A method for preparing a carbon nanotube heterojunction device and applications thereof

Abstract

The application discloses a preparation method of a carbon nanotube heterojunction device, which comprises the following steps: preparing a pre-patterned Au / Ni thin film metal electrode on a silicon wafer substrate containing silicon dioxide by using a sputtering and stripping process, wherein the electrode spacing is 1.5-2 mu m, and the electrode width and spacing ratio is 15-30; and assembling carbon nanotubes between the electrodes by using a dielectrophoresis technology, wherein a heterojunction node array is formed between the carbon nanotubes and the electrodes, and the silicon wafer is connected to an external adapter plate by using a wire bonding technology, wherein the external adapter plate is used for inputting and collecting electric signals of the heterojunction node array. The carbon nanotube heterojunction device prepared by the application is used for constructing a reservoir pool, provides efficient reservoir pool calculation, realizes a reservoir pool with reconfigurable topology and extremely low physical implementation complexity by performing mask expansion on a single node of the carbon nanotube heterojunction, and meets the demand of high complexity tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-like computing and artificial intelligence, in particular to a method for preparing a carbon nanotube heterojunction device, constructing a carbon nanotube reservoir pool and processing data based on the reservoir pool. BACKGROUND

[0002] Reservoir computing (RC) is an efficient machine learning method, especially suitable for processing time series signals. Traditional RC relies on a large-scale randomly connected node network, which has problems such as complex hardware implementation and high power consumption. Carbon nanotube heterojunction has excellent electrical properties, which provides the possibility for constructing a small-sized, low-power reservoir pool.

[0003] Reservoir computing is a lightweight recurrent neural network, and its core idea is to use a fixed, randomly initialized reservoir to process input data, and output through a simple linear readout layer. Because the input layer weight and the reservoir weight are pre-generated and fixed, the reservoir computing network has less parameters and less computation, so it can realize low-cost, high-efficiency and low-power time series task processing. At the same time, the concept of reservoir computing can be directly mapped to a physical system, and the natural dynamics of the physical system is used to construct a physical reservoir pool, which has a complex nonlinearity that software cannot achieve, and the upper limit of the ability to process time series data is higher.

[0004] Existing physical reservoir pool nodes often use electronic, optical or mechanical nodes to realize, and lack of exploration of new nonlinear physical node implementation, and often use a physically randomly interconnected topology structure to realize, and is fixed after generation, and has no flexibility, and the physical implementation is difficult.

[0005] Carbon nanotube (CNT) is a new type of nanomaterial, which has attracted widespread attention due to its excellent electrical conductivity, mechanical properties and nonlinear characteristics. The reservoir pool based on carbon nanotube heterojunction provides a new possibility for the efficient implementation of reservoir computing due to its natural nonlinearity and easy hardware integration. By mask expansion of the strong nonlinear carbon nanotube heterojunction single node, a reservoir pool with reconfigurable topology and extremely low physical implementation complexity is realized. SUMMARY

[0006] To achieve the above purpose, the present application provides a carbon nanotube heterojunction device preparation method, comprising:

[0007] S1: using sputtering and stripping process to prepare pre-patterned Au / Ni thin film metal electrode on silicon wafer substrate containing silicon dioxide, wherein the electrode spacing is 1.5-2 μm, and the electrode width and spacing ratio is 15-30; and

[0008] S2: assembling carbon nanotubes between the electrodes by dielectrophoresis technology, wherein the carbon nanotubes form a heterojunction node array with the electrodes, and the silicon wafer is connected to an external adapter plate by wire bonding technology, wherein the external adapter plate is used to input and collect electrical signals of the heterojunction node array.

