KR102253227B1 - Method for analyzing atomic images of low dimensional materials - Google Patents

Abstract

The present invention relates to analysis technology capable of learning characteristics of an image acquired through a transmission scanning electron microscope, and acquiring structural information of a defect which a low-dimensional material can have in the image, in particular a point defect caused when an atom is not in its position or added to an atom position. According to the present invention, the analysis method includes the following steps of: (a) making training data by producing a transmission scanning electron microscope image of a low-dimensional image including a defect through a simulation; (b) deriving a prediction model through machine learning using the training data; and (c) acquiring an image with respect to the low-dimensional material through a transmission scanning electron microscope, and analyzing the defect through the prediction model. The training data is made by selecting a primitive unit cell smaller than a unit cell of the low-dimensional material, from the transmission scanning electron microscope image by the simulation.

Description

Atomic image analysis method of low-dimensional materials {METHOD FOR ANALYZING ATOMIC IMAGES OF LOW DIMENSIONAL MATERIALS}

The present invention relates to a technique for analyzing an atomic image of a low-dimensional material obtained through a scanning transmission electron microscope (STEM) based on deep learning. More specifically, the characteristics of the atomic image obtained through the scanning transmission electron microscope are learned through supervised learning, one of artificial intelligence learning methods, and defects that low-dimensional materials may have in the image, especially if the atom is empty in its original position or at the atomic position. It relates to an analysis technique that obtains structural information of point defects such as being added.

With the advancement of scanning transmission electron microscopy equipment, high-resolution images of atomic structures can be obtained, and based on this, understanding of the structure of materials has become possible. Based on this understanding, it is possible to clarify how the atomic structure of a material affects its properties and how the properties are expressed differently as the structure changes. In particular, low-dimensional materials in which atoms of atomic layer thickness are in a two-dimensional arrangement can make various changes in the stacking method and the number of stacks, and due to the large area exposed to the external environment, defects that occur naturally or artificially in the material can be observed. It provides many opportunities to examine the structural characteristics of the material and its properties, such as making changes to the material.

In a scanning transmission electron microscope, there are a high-angle annular dark field (HAADF) imaging mode and an annular bright field (ABF) imaging mode. HAADF is a mode used to image heavy atoms, and ABF is a mode used to image relatively light elements (oxygen, nitrogen, and hydrogen, etc.). It is possible to analyze a variety of possible point defects.

Research on the various defects that a material can have and the effect of the defects on the electrical properties of the material has received a lot of attention. Among them, low-dimensional materials are used in many studies for controlling electrical properties because structural deformation and defect generation are easy. For example, in the case of 2H-molybdenum ditelluride (2H-MoTe ₂ ), if the tellurium atom is empty, it becomes an n-type semiconductor in which electrons carry charge. Becomes.

In addition, electrical properties can be controlled even when an atom is chemically adsorbed from the outside.It acts as an electron donor that provides one electron when a hydrogen atom is chemically adsorbed, making it an n-type semiconductor, and an electron when an oxygen atom is chemically adsorbed. It acts as an electron acceptor to receive and becomes a p-type semiconductor.

Although such atomic-level defects and their effects on the electrical properties of low-dimensional materials have been studied a lot, it is difficult to analyze the defects. Compared to a pristine low-dimensional material with no defects, if an atom is empty or an atom is added, the brightness of the atom at that location becomes darker or brighter. In one STEM image obtained by an experiment, there are as few as a thousand and as many as tens of thousands of atoms, and it takes a lot of time and effort to determine the number or type of defects in atomic units.

To solve this problem, commercial software exists. This software sets the size of an atom as a motif (or base) to determine the location (center) of the atom as the center of mass within the area, and based on this, the atoms at other locations Also recognize. However, depending on the size of the motif set here, atoms at other positions may or may not be recognized as atoms. In the case of noise that appears in the image brightness, it can also be regarded as an atom, so there is a hassle of removing this value later.

When analyzing the number and type of defects using commercial software, it is obtained by integrating the brightness within the motif region at the reference atom, and calculating the ratio of the integral intensity at the other atomic positions. At this time, there is a disadvantage in that additional time is required because the reference value of the relative brightness intensity to determine the type of defect must also be statistically obtained through simulation of the HAADF image.

