Leaky Integrate‐and‐Fire Neuron Based on Organic Electrochemical Transistor for Spiking Neural Networks with Temporal‐Coding

Spiking neural networks (SNNs) employ discrete spikes that mimic the firing of neurons in biological systems to process and transmit information. This characteristic enables SNNs to effectively capture temporal dynamics and capitalize on the time information inherent in time‐varying inputs, such as motion, audio/video streams, and other sequential data. Currently, most hardware implementations of SNNs are designed to use rate‐coding, where information is encoded in the rate of spikes. However, it still remains challenging for the hardware implementation of temporal coding in SNNs, which allows for higher input sparsity and exploits additional dimensions such as precise spike timing and relative spike timings. This study presents hardware implementations of SNNs constructed by organic electrochemical transistors (OECTs), processing temporal‐coded information. The protic dynamics in response to electrical stimuli enable the emulation of temporal integration, reset, and leaking of membrane potential in a simple leaky integrate‐and‐fire (LIF) neuron circuit. By utilizing these features, the emulated LIF neuron can be employed to construct SNNs capable of processing temporal‐coded information in complex tasks including coincidence detection and dynamic handwriting recognition, exhibiting high performance and good tolerance even when dealing with noisy datasets.


Introduction
[3] Figure 1a represents the spatial and temporal integration of biological neurons.Spatial integration involves the summation of stimuli from different pre-neurons connecting to the post-neuron's dendrites, while temporal integration refers to the combining of stimuli over time. [4,5]This spatiotemporal integration of pre-neurons' stimuli determines whether the membrane potential of the soma of postneuron reaches the threshold for firing an action potential, also known as a spike, being transmitted down the axon and further signaling downstream neurons.This spiking behavior ensures that only preferred inputs that exceed the threshold after integration trigger the output, saving energy by avoiding the unnecessary firing of weak or irrelevant stimuli. [6]y emulating the neural spatiotemporal integration and spiking behavior, SNNs are designed to process complex information encoded by spike trains through rate (frequency)-coding and temporal-coding, [7] respectively, shown in Figure 1b,c.[28][29][30] The operation of such memristive devices relies on structural transformation, like the forming or breaking of conductive filaments within the material, in response to an applied voltage or current.The inherent randomness and intricate interplay among internal factors of the memristive materials make it challenging to precisely manipulate such physical transformation using external stimuli, consequently leading to fluctuations in the output of the device.Therefore, TS devices are primarily utilized for implementing SNNs with rate-coding, where information is encoded through the average firing rate. [31,32]However, it's still a challenge for the hardware implementation of SNN with precise temporal coding, where information is encoded through the timing of individual spikes with higher sparsity, as illustrated in Figure 1c.Compared to rate coding which relies solely on statistical representations, temporal coding has the potential to increase the amount of processed information and improve energy efficiency when dealing with large volumes of sequential data.
In this work, organic electrochemical transistors (OECTs) are investigated for building spiking neurons and SNN systems with precise temporal coding.[35][36][37] OECTs are attractive for this purpose due to their ability to couple intrinsic neuron-like ionic dynamics to the modulation of device properties.However, most of these works were unable to effectively process temporalcoded information.Here, a LIF neuron circuit has been designed, where the OECT plays a crucial role in realizing the neural dynamics, including temporal integration, reset, and leaking of membrane potential.By utilizing this LIF neuron and temporal coding, SNNs are simulated to perform tasks including coincidence detection and handwriting recognition.

