A method to quickly identify and analyze 2D materials using machine learning methods

Building a neural network through machine learning methods and using RGB&HSV information to identify two-dimensional materials, solving the problems of time-consuming and expensive equipment in traditional methods, realizing the rapid and accurate identification and classification of two-dimensional materials, improving characterization efficiency and promoting device production Application.

CN112818835BActive Publication Date: 2024-11-22NANJING UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202110124919.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-01-29
Publication Date
2024-11-22
Estimated Expiration
2041-01-29

AI Technical Summary

Technical Problem

Traditional methods for characterizing two-dimensional materials are time-consuming and expensive, making it difficult to achieve fast and efficient identification and analysis.

Method used

Using machine learning methods, optical microscopes are used to take photos of two-dimensional materials, and by building neural networks, RGB&HSV information is used to identify and classify them, including building neural network structures input to output, training and testing to improve accuracy.

Benefits of technology

It realizes rapid and accurate identification and classification of two-dimensional materials, improves characterization efficiency, average accuracy rate reaches 91%, and maintains accuracy under different light intensity conditions, promoting the application of two-dimensional materials in device production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112818835B_ABST
    Figure CN112818835B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for quickly identifying and analyzing materials using a machine learning method, comprising the following steps: step 1, taking an optical microscope picture of a two-dimensional material; step 2, extracting RGB&HSV information of 50×50 pixels of the picture, and sending it to a constructed neural network to analyze the composition and number of layers; in step 2, the constructed neural network includes an input layer, a hidden layer, and an output layer connected in sequence from input to output, the input data is six parameters of RGB&HSV, the number of neurons in the four hidden layers is 256, 1024, 1024, 256 respectively, and each neuron in each hidden layer is the result obtained by adding the weights of all neurons in the previous layer and then passing through a nonlinear activation function, and then the result is passed to the next layer for calculation. This method of quickly identifying and analyzing two-dimensional materials using a machine learning method can realize accurate and rapid identification of various types of two-dimensional materials.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence application technology, and in particular relates to a method for quickly identifying and analyzing two-dimensional materials using a machine learning method. Background Art

[0002] Since Professor Geim and Professor Novoselov used tape to mechanically peel off graphene in 2004, researchers have begun to study the unique properties of two-dimensional materials. This type of material, which can be as thin as an atomic layer, exhibits excellent optical, electrical, thermal, and mechanical properties, and they often need to be integrated and combined to be applied to optoelectronic devices. Therefore, it is of great significance to effectively identify and characterize them on the substrate. Traditional methods for characterizing two-dimensional materials require a lot of time, and the characterization instruments are also very expensive, such as Raman spectroscopy, atomic force microscopy, transmission electron microscopy, etc., which need to be improved, and this case came into being. Summary of the invention

[0003] The purpose of the present invention is to provide a method for quickly identifying and analyzing two-dimensional materials using machine learning methods, which can achieve accurate and rapid identification of various types of two-dimensional materials.

[0004] In order to achieve the above object, the solution of the present invention is:

[0005] A method for quickly identifying and analyzing materials using a machine learning method comprises the following steps:

[0006] Step 1, taking an optical microscope picture of the two-dimensional material using a 100x lens of an optical microscope;

[0007] Step 2: Use the open source matplotlib module of Python 3.7 to extract the RGB & HSV information of the 50×50 pixels of the image and feed it into the constructed neural network for component and layer analysis;

[0008] In the step 2, the constructed neural network includes an input layer, a hidden layer, and an output layer connected in sequence from input to output. The input data is six parameters of RGB&HSV. The hidden layer is provided with four layers, and the number of neurons in the four layers is 256, 1024, 1024, and 256 respectively. Each neuron in each hidden layer is the result of adding the weights of all neurons in the previous layer and then passing through a nonlinear activation function, and then the result is passed to the next layer for calculation.

[0009] In the above step 2, in the analysis result output by the neural network, the component is one of the following 8 types: tungsten disulfide, molybdenum disulfide, molybdenum ditelluride, tungsten diselenide, molybdenum diselenide, graphene, hexagonal boron nitride (h-BN), Bi2Sr2CaCu2O8.

[0010] In step 2 above, the specific process of building a neural network is:

[0011] S1, using a 100x lens of an optical microscope to take N groups of optical microscope images of 8 two-dimensional materials at equally spaced intervals from the darkest to the brightest light intensity, where N is a natural number;

[0012] S2, using Raman spectroscopy to determine the type and number of layers of each two-dimensional material;

[0013] S3, using Python 3.7's open source matplotlib module to extract the RGB & HSV information of 50 × 50 pixels of each material, and feed it into the neural network to train the neural network until the loss function is minimized;

[0014] S4, using the pictures in the test set to test the accuracy of the neural network, adjusting the number of training times and the training rate until the accuracy meets the predetermined requirements, thereby obtaining a constructed neural network structure.

