Device and device arrangement for artificial neural network comprising same, and methods for forming same

Abstract

According to embodiments of the present invention, a device is provided. The device includes a substrate, a plurality of carbon nanotubes arranged in a two-dimensional layer with a thickness of between about 1 nm and about 40 nm on the substrate, the plurality of carbon nanotubes being arranged with gaps between the carbon nanotubes, and a conducting polymer arranged over the plurality of carbon nanotubes and in the gaps, wherein the device is configured for, in response to an electrical signal applied to the device, formation of conductive channels in the gaps and through the conducting polymer, and in response to removal of the electrical signal, dissolution of the conductive channels. A device arrangement for an artificial neural network is also provided, having the device, and an electrode arrangement electrically coupled to the device, the electrode arrangement being configured for electrical signal measurements.

Description

[0001] DEVICE AND DEVICE ARRANGEMENT FOR ARTIFICIAL NEURAL

[0002] NETWORK COMPRISING SAME, AND METHODS FOR FORMING SAME

[0003] FIELD OF THE DISCLOSURE

[0004] The present disclosure relates to a device and a method for forming the device, a device arrangement for an artificial neural network and a method for forming the device arrangement, a system for an artificial neural network, a method for determining a weight matrix for a device arrangement for an artificial neural network, and a method for determining a resultant output for a device arrangement for an artificial neural network.

[0005] BACKGROUND

[0006] Known artificial neural networks (ANNs) are inspired by the structure and function of biological neural networks, and consist of interconnected nodes that transmit and process information through weighted connections. The two simplest ANN structures are feedforward neural networks (FNN) and recurrent neural networks (RNN). In FNN, information flows through layers of neurons in one direction. Alternatively, in RNN, information can back-propagate to previous layers of neurons resulting in a historydependent response to the input information. Both of these frameworks require extensive training of all of the weights between neurons to be able to perform information processing tasks.

[0007] Reservoir computing is a type of recurrent neural network (RNN) that includes three (simple) layers: an input, a reservoir, and an output. The role of the reservoir is to transform an input signal into a high-dimensional non-linear space that can produce the desired output using a (simple) recurrence-free linear readout. Training costs are much lower than FNNs or RNNs because the reservoir’s dynamics remain fixed while (only) the output layer is trained.

[0008] There is increasing interest in exploring physical materials for use as reservoirs in reservoir computing applications as certain materials are able to non-linearly transform electrical input signals into high-dimensional separable outputs. One type of material system that has shown promise for this application is the memristor.

[0009] Memristors are two terminal devices whose resistance changes in time depending on the current which has flowed through it. One commonly used implementation of the memristor for reservoir computing applications is the crossbar array. In this architecture, the memristive elements are located between parallel rows of bottom electrodes and parallel rows of top electrodes which are positioned perpendicular to each other. This creates a grid of single memristive switches, each connected to every other switch. While this type of geometry is easy to interface to traditional CMOS hardware, it is difficult to scale this type of network, it requires a multistep lithography to fabricate, suffers from parasitic currents, and lacks the small- world connectivity observed in biological neural networks.

[0010] Networks of stochastically arranged, self- assembled memristive elements offer a scalable solution to the challenges observed in crossbar arrays. This type of system, namely atomic switch networks, utilise highly interconnected metallic nanowires or nanoparticles surrounded by an insulating material to create atomic scale memristive switches at junctions between two or more nanowires or nanoparticles.

[0011] Memristive switching occurs due to formation and dissolution of conductive filaments (metal bridges) through the insulating material. There are multiple mechanisms of filaments formation, including the migration of metallic ions, and the migration of oxygen vacancies. The collective behavior of these systems shows memory (resistive switching), critical dynamics (1 / f noise, avalanche criticality, power-law dynamics) similar to those in biological neural networks, non-linear signal transformation and separable outputs. This type of device has been realized using sulfurized silver nanowires, silver nanowires coated with PVP, silver nanowires coated with selenide, and tin nanoparticles. However, there are challenges with known atomic switch networks. Silver is known to oxidise and degrade over time, which is unfavorable for any long-term applications of this type of device. Additionally, it is difficult to encapsulate silver nanowire devices, leaving this system vulnerable to changes in the operating environment.

[0012] Further, known methods to fabricate hardware based artificial neural networks require complex fabrication processes and suffer from exposure to the environment. For example, memristive crossbar arrays require a complicated lithography process consisting of multilayer alignment. Silver nanowire networks have issues with long-term stability. Nanoparticle networks, where switches form in air gaps between nanoparticles, have unknown operation under encapsulation conditions and could be vulnerable to variations in operation to environmental conditions.

[0013] It is an object of at least preferred embodiments of the present invention to provide one or more devices and / or one or more methods that address at least some of the problems of known techniques, and / or to at least provide the public with a useful alternative.

[0014] SUMMARY

[0015] The invention is defined in the independent claims. Further embodiments of the invention are defined in the dependent claims.

[0016] In a first aspect of the invention, there is provided a device comprising: a substrate; a plurality of carbon nanotubes arranged in a two-dimensional layer with a thickness of between about 1 nm and about 40 nm on the substrate, the plurality of carbon nanotubes being arranged with gaps between the carbon nanotubes; and a conducting polymer arranged over the plurality of carbon nanotubes and in the gaps, wherein the device is configured for: in response to an electrical signal applied to the device, formation of conductive channels in the gaps and through the conducting polymer, and in response to removal of the electrical signal, dissolution of the conductive channels. In an embodiment, the plurality of carbon nanotubes comprises semiconducting carbon nanotubes.

[0017] In an embodiment, the plurality of carbon nanotubes further comprises metallic carbon nanotubes.

[0018] In an embodiment, an amount of the semiconducting carbon nanotubes is higher than an amount of the metallic carbon nanotubes.

[0019] In an embodiment, the plurality of carbon nanotubes comprises single-walled carbon nanotubes.

[0020] In an embodiment, the conducting polymer comprises polymethyl methacrylate (PMMA) or polyvinylpyrrolidone (PVP).

[0021] In an embodiment, the conducting polymer comprises a conjugated conducting polymer.

[0022] In an embodiment, the conjugated conducting polymer comprises poly (3 -hexylthiophene) (P3HT) or poly (3, 4-ethylenedioxy thiophene) polystyrene sulfonate (PEDOT:PSS).

[0023] In an embodiment, the thickness of the two-dimensional layer is between about 1 nm and about 10 nm.

[0024] In an embodiment, a thickness of the conductive polymer is between about 300 nm and about 500 nm.

[0025] In an embodiment, a size of the gaps is between about 2 nm and about 50 nm.

[0026] In a second aspect of the invention, there is provided a device arrangement for an artificial neural network comprising: the device described herein; and an electrode arrangement electrically coupled to the device, the electrode arrangement being configured for electrical signal measurements.

[0027] In an embodiment, the electrode arrangement comprises a plurality of electrodes, wherein a first electrode of the plurality of electrodes is configured to receive an input electrical signal, wherein a second electrode of the plurality of electrodes is configured to be grounded, and wherein, for each of the remaining electrodes of the plurality of electrodes, the electrode is configured for measurement of an output electrical signal in response to the input electrical signal being applied to the first electrode.

[0028] In a third aspect of the invention, there is provided a system for an artificial neural network comprising: the device arrangement described herein; and a processor configured to generate a resultant output based on the output electrical signals corresponding to at least some of the remaining electrodes of the plurality of electrodes.

[0029] In an embodiment, the processor is further configured to apply a weight matrix to the output electrical signals to generate weighted output electrical signals, and, for generating the resultant output, the processor is configured to generate the resultant output based on the weighted output electrical signals.

[0030] In a fourth aspect of the invention, there is provided a method for forming a device comprising: forming a two-dimensional layer of a plurality of carbon nanotubes with a thickness of between about 1 nm and about 40 nm on a substrate, the plurality of carbon nanotubes being arranged with gaps between the carbon nanotubes; and forming a conducting polymer over the plurality of carbon nanotubes and in the gaps, the device being configured for: in response to an electrical signal applied to the device, formation of conductive channels in the gaps an through the conducting polymer, and in response to removal of the electrical signal, dissolution of the conductive channels. In an embodiment, forming the two-dimensional layer of the plurality of carbon nanotubes comprises forming the two-dimensional layer of the plurality of carbon nanotubes on the substrate in an aqueous environment.

[0031] In an embodiment, forming the two-dimensional layer of the plurality of carbon nanotubes comprises forming the two-dimensional layer using a solution containing the plurality of carbon nanotubes.

[0032] In an embodiment, the method further comprises controlling a concentration of carbon nanotubes in the solution.