U.S. Patent Publication No. 10,176,363

An object of the present invention is to provide a method capable of analyzing an atomic image of a low-dimensional material containing point defects more efficiently and accurately than in the related art.

In order to solve the above problems, the present invention provides the steps of (a) creating training data by producing a scanning transmission electron microscope image of a low-dimensional material containing defects through simulation, and (b) machine learning using the training data. And (c) obtaining an image of a low-dimensional material using a scanning transmission electron microscope, and analyzing a defect using the predictive model, wherein the training data is obtained by the simulation. A method for analyzing defects in a low-dimensional material is provided, characterized in that a primitive unit cell, which is a unit smaller than a unit cell of the low-dimensional material, is selected and produced from a scanning transmission electron microscope image.

In the method according to the present invention, the low-dimensional material may be one in which materials having an atomic layer thickness are two-dimensionally arranged.

In the method according to the present invention, any one or both of a HAADF image and an ABF image may be used as the scanning transmission electron microscope image of the low-dimensional material containing the defect.

In the method according to the present invention, the machine learning is preferably a CNN (Convolution Neural Network) method suitable for image learning, but is not necessarily limited thereto, and any technique capable of efficient machine learning for atomic images is particularly limited. It doesn't work.

In the above step (c) of the method according to the present invention, obtaining an image of a low-dimensional material using a scanning transmission electron microscope is preferably, by dividing an odd-numbered layer and an even-numbered layer from top to bottom in a planar image. Can be done.

In the method according to the present invention, the selection of the primitive unit cell may be made by a method of selecting at least two primitive unit cells having different positions of atoms. When learning is performed using two or more primitive unit cells with different positions as described above, it may be advantageous to derive a more accurate predictive model for defects contained in low-dimensional materials compared to a method of creating training data in units of unit cells. have.

In the method according to the present invention, the low-dimensional material including defects through the simulation may be made of a monolayer or a bilayer, but is not limited thereto.

In the method according to the present invention, the defect may include vacancy, substitutional defect, interstitial defect, or chemisorption of atoms in vacancy. .

In the method according to the present invention, producing a scanning transmission electron microscope image of a low-dimensional material containing defects through the above simulation means that when the position where the atom is normally located is free, another atom is substituted instead of the atom that should be normally located. In this case, it may be fabricated to include a case where another atom is added to and bonded to an atom that normally exists, a case where an atom is normally located in an empty space, and a case where an external atom is chemically adsorbed at a location where the atom is empty.

In the method according to the present invention, the training data created through the simulation can be increased preferably through data augmentation.

The method according to the present invention relieves humans from having to analyze each atomic unit, and provides reliability that can supplement the low reliability of analysis of commercial software developed up to now.

In addition, if only training images are secured, it can be applied to structural analysis of materials other than low-dimensional materials that have not been previously revealed.

1 schematically shows a process of an atomic image analysis method according to an embodiment of the present invention.
_{FIG. 2 schematically shows the structure of 2H-MoTe 2} , which is a low-dimensional material used in an embodiment of the present invention, viewed from the top and cross section.
3 shows an example of generating training data with a change in the number of layers by selecting two primitive unit cells having different atomic positions through simulation.
4 shows an example of creating training data including public, substitutional defects, and intrusive defects through simulation.
5 shows labeled training data.
6 shows image convolution.
7 shows a convolutional layer and a neural network.
8 shows a training method among CNN architectures.
9 is for explaining a method of measuring the loss of CNN.
10 is for explaining a method of measuring the accuracy of the CNN.
11 shows the training process and results of CNN.
12 shows a process of extracting an image for CNN application from a scanning transmission electron microscope test image.
13 shows the result of applying the simulated test data to the CNN.
14 shows a reliability test of CNN for applying an actual scanning transmission electron microscope image to a CNN and a result of applying the CNN to an actual image.

Hereinafter, with reference to the accompanying drawings with respect to an embodiment of the present invention will be described the configuration and operation.

In describing the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted. In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

Create training data

In an embodiment of the present invention, as a method of analyzing defects in an atomic image of a scanning transmission electron microscope of a low-dimensional material, a predictive model capable of analyzing defects included in an atomic image is derived through CNN, and Use the method of analyzing atomic images. As shown in Fig. 1, this method includes the steps of creating training data, learning the training data through a CNN architecture to derive a prediction model, and applying an actual scanning transmission electron microscope to the derived prediction model. It is made including steps.