[0009] Preferably, the step S2 further comprises:

[0010] The original semiconductor carbon nanotube solution is dried by extraction, and the residual solid is mixed with ethanol and ultrasonicated to obtain a carbon nanotube solution;

[0011] Phosphomolybdic acid particles are mixed with ethanol and ultrasonicated to prepare a phosphomolybdic acid solution with a preset concentration, wherein the preset concentration of the phosphomolybdic acid solution is 250-400 μg / ml;

[0012] The carbon nanotube solution is mixed with the phosphomolybdic acid solution and ultrasonicated to obtain a phosphomolybdic acid-modified carbon nanotube solution; and

[0013] Probes are placed at both ends of the Au / Ni thin film electrode, the phosphomolybdic acid-modified carbon nanotube solution is dropped, and a sinusoidal alternating voltage signal is generated by a signal generator to perform the dielectrophoresis operation on the carbon nanotubes through the probes.

[0014] The application also discloses a method for constructing a reservoir pool based on a carbon nanotube heterojunction device, comprising:

[0015] A node in the heterojunction node array of the carbon nanotube heterojunction device is selected, and an output signal of the node is collected; and

[0016] A mask matrix is generated, and the collected output signal is mask expanded, comprising:

[0017] The output signal of each time step is mask expanded, specifically including multiplying the output signal with each column of the mask matrix one by one, and splicing by column to obtain a mask expansion matrix;

[0018] The mask expansion matrix is reorganized according to time steps, the output signal is mapped to a high-dimensional state space, a reservoir pool state matrix is constructed, and the construction of the reservoir pool is completed.

[0019] Preferably, the selection of the node further comprises: analyzing the voltammetric characteristics of a plurality of heterojunction nodes, and selecting a strong nonlinear node from the plurality of nodes as a collection node, when a nonlinear index 0<λ≤0.25, the node is a strong nonlinear node, when 0.25<λ<0.5, the node is a weak nonlinear node, and λ takes a value in the range of 0-0.5.

[0020] Preferably, the mask matrix is a matrix randomly generated in a uniform distribution between [-1, 1] or the mask matrix is composed of 1 and -1, wherein the number of 1 and -1 in each row of the mask matrix is a preset ratio, and the mask matrix is randomly generated in a hypergeometric distribution.

[0021] Preferably, constructing the reserve pool state matrix comprises: arranging the signals extended by the output signals of the same time step in the mask expansion column by column, and then arranging and splicing them in the time step sequence.

[0022] Preferably, each row in the mask matrix represents an independent mask sequence, and the dimension of the mask matrix matches the dimension of the high-dimensional state space.

[0023] The application further discloses a method for processing data by a reserve pool constructed based on a carbon nanotube heterojunction device, comprising:

[0024] selecting a node in a carbon nanotube heterojunction node array of the reserve pool constructed based on the carbon nanotube heterojunction device;

[0025] setting parameters for constructing the reserve pool, and loading a data set, wherein the data set comprises a training set and a test set;

[0026] inputting the training set to the selected node, and collecting a voltage signal of the selected node;

[0027] masking and expanding the voltage signal at each time step, mapping the collected voltage signal from a low-dimensional space to a high-dimensional space, and completing the construction of the reserve pool;

[0028] inputting a reserve pool state constructed by the training set to a readout layer, and training the readout layer; and

[0029] mapping the test set to the high-dimensional space through the reserve pool construction, inputting a reserve pool state of the test set into the trained readout layer to obtain target data.

[0030] Preferably, the data set is one of time series data, speech sequence data and motion state sequence data, and when the data set is speech sequence data, the target data is a speech classification result obtained in a cross-validation manner.

[0031] Preferably, the interface of the reserve pool constructed based on the carbon nanotube heterojunction device is built on a MATLAB platform of a host computer, wherein the interface comprises a parameter configuration module, a result display module and a saving module, and the interface is used for setting parameters for constructing the reserve pool, displaying a reserve pool state, a reserve pool evaluation index and a result of the target data.