[0033] In an embodiment, forming the two-dimensional layer comprises drop-casting the solution on the substrate.

[0034] In an embodiment, forming the two-dimensional layer of the plurality of carbon nanotubes comprises: depositing initial carbon nanotubes on the substrate; removing, from the initial carbon nanotubes that are deposited, carbon nanotubes that are unattached to the substrate; and controlling a time between depositing the initial carbon nanotubes on the substrate and removing the carbon nanotubes that are unattached to the substrate to form the two-dimensional layer of the plurality of carbon nanotubes.

[0035] In an embodiment, the time is less than 10 minutes.

[0036] In an embodiment, the plurality of carbon nanotubes comprises semiconducting carbon nanotubes.

[0037] In an embodiment, the plurality of carbon nanotubes further comprises metallic carbon nanotubes. In an embodiment, an amount of the semiconducting carbon nanotubes is higher than an amount of the metallic carbon nanotubes.

[0038] In an embodiment, the method further comprises controlling a ratio of the semiconducting carbon nanotubes to the metallic carbon nanotubes.

[0039] In an embodiment, the plurality of carbon nanotubes comprises single-walled carbon nanotubes.

[0040] In an embodiment, the conducting polymer comprises polymethyl methacrylate (PMMA) or polyvinylpyrrolidone (PVP).

[0041] In an embodiment, forming the conducting polymer comprises forming a conjugated conducting polymer over the plurality of carbon nanotubes and in the gaps.

[0042] In an embodiment, the conjugated conducting polymer comprises poly (3 -hexylthiophene) (P3HT) or poly (3, 4-ethylenedioxy thiophene) polystyrene sulfonate (PEDOT:PSS).

[0043] In an embodiment, the thickness of the two-dimensional layer is between about 1 nm and about 10 nm.

[0044] In an embodiment, a thickness of the conductive polymer is between about 300 nm and about 500 nm.

[0045] In an embodiment, a size of the gaps is between about 2 nm and about 50 nm.

[0046] In a fifth aspect of the invention, there is provided a method for forming a device arrangement for an artificial neural network, the method comprising electrically coupling an electrode arrangement to the device described herein, the electrode arrangement being configured for electrical signal measurements. In a sixth aspect of the invention, there is provided a method for determining a weight matrix for a device arrangement for an artificial neural network, the method comprising: applying a plurality of input electrical signals to the first electrode of the device arrangement described herein, wherein, for each input electrical signal of the plurality of input electrical signals, the input electrical signal is representative of respective known information; grounding the second electrode of the device arrangement; measuring, for the each input electrical signal and at each of at least some of the remaining electrodes of the plurality of electrodes of the device arrangement, an output electrical signal; and generating a weight matrix for the device arrangement based on the output electrical signals measured for the plurality of input electrical signals and respective known resultant outputs corresponding to the plurality of input electrical signals.

[0047] In an embodiment, for each of the plurality of input electrical signals, the input electrical signal comprises a time series of voltage pulses.

[0048] In a seventh aspect of the invention, there is provided a method for determining a resultant output for a device arrangement for an artificial neural network, the method comprising: applying an input electrical signal to the first electrode of the device arrangement described herein, wherein the input electrical signal is representative of unknown information; grounding the second electrode of the device arrangement; measuring, at each of at least some of the remaining electrodes of the plurality of electrodes of the device arrangement, an output electrical signal; applying a predetermined weight matrix associated with the device arrangement to the output electrical signals to generate weighted output electrical signals; and generating a resultant output based on the weighted output electrical signals.

[0049] In an embodiment, the method further comprises identifying the unknown information based on the resultant output. In an embodiment, the input electrical signal comprises a time series of voltage pulses.

[0050] BRIEF DESCRIPTION OF THE DRAWINGS

[0051] These and other features, aspects, and advantages of the present disclosure are described with reference to the drawings of certain embodiments, which are intended to schematically illustrate certain embodiments and not to limit the disclosure.

[0052] Figures 1 to 8 show atomic force microscope scans showing carbon nanotube network layers with tunable density.

[0053] Figures 9 to 12 show schematic views of device operation, according to various embodiments.

[0054] Figure 13 shows a schematic side view of a device, according to various embodiments. Figures 14 and 15 show schematic views of device designs.

[0055] Figure 16 shows a device arrangement for reservoir computing applications.

[0056] Figures 17 and 18 show plots of results illustrating performance of the device of various embodiments.

[0057] Figures 19 to 24 show results illustrating property or evidence of criticality in the device of various embodiments.

[0058] Figure 25 shows a plot of power spectral density for the device of various embodiments. Figures 26 and 27 respectively show a schematic side view and a perspective view of a device, according to various embodiments.

[0059] DETAILED DESCRIPTION

[0060] Although certain examples are described below, those of skill in the art will appreciate that the disclosure extends beyond the specifically disclosed examples and / or uses and obvious modifications and equivalents thereof. Thus, it is intended that the scope of the disclosure herein disclosed should not be limited by any particular examples described below. Various embodiments relate to carbon nanotubes and polymer artificial neural network (ANN) devices.

[0061] Various embodiments provide one or more atomic switch networks utilising sparse networks of carbon nanotubes interfaced with a conducting polymer. In contrast to known atomic switch networks, conductive filament formation occurs in small gaps between the carbon nanotubes due to a phase change in the conductive polymer between amorphous and crystalline states. These networks show similar favorable electrical behavior similar to biological neural networks as described herein, and may allow facile fabrication and scalability.

[0062] Various embodiments provide one or more material systems for devices, e.g., artificial neural network (ANN) devices. The material systems employ carbon nanotubes (CNTs), e.g., in the form of a CNT film, provided on a substrate (or support material) and beneath a layer of conducting polymer (e.g., conjugated conducting polymer). In this way, the CNT film is covered by the layer of (conjugated) conducting polymer. The CNTs may be attached to the substrate. The CNTs may be in contact with the substrate. A (conductive or electrical) device employing the material system of CNTs and conducting polymer may be provided.

[0063] In various embodiments, the CNT layer is a two-dimensional (2D) layer of CNTs. This means that the CNTs are arranged or distributed substantially two dimensionally on the substrate, along or across at least part of the width and length of the substrate. As a nonlimiting example, a single layer of CNTs may be deposited or provided on the substrate.

[0064] The CNTs are arranged with spacings or gaps between the CNTs, including, for example, gaps between ends or end regions of two (nearby or close proximity) CNTs. To create a CNT layer or network with gaps between the CNTs, the deposition time of the CNTs on the substrate may be controlled such that the resulting thin film of CNTs lies close to or below the percolation threshold, meaning an electrically conductive path of CNTs does not form in the 2D layer of CNTs on the substrate. The gaps (or separation distances) between the CNTs may be a few nanometers.

[0065] The conducting polymer (e.g., conjugated conducting polymer) may be provided or deposited over the 2D layer of CNTs, for example, as a layer of conducting polymer on top of the carbon nanotube layer. As a non-limiting example, a single layer of conducting polymer may be deposited or provided over the 2D layer of CNTs. The conducting polymer also fills the gaps between the CNTs. This may mean that the device of various embodiments may include a two-dimensional array or layer of CNTs interfaced with the conducting polymer (e.g., P3HT) layer, where the conducting polymer layer (only) penetrates the carbon nanotube layer in the spaces or gaps between the CNTs. As the CNTs are arranged in a 2D layer on the substrate, there are no CNTs suspended in the layer of conducting polymer, away from the substrate.

[0066] Figures 26 and 27 respectively show a schematic side view and a corresponding perspective view of a device 2000, according to various embodiments. The device 2000 includes a 2D network or 2D layer of CNTs (e.g., single-walled carbon nanotubes (SWCNTs)) 2004 on a substrate 2030, with gaps 2008 in between the CNTs 2004. A conducting polymer layer 2006 (e.g., poly (3 -hexylthiophene) (P3HT)) is arranged over the 2D layer of CNTs 2004, including in the gaps 2008.

[0067] In the context of various embodiments, and referring to Figures 26 and 27 as non-limiting examples, a “two-dimensional layer” or “2D layer” of CNTs 2004 means that the CNTs 2004 are arranged or extend substantially along or across the width and length of the substrate 2030, rather than extending upwards into the polymer matrix. There may, however, be areas of the CNT network where the CNTs 2004 may overlap with each other and be partially lifted (see, for example, portion of the device 2000 labelled “A” in Figure 26) or fully raised off the substrate 2008 (for example, in the case of a CNT situated on top of two CNTs; see portion of the device 2000 labelled “B” in Figure 26), and there may be charge transport between these overlapped CNTs 2004 and between the CNTs 2004 and the polymer 2006 arranged above the CNTs 2004. Some CNTs 2004 may, therefore, be situated away from the substrate 2030, but not both away from the substrate 2030 and out of contact with any other CNT 2004. Therefore, there are no CNTs 2004 that are suspended in the layer of conducting polymer 2006, away from the substrate 2030. In other words, the device 2000 is free of standalone CNTs 2004 that are suspended away from and above the substrate 2030 and other CNTs 2004 in the conducting polymer 2006.

