Single organic electrochemical neuron capable of anticoincidence detection

Abstract

The present invention relates to a dendritic organic electrochemical neuron (d-OECN) for classifying linearly inseparable data, said dendritic organic electrochemical neuron comprising: an apical dendrite comprising a dendritic nonlinear device comprising a semiconducting channel material and having ion-mediated antiambipolar behaviour, and a soma component comprising a spiking neuron model.

Description

[0001] SINGLE ORGANIC ELECTROCHEMICAL NEURON CAPABLE OF ANTICOINCIDENCE DETECTION

[0002] TECHNICAL FIELD

[0003] The present invention relates to a dendritic organic electrochemical neuron (d-OECN) for classifying linearly inseparable data and to a method for operating such a d-OECN.

[0004] BACKGROUND OF THE INVENTION

[0005] Leveraging neuromorphic sensing and processing, inspired by the intricate workings of biological systems, holds great potential for the development of cutting- edge bioelectronic and robotic systems. Central to these neuromorphic systems are artificial neurons. For these artificial neural circuits to be effective and biointegrable, they require a design that is both simple and operates on ion-based mechanisms, mirroring biological processes. Unfortunately, artificial neurons made from silicon and other inorganic materials fall short in these areas. In contrast, the recently developed organic electrochemical neurons (OECNs) mark a significant step forward in emulating key neural features, including neurotransmitter and ion-based modulation, which traditional silicon-based neurons cannot achieve. Despite these advancements, OECNs as well as other traditional neuron technologies have yet to fully mirror the complete range of neural functionalities. A primary unresolved issue is their inability to classify linearly inseparable data, such as the exclusive-OR (XOR) function. Currently, no artificial neuron technology can execute this function within a single neuron using a limited array of electronic elements.

[0006] XOR (also known as anticoincidence detection), which outputs true only when its two binary inputs differ, is representative of real-world linearly inseparable problems and is often used as a benchmark for nonlinearity in artificial neural networks. Performing anticoincidence detection requires nonlinear activation functions, which are difficult to achieve with the linear summation and thresholding typically seen in conventional single-compartment artificial neuron models. This limitation has traditionally necessitated multi-layered networks with hidden layers even for basic nonlinear tasks increasing complexity and posing challenges for efficient neuromorphic design.

[0007] This stands in contrast to biology, where a recent groundbreaking discovery in neuroscience shows that single human cortical neurons can solve this challenge. In this discovery, layer 2 and 3 human cortical pyramidal neurons were found to perform anticoincidence detection within their dendritic compartments using dendritic calcium spikes. The key to this operation lies in voltage-gated calcium channels in their dendritic compartments. These channels mediate local dendritic calcium action potentials (dCaAPs) with a non-monotonic amplitude dependence where the amplitude peaks at a specific preferred input strength and declines for weaker or stronger inputs. This forms a Gaussian-tuned nonlinear activation, unlike more linear relationships between input and output in conventional neurons. As a result, when two separate inputs of the preferred high amplitude (1 ,1) are applied simultaneously to different branches of the dendrite, the resulting sum of the inputs is too high, resulting in a diminished calcium spike amplitude (0). In contrast, when a preferred high input is applied to only one branch with no or low input to the other branch (corresponding to (1 ,0) or (0,1)), the sum is a preferred value which coincides with the peak of the Gaussian (dCaAP rheobase) resulting in a high final calcium spike amplitude (1). This implements a version of anticoincidence detection - the essence of XOR logic.

[0008] Prior to and since these discoveries, various studies, including experimental and mathematical models, have sought to emulate similar mechanisms for nonlinear computation in single neurons and neural networks. Neural network models like the Radial Basis Function (RBF), Probabilistic Neural Networks, and Support Vector Machines (SVM) also utilize such nonlinear activation functions, though in a different manner than biological neurons, to improve learning capabilities. There have also been recent efforts to emulate the nonlinear dendritic behaviour using sophisticated antiambipolar transistors made of 2D chalcogenide heterojunctions to enhance the learning capabilities of spiking neural networks (SNNs). However, hardware implementation of all these approaches, which are primarily based on inorganic materials, often requires intricate fabrication strategies to replicate the Gaussian activation profile and is not fully integrated with artificial spiking neurons. Moreover, their lack of responsiveness to ions and lack of biocompatibility limit their biointegration capabilities.

[0009] Emulating complex neural computations like solving linearly inseparable data within single artificial neurons has remained an elusive goal in neuromorphic engineering. In view of the above, there is a need to provide an OECN mimicking biological neurons and networks and being able to perform complex neural computations.

[0010] SUMMARY OF THE INVENTION

[0011] In view of the above, the present invention aims to solve the problems of the prior art. To this end, the present invention relates to a dendritic organic electrochemical neuron (d-OECN) for classifying linearly inseparable data. The linearly inseparable data may be anticoincidence detection (XOR), as will be explained in greater detail below. The dendritic organic electrochemical neuron according to the present invention comprises an apical dendrite comprising a dendritic nonlinear device comprising a semiconducting channel material and having ion-mediated antiambipolar behaviour. The term “non-linear” in the context of the present invention means a dendritic device having a nonlinear response to input.

[0012] Recently, different types of switching behaviour were reported in p-n heterointerfaces, in which heterogeneous p- and n-type materials form a junction only in the channel region. In these heterointerface devices, l-V characteristics do not behave monotonically. In particular, the current increases and decreases with increasing applied voltage bias. This behaviour refers to antiambipolar characteristics, as antiambipolar charge transport characteristics exhibit behaviours opposite to the V- shaped ambipolar characteristics. Antiambipolar devices provide negative differential transconductance behaviour, showing a peak current (lpeak) in the mid-range of the applied gate voltage bias ( G).