[0032] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0033] The carbon nanotube heterojunction node array in the present application is a multi-node architecture, different nodes are selected according to task requirements, improving the flexibility of selection, and the node redundancy provides greater fault tolerance for sample production, even if individual nodes fail, it does not affect the effectiveness of the overall reserve pool calculation, and the multi-node design of the carbon nanotube heterojunction node array also facilitates the expansion of the number of nodes, supports a larger scale of reserve pool state space, and meets the needs of high complexity tasks.

[0034] Compared with the existing physical reserve pool, the hardware complexity is low, the state of the single node is expanded through the mask to build the reserve pool, and the flexible adjustment of the reserve pool topology structure is supported, which adapts to different computing task requirements. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 Preparation flowchart of carbon nanotube heterojunction according to embodiments of the present application;

[0036] Figure 2 Physical diagram of carbon nanotube heterojunction node array according to embodiments of the present application;

[0037] Figure 3 Typical nonlinear I-V curve diagram of carbon nanotube heterojunction node according to embodiments of the present application;

[0038] Figure 4 Single-node reserve pool construction flowchart based on mask expansion according to embodiments of the present application;

[0039] Figure 5 Topology diagram of reserve pool based on nanotube heterostructure according to embodiments of the present application;

[0040] Figure 6 Overall operation flowchart of the reserve pool constructed according to embodiments of the present application to complete the digital speech classification task;

[0041] Figure 7 Result diagram of the reserve pool constructed according to embodiments of the present application to complete the digital speech classification task;

[0042] Figure 8 Result diagram of the reserve pool constructed according to embodiments of the present application to complete the NARMA10 time series prediction task. DETAILED DESCRIPTION

[0043] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0044] In the present application, the terms "first", "second", etc. (if any) in the present application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0045] The preparation of the carbon nanotube heterojunction uses single-walled carbon nanotube (SWCNT) material modified by phosphomolybdic acid (POM), combines the excellent nonlinear electrical characteristics thereof, constructs a heterojunction structure, and realizes dynamic nonlinear response of a single-node reserve pool.

[0046] Embodiment 1:

[0047] Figure 1 A flowchart for preparing a carbon nanotube heterojunction according to an embodiment of the present application. The preparation of the carbon nanotube heterojunction includes preparing Au / Ni thin film metal electrodes, assembling carbon nanotubes between the electrodes, and connecting the electrodes to a signal switching PCB board. The specific flowchart for preparing the metal electrode pair and assembling the carbon nanotubes is shown in FIG. 1, which includes the following steps: Figure 1

[0048] S1: Using sputtering and stripping process to prepare pre-patterned Au / Ni thin film metal electrodes on a silicon wafer substrate containing silicon dioxide, wherein the electrode spacing is 1.5-2 μm, and the electrode width and spacing ratio is 15-30; and

[0049] S2: Assembling carbon nanotubes between the electrodes using dielectrophoresis technology, wherein the carbon nanotubes form a heterojunction node array between the electrodes, and the silicon wafer is connected to an external switching board through wire bonding technology, wherein the external switching board is used to input and collect electrical signals to the heterojunction node array.

[0050] Specifically, the stripping process is to immerse the silicon wafer substrate in a photoresist remover to dissolve the photoresist and obtain the thin film pattern deposited on the exposed substrate area.

[0051] After stripping, the silicon wafer containing the Au / Ni thin film metal electrodes is annealed, such as annealing the silicon wafer in a vacuum environment at 400°C for 4 hours, to enhance the adhesion between the metal thin film and the substrate, improve the crystal structure of the thin film, and improve the conductivity and stability of the electrode. The annealing time here is not limited to 4 hours.​

[0052] Furthermore, in step S2, carbon nanotubes are assembled between the electrodes using dielectric electrophoresis. Before assembling the carbon nanotubes, the carbon nanotubes are treated to obtain a phosphomolybdic acid-modified carbon nanotube solution. The treatment process is as follows:

[0053] First, the original semiconductor carbon nanotubes were incorporated into toluene solvent, with PCz (poly[9-(1-octylnonyl)-9H-carbazole]) as the dispersant. Then, the solution containing the original semiconductor carbon nanotubes was extracted, dried, and the residual solid was mixed with ethanol and sonicated to obtain a carbon nanotube solution. Next, phosphomolybdic acid particles were mixed with ethanol and sonicated to prepare a phosphomolybdic acid solution of a preset concentration of 250-400 μg / ml. Finally, the carbon nanotube solution and the phosphomolybdic acid solution were mixed and sonicated to obtain a phosphomolybdic acid-modified carbon nanotube solution.