[0068] Referring to Figure 27, accordingly, the CNTs 2004, as arranged in a 2D layer, extend substantially in a plane that is at least substantially parallel to an upper or top surface (or major surface) of the substrate 2030. This means that the 2D layer of CNTs 2004 extends substantially in the x and y directions shown in Figure 27, rather than or minimally in the z direction. The 2D CNT layer may also be referred to as a quasi-2D layer.

[0069] The 2D layer of CNTs may have a thickness of between about 1 nm and about 40 nm, for example, between about 1 nm and about 10 nm, between about 1 nm and about 8 nm, between about 1 nm and about 5 nm, between about 1 nm and about 3 nm, between about 3 nm and about 10 nm, between about 5 nm and about 10 nm, between about 3 nm and about 8 nm, or between about 3 nm and about 5 nm.

[0070] In one embodiment, the two-dimensional layer of CNTs 2004 has a thickness of between about 1 nm and about 10 nm. In this embodiment, the two-dimensional layer of CNTs 2004 comprises single CNTs and / or small bundles of CNTs that are each two or three tubes high. The diameter of a single CNT is between 1.2 nm to 1.7 nm. In the case where two small bundles cross, each bundle being three tubes high, considering the possible range of diameters, the height of that junction can be between 7.2 nm to 10.2 nm. In an area of the CNT network where there is a single CNT, the height can be 1.2 nm.

[0071] In another embodiment, the two-dimensional layer of CNTs 2004 has a thickness of between about 1 nm and about 40 nm. In this embodiment, the two-dimensional layer of CNTs 2004 comprises a CNT network as described above comprising one or more larger bundles of CNTs. The CNT network itself can still be sparse and arranged in the x-y plane in the same manner as a single tube network, but instead each element of the network is a bundle of CNTs up to approximately 11 tubes high. In the case where a junction forms between two bundles, each 11 tubes high, the junction can be between 26.4 nm to 37.4 nm tall. If the bundles are smaller, the size of junction will be smaller accordingly.

[0072] The layer of (conjugated) conducting polymer may have a thickness of between about 300 nm and about 500 nm, for example, between about 300 nm and about 450 nm, between about 300 nm and about 400 nm, between about 300 nm and about 350 nm, between about 350 nm and about 500 nm, between about 400 nm and about 500 nm, between about 350 nm and about 450 nm, or between about 350 nm and about 400 nm.

[0073] CNTs provide the desired properties or characteristics in terms of their robustness, reliability and the actual physical size where these properties may be obtained. This may be linked to the inter-CNT gaps being created and the uniformity across the network.

[0074] The CNT platform is easy to make, for example, on plastic and / or may potentially be stretchy or bendy.

[0075] In the context of various embodiments, the plurality of CNTs may include or may be single-walled carbon nanotubes (SWCNTs), double-walled carbon nanotubes (DWCNTs), multi-walled carbon nanotubes (MWCNTs), or any combination thereof.

[0076] Preferably, the plurality of CNTs include or are SWCNTs. The SWCNTs and their sparsity in the device or system allow for good conduction and frequent small gaps of expected “similar” behaviour, e.g., there may be substantially one type of junction in the device. SWCNTs, with a diameter of about 1.5 nm, are relatively smaller than other metal or semiconducting nanowires (e.g., having diameters over 10 nm). The smaller diameter of SWCNTs may help with the morphology as well as the behaviour.

[0077] In the context of various embodiments, the conducting polymer may include at least one of polymethyl methacrylate (PMMA) or polyvinylpyrrolidone (PVP). Other suitable conducting polymers may also be employed.

[0078] In the context of various embodiments, the conducting polymer may include a conjugated conducting polymer. The conjugated conducting polymer may include at least one of poly(3 -hexylthiophene) (P3HT) or poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS). Other suitable conjugated conducting polymers may also be employed.

[0079] Preferably, the conducting polymer includes or is P3HT, because of its crystallinity and suitability for doping.

[0080] As a non-limiting example, one material system that may be employed includes a singlewalled carbon nanotube (CNT) thin film prefabricated beneath a layer of polymer material poly (3 -hexylthiophene) (P3HT).

[0081] In the context of various embodiments, the substrate may be or may include a silicon- based substrate, e.g., a silicon wafer with a 300 nm oxide layer that has been treated with a chemical linker poly-l-lysine. Other suitable substrates may be employed, for example, polyimide substrate (e.g., Kapton), glass, any plastic film material useful for electronic device platforms, including polydimethylsiloxane (PDMS), and other flexible / stretchable materials.

[0082] The carbon nanotube network according to the techniques disclosed herein is a layer that is formed such that there are gaps between the CNTs (e.g., SWCNTs) and the device (having the CNTs) is not electrically conductive prior to the deposition of the (conjugated) conducting polymer (e.g., P3HT layer). When an appropriate electrical (bias) signal (e.g., a voltage) is applied to the device, atomic scale memristive switches form in the gaps between the CNTs. Without wishing to be bound by any theory, the forming of the memristive switches in the inter-CNT gaps may be due to a structural phase change between ordered and disordered crystalline states that occurs in the conducting polymer (e.g., P3HT) in the inter-CNT gaps. The devices of various embodiments, therefore, may exhibit memristive behavior.

[0083] It follows from the abovementioned structural phase change in the conducting polymer in the inter-CNT gaps that the morphology of the CNT network is critical to the fabrication of a device that is useful for reservoir computing. Morphology statistics of networks can be manually calculated by counting the total number of CNTs and measuring the length of each CNT in an atomic force microscopy (AFM) scan. The number of CNTs per square micron is used as a measure of the density of the CNT network. The average length of the CNTs in each AFM scan is calculated from such measurements. These statistics can then be used to calculate the percolation threshold, or the density in the number of CNTs per square micron, required to form a fully connected path of CNTs across the device. From percolation theory, pth = 4.2362 n / L2where pth is the percolation threshold and L2is the average tube length in microns. The measured tube density can be compared to the percolation threshold to determine whether the CNT networks are percolating or not.

[0084] Figures 1 to 8 show AFM scans. Figures 1-6 are from devices that have shown memristive behaviour. In contrast, Figures 7 and 8 are from devices where conductive filament formation did not occur and no memristive switching was observed. The average CNT length, calculated percolation threshold, and measured CNT density for each scan are shown in Table 1 below. Devices which did not show memristive behaviour, and only showed semiconducting behaviour are greater than one standard deviation below and above the percolation threshold. For all devices which show memristive switching, the CNT density is below the percolation threshold, but within one standard deviation of the calculated percolation threshold value. These specific parameters are critical for the device to show memristive switching behaviour. Table 1 Average CNT length, calculated percolation threshold, and measured CNT density for each scan shown in Figures 1 to 8

[0085] Calculation percolation

[0086] Figure Average CNT CNT density (number threshold (number of number length (pm) of CNTs per pm2)

[0087] CNTs per pm2)

[0088] 1 0.48 + / - 0.14 25 + / - 7 24

[0089] 2 0.44 + / - 0.26 28 + / - 16 19

[0090] 3 0.42 + / - 0.26 33 + / - 20 17

[0091] 4 0.46 + / - 0.27 27 + / - 15 12

[0092] 5 0.33 + / - 0.21 53 + / - 32 23

[0093] 60.42 + / - 0.29 33 + / - 22 25

[0094] 70.50 + / - 0.30 23 + / - 13 35

[0095] 80.32 + / - 0.21 55 + / - 36 16

[0096] Additionally, the heights of 30 CNTs in each scan were sampled and showed a range of 1 nm to 4 nm. As the height range of single CNTs (available from Nanolntegris) used was 1.2 nm to 1.7 nm, this implies the network is composed of one single layer of either single CNTs or small bundles of two or three CNTs high. While some CNTs may be crossed over each other, it will be appreciated that this is still a single layer of a CNT network, not multiple layers of CNTs or a film of CNTs that is relatively thin compared to the other proportions of the device as in known devices. In the context of various embodiments, a size (or dimension) of the gaps between the CNTs may be between about 2 nm and about 50 nm for memristive switches to form in the gaps. In gaps smaller than 2 nm, conduction may occur via tunneling, so the formation of switches may not take place. The gap size may be up to about 50 nm as may be determined by estimating the power dissipation required to heat the conducting polymer (e.g., P3HT) to the molten state when the conducting polymer is in between two carbon nanotubes at a potential difference of about 10 V. A conducting polymer in gaps larger than 50 nm may not be able to melt (for structural phase change) at a potential difference equal to or lower than 10V. As a non-limiting example, for gaps between the CNTs that are filled with P3HT, the gaps may be about 2 nm to about 50 nm.