[0013] By the term “antiambipolar” in the context of the present invention is thus understood a behaviour wherein a steep increase and decrease in drain current is exhibited within a certain range of gate bias voltage. In otherwords, the current-voltage function of a device having antiambipolar behaviour is a Gaussian-shaped function.

[0014] The dendritic nonlinear device referred to above may comprise either a dendritic organic electrochemical transistor (d-OECT) or an organic electrochemical transistor (OECT)-based inverter.

[0015] In particular, the d-OECT may comprise at least two electrodes; the semiconducting channel material being arranged in contact with the electrodes; an electrolyte layer arranged in contact with the semi-conducting channel material; and a gate. The d-OECT may further comprise a carrier layer and an insulating layer. The carrier layer may be a substrate arranged for accommodating the components of the d-OECT.

[0016] The components of the d-OECT may be arranged horizontally or vertically. In a horizontal arrangement, the electrodes of the d-OECT may be arranged on a substrate, the semiconducting channel material is arranged on top of the electrodes, the electrolyte is arranged on the semiconducting channel material, and the gate is arranged on top of the electrolyte. In a vertical arrangement, the components are arranged side by side in contact with each other. In such a vertical configuration, each of the components may be arranged on the substrate. In particular, the substrate may be a glass wafer.

[0017] The electrolyte may be NaCI, and may have an ionic concentration in the range from 0.05 mM to 5000 mM.

[0018] The gate of the d-OECT may comprise or consist of at least one electrode. The gate electrode may be non-polarizable. Examples of non-polarizable gate electrodes are Ag / AgCl-coated gold electrodes or pseudo-reference electrodes (e.g., organic semiconducting material-coated electrode). When the d-OECT comprises a single gate electrode, such an electrode may comprise gold coated with Ag / AgCl. In some configurations, the d-OECT may comprise two gate electrodes connected by an organic semiconducting material, with the gate electrodes typically made of gold. The organic semiconducting material on the gate can be the same as or different from the semiconducting channel material. To ensure efficient gating, the gate electrode may generally have a capacitance greater than the capacitance of the semiconducting channel material.

[0019] The semiconducting channel material may be one single compound having the antiambipolar or Gaussian type behaviour according to the above. In particular, the semiconducting channel material may be a polymer, or a combination of two polymers, e.g. one n-type conducting polymer and one p-type conducting polymer.

[0020] When the dendritic nonlinear device comprises an organic electrochemical transistor (OECT)-based inverter, such OECT-based inverter may comprise a p-type organic electrochemical transistor with a load, where the p-type OECT functions as the driving transistor and the load can be another OECT or resistive component. Alternatively, the OECT-based inverter may comprise an n-type organic electrochemical transistor with a load, where the n-type OECT functions as the driving transistor and the load can be another OECT or resistive component. Further, the OECT-based inverter may comprise a pair of p-type and n-type organic electrochemical transistors. The p-type and n-type organic electrochemical transistors comprise a semiconducting channel material transporting holes and electrons, respectively.

[0021] In particular, the semiconducting channel material may be a rigid conjugated polymer, e.g. a ladder-type polymer. The rigid conjugated polymer may have a dihedral angle from 0° to 20°. The rigid conjugated ladder-type polymer may be poly(benzimidazobenzophenanthroline) (BBL).

[0022] The term “rigid” in the context of the present invention means a conjugated polymer having a dihedral angle from 0° to 20°, preferably below 10°. The term “dihedral angle” in the context of the present invention is the angle between repeating units of the conjugated polymer. As mentioned above, the rigidity of the conjugated polymers of the present invention is a prerequisite for excellent charge transport ability combined with high stability, since torsional defects partially break the conjugation along the polymer backbone, resulting in decreased electronic delocalization, widened band gaps, increased numbers of trapped charges, and less effective intermolecular coupling.

[0023] The rigid conjugated polymer of the present invention may be an n-type rigid conjugated polymer.

[0024] The rigid conjugated polymers in the context of the present invention may have lowest unoccupied molecular orbital (LIIMO) energy level ELUMO below -3.9 eV. It should be understood that the term “below” in relation to a negative value is a negative value having a greater absolute value. In other words, the term “below” in the context of the present invention implies values being positioned to the left from -3.9 on the number line, e.g. -4.2, -5.8 and so forth.

[0025] Rigid conjugated polymers, in which all the backbone units on the polymer main-chain are TT-conjugated and fused, have been recognized as being particularly suitable in optical and electronic applications. These polymers have attracted great interest owing to their intriguing properties, remarkable chemical and thermal stability, and potential suitability as functional organic materials. In addition, they are distinct from conventional conjugated polymers in that the fused-ring constitution restricts the free torsional motion between the aromatic units along the backbone. Because of the diminished torsional defects, rigid conjugated polymers with fully coplanar backbones provide coherent TT-conjugation, fast intra-chain charge transport, long exciton diffusion length, and strong TT-TT stacking interactions.

[0026] Since rigid conjugated polymers possess planar backbones with optimum TT- electron delocalization and are free of torsional defects, they may be considered to be analogous to graphene nanoribbons, which combine the excellent charge transport property of graphene with opened band gaps as high-performance semi-conductive materials. Furthermore, rigid conjugated polymers display potentially high thermal and optical stability as well as high resistance to chemical degradation. Such combination of unique properties of rigid conjugated polymers makes them promising candidates for a wide range of applications.

[0027] Particularly suitable type of rigid conjugated polymers according to the present invention is conjugated ladder or ladder-type polymers. In general, ladder polymers are multiple stranded polymers with periodic linkages connecting the strands, resembling the rails and rungs of a ladder, and giving an uninterrupted sequence of adjacent rings that share two or more atoms. Conjugated ladder polymers are a specific subtype of ladder polymers in which all the fused rings in the backbone are TT-conjugated. In addition, they are distinct from conventional conjugated polymers in that the fused-ring constitution restricts the free torsional motion in between the aromatic units along the backbone.