[0054] In the embodiments of the present invention, the carbon nanotubes are single-walled carbon nanotubes with a length of 0.8 to 3.2 μm and a diameter of 1.2 to 1.8 nm. The above are merely embodiments of the present invention and are not limitations thereof.

[0055] Furthermore, step S2 also includes electrophoretic treatment of carbon nanotubes assembled between Au / Ni thin film metal electrodes. In an embodiment of the present invention, probes are first placed at both ends of the Au / Ni thin film electrodes, and a phosphomolybdic acid-modified carbon nanotube solution is dripped into the center of the electrode. A signal generator generates a sinusoidal AC voltage signal, which is transmitted to both ends of the Au / Ni thin film electrode through the probes, thereby performing electrophoresis on the carbon nanotubes. For example, the peak-to-peak value of the sinusoidal AC voltage signal is 16Vpp and the frequency is 2MHz. Here, the center of the electrode contains a phosphomolybdic acid-modified carbon nanotube solution.

[0056] In one optional embodiment, a silicon wafer with a carbon nanotube heterojunction node array is connected to a custom-designed signal transfer PCB board via wire bonding. The signal transfer PCB board is connected to an interface of a host computer to achieve signal transmission. Specifically, the PCB board has two rows of electrodes, namely an input terminal and an output terminal. The PCB electrodes are connected to Au / Ni thin-film metal electrodes via wire bonding. The input terminal electrodes are used to receive timing signals from the host computer, while the output terminal electrodes are connected in series with a grounding resistor to convert the electrical response of the carbon nanotube nodes into a voltage signal, which is then acquired by the host computer. A physical diagram of the PCB board connected to the carbon nanotube heterojunction node array is shown below. Figure 2 As shown.

[0057] Specifically, Figure 2The left input electrode is used to receive timing signals from the host computer, while the right output electrode is connected in series with a grounding resistor, such as a 100Ω resistor, to convert the electrical response of the carbon nanotube node into a voltage signal, such as a 0-1V voltage signal, which is then acquired by the host computer. Figure 3 This is a typical nonlinear IV curve of the carbon nanotube heterojunction node in Example 1 of the present invention. Figure 3 The current-voltage characteristic curves of nodes 12, 16 and 29 in sample No. 2 of this invention are shown respectively.

[0058] In one optional embodiment, carbon nanotube heterojunction nodes are classified into strongly nonlinear nodes and weakly nonlinear nodes. The distinction between strongly and weakly nonlinear nodes is as follows: first, the current-voltage characteristic curve of the node is normalized; then, the area of ​​the curve integrated with respect to the x-axis is calculated and denoted as λ, where λ is the nonlinearity exponent, ranging from 0 to 0.5, used to measure the degree of nonlinearity of the node. When 0 < λ ≤ 0.25, the node is strongly nonlinear; when 0.25 < λ < 0.5, the node is weakly nonlinear. The method for distinguishing between strongly and weakly nonlinear nodes also includes referencing the current-time curve and spectral characteristics of the node. The setting of the nonlinearity exponent λ is merely one embodiment of the present invention and should not be considered a limitation thereof.