[0097] In various embodiments, the device may include a pair of electrodes to which an electrical (bias) signal may be applied. The pair of electrodes may sandwich the 2D layer of CNTs and the conducting polymer. Alternatively, the electrodes may be electrically coupled to the 2D layer of CNTs. The electrodes may, for example, form or be part of an electrode arrangement (e.g., 1060, Figure 16) discussed further below.

[0098] When an electrical (bias) signal is applied to the device, it will either operate in one of two modes: as a semiconductor (due to the nature of the conducting polymer which spans the electrodes of the device), or as a network of memristors (due to physical changes in the structure of conducting polymer which creates a conductive filament). The mode of operation depends on the device’s morphology and the magnitude of the electrical bias applied. The device will operate as a semiconductor if the density of the network of CNTs is far below the percolation threshold (greater than one standard deviation) or greater than the percolation threshold, and if the magnitude of the electrical bias applied to the device is too low. In both cases, the conditions for forming conductive filaments will not be met. If the density of the CNT network is below the percolation threshold within one standard deviation, and the electrical bias applied is of a high enough magnitude, conductive filaments will form in the device and the device will operate as a network of memristors. Accordingly, before the polymer is applied (CNT network only), the device will not electrically conduct, and after the polymer is applied the device will conduct in one of the two abovementioned modes (semiconducting or memristive).

[0099] A respective memristive switch may be formed in a local region due to the local structural phase change of the conducting polymer in that region. The changes in the polymer may occur at many or different regions (or places) within the material system, leading to formation of a plurality of memristive switches. The memristive switches may be independent of one another. There may be varied magnitude of switching events due to various or different switches forming in the material network.

[0100] Without wishing to be bound by any theory, the memristive switching mechanism observed in the devices of various embodiments may be as follows, using the non-limiting examples of a carbon nanotube / P3HT system. Conduction through the carbon nanotube / P3HT system may be due to pinch points which occur at gaps in the conduction path of the underlying carbon nanotube network. Memristive switches form at the gaps between carbon nanotubes through the P3HT inside of these gaps.

[0101] As P3HT is a semiconducting polymer, and with P3HT being a conjugated polymer whose electrical properties are strongly dependent on the crystallisation of the polymer, memristive switching may be due to a structural phase change between ordered and disordered crystalline states that occurs in the polymer in the inter-CNT gaps.

[0102] With the application of an electrical (bias) signal (e.g., a voltage bias) across the device, the electric field in the inter-CNT gaps is expected to be very large due to the small size of the gaps of a few nanometers on average. This local field causes Joule heating and the subsequent increase of the temperature of the polymer leading it to melt in this region (Wang, B., et al., Macromolecules, 51(19), 7692-7698 (2018)). This high electric field across the P3HT in the inter-CNT gap, now in its molten state, may cause low levels of crystallization (Lee, S. W., et al., Electronic Materials Letters, 9(4), 471-476 (2013)). Because short-range order in conjugated polymer systems is sufficient to provide effective charge transport (Noriega, R., et al., Nature Materials, 12(11), 1038-1044 (2013)), when the polymer in this region is semi-crystallised, the charge transport is increased, and the switch transitions into the “ON” state. Conductive channels (or conduction channels) may, therefore, be formed in the inter-CNT gaps as a result of the (localised) semi-crystallised or molten state of the conducting polymer in the gaps. This allows electrical conduction between CNTs in the gaps and through the conducting polymer via the conductive channels.

[0103] The removal of the bias signal causes the polymer to recrystallise, leading to the dissolution of the conductive channels. Due to the thickness of the polymer layer (which may be about 300 nm), it is likely that (only) the layers of polymer close to or in plane with the carbon nanotube layer may undergo the melting process. The surrounding polymer layers that are unchanged by the melting process may act as seed regions, and the melted region may recrystallise in independent smaller crystalline regions in order with these seed regions (Wang, B., et al., Macromolecules, 52(16), 6088-6096 (2019)). This means the recrystallised polymer in the inter-CNT gaps may not form a single highly crystalline region, but rather forms back into a disordered state having smaller crystalline regions with inter-crystalline defects and potential amorphous regions. Thus, the charge transport through the polymer may be limited as there may no longer be short-range order and the switch transitions back into an “OFF’ state.

[0104] This process may be reversible and stable as the melting and reordering process may be repeated and the same mechanism may be achievable with either a positive or negative bias signal (e.g., positive or negative voltage bias).

[0105] In contrast, in known material systems employing insulating polymers where dielectric burnout occurs, memristive switching is due to the creation and dissolution of carbon- rich conductive filaments in the polymer between carbon nanotubes resulting from the breakdown and subsequent pyrolysis of the polymer creating voids inside the polymer matrix.

[0106] Accordingly, in various embodiments, formation of conductive channels in the inter-CNT gaps through the conducting polymer may occur in response to the application of an electrical bias signal of a first polarity (e.g., positive voltage bias) to the device. Formation of the conductive channels may also occur in response to the application of another electrical bias signal of a second polarity opposite to the first polarity (e.g., negative voltage bias) to the device. With the formation of the conductive channels, the switch is in or transitions into a first state (e.g., the “ON” state). Further, the device is in a first resistance state (e.g., a low resistance state (LRS)).

[0107] Dissolution or termination of the conductive channels in the inter-CNT gaps occurs when the applied electrical bias signal is removed. The switch, then, is in or transitions into a second state (e.g., the “OFF” state). Further, the device is in a second resistance state (e.g., a high resistance state (HRS)).

[0108] Further, the techniques disclosed herein may provide for the ability to tune the carbon nanotube thin films in terms of density and / or connectivity to create an interconnected network of memristive switches. With the fabrication method disclosed herein, the conductivity (e.g., components of metallic versus semiconducting CNTs) may also be tuned.

[0109] Various embodiments may further provide hardware based atomic switch networks for use in artificial neural network (ANN) applications such as reservoir computing. As nonlimiting examples, the devices of various embodiments may be employed as or may function as an ANN device, e.g., a physical ANN.

[0110] Various embodiments may further provide a more robust and cost effective atomic switch network using organic materials compared to known networks. Local structural phase change of the organic or polymer system caused by heating (e.g., Joule heating) can result in the formation of one or more (conductive) channels (or conductive pathways). Such channel(s) may be created and destroyed with the application of an appropriate electrical (bias) signal (e.g., a voltage bias), leading to memristive behavior. As a non-limiting example, a network of memristive switches may be provided or created with a sparse thin film of single-walled carbon nanotubes coated with a layer of P3HT, where switches form (via or as a result of formation and dissolution of conductive channels) through the polymer between the carbon nanotubes.

[0111] The devices of various embodiments employing a two-dimensional (2D) layer of CNTs and with a conjugated conducting polymer over the 2D layer and in between gaps of the CNTs as described herein may be a memory device, a switching device, a memristive device, e.g., a memristive switching device. The device may switch between a high resistance state (HRS) and a low resistance state (LRS) (see, for example, Figure 18 discussed further below), and, thus, may be able to provide memory functionality.

[0112] Various embodiments may further provide one or more methods to create a robust material system that may be easily fabricated.

[0113] As a non-limiting example, a fabrication process for creating single-walled carbon nanotube and poly(3-hexylthiophene) devices to be used in artificial neural network applications such as reservoir computing is provided. The fabrication process may be carried out in two separate steps: first, a CNT layer may be formed or deposited on a substrate or support material, and, second, a P3HT layer may be formed or deposited over the CNT layer. Carrying out such a two-step process allows for control over the morphology of the CNT layer and for the creation of a CNT network with gaps between the CNTs. With the addition of a P3HT layer over the entire CNT network, nano-scale memristive switches are now able to form in the gaps between the CNTs with the application of an appropriate electrical (bias) signal, e.g., a voltage. The presence of memristive switches contribute to collective electrical behaviors of the device similar to biological neural networks which are shown to be favorable for information processing. This along with non-linear signal transformation and memory properties make the material system of the techniques disclosed herein a suitable candidate for reservoir computing applications.