[0028] Stemming from the fused backbone, conjugated ladder polymers exhibit extraordinary thermal, chemical, and mechanical stability. Because of the diminished torsional defects, conjugated ladder polymers with fully coplanar backbones provide coherent TT-conjugation, fast intra-chain charge transport, long exciton diffusion length, and strong TT-TT stacking interactions.

[0029] As mentioned above, the rigid conjugated polymer of the present invention may be a conjugated ladder-type polymer. In particular, the rigid conjugated polymer of the present invention may be BBL, comprising from 2 to 10000, preferably 2 to 200, more preferably 50 to 180 repetitive units (n).

[0030] Poly(benzimidazobenzophenanthroline), BBL

[0031] The recently reported antiambipolar behaviour in the n-type ladder polymer poly(benzimidazobenzophenanthroline) (BBL) is highly relevant for the present invention. Previously, this behaviour was used to simulate the ion channels of biological neurons (Na+and K+), resulting in the creation of bio-realistic conductancebased OECNs according to the Hodgkin-Huxley neural model.

[0032] According to the present invention, this phenomenon is employed to mimic dCaAP-like functions generated by voltage-gated calcium channels found in dendrites. To this end, the d-OECN according to the present invention comprises a soma component comprising a spiking neuron model. Such a spiking neuron model may exhibit either a class I or class II spiking behaviour. The functions of the soma component can be adjusted both ionically and electrically. By integrating with a spiking neuron model, a new dendritic organic electrochemical neuron (d-OECN) is developed, which is capable of classifying linearly inseparable data, such as the XOR problem. This classification problem necessitates multiple decision boundaries and is beyond the capability of conventional single-layer perceptrons. Furthermore, the d-OECN enables modifying the decision boundaries for such complex inputs through straightforward modulation of the neuron's threshold voltages and ionic concentrations, an achievement not possible with single artificial neurons made from any other technology.

[0033] The soma component of the d-OECN of the present invention may comprise a first OECT representing the sodium ion channel (Na-OECT) and a second OECT representing the potassium ion channel (K-OECT), based on the Hodgkin-Huxley (HH) neuron model. The Na-OECT may be connected to a first voltage source EN3, and the K-OECT may be connected to a second voltage source EK.

[0034] In an alternative configuration, the dendritic compartment of the d-OECN can be integrated with a leaky integrate-and-fire (LIF)-type OECN functioning as the soma. This integration demonstrates the flexibility and adaptability of the approach across various neuron models, which also supports its potential for scalability. The present invention further relates to a method for operating a dendritic organic electrochemical neuron (d-OECN) as described above. Such a method comprises the steps of: a) providing a range of (x, y) input current pairs to the dendritic nonlinear device; wherein the range of (x, y) input current pairs comprises at least one target (x, y) combination for XOR logic; b) converting the (x, y) input current pairs to an input voltage Vd, thus providing an output drain current Id; wherein the output drain current Id is above a predefined threshold when at least one target (x, y) combination is received; c) providing the output drain current Id to the soma component; d) generating a spike by the soma component when the output drain current Id is above the predefined threshold.

[0035] Step b) may be performed by processing the input current pairs through a non-monotonic activation function to produce a unique output current indicative of XOR logic. According to the method of the present invention, the threshold and response characteristics of the soma component may be adjusted based on the output drain current from the dendritic non-linear device. Step d) generates a spiking output when the XOR condition is met by the (x, y) input current pair.

[0036] While XOR functionality for a single combination of inputs x and y have been shown, real-world scenarios necessitate the generalization of this functionality to classify a broader spectrum of inputs. Therefore, the tunability of the d-OECT is desirable. To this end, the method according to the present invention may further comprise at least one of the steps of: e) altering the ionic concentration of the electrolyte in the dendritic nonlinear device; f) injecting a control current, lc into the gate of the d-OECT ; g) varying the drain voltage of the d-OECT. The steps e)-g) may occur before step a). Alternatively, in case of practical applications, the output of the neuron may be monitored for a preferred (x, y) input current pair and then the soma may be tuned accordingly to generate a spike, e.g. in training of neural networks.

[0037] Lowering the electrolyte concentration shifts the peak of the Gaussian towards higher gate voltages following a log relationship, as will be described below. This shift happens because of the voltage drops occurring at the gate / electrolyte and polymer / electrolyte interfaces according to Nernstian relationships.

[0038] An alternative method is the electrical approach, using a dual-gate configuration of OECTs. In this setup, the single gate electrode is split into a first gate electrode and a second gate electrode, wherein the two gate electrodes are connected by a semi-conducting channel material to form dual gate configuration. The first gate electrode and the second gate electrode may be gold electrodes connected by the semi-conducting channel material. Injecting a control current, lc, into the second gate electrode results in an additional voltage equivalent to lc multiplied by the resistance of the semi-conducting channel material on top of the original gate voltage VG. This adjustment shifts the peak position depending on the value of lc injected, hence changing the mean of the Gaussian.

[0039] It is also possible to alter both the amplitude and standard deviation of the Gaussian by varying the drain voltage of the d-OECT. Increased drain voltages lead to a Gaussian distribution with a wider full width at half maximum, attributed to varying doping levels at the drain and source electrodes.

[0040] The present invention thus demonstrates an unprecedented dendritic organic electrochemical neuron (d-OECN) capable of achieving anticoincidence detection by classifying the exclusive-OR (XOR) problem - a quintessential linearly inseparable task - within an individual neuron. Inspired by human cortical neurons that perform XOR through dendritic calcium spikes, the d-OECN of the present invention leverages ion-tunable antiambipolarity in mixed ionic-electronic conducting polymers to mimic voltage-gated dendritic calcium dynamics. Integrating this dendritic component with a tunable spiking neuron soma, the d-OECN successfully solves XOR via its inherent nonlinear activation profile. Crucially, the decision boundaries of the classification are ionically / electrically tunable. This d-OECN, replicating sophisticated biological intelligence, constitutes a major neuromorphic advance paving the way for nextgeneration bioelectronics and robotics requiring complex neural computation.