[0015] In the above step S1, the value of N is 11.

[0016] In the above step S1, when taking optical microscope photos, the eight two-dimensional materials are placed under N groups of light intensities that vary from the darkest to the brightest with equal intervals for photographing.

[0017] In the above step S3, the open source matplotlib module of Python 3.7 is used to extract the RGB&HSV information of 50×50 pixels of each material, and the training set and test set are constructed in a ratio of 4:1. The information obtained from the training set is sent to the neural network to train the neural network.

[0018] In the above step S3, the expression of the loss function L is as follows:

[0019]

[0020] in, represents the predicted value of the neural network, Y i Indicates the value represented by the true result.

[0021] After adopting the above scheme, the present invention combines artificial neural network with material science, uses the color space information (RGB & HSV) related to the reflectance spectrum in the pixels of the optical microscope photos of two-dimensional materials as input data, and obtains a stable neural network that can be used to classify optical microscope photos of different two-dimensional materials through computer training and learning, with an average accuracy rate of up to 91%; in particular, various types of two-dimensional materials can be accurately identified even under different shooting light intensities. The present invention improves the characterization efficiency of two-dimensional materials and is also of great help in promoting the production of devices for the use of two-dimensional materials in basic research. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of training and testing ANN in the present invention;

[0023] Figure 2 is a neural network structure diagram in the present invention;

[0024] Figure 3 is a schematic diagram of experimental results of an embodiment of the present invention;

[0025] Among them, (a) is an optical microscope image of the two-dimensional material used for testing, and (b) is the ANN Figure 3 (a) is the recognition result, (c) is the recognition accuracy of ANN for various two-dimensional materials;

[0026] Figure 4 is a schematic diagram of the present invention for identifying a two-dimensional material heterojunction;

[0027] Among them, (a) is the optical microscope image of the heterojunction used for testing and the ANN prediction result, and (b) is the Raman intensity integral diagram of the WS2 characteristic peak and the MoS2 characteristic peak. DETAILED DESCRIPTION

[0028] The technical solutions and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] The invention provides a method for rapidly identifying and analyzing two-dimensional materials by using a machine learning method. The method mainly combines a deep learning algorithm, namely an artificial neural network, with material science. The chromaticity values ​​Red (R), Green (G), Blue (B) and Hue (H), Saturation (S), and Value (V) of pixel points of an optical microscope picture of the two-dimensional material are used as input neurons, and nearly 1000 optical microscope pictures of different two-dimensional materials are used as data sets. An artificial neural network algorithm (ANN) is trained to distinguish eight typical single- and double-layer two-dimensional materials with different functions (tungsten disulfide (WS2), molybdenum disulfide (MoS2), molybdenum ditelluride (2H-MoTe2), tungsten diselenide (2H-WSe2), molybdenum diselenide (2H-MoSe2), graphene, hexagonal boron nitride (h-BN), and Bi2Sr2CaCu2O8 (BSCCO-2212)) which are prepared on a silicon substrate coated with 285 nanometer silicon oxide and have different properties by using optical microscope pictures, thereby realizing rapid and accurate identification and classification of the two-dimensional materials.

[0030] The core of the present invention is the construction of an artificial neural network structure, and its construction process is as follows: Figure 1 As shown, the following steps are included:

[0031] S1, using an optical microscope (WITEC 300R Alpha) with a 100x lens to take optical microscope images of 8 two-dimensional materials under N groups of light intensities. In this embodiment, 11 groups of light intensities are used, and optical microscope images of 8 two-dimensional materials are taken at light intensities that change quantitatively from the darkest to the brightest with equal intervals;

[0032] S2, using Raman spectroscopy to determine the type and number of layers of each two-dimensional material;

[0033] S3, using Python 3.7's open source matplotlib module to extract the RGB & HSV information of 50 × 50 pixels of each material, of which 2000 points constitute the training set and 500 points constitute the test set. The training information is sent to Figure 2 In the neural network shown, the neural network is built using the pytorch module, including an input layer, a hidden layer, and an output layer connected in sequence from input to output. The input data is six parameters of RGB&HSV. There are four hidden layers, and the number of neurons in the four hidden layers is 256, 1024, 1024, and 256 respectively. Each neuron in each hidden layer is the result of adding the weights of all neurons in the previous layer and then passing through a nonlinear activation function (ReLu function). The formula is as follows:

[0034] f(x)=max(0,A j )

[0035] Where f(x) is the value of a neuron;

[0036] A j is the sum of the weights of all neurons in the previous layer of the neuron, expressed as follows:

[0037]

[0038] Among them, x i represents the i-th neuron in the previous layer, w ji Represents the weight of the i-th neuron, and the initial value of the weight can be any value.