[0114] By carrying out two separate steps for depositing the CNTs and P3HT respectively, a separate CNT layer may be deposited. With the separate CNT layer, the area where memristive switches may form is isolated from the environment. This may improve the stability of the memristive switches over time to prevent environmental degradation. Further, by depositing the CNTs prior to depositing the polymer layer, the morphology of the CNT layer may be controlled.

[0115] The device of various embodiments may be used for artificial neural network (ANN) applications. An electrode arrangement may be electrically coupled to the device to define a device arrangement that may be configured for or as an artificial neural network (ANN). The electrode arrangement may be, for example, electrically coupled to the two- dimensional (2D) layer of CNTs. This may mean that the electrode arrangement may be electrically coupled to the plurality of CNTs.

[0116] A non-limiting example of device operation is as follows. The CNT / P3HT network of the device may be connected with or electrically coupled to an electrode arrangement, e.g., having 16 electrodes. The electrodes may include metal electrodes, e.g., titanium / gold electrodes.

[0117] An electrical (bias) signal, e.g., an AC (alternating current) or DC (direct current) voltage bias may be provided or inputted at or to one (first) electrode. A second electrode may be grounded. Floating voltages may be measured at the remaining electrodes (e.g., 14 electrodes of a 16-electrode arrangement). The input signal is non-linearly transformed by the CNT / P3HT system to form separable outputs at the 14 electrodes measuring the floating voltages. A final output (or resultant output) may then be created by computing a linear weighted sum of the 14 floating voltage measurements. The weights may be calculated by training the system with data where the final output is already known. This forms the basis of the computing framework of reservoir computing done using hardware as a physical reservoir.

[0118] The floating voltage outputs are separable outputs as the input signal is nonlinearly transformed by the reservoir (CNT / P3HT system) so that the voltage signal for each output or at each of the remaining 14 electrodes is unique. The voltage signal for a given output is not just a linear multiple (e.g., scaled up or scaled down version) of another output. The voltage signal for a given output (at a respective electrode measuring a floating voltage) changes in time in a different manner than the voltage signal for another output (at another respective electrode measuring another floating voltage).

[0119] Depending on the applications, the floating voltage outputs from less than 14 electrodes may be sufficient for some applications. For example, a chaotic time series prediction task may be sufficiently carried out or completed using less than 14 floating voltage outputs (e.g., 4 to 6 outputs).

[0120] It should be appreciated that the electrode arrangement may include any suitable number of electrodes, e.g., 3, 4, 5, 6 or any higher number of electrodes. Further, it should be appreciated that for an electrode arrangement having a certain number of electrodes, not all of the electrodes are required to be used at any one time or for any one application.

[0121] The device arrangement is symmetrical. Any one of the electrodes of the electrode arrangement may act as the electrode where the input signal (which may be or may act as a bias signal) is applied to. Any one of the electrodes may be grounded. One or more of the remaining electrodes after determining the electrodes for the input signal and for grounding may be used for floating voltage measurements. The behaviour of the device arrangement or the result derived from the device arrangement, for example, in terms of the final output that is obtained, may depend on which one of the electrodes of the electrode arrangement is used for receiving the input signal.

[0122] While the fabrication process and device are described using the material system of SWCNTs and P3HT, this serves as a non-limiting example, and other types of carbon nanotubes (e.g., DWCNTs, MWCNTs), and / or other conducting polymers (e.g., PMMA, PVP, PEDOT:PSS) may be employed for the fabrication process and device.

[0123] Various embodiments or techniques will now be further described in detail by way of the following non-limiting examples and with reference to the figures.

[0124] The fabrication process for creating single-walled carbon nanotube and poly(3- hexylthiophene) devices will now be described.

[0125] For SWCNT deposition, an aqueous solution of presorted carbon nanotubes (e.g., presorted aqueous CNTs by Nanointegris) with a tunable metallic to semiconducting ratio is drop-cast onto a substrate (or support material) and let sit in or exposed to an aqueous environment (e.g., steam) for a set (or predetermined) amount of time (e.g., less than 10 minutes, between 4 minutes and 8 minutes) before being rinsed off the support material with de-ionized water. For example, the CNT solution may be drop-cast onto a silicon wafer with a 300 nm oxide layer that has been treated with a chemical linker poly-l-lysine.

[0126] The aqueous environment helps to minimise or prevent the drop-cast CNT solution from drying over the course of the deposition time. Drying may cause the resulting film of CNTs to be inhomogenous due to migration of solid material in a liquid during the drying process (coffee ring effect). The rinsing step helps to remove (all) CNTs that are not yet attached or are unattached to the substrate. Rinsing effectively helps to halt the CNT deposition process leaving (only) the CNTs that have stuck or attached to the substrate.

[0127] The deposition time or the amount of time the CNTs are left on the substrate, prior to rinsing, or effectively the amount of time between the aqueous CNT solution being dropcast onto the substrate and the remaining CNT solution being rinsed off of the substrate, may determine or dictate the density of the carbon nanotube network, e.g., the amount of CNTs in a 2D layer. As a non-limiting example, the deposition time may be chosen or determined such that the resulting thin film of CNTs lies close to or below the percolation threshold, meaning an electrically conductive path of carbon nanotubes does not form across the eventual fabricated device (e.g., across a surface of the substrate on which the CNTs are deposited) due to formation of gaps between the CNTs. In other words, the CNTs may not form a continuous conductive path in the 2D layer on the substrate as a result of gaps between the CNTs, e.g., gaps between the ends of the CNTs.

[0128] Devices of various embodiments, possessing electrical properties described herein, may include approximately one to four gaps of a suitable size (e.g., 2 nm - 50 nm) per square micron of carbon nanotube network. The gaps of the devices may be visually observed or imaged, for example, using microscopy techniques such as atomic force microscopy (AFM). The gaps in between isolated areas of the CNT network may be measured and the number thereof may be counted. Additionally or alternatively, a device may be determined to have a sufficient number of gaps and suitably sized gaps if there is no continuous conductive path through the CNTs (i.e., device is not electrically conductive) and memristive switching is observed in the operation of the final device (after the conducting polymer layer has been deposited).

[0129] Aqueous CNTs solutions with defined semiconducting to metallic CNT ratios may be used. It is preferred that there is a higher content or amount of semiconducting CNTs compared to metallic CNTs which may result in better device performance. As non- limiting examples, 99.9%, 98% and 90% semiconducting CNT (e.g., SWCNT) solutions may be used. The percentage refers to the number of semiconducting CNTs to metallic CNTs. For example, a 99.9% CNT solution contains 99.9% semiconducting carbon nanotubes and 0.1% metallic carbon nanotubes.

[0130] The deposition time for a resulting thin film of CNTs that is close to or below the percolation threshold may be chosen or determined based on the metallic / semiconducting CNT ratio and / or the concentration of CNTs in the aqueous solution.

[0131] The deposition time may depend on the metallic / semiconducting CNT ratio as metallic CNTs and semiconducting CNTs have different deposition dynamics.

[0132] In terms of the concentration of CNTs in the aqueous solution, an initial 10 ug CNT / 1 mL surfactant solution may be diluted with additional surfactant to a concentration of either 4 ug / mL or 5 ug / mL, and the deposition time may be between 4 minutes and 8 minutes depending on the metallic to semiconducting CNT ratio. In various embodiments, the deposition time (i.e., between deposition of CNTs and prior to rinsing) may be less than 10 minutes, for example, less than 8 minutes, or less than 5 minutes. The additional surfactant used for dilution may be the same surfactant that the CNTs are suspended in in the aqueous solution of CNTs.

[0133] In various embodiments, a concentration of carbon nanotubes in the aqueous solution and / or the metallic / semiconduction CNT ratio may be controlled or adjusted.

[0134] Figures 1 and 2 show atomic force microscope (AFM) scans showing single-walled carbon nanotube network layers of different densities. For the density shown in Figure 1, 100 uL of undiluted (1 mg / lOOmL) 99.9% semiconducting carbon nanotube solution was deposited on the substrate for 10 minutes before rinsing, while for the density shown in Figure 2, 100 uL of 40% diluted (0.4 mg / lOOmL) 99.9% semiconducting carbon nanotube solution was deposited on the substrate for 6 minutes before rinsing. Alternative to drop-casting, the substrate or substrate may be immersed or submerged in the aqueous CNT solution to form a homogenous CNT film on the substrate. Time and / or CNT solution concentration may be controlled to obtain the desired CNT density.