[0041] BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Embodiments of the invention will now be described by way of example with reference to the accompanying drawings, of which:

[0043] Fig. 1a shows the structure of cortical layer 2 / 3 pyramidal neurons and their corresponding circuit representation;

[0044] Fig. 1 b depicts the d-OECT device according to the present invention;

[0045] Fig. 1c illustrates the graphical representation of XOR logic operation with multiple decision boundaries;

[0046] Fig 1d depicts the relationship between dendritic calcium action potential (dCaAP) amplitude and the input current to the dendrite;

[0047] Fig. 1e shows the anti-ambipolar behaviour demonstrated in poly(benzimidazobenzophenanthroline) (BBL);

[0048] Fig. 2a illustrates the artificial neuron circuit incorporating dendritic XOR functionality;

[0049] Fig. 2b the spiking pattern of the neuron with various input combinations;

[0050] Fig. 2c depicts tuning the peak position (mean) of the Gaussian function using ionic concentration;

[0051] Fig. 2d shows adjusting the peak position of the Gaussian using a dual-gate configuration;

[0052] Fig. 2e depicts modulating the standard deviation and amplitude of the Gaussian function using the drain voltage;

[0053] Fig. 3a and 3b illustrate the artificial neuron circuit with the capability for XOR classification featuring tunable decision boundaries;

[0054] Figs. 3c to 3e depict narrowing the width of the decision boundary by adjusting the potassium channel threshold (electrolyte concentration = 100 mM); Figs. 3f to 3h illustrate shifting the decision boundary by modifying the mean of the Gaussian through changes in ion concentration (EK = -0.71 V);

[0055] Fig. 4a shows Schematic of the Hodgkin-Huxley neuron model depicting sodium (Na+) and potassium (K+) ion channels;

[0056] Fig. 4b is a graphical representation of a biological action potential with the activity profiles of Na+and K+ ion channels;

[0057] Fig. 4c illustrates the comprehensive schematic of the c-OECN (soma component) with corresponding components;

[0058] Fig. 4d depicts a detailed circuit diagram of the NMOS-based inverting amplifier utilized within the soma circuit;

[0059] Fig. 5 depicts the structure of an OECT device used in the soma component of the d-OECN according to the present invention;

[0060] Figs. 6-8 depict different types of OECT-based inverters.

[0061] Fig. 9a shows circuit diagram illustrating the XOR-based edge detection setup with two resistive pressure sensors (RX and RY) as inputs to the d-OECN;

[0062] Fig. 9b illustrates resistance change of the pressure sensor as a function of different pressure levels;

[0063] Fig. 9c depicts experimental demonstration of edge detection: sequential photos show the contact of two sensors with an object;

[0064] Fig. 9d shows emulation of a slowly adapting Merkel-type mechanoreceptor: continuous spiking in response to sustained input pressure;

[0065] Fig. 9e depicts emulation of a rapidly adapting Meissner-type mechanoreceptor: spiking in response to dynamic pressure changes.

[0066] Fig. 10 illustrates electrical characteristics of the pressure sensor;

[0067] Fig. 11 illustrates simulated XOR spiking behaviour of the d-OECN model;

[0068] Fig. 12 depicts effective VGS and VDS in d-OECNs;

[0069] Fig. 13 shows d-OECN incorporating a leaky integrate-and-fire (LIF) soma. Fig. 14 depicts simulated decision boundaries;

[0070] Fig 15 shows simulated XOR classification.

[0071] DETAILED DESCRIPTION OF THE INVENTION

[0072] As mentioned above, the present invention provides a dendritic organic electrochemical neuron (d-OECN) for classifying linearly inseparable data. The linearly inseparable data may be anticoincidence detection (XOR), as will be explained in greater detail below. The dendritic organic electrochemical neuron according to the present invention comprises an apical dendrite comprising a dendritic nonlinear device having ion-mediated antiambipolar behaviour.

[0073] Fig. 1a illustrates the structure of cortical layer 2 / 3 pyramidal neurons and their corresponding circuit representation.

[0074] As mentioned above, the dendritic nonlinear device referred to above may comprise either a dendritic organic electrochemical transistor (d-OECT) or an organic electrochemical transistor (OECT)-based inverter.

[0075] As may be seen in Fig. 1 b, the d-OECT comprises two electrodes; a semiconducting channel material being arranged in contact with the electrodes; an electrolyte layer arranged in contact with the semi-conducting channel material; and a gate. The components of the d-OECT are arranged horizontally.

[0076] The gate of the d-OECT may comprise at least one electrode (single gate).

[0077] In the embodiment illustrated in Fig. 1 b and 1e, the semiconducting channel material is a rigid conjugated poly(benzimidazobenzophenanthroline) (BBL) having antiambipolar behaviour. As clearly seen in Fig. 1e, BBL demonstrates a peak current in the mid-range of the applied gate voltage bias.

[0078] The recently reported antiambipolar behaviour in the n-type ladder polymer poly(benzimidazobenzophenanthroline) (BBL) is highly relevant for the present invention. Previously, this behaviour was used to simulate the ion channels of biological neurons (Na+and K+), resulting in the creation of bio-realistic conductancebased OECNs according to the Hodgkin-Huxley neural model. According to the present invention, this phenomenon is employed to mimic dCaAP-like functions generated by voltage-gated calcium channels found in dendrites (Fig. 1d).

[0079] Fig. 4c shows a schematic of the soma component comprising a spiking neuron model.