Results and Discussion
Figure 2a shows the schematic structure and circuit diagram of our poly-[2,5-bis(2-octyldodecyl)−3,6-di(thiophen-2-yl)pyrrolo [3,4-c]pyrrole-1,4(2H,5H)-dionel-alt-thieno [3,2b]thiophene] (DPPT-TT)-based OECT, gated by chitosan-based electrolyte dielectric.The fabrication process is illustrated in Figure S1 (Supporting Information).Figure 2b,c shows the transfer and output characteristics of the DPPT-TT transistor, indicating its typical p-type behavior.The operating voltage is relatively low compared to that of the previous research, [38,39] due to the electric-double-layer (EDL) capacitive behavior of the chitosan-based protic electrolyte dielectric. [40]The EDL is formed at the interface between the electrolyte and the channel, whereupon gate-to-channel voltages the depletion/accumulation of interfacial protons induce/reduce holes in the p-type DPPT-TT channel, turning on/off the device, as illustrated in Figure S2 (Supporting Information).Figure 2d,e shows transient I DS responses upon V GS pulses with varied amplitudes (from −1.0 to −5.0 V) and varied pulse widths (from 1.0 to 7.0 s), respectively, at V DS of −5 V.Each I DS response includes an increase caused by the proton depletion at the electrolyte/channel interface under a negative V GS pulse, followed by a decrease after the pulse because of the spontaneous recovery from proton-depleted state, also explained in Figure S2 (Supporting Information).This transient response usually emulates excitatory postsynaptic current (EPSC) which is typical in biological neural systems. [41]EPSC is generated in the post-neuron when it receives an excitatory input from the pre-neuron, caused by the release of neurotransmitters.Similar to the biological EPSC, the amplitude of EPSCs (I DS responses) in the OECT can be modulated. [42]Figure 2d,e shows that the V GS pulse with a larger amplitude or width can result in a higher peak of EPSC.A larger negative V GS generates a stronger channel-to-gate electric field, thereby enhancing the proton depletion at the interface.As the duration of the negative V GS increases, so does the extent of proton depletion.Therefore, the EPSC peak increases with both the amplitude and width of the negative V GS pulse.
The neural response would be shortly potentiated or depressed due to the synaptic plasticity, [43] which is essential for neural computation.Paired-pulse facilitation (PPF) is a typical form of synaptic plasticity, whereby the synaptic weight between pre-and postneuron can be transiently strengthened following two consecutive stimuli. [44]In the OECT, by applying two consecutive V GS pulses, the EPSC triggered by the second pulse is greater than that triggered by the first, mimicking the PPF behavior, as illustrated in the inset of Figure 3a.The PPF index (A 2 /A 1 , where A 1 and A 2 represent the EPSC peak amplitudes triggered by the first and second V GS pulses, respectively) decreases with the pulse interval, as shown for the black dots in Figure 3a.It can be explained as follows.The proton depletion at the electrolyte/channel interface caused by the first V GS pulse does not have enough time to fully recover when the interval between two pulses is very short.This results in a higher peak of EPSC for the second pulse.It can be inferred that this effect weakens as the pulse interval increases.The PPF index decay can be described by PPF index = 1 When a neuron receives a sequence of continuous stimuli, the way it responds depends on several factors including stimulus intensity, duration, and frequency. [45]In terms of biological significance, these effects allow neurons to encode different types of sensory information from the environment and convert them into spiking patterns that suit neural circuits, providing the basis for sensory perception and cognition. [43]To emulate such effects, Figure 3b-d depicts the change in the peak amplitude of EPSC induced by negative V GS pulses, with varied pulse amplitudes, pulse widths, and pulse intervals, respectively.All the results show a trend of potentiation, where the EPSC peak amplitude exhibits a gradual increase with increased pulse numbers.Furthermore, the EPSC peak amplitude increases with the amplitude of the negative V GS pulse, because of the enhanced proton depletion at the electrolyte/channel interface under a stronger channel-to-gate electric field.When increasing pulse width or decreasing pulse interval, EPSC peak amplitude increases due to the larger ratio between the proton depletion during the pulse and recovery during the interval.Figure 3e,f demonstrates depression and potentiation of the EPSC peak amplitude, realized by applying negative and positive V DS pulses, respectively.Such depression/potentiation can be attributed to the gradual accumulation/depletion of interfacial protons under a positive/negative gate-to-channel electric field, also described in Figure S2 (Supporting Information), which is generated by negative/positive V DS pulses.When decreasing the pulse interval, the depression/potentiation effect can also be enhanced because of the increased ratio between the proton accumulation/depletion during the pulse and recovery during the interval.The emulated potentiation and depression represent important mechanisms in neural systems, which also hold significance for the following design of the LIF neuron.
Among various neuron models, the LIF neuron is often used as the basic building block in SNNs due to its biologically plausible properties.In SNN, LIF neurons integrate temporal inputs to increase the membrane potential, which leaks away over time spontaneously.Once the membrane potential exceeds a certain threshold, an output spike is fired, which quickly resets the membrane potential to a lower value called the reset potential.Due to the above-demonstrated features of the OECT, a neuromorphic circuit is designed to realize a LIF neuron by simulation, as shown in Figure 4a.The OECT is behaviorally modeled based on the ionic accumulation/relaxation model and compact model of thin-film transistors, [46][47][48][49] which is detailed in Figure S3 (Supporting Information).The device simulation of short-term plasticity based on this model can be found in Figure S4 (Supporting Information).The LIF neuron circuit in Figure 4a differs from the traditional design of spiking neurons, which typically use capacitor charging/discharging to emulate integration/leaking of membrane potential and use transistors as reset paths. [50,51]Our circuit includes only an OECT, an operational amplifier, and resistors.The protic dynamics within the device are used to emulate the evolution of the membrane potential in LIF neurons.The OECT can also be used as a reset path.The operational details are as follows.1) Integration: corresponding to the "proton accumulation" of the OECT in Figure S2 (Supporting Information), when stimulating the input terminal, i.e., the drain electrode with negative voltage pulses, the produced gate-to-channel electric field will gradually accumulate protons at the electrolyte/channel interface and thus reducing the OECT's channel conductance.Consequently, the potential V MEM extracted from the source electrode increases with the pulse inputs and emulates a typical temporal integration of the membrane potential in LIF neurons, as shown for black lines in the middle panels of Figure 4b,c.
2) Leaking: corresponding to the "recovery from accumulation" of the OECT in Figure S2 (Supporting Information), during the interval of pulse inputs, the "integrated" V MEM "leaks" due to the spontaneous recovery of the proton accumulation state.