[0059] like Figure 3 As shown, due to the inherent dispersion of the nodes, there are differences in the dynamic response of each node. In order to analyze the nonlinearity of each carbon nanotube heterojunction node, the electrical characteristics of the device samples at both ends were tested and analyzed under room temperature and atmospheric conditions. The IV characteristics of the corresponding nodes were obtained by cyclic testing of the device samples using a semiconductor analyzer B1500. The input bias of each measurement cycle was scanned from 1V to -1V and then back from -1V to 1V. The bias voltage pulse width was about 124ms. The IV characteristics of nodes 12, 16 and 29 were measured. Although the dynamic response of each node is different, they all showed obvious nonlinear characteristics. For example, the amplitude and growth trend of the current response are not completely symmetrical in the positive and negative voltage ranges, and voltage pulses appeared.

[0060] Example 2:

[0061] This invention discloses a method for constructing a reservoir based on the carbon nanotube heterojunction device in Example 1, comprising the following steps:

[0062] One node in the heterojunction node array of the carbon nanotube heterojunction device in Example 1 above is selected, and the output signal of the node is acquired. In an optional embodiment of the present invention, the selected node is a strongly nonlinear node. The specific method for determining a strongly nonlinear node is described in Example 1, and will not be repeated here to avoid redundancy.

[0063] Secondly, a mask matrix is generated, and the collected output signal is mask expanded. In an optional embodiment of the present application, the mask matrix is a matrix randomly generated in a uniform distribution between [-1, 1] or composed of elements 1 and -1, wherein the number of 1 and -1 in each row of the mask matrix is a preset ratio, and the mask matrix is randomly generated in a hypergeometric distribution. For example, assuming that the mask matrix size is [10, 8], composed of elements 1 and -1, and the positive and negative ratio is 4:1, the mask matrix is as follows:

[0064]

[0065] The above is only one embodiment of generating a mask matrix in the present application, and the idea of the present application is not limited thereto, and other ways of generating a mask matrix can also be selected.

[0066] Further, the collected output signal is mask expanded. Firstly, the output signal of each time step is mask expanded, specifically including multiplying the output signal with each column of the mask matrix one by one, and splicing by column to obtain a mask expansion matrix. Secondly, the mask expansion matrix is reorganized according to the time step, the output signal is mapped to a high-dimensional state space, a reservoir state matrix is constructed, and the construction of the reservoir is completed. Specifically, the signals expanded by the output signal of the same time step in the mask expansion are arranged by column, and then arranged and spliced in the time step order. And each row in the mask matrix represents an independent mask sequence, and the dimension of the mask matrix matches the dimension of the high-dimensional state space. Further, in order to better understand the mask expansion process, a data set with a sample number of I is used for illustration:

[0067] (1) The sample number of the input data set (including the training set or the test set) is I, and the step length of each sample is J. Therefore, the collected output signal is a sequence S with a length of I x J.

[0068] (2) Assuming that the mask matrix size is [N, ML], N is the number of expanded virtual neurons, and ML is the virtual neuron state number.

[0069] (3) Each time step of the sequence S is multiplied with each column of the mask matrix [N, ML] one by one, and spliced by column to obtain a mask expansion matrix E with a size of [N, ML x I x J].

[0070] (4) The mask expansion matrix E is reorganized, the elements of each row of E are extracted according to the interval ML, and then spliced by column to obtain a sequence Pi (i = 1, 2, 3…ML), and Pi is spliced by row to obtain an expanded state matrix Qj (j = 1, 2, 3…N), Qj size is [ML, I x J].

[0071] (5) concatenate Qj in row to obtain a complete extended size of [N x ML, I x J] reservoir pool state matrix R. In embodiment 2 of the present application, the collected single node state is expanded, the dynamic characteristics of the strong nonlinear carbon nanotube heterojunction single node are fully utilized, the input signal is mapped to a high-dimensional state space, and is used to realize complex calculation tasks, such as Figure 4 Figure 2 shows a flowchart of constructing a reservoir pool according to embodiment 2 of the present application.