As a first step for this, for supervised learning, a HAADF image containing possible defects in low-dimensional materials is obtained through simulation. The simulation of the HAADF image consists of selecting a primitive unit cell and introducing defects to the selected portion.

_{FIG. 2 schematically shows the structure of 2H-MoTe 2} , which is a low-dimensional material used in an embodiment of the present invention, viewed from the top and cross section, and FIG. 3 is a simulation by selecting two primitive unit cells with different atomic positions. This is an example of creating data for use.

_{As shown in FIG. 2, 2H-MoTe 2} , a low-dimensional material for obtaining training images, shows a shape in which a molybdenum atom and a tellurium atom form a hexagonal shape when viewed from the top (top view), and when viewed in a cross section sectional view), shows a structure in which molybdenum and tellurium are arranged at a predetermined distance in the thickness direction.

In the planar HAADF image of the low-dimensional material of 2H-MoTe ₂ , the training image is a primitive unit cell composed of a molybdenum atom and a tellurium atom in the uppermost layer, as shown in FIG. 3. Make it the object. In this case, as shown in FIG. 3, the primitive unit cell in which molybdenum is located in the upper left and tellurium is located in the lower right, whereas the primitive unit in which molybdenum is located in the upper right and tellurium is located in the lower left. A cell can be selected, and this method is desirable to increase the training efficiency of machine learning since it is possible to flexibly select a primitive unit cell according to an image obtained in an experiment.

FIG. 4 shows an example of creating training data including vacancy, substitutional defects, and intrusive defects through simulation as a monolayer 2H-MoTe ₂ . In FIG. 4, V _Mo denotes a state in which one Mo atom is empty, and in _{the case of V Te1} , when one Te atom is empty, V _Te2 denotes a state in which both Te atoms are empty. In addition, Te _ad refers to the case where the Te atom is originally located on the Te atom, Mo _Te refers to the case where Mo is located at Te _{, and Te Mo} refers to the case where Te is located at Mo. Through this method, a training image having a total of 6 points of defects can be created, and various defects according to various number of layers can be additionally created according to the experimental results and the purpose of analysis.

5 is a labeling of training data in FIG. 4. Supervised learning is performed by labeling 0 to n times on each image including'pristine' without defects obtained by the above simulation. In addition, the number of images to be trained is increased by using data augmentation for each labeling image. The data augmentation has the advantage of securing the number of training images required for supervised learning, and enabling training even when deviating from an ideal state of defects to some extent. In addition, subsampling such as adding noise to the obtained image can be added to create training data.

Training by CNN

CNN is a compound word of convolution and neural network. In the convolution step, as shown in FIG. 6, the training image of a low-dimensional material containing defects is convolved using a filter. At this time, the matrix size of the filter to convolve the image, the number of filters, etc. become variables.

A plurality of filters can extract or maintain various features from one training image while convolving the image, and as a result, a convolutional layer 1 is formed as shown in FIG. 7. After convolution, a method of subsampling including a max pooling layer that enhances the features of the image can also be used.

In the next convolution (convolutional layer 2) step, the image is convoluted using a plurality of new filters, and finally, the convolutional layers of the convolution step are made into a single column or row vector to be an input variable of the neural network.

The neural network means that nerves are connected like a net, as shown in FIG. 8, and input variables are sent to each nerve (or node). At each node, a weight is multiplied and a bias is added to add a weight to the incoming input variable.

Each input variable comes into nodes, and the nodes at the same level form a layer, which is called a hidden layer. The weights and biases of each node existing in the hidden layer are grouped, and each is represented by a simple matrix.

Data to which weights are added through the hidden layer are finally classified by predicting defects through a fully connected layer. At this time, values of training accuracy, training loss, validation accuracy, and validation loss are used to determine whether training is successful.

As shown in FIG. 9, a method of measuring the loss value is obtained by using the predicted probability for each label by a true label and a CNN model, using the equation of FIG. 9.