[0080] The operation of the soma component in the d-OECN of the present invention utilizes a conductance-based model, incorporating sodium and potassium channels that mimic the Hodgkin-Huxley model (Fig. 4a). The circuit comprises two main transistors: a sodium ion-based transistor (Na-OECT) and an NMOS transistor TK, each connected to two distinct voltage sources, ENa(500 mV) and EK (ranging from - 0.71 to -0.75 V, Fig. 4c). It is noteworthy that NMOS was used to mimic the K channel, but an OECT can also be used for the K channel, so the NMOS can be replaced by the K-OECT. An exemplary structure of a suitable OECT is shown in Fig. 5.

[0081] These components are parallel to the ion channels and batteries in the original Hodgkin-Huxley model. Voltage ENais applied to the drain of Na-OECT and EK to the source of TK.

[0082] In this setup, an input current lin, varying between 2-15 pA, is fed into the circuit. This current is accumulated by the membrane capacitance Cmem, causing the membrane potential Vmem to rise from its baseline. Concurrently, this rise in potential drives the gate voltage of Na-OECT from approximately 1.2 V to 0.8 V via an n-type metal-oxide-semiconductor (NMOS) based inverting amplifier (Fig. 4d).

[0083] This adjustment enables the device to navigate through the peak of the antiambipolar transfer curve, initiating a spike in the current which further charges the capacitor, thereby sharply increasing Vmem (depolarization phase).

[0084] Following this, the NMOS transistor TK activates after a brief delay, controlled by resistor RK and capacitor CK, and achieves its maximum current subsequent to the peak current of Na-OECT. The capacitor discharges through TK, leading to a decrease in Vmem (repolarization phase) and returning it towards its original level. Due to the higher and longer-lasting current of the K-OECT, the voltage momentarily dips below the baseline (hyperpolarization phase). This sequence of events recurs cyclically with a constant input current, thereby sustaining the generation of action potentials. Adjustments to the threshold TK can be made by modifying EK from approximately -0.71 to -0.75 V, optimizing the operational range of the soma necessary for effective integration within the d-OECN system.

[0085] Both TK and TN are NMOS transistors (Infineon BSP295 model). Components RK and CK introduce the necessary delay for emulating the dynamics of the potassium channel.

[0086] In Figure 2a, the architecture of the d-OECN according to the present invention is shown. The d-OECN is delineated into two integral components: an apical dendrite comprising a dendritic nonlinear device having ion-mediated antiambipolar behaviour, and a soma component comprising a HH spiking neuron model. The apical dendrite executes nonlinear computations crucial for the XOR logic function, and the soma applies a thresholding mechanism, activating a spike only when stimuli exceed a certain threshold. A BBL-based d-OECT was employed to simulate the dendrite's function. The soma's threshold behaviour is mimicked by the HH-based c-OECN shown in Fig. 4c. This adjustment involves calibrating the potassium channel threshold to incorporate an appropriate leakage current that forces the neuron to only spike past a predefined threshold.

[0087] Additionally, the dendritic compartment of the d-OECN can be integrated with a LIF-type OECN functioning as the soma. Fig. 11 illustrates SPICE simulation of XOR spiking behaviour in d-OECNs with (A) an HH and (B) a LIF neuron. In the SPICE simulation, the d-OECN circuits consist of a d-OECT with a voltage-integrating resistor at the gate, paired with either the HH or LIF OECN.

[0088] Fig. 12 shows simulated VGS and VDS at different operational states of (A) an HH d-OECN and (B) a LIF d-OECN. The source terminal of the d-OECT is connected to the Vmem node. Therefore, the effective VGS and DS have different baselines under different input conditions and would fluctuate when the neuron spikes.

[0089] After this fine-tuning, the artificial neuron's functional range for input currents was established (typically 2-15 pA). The d-OECT is then chosen to deliver the current range required for the neuron to respond to preferred input combinations for XOR operations- these are (x, y) pairs that will result in spiking, here specifically (30 pA, 15 pA) or (15 pA, 30 pA), in the context of XOR logic (Fig. 2b). These (x, y) currents are converted to voltages and summed at the d-OECT gate via a resistor (RD = 18 kQ) (Figure 2a). This setup ensures that when a target (x, y) combination for XOR logic is received, the neuron fires. If, however, the inputs are outside the targeted XOR range, like (15 pA, 15 pA) or (30 pA, 30 pA), the input voltage of the d-OECT falls on the extremes of the Gaussian curve and the resulting output drain current is too low, thereby failing to provide ample input current to produce a spike (Figure 2b). Thus, the neuron effectively delineates the XOR logic's decision boundaries, utilizing the Gaussian function to distinguish between the linearly inseparable combinations of inputs exactly as what is happening in the layer 2 / 3 pyramidal neurons, as discussed previously.

[0090] As mentioned above, tunability of the d-OECT is of paramount importance. Beyond this specific XOR application, the ability to adjust the mean, standard deviation, and amplitude of the Gaussian function is also advantageous for incorporating d-OECTs into alternative neural network models like probabilistic neural networks.

[0091] One method to modify the mean of the Gaussian is to alter the ionic concentration of the electrolyte. Lowering the electrolyte concentration shifts the peak of the Gaussian towards higher gate voltages following a log relationship (Fig 2c). This shift happens because of the voltage drops occurring at the gate / electrolyte and polymer / electrolyte interfaces according to Nernstian relationships.

[0092] An alternative method is the electrical approach, using a dual-gate configuration of OECTs (Fig. 2d inset). In this setup, the single gate electrode is split into a first gate electrode and a second gate electrode, wherein the two gate electrodes are connected by a semi-conducting channel material to form dual gate configuration. The first gate electrode and the second gate electrode may be gold electrodes connected by the semi-conducting channel material. Injecting a control current, lc, into the second gate electrode results in an additional voltage equivalent to lc multiplied by the resistance of the semi-conducting channel material on top of the original gate voltage VG. This adjustment shifts the peak position depending on the value of lc injected, hence changing the mean of the Gaussian (Fig. 2d).