[0072] In summary, the single node state is mapped to a high-dimensional nonlinear space by mask expansion, a high-dimensional reservoir pool state matrix is generated to complete the construction of the reservoir pool, and the time sequence state of the original signal is maintained through splicing, rearrangement and other operations to enhance the calculation ability and dynamic characteristics of the system.

[0073] Embodiment 3

[0074] The present application discloses a method for processing data based on a reservoir pool constructed by a carbon nanotube heterojunction device, comprising the following steps:

[0075] First, select a node in the carbon nanotube heterojunction node array of the above-mentioned reservoir pool constructed by the carbon nanotube heterojunction device; then set the parameters for constructing the reservoir pool, load the data set, wherein the data set includes a training set and a test set; further, input the training set to the selected node and collect the voltage signal of the selected node; mask expansion is performed on the signal at each time step of the voltage signal, the collected voltage signal is mapped from a low-dimensional space to a high-dimensional space, and the construction of the reservoir pool is completed.

[0076] Secondly, input the reservoir pool state constructed by the training set to the readout layer and train the readout layer; and map the test set to the high-dimensional space through the reservoir pool construction, input the reservoir pool state of the test set into the trained readout layer to obtain the target data result. The reservoir pool constructed by the present application can process data types including time series data prediction, speech data classification and recognition, and motion state control, etc.

[0077] For example, when predicting the trend of the target data using the reservoir pool constructed by the present application, the target data here refers to the prediction result of the target data; when processing the data set of speech sequence data, the target data here refers to the classification or recognition of the speech sequence data; when processing the data set of motion state sequence data, the target data is the future dynamic control data of the dynamic object. According to the type of the processed data set, the target data here is also different. The above are only some examples of the present application, and are not limitations of the present application. Therefore, according to embodiment 3 of the present application, the single node electrical signal in the carbon nanotube heterojunction node array is collected, the reservoir pool is constructed by mask expansion, and the processing of time series data is realized, including data prediction, speech classification and recognition, dynamic object control, etc.

[0078] In addition, in order to increase the convenience of actual operation, more conveniently change the parameters of the reserve pool construction and more intuitively view the results of the reserve pool calculation, in the process of actually constructing the reserve pool and performing the task, the reserve pool construction interface is built on the MATLAB platform of the host computer to realize, including a parameter configuration module, a result display and saving module; when the reserve pool is constructed, various parameters of the reserve pool construction are set in the interface, and a static two-dimensional graph, a dynamic three-dimensional graph of the state of the reserve pool, and calculation results of reserve pool evaluation indexes MC (Memory Capacity, memory capacity), KR (Kernel Quality, kernel quality) and GR (Generalization Rank, generalization rank) and task calculation results are displayed in the interface. For specific reserve pool construction interface, carbon nanotube node array and connection relationship between the reserve pool, please refer to Figure 5 .

[0079] Further, in order to better explain the core idea of the present application, the data of node 16 in the collected sample 2 is taken as an example to illustrate embodiment 3 of the present application, and those skilled in the art should understand that the following examples are only for explaining the process of data processing of the present application, and are not a limitation of the present application. First, a node is selected on the PCB signal adapter plate of the carbon nanotube heterojunction node array with good wire bonding, such as node 16 in Figure 3 , then the parameters of the reserve pool construction are set in the host computer, and the digital speech data set is loaded. In the present application, the digital speech data set is taken as an example for illustration, and the data set here can also be other types of data sets. For the processing of speech data, in addition to the method disclosed in the above embodiment 3, the speech data processing process also includes cross-validation of the speech data set, Figure 6 , which is taken as an example to illustrate ten-fold cross-validation. Here, the selection of ten-fold cross-validation is only one implementation manner, and is not limited thereto. For specific ten-fold cross-validation process, please refer to Figure 6 , where variable k represents cross-validation, starting from the first fold cross-validation.