3) Reset: corresponding to the "proton depletion" of the OECT in Figure S2 (Supporting Information), when the "integrated" V MEM reaches the threshold V ref which equals to (V OUT +V 2 )/2, it will generate at the output of the operational amplifier a spike (V OUT = −5 V, as shown for black lines in the right panel of Figure 4b,c), which is immediately fed back to the gate of the OECT.The produced channel-to-gate electric field will deplete the accumulated interfacial protons, thus increasing the OECT's channel conductance and "resetting" the V MEM .
Overall, the LIF neuron's temporal integration, leaking, and reset are respectively represented by the accumulation, spontaneous recovery, and depletion of interfacial protons within our OECT, which can be electrically demonstrated by the evolution of the V MEM .It should be noted that the operational amplifier is only used for generating binary outputs, and the feedback in the reset process.Like the LIF neuron model, higher-frequency inputs can trigger output spikes at higher firing rates, as compared to Figure 4b,c.It is because when the interval is long, there is enough time for the recovery (leaking) of interfacial protons before receiving another input.Therefore, it takes more inputs for the V mem to reach the firing threshold and the firing at V OUT .The conditions for varied pulse amplitudes/widths can be found in Figure S5 (Supporting Information).Figure 4b,c also shows the fitting results, where the red lines represent the results of the LIF neuron model.The good agreement further confirms that our LIF neuron circuit is reasonable.The fitting is subsequently used to model the LIF neuron and construct the SNN in the following simulations.To emulate the spatial integration, the LIF neuron circuit with multiple weighted inputs is designed and simulated in Figure S6 (Supporting Information).
Coincidence detection refers to the ability of neurons to detect and amplify synchronous input signals, which otherwise may go unnoticed.By detecting coincident input patterns, neurons can exploit the temporal correlation of multiple inputs arriving within close proximity in time, allowing more efficient detection of meaningful patterns and correlations while reducing irrelevant noise interference.In our simulation, the SNN contains 1500 excitatory input nodes (connected with positive weight) and 500 inhibitory input nodes (connected with negative weight), feeding into one LIF neuron as the output node.When all input nodes receive asynchronous spikes (uniformly distributed in time), treated as noise in the left panel of Figure 5a, only one spike would be generated by the output neuron as shown in Figure 5b, due to the random distribution of inputs.However, when 10 randomly selected excitatory input nodes start receiving synchronous spikes with a mean frequency of 4 Hz, schematically shown in the right panel of Figure 5a, coincidence detection occurs and triggers the output neuron to fire up to 33 spikes as shown in Figure 5c, despite the persistent background noise (asynchronous spikes).This happens because by spatiotemporally integrating synchronous spikes from multiple input nodes over a short period, it becomes easier for the output LIF neuron to reach the firing threshold.Therefore, this implemented SNN using binary spiking events for temporal-coding of information, has the ability to detect relevant patterns from noisy data streams, making it particularly useful for applications such as audio and video processing.
Although ANNs are commonly used for image classification tasks, [52] the focus is mainly on identifying distinct visual features or patterns in the static images.In contrast, dynamic tasks, like handwriting recognition, require the ability to recognize handwritten strokes that occur over time, which can be more complex and variable than static images.SNNs are well-suited for such time-varying tasks due to their ability to effectively process spatiotemporal information, which can be encoded by binary spike events that are biologically plausible and energy efficient.Here we further develop the method that encodes the projection of pen strokes on a 2D surface into binary spike trains.Specifically, as shown in Figure 6a, for a Chinese character " " written in a 5 × 5 grid, each row or column corresponds to an input node, resulting in 10 input nodes for the SNN.For each timestep, we suppose that the handwritten stroke moves to a new cell of the grid and stimulates the input nodes corresponding to the current cell's row and column simultaneously.For example, at the fifth time step, the stroke moves to the cell corresponding to the second row and fifth column, as shown for the location of the black solid pen in Figure 6a.At this timestep, the input nodes corresponding to the second row (red) and fifth column (purple) receive a spike.In this way, the pen stroke trace of each 5 × 5 handwritten Chinese characters can be encoded into 10 spike trains (detailed in Figure S7, Supporting Information) sent to the 10 input nodes of the SNN, where the modeled LIF neurons act as the output nodes.The precise-spike-driven (PSD) supervised learning rule is employed to adjust the weights within the SNN: [53,54] the output spikes of each output node are compared to the teacher spikes (desired output spikes), and the weights between input and output nodes are adjusted to minimize their difference.Specifically, if an input spike is shortly before the output and teacher spike, the positive/negative difference between the two spikes will cause the potentiation/depression of weight corresponding input and output nodes (detailed in Figure S8, Supporting Information), and thus advancing/postponing the firing of output spike.Over time, the firing time of the output spike will gradually approach that of the teacher spike, allowing the network to learn to fire the spikes at the desired time.For an example of training one Chinese character " ", we apply a single teacher spike exclusively to the corresponding labeled output neuron.Following 20 epochs of training, the input spike trains representing this character can trigger the firing of the labeled output neuron at the timing of the teacher spike, as depicted in Figure 6b.The evolution of weight in the SNN and the detailed record of the output spike toward the teacher spike can be found in Figure S9 (Supporting Information).
To demonstrate the recognition of the noisy dataset, we introduce 10% noise to the original input spike trains representing handwritten Chinese characters " " and " ", by randomly removing each individual spike with a probability of 10%, which can also simulate sensor failure in practical use, schematically shown in Figure 6c.This generates a noisy dataset of 1000 samples that is used for testing.During training, multiple teacher spikes, uniformly distributed in the time domain, are applied to the corresponding labeled output neuron.During recognition, the predicted class is determined by identifying the output neuron with the highest frequency of firing spikes.After 70 epochs of training by original spike trains, the SNN can be well adjusted such that it can achieve up to 98.5% accuracy when classifying the noisy dataset, as shown in Figure 6d, demonstrating both good tolerances under real-world conditions and high computational efficiency.More recognition tasks can be found in Figure S10 (Supporting Information).The significant advantage of employing temporal coding in such hardware SNN is that it can reduce the complexity of the input data, by reducing multiple frames of 5 × 5 images into only 10 spike trains, which allows easier convergence and lower power consumption for the network.