[0093] It is also possible to alter both the amplitude and standard deviation of the Gaussian by varying the drain voltage of the d-OECT. Increased drain voltages lead to a Gaussian distribution with a wider full width at half maximum, attributed to varying doping levels at the drain and source electrodes (Fig. 2e). Fig. 13 illustrates d-OECN incorporating a LIF-based soma. In Fig. 13A, the schematic circuitry layout of the d-OECN incorporating a LIF-OECN as the soma is shown. Circuit parameters are as follows: RD = 18 kilohms, ED = 0.5 V, Rieak = 100 kilohms, Cf = Cmem = 1 pF, VDD = 0.6 V. In Fig. 13B, XOR response of the d-OECN with LIF-OECN-soma is illustrated.

[0094] The ability of artificial neurons to process a range of (x, y) input pairs for XOR tasks, rather than just discrete pairings, is important, mirroring the complexities of inputs in real-life situations. Hence, the artificial neuron is exposed to a broad array of input currents, x and y, to gauge its response (Figure 3a). It is observed that the neuron typically spikes in response to a continuum of (x, y) inputs, not merely specific pairs. This behaviour occurs because the soma responds to a breadth of input currents, where the input current is a function of both x and y. The tails of the Gaussian function delineate the decision boundaries. Therefore, altering the ion concentration, and thus shifting the mean of the d-OECT, repositions the decision boundaries in tandem with the Gaussian shift (Figure 3f-h). It was also possible to fine-tune the width of the decision region by adjusting the potassium channel threshold (Figure 3c-e). Solid dots represent combinations that trigger neuron spiking, while hollow dots represent nonspiking behaviour. When EK, the voltage source associated with the K-channel, is shifted towards more negative values, more input current is allowed to leak through the K-channel, effectively raising the minimum input current required for the neuron to spike. Such tunability in decision boundaries provides artificial neurons the flexibility to adjust to novel patterns, akin to biological neurons that modify synaptic strengths for learning and memory. This flexibility is crucial during training, as the system refines its data classification, and is equally important in real-world applications where discerning and responding to a diverse array of stimuli is necessary.

[0095] Fig. 14 illustrates simulated decision boundaries of the d-OECN using an HH neuron model (A) or a LI F neuron model (B) as a function of the ionic concentration. Fig. 15 shows simulated XOR classification for a randomly generated dataset as a function of the ionic concentration for the HH d-OECN (A-B) and LIF d-OECN (C-D). In the simulation, the dataset X ~ N(n, 0) consists of two clusters, each with two data points. The cluster means are z1 = (-0.2, 0.2) and z2 = (0.2, -0.2), representing class 1 (red) and 0 (black), respectively. The X, Y input currents are mapped to this dataset.

[0096] For all the measurements, the electrolyte is NaCI in water. Materials such as Naphthalenetetracarboxylic dianhydride (NDA), 1 , 2,4,5- tetraaminobenzene tetra hydrochloride (TABH), poly(phosphoric acid) (PPA), methanesulfonic acid (MSA), chloroform, 1 ,2-dichlorobenzene, ethylene glycol, (3- glycidyloxypropyl)trimethoxysilane, and 4-dodecylbenzenesulfonic acid were sourced from Sigma-Aldrich. BBL, with a viscosity of 6.3 dL g-1in MSA at 30°C and a molecular weight of 35 kDa, was synthesized through the polycondensation of NDA and TABH in PPA under high-temperature conditions.

[0097] For thin film casting, BBL is initially dissolved in methanesulfonic acid (MSA) and heated at 100°C for 12 hours. After cooling to ambient temperature, the resulting BBL-MSA solution is applied to OECT substrates using a spin-coating process (1000 rpm for 60 seconds, with an acceleration of 1000 rpm s-1). Any remaining MSA in the film is then removed by soaking in deionized water, followed by drying under a nitrogen stream.

[0098] The SPICE models of the d-OECT and d-OECN were developed using B2 SPICE (EMAG Technologies). These models are designed to simulate the spiking behaviour of the d-OECN as well as the modulation of decision boundaries in XOR classifications.

[0099] The fabrication and testing of OECTs begins with thorough cleaning of 4-inch glass wafers using acetone, deionized water, and isopropyl alcohol (IPA), followed by drying with nitrogen. The electrodes, comprising 5 nm of chromium and 50 nm of gold, are deposited by thermal evaporation and patterned using photolithography. A 1 pm parylene C layer is applied over the electrodes in the presence of 3- (trimethoxysilyl)propyl methacrylate (A-174 Silane) to enhance adhesion and serve as an insulating barrier, reducing capacitive interference at the electrode-electrolyte boundary.

[0100] Subsequently, a 2% Micro-90 industrial surfactant is spin-coated as an antiadhesive layer. Over this, a sacrificial layer of parylene C (2 pm thick) is deposited, followed by a 5 pm thick layer of AZ10XT520CP positive photoresist. This setup protects the underlying layers during the plasma reactive ion etching (RIE) step (150 W, O2 = 500 seem, CF4 = 100 seem, for 380 seconds). Photolithography is again employed to delineate the contact pads and the OECT channel, after which the photoresist is developed with AZ developer. The plasma RIE step then removes the photoresist and sacrificial parylene, exposing the OECT channel and contact pads while preserving the other areas with parylene layers. The channels are patterned to specific Width / Length ratios: 40 pm / 6 pm for Na-OECTs. The BBL-MSA solution is then spin-coated to achieve a film thickness of 20 nm (Na-OECT). The sacrificial parylene layer is peeled off, leaving behind semiconductor film confined to the wells and connecting the source / drain electrodes of the OECT. Ag / AgCI paste is used to form a gate electrode (1 pm thick, 9 mm2).