[0080] Specifically, the training set data is input to the node and the node voltage signal, i.e., the node state, is collected, the node voltage signal is transmitted to the upper computer, and after the signal collection is completed, the reserve pool construction is started on the MATLAB platform of the upper computer. The specific process of reserve pool construction is described in the above embodiments 2 and 3, which will not be repeated here. After mask expansion, the reserve pool state is input to the readout layer, and the readout layer is trained and optimized until the training is completed. The training algorithm of the readout layer includes: ridge regression algorithm, linear regression algorithm, etc. After the readout layer training is completed, the test set data is input to the readout layer, and after the collection is completed, it is transmitted to the upper computer for mask expansion processing, and the classification result is calculated, such as the classification and recognition result of speech data. After the ten-fold cross-validation is completed, the average recognition rate of the speech data is further calculated and the confusion matrix is drawn.

[0081] The confusion matrix of the digital speech classification result through ten-fold cross-validation is shown in FIG. 6. Figure 7 The experimental results show that the classification and recognition rate of the model constructed by the reserve pool based on the single node of the carbon nanotube heterojunction on the TI 46-Word data set reaches 94%. This high recognition rate fully shows that the single node reserve pool based on the carbon nanotube heterojunction has a significant advantage in processing speech data, and its strong nonlinear characteristics and dynamic response capability can effectively improve the performance of the speech classification task.

[0082] Further, in order to verify the excellent effect of the reserve pool constructed by the application in time series prediction, the application takes the NARMA10 time series data set as an example for verification, and calculates the normalized mean square error (Normalized Mean Square Error, NMSE) of NARMA10 time series prediction. The prediction result line chart of NARMA10 time series and the NMSE of the prediction result are shown in FIG. 7. Figure 8 The experimental results show that the NMSE of the model constructed by the reserve pool based on the single node of the carbon nanotube heterojunction on the NARMA10 time series prediction task reaches 0.057732. This low error fully shows that the single node reserve pool based on the carbon nanotube heterojunction has a significant advantage in processing time series signals, and its strong nonlinear characteristics and dynamic response capability can effectively improve the performance of the time series prediction task.

[0083] To sum up, in the optional embodiment of the application, the carbon nanotube heterojunction node array is a 32-node architecture, different nodes are selected according to task requirements, the flexibility of selection is improved, and the node redundancy provides greater fault tolerance for sample production, that is, even if individual nodes fail, the effectiveness of the overall reserve pool calculation is not affected, the multi-node design of the carbon nanotube heterojunction node array also facilitates the expansion of the number of nodes, supports a larger scale of the reserve pool state space, and meets the demand of high complexity tasks.

[0084] Those skilled in the art will easily understand that the above description is only the preferred embodiment of the application and is not intended to limit the application, and any modifications, equivalent replacements and improvements made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for constructing a reservoir based on a carbon nanotube heterojunction device, characterized in that, include: Select one node in the heterojunction node array of the carbon nanotube heterojunction device and collect the output signal of the node; as well as Generate a mask matrix and perform mask expansion on the acquired output signal, including: The output signal at each time step is masked and expanded, specifically by multiplying the output signal by each column of the mask matrix one by one and concatenating them column by column to obtain the mask expanded matrix. The mask extension matrix is ​​renormalized according to the time step, and the output signal is mapped to a high-dimensional state space to construct the reservoir state matrix, thus completing the construction of the reservoir. The carbon nanotube heterojunction device is fabricated using the following method: S1: Pre-patterned Au / Ni thin-film metal electrodes are fabricated on silicon dioxide-containing substrates using sputtering and lift-off processes, wherein the electrode spacing is 1.5–2 μm and the electrode width to spacing ratio is 15–30; and S2: Carbon nanotubes are assembled between the electrodes using dielectric electrophoresis, wherein the carbon nanotubes and the electrodes form a heterojunction node array. The silicon wafer is connected to an external adapter plate via wire bonding technology, wherein the external adapter plate is used to input and acquire electrical signals to the heterojunction node array. Step S2 further includes: The original semiconductor carbon nanotube solution was extracted, dried, and the residual solid was mixed with ethanol and sonicated to obtain a carbon nanotube solution. The original semiconductor carbon nanotubes were used with toluene as solvent and poly[9-(1-octylnonyl)-9H-carbazole] as dispersant. Phosphomolybdic acid particles are mixed with ethanol and ultrasonically treated to prepare a phosphomolybdic acid solution of a preset concentration, wherein the preset concentration of the phosphomolybdic acid solution is 250-400 μg / ml. The carbon nanotube solution and the phosphomolybdic acid solution were mixed and sonicated to obtain a phosphomolybdic acid-modified carbon nanotube solution; and The probes are placed at both ends of the Au / Ni thin film electrode, and the phosphomolybdic acid-modified carbon nanotube solution is dripped in. A signal generator generates a sinusoidal AC voltage signal, which is then passed through the probes to perform the dielectric electrophoresis operation on the carbon nanotubes.