The method of measuring the accuracy, as shown in FIG. 10, compares a true label and a predicted label, and gives 1 if the true label and the predicted label are the same, and gives 0 if the prediction is incorrect. This is done by calculating the average of the obtained 1s and 0s. In this case, the predicted label is a one-hot encoding method in which the highest probability value among the probabilities of the labels predicted by the model by CNN, which is used for loss, is 1, and all remaining probabilities are 0. use. In the embodiment of the present invention, 70% of the prepared data was used for training, and the remaining 30% was used for testing to evaluate training accuracy and verification loss.

Meanwhile, in the embodiment of the present invention, training (learning) through CNN means maximum accuracy by adjusting the weight and bias of each hidden layer constituting the CNN architecture by the CNN model. It means that it should have a minimum loss value.

11 shows a training process and results using a CNN architecture. Once training is completed for the entire training data, the training epoch is completed once.

Application of CNN prediction model

The application of the CNN prediction model is performed through a new image not used for training. In the embodiment of the present invention, an image for testing through simulation and an actual scanning transmission electron microscope image were used.

Specifically, from a scanning transmission electron microscope image of a low-dimensional material obtained through an experiment or made for testing, an image is taken in units of primitive unit cells in the same manner as the trained image, inputted into the CNN model, and derived by the predictive model. At the same time as marking the defect in the applied image, the final result is obtained until the statistical analysis of the point defect through the extraction of the coordinates of the imported primitive unit cell.

In this case, as shown in FIG. 12, it is preferable to divide the odd-numbered and even-numbered layers from the top to the bottom, and to obtain the images to be input to the prediction model differently from each other and apply them as shown in FIG. 12.

13 shows the results of applying the simulated test data to the CNN using the training data of FIG. 4. As shown in FIG. 13, all images with defects obtained by simulation are accurately predicted, and in particular, defects that are difficult to distinguish with the naked eye are also accurately predicted.

FIG. 14 shows the results of applying the CNN and confirming the reliability of point defect prediction before applying the actual scanning transmission electron microscope image to the CNN. As shown in Fig. 14, all images with defects obtained by simulation in the CNN reliability test are accurately predicted, and the scanning transmission electron microscope image of _{the 2H-MoTe 2 bilayer material obtained through the actual experiment is applied to the prediction model.} And statistical analysis was possible.

Claims (10)

(a) creating training data by producing a scanning transmission electron microscope image of a low-dimensional material containing defects through simulation;
(b) deriving a prediction model through machine learning using the training data; And
(c) obtaining an image of a low-dimensional material using a scanning transmission electron microscope, and analyzing a defect using the predictive model; and
The training data is created by selecting a primitive unit cell, which is a unit smaller than a unit cell of the low-dimensional material, from the scanning transmission electron microscope image by the simulation,
The selection of the primitive unit cell is performed by a method of selecting at least two primitive unit cells having different atomic positions. The method of claim 1,
In the low-dimensional material, materials having an atomic layer thickness are arranged two-dimensionally. The method of claim 1,
The scanning transmission electron microscope image of the low-dimensional material containing the defect is a HAADF image or an ABF image, an atomic image analysis method of a low-dimensional material. The method of claim 1,
The machine learning is based on a CNN (Convolution Neural Network) method, a method of analyzing an atomic image of a low-dimensional material. The method of claim 1,
In step (c), obtaining an image of a low-dimensional material using a scanning transmission electron microscope is an atomic image analysis of a low-dimensional material in which an odd-numbered layer and an even-numbered layer are separately obtained from the top to the bottom in the planar image. Way. delete The method of claim 1,
The low-dimensional material is composed of a monolayer or a bilayer, atomic image analysis method of a low-dimensional material. The method of claim 1,
The defect includes vacancy, substitutional defect, interstitial defect, or chemisorption of atoms in vacancy. The method of claim 1,
To produce a scanning transmission electron microscope image of a low-dimensional material containing defects through the above simulation, when the position where the atom is normally located is empty, when another atom is substituted instead of the atom that should be normally located, another atom is replaced on another atom. A method of analyzing an atomic image of a low-dimensional material, which is produced by including the case where an atom is added and bonded, when an atom is normally located in an empty space, or a case where an external atom is chemically adsorbed in an empty space. The method of claim 1,
The training data created through the simulation is increased through data augmentation, a method of analyzing an atomic image of a low-dimensional material.

Images