[0101] All OECT measurements are conducted in a 0.1 M NaCI aqueous solution, unless specified otherwise. The devices are characterized using a Keithley 4200A- SCS.

[0102] When the dendritic nonlinear device comprises an organic electrochemical transistor (OECT)-based inverter, such OECT-based inverter may comprise a p-type organic electrochemical transistor with a load (Fig. 7), where the p-type OECT functions as the driving transistor and the load can be another OECT or resistive component. Alternatively, the OECT-based inverter may comprise an n-type organic electrochemical transistor with a load (Fig. 8), where the n-type OECT functions as the driving transistor and the load can be another OECT or resistive component. Further, the OECT-based inverter may comprise a pair of p-type and n-type organic electrochemical transistors (Fig. 6). The p-type and n-type organic electrochemical transistors comprise a semiconducting channel material transporting holes and electrons, respectively.

[0103] It has been shown that d-OECN according to the present invention could be used for tactile sensing enabling edge detection. In biological systems, tactile edge detection arises from neurons responding to contrasts in mechanical input across their receptive fields. Mechanoreceptors, such as Merkel cells and Meissner corpuscles, play pivotal roles in this process. Merkel cells, which are slowly adapting (SA), respond to sustained pressure, while Meissner corpuscles, which are rapidly adapting (RA), detect dynamic changes in pressure. These mechanoreceptors work together to identify tactile features such as edges, where contrasting inputs — high pressure on one side and low pressure on the other — are detected by neighbouring sensors. Tactile edge detection is further refined by lateral inhibition, a mechanism in which strongly activated neurons suppress the activity of neighbouring neurons with similar input levels. This process sharpens contrasts and enhances the perception of boundaries, functioning in a manner analogous to XOR logic, which detects mismatched inputs through anticoincidence detection.

[0104] Inspired by these biological principles, it has been demonstrated that the d- OECN according to the present invention was capable of edge detection. The experimental setup, illustrated in Fig. 9a, consists of two resistive pressure sensors (RX and RY, with pressure characteristics shown in Fig. 10) connected to the d-OECN. These sensors generate input currents for the neuron based on their resistance and the applied voltage (EX, EY). When a resistive sensor is pressed against an object, its resistance decreases (Fig. 9b), resulting in an increase in current passing through the sensor.

[0105] Fig. 10A illustrates l-V characteristics of the pressure sensor under different pressure inputs. In Fig. 10B, continuous recording of the pressure sensor’s resistance changes in response to 4 consecutive pressing inputs applied using a finger is shown.

[0106] The circuit enables the d-OECN to execute an XOR operation, producing spikes only when one sensor is active (high pressure on one sensor and low pressure on the other). The d-OECN remains silent when both sensors are either inactive or simultaneously active, effectively replicating the principle of anticoincidence detection observed in biological systems.

[0107] In Fig. 9c, the edge detection functionality of the d-OECN was validated. Sequential contact of the sensors, mounted on a glove, with an object is recorded alongside the corresponding neuron’s Vmem. The neuron exhibits spiking behaviour exclusively under XOR conditions (0,1) or (1 ,0), while it remains silent under non-XOR conditions (0,0) or (1 ,1). When the finger is positioned entirely outside the object’s surface, both sensors are inactive, providing an input of (0,0) to the neuron. At the edge of the object, the input is (0,1) or (1 ,0), as one sensor contacts the surface while the other remains outside, satisfying the XOR condition.

[0108] When both sensors are fully on the surface of the object, both inputs are active (1 ,1). This functionality allows the d-OECN to effectively detect the edge of the object.

[0109] In combination with XOR-based operations, the d-OECN can also emulate the behaviour of biological mechanoreceptors, as demonstrated in Fig. 9d-e. When configured to mimic SA mechanoreceptors, such as Merkel cells, the d-OECN produces continuous spikes in response to sustained input pressure (Fig. 9d). Conversely, when configured to mimic RA mechanoreceptors, such as Meissner corpuscles, it generates spikes only during dynamic pressure changes (Fig. 9e), while simultaneously executing XOR-based edge detection. This reconfiguration is achieved by simply tuning the K-channel voltage of the soma compartment. The mechanoreceptor functions are emulated by adjusting the potassium channel voltage (EK) in the HH neuron model. With EK at -0.7 V, a continuous, slow adapting (Class 1) spiking behaviour is observed, similar to slowly adapting mechanoreceptors. Lowering EK to -0.75 V produces a fast-adapting (Class 3 / phasic) response, where spiking occurs mainly at the onset of the input. These combined functionalities underscore the versatility of the d-OECN in real-time sensory processing.

[0110] Historically, the XOR function, emblematic of linearly inseparable problems, has remained elusive for single-neuron models due to its demand for nonlinear activation functions. The approach of the present invention, inspired by the discovery of dendritic calcium spikes in human cortical neurons, utilizes a d-OECT modulated by ionic and electrical means to emulate this biological phenomenon. By successfully integrating dendritic calcium channel behaviour within a conductance-based OECN, the inventors have not only replicated a key aspect of neuronal processing but also enabled the classification of XOR (anticoincidence detection) and similar linearly inseparable data within a single neuron setup. This advancement is pivotal, offering a simpler, more efficient pathway for neuromorphic design and biointegration, thus bringing artificial neurons a step closer to their biological counterparts.

[0111] One promising avenue of the present invention could be enhancing the functionality of silicon-based neuromorphic processors for edge applications. OECNs are particularly suited for event-based sensing applications due to their inherent ability to detect multiple sensory modalities, including ions, biomolecules, pressure, temperature, and light - capabilities not inherently present in traditional silicon-based neuronal circuits. Having the added ability to perform nonlinear classification could significantly enhance active processing at the edge with fewer neurons.