[0039] After several layers of calculation, an output value can be finally obtained. After a lot of training, the neural network adjusts the weight parameters in the network to make the predicted value consistent with the true value. This can be observed through the loss function (mean square error). The formula is as follows:

[0040]

[0041] in, represents the predicted value of the neural network, Y i Indicates the value represented by the true result. When the loss function reaches the minimum, it proves that the network is stable and usable.

[0042] S4, then use the test set to test the network accuracy. If the accuracy is not high, the accuracy can be improved by adjusting the number of network training times and training rate, and repeating the experiment until it meets the predetermined requirements, thereby obtaining the neural network structure adopted by the present invention.

[0043] The specific process of the present invention for identifying and classifying an unknown two-dimensional material is:

[0044] (1) Use an optical microscope (WITEC 300R Alpha) with a 100x lens to take optical microscope images of two-dimensional materials;

[0045] (2) The open source matplotlib module of Python 3.7 was used to extract the RGB and HSV information of the 50×50 pixels in the image and feed it into the neural network constructed above for component and layer analysis.

[0046] Example: Optical microscope images were taken of several two-dimensional materials used for testing, such as Figure 3 (a) Using a trained stable neural network to recognize images, the results are shown below Figure 3 As shown in (b), we can see that each pixel of the optical image is accurately classified and the recognition time is very short, which proves that this method is accurate and fast for characterizing two-dimensional materials. Figure 3 (c) is the classification accuracy of each material. The elements on the diagonal are the accuracy rates of various materials being accurately identified, and the elements on the off-diagonal are the probabilities of the material being confused with other materials. It can be seen that the accuracy rates of most materials are above 85%.

[0047] In addition, we can also use existing datasets to identify the composition information of 2D material heterojunctions, such as Figure 4 As shown, Figure 4 (a) An optical microscope image of the two-dimensional material heterojunction molybdenum disulfide / tungsten disulfide (MoS2 / WS2) and the composition prediction results using ANN. Figure 4 (b) is the real composition analysis of the heterojunction obtained by Raman spectroscopy. It can be seen that the analysis of ANN and Raman spectroscopy is basically consistent. This method of quickly identifying the composition of heterojunctions has great significance for the application of two-dimensional material heterojunctions. It shortens the time required for sample characterization, thereby quickly screening out the heterostructures required for device manufacturing.

[0048] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for rapidly identifying and analyzing materials using machine learning methods, characterized in that: The steps include: Step 1, taking an optical microscope photo of the two-dimensional material using a 100x lens of an optical microscope; Step 2: Use Python 3.7’s open source matplotlib module to extract RGB & HSV information of 50×50 pixels of the photo and feed it into the constructed neural network for component and layer analysis; In the step 2, the constructed neural network includes an input layer, a hidden layer, and an output layer connected in sequence from input to output, the input data is six parameters of RGB&HSV, the hidden layer is provided with four layers, and the number of neurons in the four layers is 256, 1024, 1024, and 256 respectively, wherein each neuron in each hidden layer is the result obtained by adding the weights of all neurons in the previous layer and then passing through a nonlinear activation function, and then the result is passed to the next layer for calculation; In step 2, the specific process of building a neural network is: S1, using a 100x lens of an optical microscope to take N groups of optical microscope photos of 8 two-dimensional materials at equally spaced intervals from the darkest to the brightest light intensity, where N is 11; When taking optical microscope photos, the eight two-dimensional materials were placed under N groups of equally spaced light intensities varying from the darkest to the brightest. S2, using Raman spectroscopy to determine the type and number of layers of each two-dimensional material; S3, using Python 3.7's open source matplotlib module to extract the RGB & HSV information of 50 × 50 pixels of each material, and feed it into the neural network to train the neural network until the loss function is minimized; The open source matplotlib module of Python 3.7 was used to extract the RGB&HSV information of 50×50 pixels of each material, and the training set and test set were constructed in a 4:1 ratio. The information obtained from the training set was fed into the neural network to train the neural network. The expression of the loss function is as follows: in, represents the predicted value of the neural network, Y i Indicates the value represented by the true result; S4, using the photos in the test set to test the accuracy of the neural network, adjusting the number of training times and the training rate until the accuracy meets the predetermined requirements, thereby obtaining a constructed neural network structure.

2. The method according to claim 1, characterized in that In the step 2, in the analysis result output by the neural network, the component is one of the following eight types: tungsten disulfide, molybdenum disulfide, molybdenum ditelluride, tungsten diselenide, molybdenum diselenide, graphene, hexagonal boron nitride (h-BN), Bi2Sr2CaCu2O8.

Citation Information

Cited By

  • Automatic configuration of pump attributes of a raman amplifier to achieve a desired gain profile

    US20230085524A1