[0135] While deposition of the CNTs has been described in terms of using an aqueous solution of carbon nanotubes, the CNTs may be deposited using any suitable (liquid) solution of CNTs, including, for example, CNTs dispersed in any suitable solvents. Thus, a similar morphology of the CNT network may also be achieved with a solvent deposition method. Non- limiting examples of such solvents may include acetone, ethanol, DCB (1,2- dichlorobenzene), NMP (N-Methyl-2-pyrrolidone), and DMF (Dimethylformamide).

[0136] For P3HT deposition, a solution of powder form P3HT was dissolved in a solvent (dichlorobenzene @ 10 mg / mL) and sonicated for 16 hours. The resulting solution was refrigerated for at least 1 day. 75 microlitres of the P3HT solution was then spin-coated at 500 RPM for 90 seconds onto the substrate with the deposited CNT layer. The excess P3HT was wiped from the edges of the device with a cleanroom wipe wetted with acetone. The device (having the CNT layer and P3HT) was annealed for 20 minutes at 110°C.

[0137] It should be appreciated that the solution of powder form P3HT may be dissolved in other suitable solvents including but not limited to chloroform, chlorobenzene, toluene, xylene, hexane, tetrahydrofuran (THF).

[0138] The two-step fabrication process allows the morphology of the carbon nanotube layer to be tuned in order to create a CNT network with gaps between the CNTs. With the addition of the P3HT layer over the entire CNT network, nano-scale memristive switches are able to form in the gaps between the CNTs with the application of a suitable electrical (e.g., voltage) bias that is sufficient or high enough to cause Joule heating and a resulting structural phase change of the polymer in the gaps. The process or state of the polymer is reversible. As non-limiting examples, the electrical (e.g., voltage) bias may be between about 0.1 V and about 10 V, for example, between about 0.1 V and about 5 V, between about 0.1 V and about 2 V, between about 1 V and about 10 V, between about 1 V and about 5 V, between about 5 V and about 10 V, or between about 3 V and about 5 V.

[0139] The schematic views of Figures 9 to 12 outline the operation of the device and highlight the necessity of the gaps between the CNTs for the switching process to occur. In response to application of a bias signal to the device, there is formation of conductive channels in the gaps between the CNTs and through the conducting polymer in the gaps to enable a current to flow through the conductive channels and the CNTs interconnected to one another via the conductive channels. As such, there is a conduction pathway defined through at least some of the CNTs and the conductive channels interconnecting these CNTs. In response to removal of the bias signal, there is dissolution of or disruption to the conductive channels, thus, breaking the previously formed conduction pathway, and consequently, interrupting the flow of current through the CNTs.

[0140] Figure 9 shows a schematic view of a (pristine) device 300 having a carbon nanotube network 302 on a substrate (not shown). The carbon nanotube network 302 includes a plurality of single-walled carbon nanotubes (CNTs) (shown in white with two CNTs labelled as 304) coated with poly (3 -hexylthiophene) (P3HT) 306. The CNT network 302 is formed or deposited as a 2D layer on the substrate. As may be observed in Figure 9, the CNTs 304 are spaced apart from one another with gaps 308 between the CNTs 304, for example, there are gaps 308 in between the ends or end regions of the CNTs 304. In other words, the CNT network 302 does not form a connected path across the device 300. For example, the CNTs 304 do not form a continuous (conductive) path from one side (or end) of the device 300 to another side (e.g., opposite side or end) of the device 300. P3HT 306 may be formed as a layer over the CNT network 302 or the CNTs 304. P3HT 306 is provided or deposited in the gaps 308. Figure 10 shows a schematic view of the device 300 where a voltage bias 310 (e.g., AV) is applied across the device 300, causing heating (e.g., Joule heating) in the gaps 308 between the carbon nanotubes 304 through the polymer 306. The arrows 314 represent the localised region of high electric field that occurs in the gaps 308 between the carbon nanotubes 304 upon the application of the voltage bias 310.

[0141] Referring to Figure 11, (Joule) heating causes a (localised) structural phase change of the polymer 306 in the gaps 308, leading to formation of conductive channels 316 in the gaps 308. The conductive channels 316 are formed through the P3HT 306. The conductive channels 316 may be formed between ends or end regions of two CNTs 304. Current is now able to flow through the paths of some of the carbon nanotubes (shown in dark with white boundaries with five such CNTs labelled as 305). In this way, the CNTs 305, together with the conductive channels 316, form or define a continuous electrical or conductive path through the device 300. The device 300 is in a low resistance state (LRS).

[0142] Referring to Figure 12, the removal of the voltage bias 310 causes the conductive channels 316 to rupture, as a result of recrystallisation of P3HT 306 in the gaps 308, bringing the device 300 into a high resistance state (HRS). In other words, there is dissolution of the conductive channels 316 thereby breaking or interrupting the previously formed continuous conductive path through the CNTs 305.

[0143] The polymer coating 306 not only enables the creation of (memristive) switches between the CNTs 304 (e.g., via forming and dissolution of conductive channels 316), but additionally acts as an encapsulation layer protecting the switching area from the surrounding environment.

[0144] Figure 13 shows a schematic side view of a device 700. The device 700 may operate as described in the context of the device 300. The device 700 includes a support material or substrate 730 on which a 2D layer 702 of carbon nanotubes (CNTs) 704 (e.g., SWCNTs) is deposited. There are gaps 708 in between the CNTs 704, defining switching areas where conductive channels (not shown) may be formed and ruptured. The device 700 further includes P3HT 706 over the CNTs 704 and in the gaps 708 between the CNTs 704. The device 700 may further include electrodes 740, 742 formed to electrically couple to the network of CNTs 704. The relative heights of the CNT and P3HT layers 702, 706 may be as shown in Figure 13, where the height of the CNT layer 702 may be about 1 nm to about 40 nm, such as about 1 nm to about 10 nm, while the height of the P3HT layer 706 may be about 300 nm to about 500 nm. The polymer layer 706 may also be compatible with other standard encapsulation material protocols (e.g., those used for organic photo voltaics), meaning that device longevity can likely be maintained.

[0145] The devices of various embodiments, including devices 300, 700, may be used for reservoir computing. Figures 14 and 15 show schematic views of a non-limiting example of a device design for reservoir computing.

[0146] Figure 14 shows a schematic top view of an area of a carbon nanotube layer on a device 800, while Figure 15 shows a schematic top view of an electrode arrangement or electrode design 960 for or on the device 800 for electrically coupling to the device 800 so as to define a device arrangement. As a non-limiting examples, the electrode arrangement 960 may include 16 electrodes 962, e.g., titanium / gold electrodes. A device arrangement having the device 800 (with CNT and conjugated conductive polymer layers) and having the electrode arrangement 960 electrically coupled to the device 800 may define an atomic switch network device for reservoir computing. A central area of such an atomic switch network device may be as shown in Figure 13.

[0147] Figure 16 shows a non-limiting example of a device arrangement (e.g., atomic switch network device) 1070 for ANN applications (e.g., reservoir computing applications), illustrating a geometry of the device arrangement 1070. The device arrangement 1070 includes a (switching) device 1000 having a 2D layer of CNTs (e.g., SWCNTs) and conducting polymer (e.g., P3HT) as described herein, and an electrode arrangement 1060 electrically coupled to the device 1000 or the CNT layer. The electrode arrangement 1060 includes a plurality of electrodes (six electrodes are labelled as 1061, 1062, 1063, 1064, 1065, 1066), e.g., 16 electrodes. A central area of the atomic switch network device 1070 may be as shown in Figure 13.

[0148] Also shown in Figure 16 is the device measurement for use in reservoir computing. A (input) voltage bias, for example, an AC or DC voltage (e.g., Vm) is provided or supplied to one electrode (e.g., electrode 1061), and a second electrode (e.g., electrode 1062) is grounded (GND). Floating voltages may be measured or acquired at some or all of the remaining 14 electrodes to provide floating voltage measurements, Outputs 1 to 14 (e.g., Vi to Vw). For example, Output 1 (e.g., Vi) is a floating voltage measurement at electrode 1063, Output 8 (e.g., Vs) is a floating voltage measurement at electrode 1065, while Output 14 (e.g., V14) is a floating voltage measurement at electrode 1065.

[0149] The input signal (provided to the electrode 1061) is transformed by the CNT / P3HT system of the device 1000 to form separable outputs (e.g., Vi, V2, . .., V14) at the 14 electrodes measuring floating voltages. A (predetermined) weight matrix may be applied to the outputs (e.g., Vi, V2, ..., V14) to generate weighted outputs. A weighted sum of these separable outputs is then computed to create or produce a final output (or resultant output) 1080. For example, the final output 1080 is a linear weighted sum of the 14 floating voltage measurements.