[0112] Furthermore, organic semiconductor-based artificial synapses have shown significant progress in scalability and reliability, offering promise for future integration with d-OECNs to create efficient neural microcircuits for edge processing. By leveraging the capabilities of d-OECNs, we can emulate biological sensory processing more efficiently, leading to simpler architectures for event-based sensors and improved feature extraction at the sensor level by augmenting the functionality of silicon based neuromorphic processors. In addition to such applications, the study also encourages further exploration into multi-compartmental neuron models using OECTs to emulate various neuron sections, including the potential for mimicking additional dendritic functionalities, such as spatial-temporal summation. These prospects underscore the wide range of opportunities for advancing our understanding and utilization of OECNs in mimicking biological neurons and networks.

[0113] While immediate applications of the XOR-capable d-OECNs may not be fully defined, it is common for the true potential of emerging technologies to unfold over time. One promising avenue could be enhancing the functionality of silicon-based neuromorphic processors for edge applications. OECNs are particularly suited for event-based sensing and processing applications due to their inherent ability to detect multiple sensory modalities, including ions, biomolecules, pressure, temperature, and light — capabilities not inherently present in traditional silicon-based neuronal circuits. Having the added ability to perform nonlinear classification could substantially enhance active processing at the sensor level with fewer neurons. This is demonstrated by our proof-of-concept implementation of edge detection in a tactile sensing system using a single neuron.

[0114] This capability holds strong relevance for applications such as artificial skin in robotics and prosthetics, where tactile edge detection allows systems to identify object boundaries, surface contours, and texture gradients — key for manipulation and feedback. Similarly, in neuromorphic vision systems, edge detection plays a central role in interpreting images. The anticoincidence detection behaviour exhibited by the d-OECN is analogous to the function of lateral inhibition circuits found in the retina (47), which are responsible for enhancing edge contrast and sharpening visual perception. Our demonstration that a single d-OECN can detect tactile edges and dynamically reconfigure to mimic distinct mechanoreceptor behaviours provides a proof-of-concept for embedding such bioinspired computation directly at the sensory interface. This approach is particularly advantageous in event-based edge computing, where real-time, low-latency processing at the sensor level reduces the data load on downstream processors and supports efficient operation. Although the present invention has been described with reference to various embodiments, those skilled in the art will recognize that changes may be made without departing from the scope of the invention. It is intended that the detailed description be regarded as illustrative and that the appended claims including all the equivalents are intended to define the scope of the invention.

Claims

CLAIMS1. A dendritic organic electrochemical neuron (d-OECN) for classifying linearly inseparable data, said dendritic organic electrochemical neuron comprising: an apical dendrite comprising a dendritic nonlinear device comprising a semiconducting channel material and having ion-mediated antiambipolar behaviour, and a soma component comprising a spiking neuron model.

2. The d-OECN according to claim 1 , wherein said dendritic nonlinear device comprises either a dendritic organic electrochemical transistor (d-OECT) or an organic electrochemical transistor (OECT)-based inverter.

3. The d-OECN according to claim 2, wherein said d-OECT comprises: at least two electrodes; said semiconducting channel material being arranged in contact with said electrodes; an electrolyte arranged in contact with said semiconducting channel material; a gate arranged in contact with said electrolyte.

4. The d-OECN according to claim 3, wherein the d-OECT comprises said electrolyte having an ionic concentration in the range from 0.05 mM to 5000 mM.

5. The d-OECN according to claim 3 or 4, wherein said gate of said d-OECT comprises a first gate electrode and a second gate electrode connected by said semi-conducting channel material.

6. The d-OECN according to claim 2, wherein said OECT-based inverter comprises: a pair comprising a p-type organic electrochemical transistor (p-OECT) and an n-type organic electrochemical transistor (n-OECT), or a p-OECT and a load, oran n-OECT and a load, wherein said p-OECT and said n-OECT comprise said semi-conducting channel material.

7. The d-OECN according to any one of the preceding claims, wherein said semiconducting channel material comprises a rigid conjugated polymer.

8. The d-OECN according to claim 7, wherein said rigid conjugated polymer is a ladder-type polymer.

9. The d-OECN according to claim 8, wherein said conjugated ladder-type polymer is poly(benzimidazobenzophenanthroline) (BBL).

10. The d-OECN according to claim 9, wherein said BBL comprises from 2 to 10000, more preferably from 2 to 100, most preferably from 30 to 50 repetitive units.

11. The d-OECN according to any one of the preceding claims, wherein said spiking neuron model is a Hodgkin-Huxley (HH) neuron model or a leaky integrate-and-fire (LIF) neuron model.

12. The d-OECN according to any one of the preceding claims, wherein said linearly inseparable data is anticoincidence detection (XOR).

13. The d-OECN according to any one of the preceding claims, wherein said soma component comprises a first OECT representing sodium ion channel (Na- OECT) and a second OECT representing potassium ion channel (K-OECT), wherein said Na-OECT is connected to a first voltage source EN3, and the K- OECT is connected to a second voltage source Ek.

14. A method for operating a dendritic organic electrochemical neuron (d-OECN) according to claims 1-13, said method comprising the steps of: a) providing a range of (x, y) input current pairs to the dendritic nonlinear device; wherein said range of (x, y) input current pairs comprises at least one target (x, y) combination for XOR logic; b) converting said (x, y) input current pairs to an input voltage Vd, thus providing an output drain current Id; wherein said output drain currentId is above a predefined threshold when at least one target (x, y) combination is received; c) providing said output drain current Id to said soma component; d) generating a spike by said soma component when said output drain current Id is above said predefined threshold.

15. The method according to claim 14, where said method further comprises at least one of the steps of: e) altering the ionic concentration of the electrolyte in the dendritic nonlinear device; f) injecting a control current, lc into the gate of the d-OECT ; g) varying the drain voltage of the d-OECT.

Images