2. The method for constructing a reservoir based on a carbon nanotube heterojunction device according to claim 1, characterized in that, The selection of the node also includes: analyzing the current-voltage characteristics of multiple nodes, selecting a strongly nonlinear node from the multiple nodes as the acquisition node. When the nonlinearity index 0 < λ ≤ 0.25, the node is a strongly nonlinear node, and when 0.25 < λ < 0.5, the node is a weakly nonlinear node, with λ ranging from 0 to 0.

5.

3. The method for constructing a reservoir based on a carbon nanotube heterojunction device according to claim 1, characterized in that, The mask matrix is ​​a matrix randomly generated in a uniform distribution between [-1, 1] or the mask matrix is ​​composed of 1 and -1, wherein the number of 1 and -1 in each row of the mask matrix is ​​a preset ratio, and the mask matrix is ​​randomly generated in a hypergeometric distribution.

4. The method for constructing a reservoir based on a carbon nanotube heterojunction device according to claim 1, characterized in that, Constructing the state matrix of the reservoir includes: arranging the signals extended by the output signals of the same time step in the mask extension by columns, and then arranging and splicing them in the order of the time steps.

5. The method for constructing a reservoir based on a carbon nanotube heterojunction device according to claim 3, characterized in that, Each row in the mask matrix represents an independent mask sequence, and the dimension of the mask matrix matches the dimension of the high-dimensional state space.

6. A method for processing data in a reservoir constructed based on a carbon nanotube heterojunction device, characterized in that, include: Choose one node from the carbon nanotube heterojunction node array on the reservoir constructed based on the carbon nanotube heterojunction device as described in any one of claims 1-5; Set the parameters for constructing the reserve pool and load the dataset, which includes the training set and the test set; The training set is input into the selected node, and the voltage signal of the selected node is acquired; The voltage signal is masked at each time step to map the acquired voltage signal from a low-dimensional space to a high-dimensional space, thereby completing the construction of the reservoir. The state of the reservoir constructed from the training set is input into the readout layer, and the readout layer is trained. as well as The test set is mapped to a high-dimensional space through a reservoir, and the reservoir state of the test set is input into the trained readout layer to obtain the target data.

7. The method for processing data in a reservoir constructed based on a carbon nanotube heterojunction device according to claim 6, characterized in that, The dataset is one of time series data, speech sequence data, and motion state sequence data. When the dataset is speech sequence data, it also includes obtaining the target data by cross-validation, and the target data is the speech classification result.

8. The method for processing data in a reservoir constructed based on a carbon nanotube heterojunction device according to claim 6, wherein the interface of the reservoir constructed by the carbon nanotube heterojunction device is built on a MATLAB platform of a host computer, wherein the interface includes a parameter configuration module, a result display module, and a saving module, wherein the interface is used to set the parameters for constructing the reservoir, display the reservoir status, the reservoir evaluation index, and the results of the target data.

Images