[0150] As a non-limiting example, a processor may be used to receive the separable outputs, apply a weight matrix to the separable outputs, and produce the final output 1080 based on a weighted sum of the separable outputs. Weights may be applied to one or more of the separable outputs. A respective weight may be applied to a respective separable output. Different weights may be applied to different separable outputs.

[0151] As a non-limiting example, a processor or processing module based on the Python programming language may be used to process the floating voltage measurements, calculate the appropriate weights (or determine the weight matrix), and compute the final output of the system.

[0152] For computing tasks, the process of using the device arrangements of various embodiments may be separated into two stages: learning (or training), and task performance. Using the device arrangement 1070 as a non-limiting example, to begin, some type of labeled or known or identified data (or information) (e.g., a pattern, image, audio file, sequence of numbers, etc.) may be encoded as a time series of voltage pulses and input into the device arrangement 1070. The 14 floating voltages are then measured. For the learning stage, the final output of the device 1070 is known, corresponding to the labeled or identified input data. The final output and the reservoir response (floating voltage measurements) to that specific input are used to calculate the weight matrix. This may be repeated for a large set of labelled information (e.g., known data) to optimise the weight matrix.

[0153] As a non-limiting example, the weight matrix may be determined using the following equation,

[0154] W = (XTX + aI)-1XTY, where W is the weight matrix, X is the reservoir response (floating voltage measurements) matrix, Y is the final output matrix, I is the identity matrix, and a is the hyperparameter. T refers to the transpose of a matrix, which is an operator which switches the indices of the row and column of the matrix it is applied to. In the above equation, XTis the transpose of the reservoir response matrix X. The weight matrix W may be optimised, for example, using a ridge regression algorithm.

[0155] Unlabeled data (or unlabeled information) may then be similarly encoded and input into the device arrangement 1070. Unlabeled data may be, for example, unknown data (or information), unclassified data (or information), unidentified data (or information), or data (or information) that require interpretation. The final output may then be predicted by computing a weighted sum of the 14 floating voltage measurements using the (predetermined) weight matrix found during the learning phase (or training phase).

[0156] As a non-limiting example, the final output, Y, may be determined by completing the following matrix multiplication:

[0157] WX = Y, where W is the (predetermined) weight matrix, and X is the reservoir response (floating voltage measurements).

[0158] In order to perform the type of data sorting and labelling described, the physical reservoir needs to possess certain electrical properties. As described above, it is required for the input signal to be non-linearly transformed and for the outputs to be separable. In addition to these properties, the reservoir should possess some sort of memory so that the same input results in the same output from the device. Memristive devices has the ability to switch between a high resistance state (HRS) and a low resistance state (LRS) depending on the current that has flowed through the device. This switching may be observed in swept current voltage measurements of the device (arrangement) of various embodiments.

[0159] Figures 17 and 18 show performance of the device of various embodiments, e.g., with memory property. Figure 17 shows a plot of current-voltage (IV) measurements over 50 cycles plotted from 0 to 10 volts with current shown on a log scale to emphasise the orders of magnitude change in current.

[0160] Figure 18 shows a plot illustrating the resistance stability calculated using the value of the current taken at a voltage of 3 V (where the straight line is in Figure 17) plotted as a function of the cycle number. The points or results in the region 1280 of the plot correspond to the forward sweeps (from 0 V to 10 V) while the points or results in the region 1282 of the plot correspond to the backward sweeps (from 10 V to 0 V). Clear switching between the high resistance state (HRS) and the low resistance state (LRS) may be observed.

[0161] Other properties of the network include spatio-temporal correlations and critical dynamics similar to those seen in biological neural networks. Criticality refers to the phenomenon where a system operates at the boundary between order and chaos. Studies have shown that information transmission and information capacity in the cortex are functionally optimised at criticality, and it is expected that the same follows for the devices according to the techniques disclosed herein.

[0162] In networks of memristors, criticality arises due to the inter-connectivity of the networks and the formation of recurrent connectivity loops. Criticality may be observed in the devices in power-law distributions, long-range temporal correlations, and the emergence of avalanches. Figures 19 to 24 show results illustrating property or evidence of criticality in the devices.

[0163] Figure 19 shows the conductance as a function of time while Figure 20 shows the changes in the conductance over the same period of time as in Figure 19. Changes in conductance greater than the threshold magnitude are classified as switching events. The threshold magnitude may be chosen to be over an order of magnitude greater than the noise for the electrical measurement equipment that is used.

[0164] Figure 21 shows switching events over the same period of time as in Figure 19. Consecutive switching events separated by less than the average inter-event-interval are classified as avalanches. The inter-event-interval (IEI) may be defined as the time between successive switching events. The average IEI is the average of all IEI in a given dataset or measurement taken. As a non-limiting example, the value for the IEI corresponding to the results shown in Figures 19 to 24 is about 0.05s. Probability distributions of the avalanche size, S, (see Figure 22) and lifetimes, T, (see Figure 23) and the average size as a function of lifetime (see Figure 24) follow powerlaw distributions with exponents obeying the crackling relationship indicating the system is in a critical state.

[0165] Figure 25 shows a plot of power spectral density for the device of various embodiments. The power spectral density exhibits power-law behaviour with exponent consistent with device displaying long-range temporal correlations.

[0166] The carbon nanotube and poly (3 -hexylthiophene) system fabricated using the two-step deposition method has produced devices with electrical behavior or property favourable for reservoir computing applications.

[0167] The devices of various embodiments may be used in various applications that employ ANNs, and also in the field for edge-computing applications. Applications may include biotech applications, e.g., quick analysis of optical signals and pattern recognition during vaccine production. There may also be applications for safety situations in high technology dangerous environments, such as mining and fuel production, where edge computing and quick diagnostics for changing gas compositions may be required. Other suitable applications may include applications involving security, for example, facial recognition for prevention of shoplifting, criminal activities, etc.

[0168] The processor described in connection with the examples disclosed herein may be implemented or performed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic component, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, circuit, and / or state machine. A processor may also be implemented as a combination of computing components, e.g., a combination of a DSP and a microprocessor, a number of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0169] The term ‘two-dimensional layer’ or ‘2D layer’ as used in this specification and claims means a layer of two or more carbon nanotubes (CNTs) on a substrate, the layer comprising one or more individual CNTs on the substrate and / or one or more bundles of CNTs on the substrate, each bundle comprising two or more CNTs in contact with each other, wherein at least a substantial number of the CNTs in the layer are in contact with the substrate. For example, in one embodiment, at least 50%, at least 60%, at least 70%, at least 80%, or at least 90% of the CNTs in the layer are in contact with the substrate. In one embodiment, at least a substantial portion of each CNT in contact with the substrate is contacting the substrate, for example, at least 50%, at least 60%, at least 70%, at least 80%, or at least 90% of each CNT in contact with the substrate is contacting the substrate.

[0170] The term ‘comprising’ as used in this specification and claims means ‘consisting at least in part of’ . When interpreting statements in this specification and claims which include the term ‘comprising’, other features besides the features prefaced by this term in each statement can also be present. Related terms such as ‘comprise’ and ‘comprised’ are to be interpreted in a similar manner.

[0171] It is intended that reference to a range of numbers disclosed herein (for example, 1 to 10) also incorporates reference to all rational numbers within that range (for example, 1, 1.1, 2, 3, 3.9, 4, 5, 6, 6.5, 7, 8, 9 and 10) and also any range of rational numbers within that range (for example, 2 to 8, 1.5 to 5.5 and 3.1 to 4.7) and, therefore, all sub-ranges of all ranges expressly disclosed herein are hereby expressly disclosed. These are only examples of what is specifically intended and all possible combinations of numerical values between the lowest value and the highest value enumerated are to be considered to be expressly stated in this application in a similar manner. This invention may also be said broadly to consist in the parts, elements and features referred to or indicated in the specification of the application, individually or collectively, and any or all combinations of any two or more said parts, elements or features.

[0172] To those skilled in the art to which the invention relates, many changes in construction and widely differing embodiments and applications of the invention will suggest themselves without departing from the scope of the invention as defined in the appended claims. The disclosures and the descriptions herein are purely illustrative and are not intended to be in any sense limiting. Where specific integers are mentioned herein which have known equivalents in the art to which this invention relates, such known equivalents are deemed to be incorporated herein as if individually set forth.

[0173] As used herein the term ‘(s)’ following a noun means the plural and / or singular form of that noun.

[0174] As used herein the term ‘and / or’ means ‘and’ or ‘or’, or where the context allows both.

[0175] As used herein, the phrase of the form of “at least one of A or B” may include A or B or both A and B. Correspondingly, the phrase of the form of “at least one of A or B or C”, or including further listed items, may include any and all combinations of one or more of the associated listed items.