Conclusion
In summary, this work proposes hardware SNN based on OECTs, capable of processing temporal-coded information.The OECT's working mechanism, which involves proton accumulation, depletion, and spontaneous recovery within the electrolyte dielectric, enables the device to respectively emulate the temporal integration, reset, and leaking characteristics in a LIF neuron circuit.SNN based on the proposed LIF neurons is developed in simulation, which successfully performs coincidence detection by effectively detecting synchronous inputs amidst asynchronous background noise.In addition, the SNN combined with a PSDsupervised learning algorithm can be utilized for dynamic handwriting recognition, where the pen stroke traces of 5 × 5 handwritten Chinese characters are encoded into spatiotemporally distributed binary spikes from their projections in two directions.This temporal-coding strategy significantly reduces network complexity and training convergence difficulties and enables the SNN to achieve a high recognition accuracy even in noisy datasets.

Experimental Section
Device Fabrication: The DPPT-TT-based OECTs were fabricated on p++ Si substrates, which were also used as the gate electrodes.First, chitosan solution (2 wt% in Acetic acid) was drop-casted on the substrate and baked at 90 °C for 0.5 h as the dielectric.Then, DPPT-TT solution (5 mg mL −1 in 1,2-Dichlorobenzene) was spin-coated on the chitosan film at a speed of 1500 rpm for 1 min and baked in a nitrogenfilled glove box at 150 °C for 1 h as the semiconducting layer.Finally, 50 nm thick Cu top electrodes were thermally evaporated through a metal shadow mask as the source/drain.The width of the channels is 1200 μm, while the length of the channels is 100 μm.The fabrication process is schematically shown in Figure S1 (Supporting Information).
Electrical Measurement: The electrical characteristics of the OECTs were measured using a Keysight B1500A with a Cindbest probe station.All measurements were performed under ambient air condition at room temperature.Voltage sweep and voltage list sweep were used to perform DC and transient measurements, respectively.
Simulation of the LIF Neuron and the SNNs: The LIF neuron circuit was designed and simulated in HSPICE circuit simulator.The behavioral modeling of the DPPT-TT-based OECT was built by Verilog-A language, which is detailed in Figure S3 (Supporting Information).The mathematical modeling of the LIF neuron and the simulation of the SNNs were carried out in Python 3.9 with Brian2. [55]

