile of ΔQ, whose slope (i.e. magnitude for constant pulse duration) depends on the applied V_GS,Offset. A phenomenologically consistent behavior can be observed in response to positive single pulses, albeit with an opposite sign and different magnitude (figure 3(d)). Figures 3(e) and (f) show the magnitude of the exchanged charge in the channel upon pulse administration as a function of V_GS,pulse for negative and positive voltage pulses, respectively. Here V_GS,pulse equals V_GS,Offset−0.5 V for negative pulses and V_GS,Offset+ 0.5 V for positive ones.

The non-linear trends in figures 3(e) and (f) can be rationalized assuming that the charge exchanged in the channel during the single pulse is affected by the channel doping prior to the pulse administration and by its proneness to be altered which is physically measured by the transconductance. Indeed, both are set by V_GS,Offset. A demonstration is given in figure 3(g), which shows the overlay between the trends of the transconductance, g_m , and of the absolute value of charge exchanged in the channel during pulse administration, |ΔQ_pulse |, vs V_GS,Offset. It is possible to observe a clear correlation between these two observables (figure 3(h)), especially at small transconductance (else positive V_GS,Offset values), which is progressively lost upon moving towards fully doped channels (at strongly negative V_GS,Offset). The charge variation saturates when transconductance attains its highest values.

Noticeably, as reported in figures 3(c) and (d), upon pulse removal, ΔQ does not always remain constant. Yet, it may exhibit different relaxation kinetics, which vary according to the V_GS,Offset and, hence, to g_m , before reaching a steady-state value. In particular, for high transconductances, the system exhibits an overcharging—or an overdepletion for positive pulses—upon pulse administration, which needs to be recovered in time, following the kinetics of the relevant displaced charged species (i.e. anions or cations) in the electrolyte at the interface with the channel. On the other hand, in regimes of low transconductance, pulse removal results in total cessation of charge exchange in the channel in the case of negative pulses, whereas it causes slow charge accumulation in the case of positive ones. This important difference can be ascribed to the different sign and hidro-dynamic volume of the ions which come into play, according to the pulse sign, coupled to the 'accessible' coordination sites for ions present in the channel, set by V_GS,Offset. In this sense, this kinetic processes closely mirror the behavior of transconductance and, for this reason, transconductance non-linearity is at the origin of neuromorphic tunability in EGOTs.

The applicability of such concepts to neuromorphic circuitry development is assessed in a paired-pulse experiment, investigating depressive and facilitating phenomena upon the application of paired square pulses (amplitude −500 mV) at the pre-synaptic gate terminal, while varying the inter-stimulation interval—ISI—from 20 ms up to 200s and V_GS,offset from −0.7 V to 0.7 V (figure 4(a), left). A quantitative representation of the entity of depression/facilitation can be given defining a paired-pulse plasticity index—PPP_i :

Zoom In Zoom Out Reset image size

$$\begin{equation*}{\text{PP}}{{\text{P}}_i} = 1 - \frac{{{A_2}}}{{{A_1}}}.\end{equation*}$$

In such definition, A₁ and A₂ are the differences between post-synaptic current peaks (in response to the first and the second pre-synaptic pulse, respectively) and the baseline current value before pulse administration, as shown in representative traces in figure 4(a). Positive values of PPP_i indicate a depressive response while negative values testify a facilitative behavior; in both cases, |PPP_i | measures the entity of the phenomenon.

PPP_i dependency on V_GS,offset and ISI is shown in figure 4(b). It is possible to distinguish between four main scenarios:

• (a)  
  When ISI is short and the semiconductive channel is ON (i.e. at negative V_GS,offset values), a depressive response is observed, with PPP_i > 0. This is the most commonly reported case in the literature;
• (b)  
  When ISI is short and the semiconductive channel is OFF (i.e. at positive V_GS,offset values), the system responds linearly with PPP_i ≈ 0;
• (c)  
  When ISI is long and the channel is ON, the system exhibits depressive response, although of smaller magnitude (i.e. smaller positive PPP_i values);
• (d)  
  When ISI is long and the channel is OFF, it is possible to achieve consistent potentiation of the post-synaptic current (i.e. PPP_i < 0).

Noticeably, a number of intermediate situations can be obtained while moving in between these four extremes, resulting in the possibility of arbitrarily selecting PPP_i by properly tuning ISI and V_GS,offset.

It is important to highlight the fact that, at the origin of scenarios 3 and 4, there is the same physical peculiarity of the channel. In particular, if the channel is left unperturbed for a certain amount of time and, at the same time, it is exposed to a constant V_GS,offset value, its conductivity tends to increase. This is surprising since constant V_GS,offset should imply a constant ionic environment in close proximity to the channel which should set a single conductivity value in EGOTs. The fact that this effect is more pronounced in the case of positive V_GS,offset , when cations are forced at the interface with the PEDOT:PSS channel, hints at the presence of slow rearrangement phenomena which allow the system to 'cope' to the presence of cations and to partially restore, in time, its conductivity. This explanation, which ultimately implies the presence of kinetically stored electrostatic energy in PEDOT:PSS films at the interface with electrolytes, is in good agreement with previous reports of the existence of electrochemical pseudo-inductive contributions to such interfaces [7], which give rise to facilitative plasticity responses at long ISI and widens the applicability context of neuromorphic EGOTs.

3.1. EGOT fabrication

3.2. Electrical characterization

3.3. Analysis, graphing and presentation

In this work, a comprehensive study on the critical importance of the DC offset bias at the gate electrode in determining the neuromorphic response of EGOTs is presented. V_GS,offset is shown to control the number of accessible discrete conductance states in short-term plasticity-based memories, operated both in depression and in potentiation. It is discussed the relationship between V_GS,offset, device transconductance and exchanged charge upon pulse administration and it is explained how the complex interplay between these parameters governs neuromorphism. Finally, paired-pulse plasticity is investigated as a function of V_GS,offset, leading to a tuning map for PPP_i in EGOT-based neuromorphic applications.

These results add a degree of freedom in the design of neuromorphic circuitry by endowing the operator of an additional powerful tuning parameter, whose role arises from the peculiar features of EGOTs, paving the way towards the new-generation organic neuromorphic platform for real-time signal processing and pattern recognition.

Research work leading to this publication was funded by IIT—Istituto Italiano di Tecnologia, University of Ferrara and University of Modena and Reggio Emilia (FAR 2018 Project e-MAP).

The data that support the findings of this study are available upon reasonable request from the authors.