[0176] Conditional language, such as “can,” “could,” “might,” or “may,” unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements, and / or steps. Thus, such conditional language is not generally intended to imply that features, elements, and / or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements, and / or steps are included or are to be performed in any particular embodiment.

[0177] In this specification where reference has been made to external documents, or other sources of information, this is generally for the purpose of providing a context for discussing the features of the invention. Unless specifically stated otherwise, reference to such external documents, or such sources of information, is not to be construed as an admission that such documents, or such sources of information, in any jurisdiction, are prior art, or form part of the common general knowledge in the art.

[0178] Preferred embodiments of the invention have been described by way of example only and modifications may be made thereto without departing from the scope of the invention.

Claims

CLAIMS1. A device comprising: a substrate; a plurality of carbon nanotubes arranged in a two-dimensional layer with a thickness of between about 1 nm and about 40 nm on the substrate, the plurality of carbon nanotubes being arranged with gaps between the carbon nanotubes; and a conducting polymer arranged over the plurality of carbon nanotubes and in the gaps, wherein the device is configured for: in response to an electrical signal applied to the device, formation of conductive channels in the gaps and through the conducting polymer, and in response to removal of the electrical signal, dissolution of the conductive channels.

2. The device as claimed in claim 1, wherein the plurality of carbon nanotubes comprises semiconducting carbon nanotubes.

3. The device as claimed in claim 2, wherein the plurality of carbon nanotubes further comprises metallic carbon nanotubes.

4. The device as claimed in claim 3, wherein an amount of the semiconducting carbon nanotubes is higher than an amount of the metallic carbon nanotubes.

5. The device as claimed in any one of claims 1 to 4, wherein the plurality of carbon nanotubes comprises single-walled carbon nanotubes.

6. The device as claimed in any one of claims 1 to 5, wherein the conducting polymer comprises polymethyl methacrylate (PMMA) or polyvinylpyrrolidone (PVP).

7. The device as claimed in any one of claims 1 to 5, wherein the conducting polymer comprises a conjugated conducting polymer.

8. The device as claimed in claim 7, wherein the conjugated conducting polymer comprises poly (3 -hexylthiophene) (P3HT) or poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).

9. The device as claimed in any one of claims 1 to 8, wherein the thickness of the two-dimensional layer is between about 1 nm and about 10 nm.

10. The device as claimed in any one of claims 1 to 9, wherein a thickness of the conductive polymer is between about 300 nm and about 500 nm.

11. The device as claimed in any one of claims 1 to 9, wherein a size of the gaps is between about 2 nm and about 50 nm.

12. A device arrangement for an artificial neural network comprising: the device as claimed in any one of claims 1 to 11 ; and an electrode arrangement electrically coupled to the device, the electrode arrangement being configured for electrical signal measurements.

13. The device arrangement as claimed in claim 12, wherein the electrode arrangement comprises a plurality of electrodes, wherein a first electrode of the plurality of electrodes is configured to receive an input electrical signal, wherein a second electrode of the plurality of electrodes is configured to be grounded, and wherein, for each of the remaining electrodes of the plurality of electrodes, the electrode is configured for measurement of an output electrical signal in response to the input electrical signal being applied to the first electrode.

14. A system for an artificial neural network comprising: the device arrangement as claimed in claim 13; and a processor configured to generate a resultant output based on the output electrical signals corresponding to at least some of the remaining electrodes of the plurality of electrodes.

15. The system as claimed in claim 14, wherein the processor is further configured to apply a weight matrix to the output electrical signals to generate weighted output electrical signals, and wherein, for generating the resultant output, the processor is configured to generate the resultant output based on the weighted output electrical signals.

16. A method for forming a device comprising: forming a two-dimensional layer of a plurality of carbon nanotubes with a thickness of between about 1 nm and about 40 nm on a substrate, the plurality of carbon nanotubes being arranged with gaps between the carbon nanotubes; and forming a conducting polymer over the plurality of carbon nanotubes and in the gaps, the device being configured for: in response to an electrical signal applied to the device, formation of conductive channels in the gaps and through the conducting polymer, and in response to removal of the electrical signal, dissolution of the conductive channels.

17. The method as claimed in claim 16, wherein forming the two-dimensional layer of the plurality of carbon nanotubes comprises forming the two-dimensional layer of the plurality of carbon nanotubes on the substrate in an aqueous environment.

18. The method as claimed in claim 16 or 17, wherein forming the two-dimensional layer of the plurality of carbon nanotubes comprises forming the two-dimensional layer using a solution containing the plurality of carbon nanotubes.

19. The method as claimed in claim 18 , further comprising controlling a concentration of carbon nanotubes in the solution.

20. The method as claimed in claim 18 or 19, wherein forming the two-dimensional layer comprises drop-casting the solution on the substrate.

21. The method as claimed in any one of claims 16 to 20, wherein forming the two- dimensional layer of the plurality of carbon nanotubes comprises: depositing initial carbon nanotubes on the substrate; removing, from the initial carbon nanotubes that are deposited, carbon nanotubes that are unattached to the substrate; and controlling a time between depositing the initial carbon nanotubes on the substrate and removing the carbon nanotubes that are unattached to the substrate to form the two- dimensional layer of the plurality of carbon nanotubes.

22. The method as claimed in claim 21, wherein the time is less than 10 minutes.

23. The method as claimed in any one of claims 16 to 22, wherein the plurality of carbon nanotubes comprises semiconducting carbon nanotubes.

24. The method as claimed in claim 23, wherein the plurality of carbon nanotubes further comprises metallic carbon nanotubes.

25. The method as claimed in claim 24, wherein an amount of the semiconducting carbon nanotubes is higher than an amount of the metallic carbon nanotubes.

26. The method as claimed in claim 24 or 25, further comprising controlling a ratio of the semiconducting carbon nanotubes to the metallic carbon nanotubes.

27. The method as claimed in any one of claims 16 to 26, wherein the plurality of carbon nanotubes comprises single-walled carbon nanotubes.

28. The method as claimed in any one of claims 16 to 27, wherein the conducting polymer comprises polymethyl methacrylate (PMMA) or polyvinylpyrrolidone (PVP).

29. The method as claimed in any one of claims 16 to 27, wherein forming the conducting polymer comprises forming a conjugated conducting polymer over the plurality of carbon nanotubes and in the gaps.

30. The method as claimed in claim 29, wherein the conjugated conducting polymer comprises poly (3 -hexylthiophene) (P3HT) or poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).

31. The method as claimed in any one of claims 16 to 30, wherein the thickness of the two-dimensional layer is between about 1 nm and about 10 nm.

32. The method as claimed in any one of claims 16 to 31, wherein a thickness of the conductive polymer is between about 300 nm and about 500 nm.

33. The method as claimed in any one of claims 16 to 32, wherein a size of the gaps is between about 2 nm and about 50 nm.

34. A method for forming a device arrangement for an artificial neural network, the method comprising electrically coupling an electrode arrangement to the device as claimed in any one of claims 1 to 11, the electrode arrangement being configured for electrical signal measurements.

35. A method for determining a weight matrix for a device arrangement for an artificial neural network, the method comprising: applying a plurality of input electrical signals to the first electrode of the device arrangement as claimed in claim 13, wherein, for each input electrical signal of the plurality of input electrical signals, the input electrical signal is representative of respective known information; grounding the second electrode of the device arrangement; measuring, for the each input electrical signal and at each of at least some of the remaining electrodes of the plurality of electrodes of the device arrangement, an output electrical signal; and generating a weight matrix for the device arrangement based on the output electrical signals measured for the plurality of input electrical signals and respective known resultant outputs corresponding to the plurality of input electrical signals.

36. The method as claimed in claim 35, wherein, for each of the plurality of input electrical signals, the input electrical signal comprises a time series of voltage pulses.

37. A method for determining a resultant output for a device arrangement for an artificial neural network, the method comprising: applying an input electrical signal to the first electrode of the device arrangement as claimed in claim 13, wherein the input electrical signal is representative of unknown information; grounding the second electrode of the device arrangement; measuring, at each of at least some of the remaining electrodes of the plurality of electrodes of the device arrangement, an output electrical signal; applying a predetermined weight matrix associated with the device arrangement to the output electrical signals to generate weighted output electrical signals; and generating a resultant output based on the weighted output electrical signals.

38. The method as claimed in claim 37, further comprising identifying the unknown information based on the resultant output.

39. The method as claimed in claim 37 or 38, wherein the input electrical signal comprises a time series of voltage pulses.

Images