Figure 1 .
Figure 1.Schematic of spatiotemporal integration in neural networks.a) Schematic image of integrated stimuli from pre-neurons and fired action potential in post-neuron, for biological neural system.Schematic structure of a neuro-inspired SNN model processing information b) through ratecoding and c) through temporal-coding.

Figure 2 .
Figure 2. DPPT-TT-based OECT and its measured DC/transient characteristics.a) Schematic structure and circuit diagram of the device.b) Transfer curves under V DS of −0.5 and −5 V, and c) output curves with varied V GS (−2.0/−3.0/−3.5/−4.0V).EPSC curve (I DS response) under V DS of −5.0 V in response to V GS pulses d) with fixed pulse width (1 s) and varied amplitudes (from −1.0 to −5.0 V), and e) with fixed amplitude (−5.0 V) and varied pulse width (from 1.0 to 7.0 s).

Figure 3 .
Figure 3. Short-term plasticity of the OECT by applying V GS or V DS pulses.a) PPF index (A 2 /A 1 ) as a function of the pulse interval (Δt) between applied paired V GS pulses (−5.0 V, 500 ms) under V DS of −5.0 V, where A 1 and A 2 are the EPSC peak amplitudes induced by the first and second V GS pulses, respectively, as illustrated in the inset.EPSC peak amplitudes as a function of the number of V GS pulses under V DS of −5.0 V b) with fixed pulse width/interval (150 ms, 150 ms) and varied pulse amplitudes (from −3.5 to −5.0 V), c) with fixed pulse amplitude/interval (−5.0 V, 150 ms) and varied pulse width (from 100 to 300 ms), and d) with fixed pulse amplitude/width (−5.0 V, 150 ms) and varied pulse interval (40/60/90/150/250/350 ms).EPSC peak amplitudes as a function of the number of V DS pulses e) with fixed negative pulse amplitude/width (−5.0 V, 50 ms) and varied pulse interval (50/300/500 ms) under V GS of −4.0 V and f) with fixed positive pulse amplitude/width (5.0 V, 50 ms) and varied pulse interval (50/200/500 ms) under V GS of −3.0 V.

Figure 4 .
Figure 4. Circuit realization of a LIF neuron based on the OECT.a) Circuit diagram of the proposed LIF neuron, consisting of an OECT, an operational amplifier, and resistors.Cycles of the integration/leaking and reset of emulated membrane potential V MEM (middle panel), and the fired output spike (right panel), obtained from both circuit (black lines) and LIF neuron model (red lines) in response to input spikes (left panel) with period T b) of 200 ms, and c) of 250 ms.

Figure 5 .
Figure 5. Coincidence detection simulated in a 2000-in-1-out SNN based on the modeled LIF neuron, including 1500 excitatory (positively weighted) and 500 inhibitory (negatively weighted) input nodes.a) Schematical illustration of 2000 input trains of spikes uniformly distributed in time (at a mean frequency of 0.1 Hz) without (left panel) and with (right panel) the application of synchronous spikes (at a mean frequency of 4 Hz) onto 10 randomly selected excitatory input nodes.Simulated evolution of membrane potential (blue lines) and output spikes (black lines) b) without and c) with the introduction of synchronous spikes.

Figure 6 .
Figure 6.Dynamic handwriting recognition simulated in an SNN based on the modeled LIF neuron.a) Schematic of a Chinese character written in a 5 × 5 grid, where each row or column corresponds to an input node of the SNN.b) The teacher and output spikes during 20 epochs of training for one Chinese character.c) Example of original samples and noisy test samples for two Chinese characters.d) Recognition accuracy as a function of training epochs for 1000 noisy test samples.