A method for predicting crystal structure based on improved echo state network and XRD patterns

Abstract

The present invention relates to a method for predicting crystal structure based on an improved echo state network and XRD patterns, comprising the following steps: S1, collecting XRD pattern data of various crystal structures from different sources as a sample data set, pre-processing the sample data set and then dividing it into a training set and a test set; S2, constructing an XRD space group recognition model based on the improved echo state network model, training the constructed XRD space group recognition model according to the training set and the test set of S1 to obtain a trained XRD space group recognition model, wherein the improved echo state network model is composed of a basic echo state network and a RNN neural network. The present invention overcomes the shortcomings of existing crystal structure prediction technologies and can automatically and rapidly obtain fairly accurate space groups of crystal materials on a large scale without any human intervention.

Description

Technical Field

[0001] The invention belongs to the technical field of material structure, and in particular relates to a method for predicting crystal structure based on an improved echo state network and an XRD spectrum. Background Art

[0002] The space group is a key parameter of a crystalline material's structure. It directly influences the material's electronic structure and, consequently, its performance. Successfully predicting the crystal structure will accelerate the discovery of new materials. Currently, crystal structure is primarily determined through X-ray diffraction experiments or calculated using density functional theory. However, both methods require significant investment in cost and resources, and suffer from low prediction efficiency. Summary of the Invention

[0003] Based on the existing technology, the present invention overcomes the shortcomings of low efficiency and improves the accuracy of crystal structure prediction. It proposes a method for quickly predicting crystal structure using XRD patterns. Without any human intervention, it can quickly and automatically obtain a fairly accurate space group to which the crystal material belongs, and roughly obtain the relative content relationship of multiple space group components.

[0004] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0005] A method for predicting crystal structure based on an improved echo state network and XRD patterns comprises the following steps:

[0006] S1. Collect XRD pattern data of various crystal structures from different sources as a sample data set, pre-process the sample data set, and divide it into a training set and a test set;

[0007] S2. Constructing an XRD space group recognition model based on the improved echo state network model, and training the constructed XRD space group recognition model according to the training set and test set of S1 to obtain a trained XRD space group recognition model, wherein the improved echo state network model is composed of a basic echo state network and a RNN neural network;

[0008] S3. For the crystal material to be predicted, input the XRD spectrum data of the crystal material to be predicted into the trained XRD space group recognition model, and output the crystal structure of the material to be predicted.

[0009] As a further optimization solution of the present invention, the different sources in S1 are specifically experimental data, database data, and published paper data.

[0010] As a further optimization solution of the present invention, the XRD spectrum data specifically includes diffraction angle and diffraction peak intensity data.

[0011] As a further optimization solution of the present invention, the preprocessing in S1 is specifically as follows:

[0012] Step 1.1, performing the first step of data enhancement processing on the sample data set: setting a horizontal shift scaling factor a and a diffraction peak scaling factor b, setting step sizes c and d to horizontally shift and vertically scale the diffraction peaks while ensuring that the spectral data is not distorted, thereby expanding the sample data set;

[0013] Step 1.2: Represent the expanded sample dataset as a set of time series vectors as the original dataset D. Perform the second step of enhancement processing on the original dataset using the pre-built smote algorithm.

[0014] Step 1.2.1: Filter out all minority class samples in the original dataset D to form the minority class sample set D min , the remaining samples constitute the majority class sample set D maj ;

[0015] Step 1.2.2, for Minority class samples , calculate it to The distance of all samples in Neighbor, from Take any sample from the nearest neighbors and The distance between The calculation formula is as follows:

[0016] ;

[0017] Where: Indicates 2 Data samples of dimension attributes;

[0018] Step 1.2.3. Calculate samples and The difference between the attribute values ​​of each corresponding attribute, multiply the difference by the interval A random number within, plus the sample Each corresponding attribute value can generate a new minority class sample , the specific calculation formula is as follows:

[0019] ;

[0020] Where: rand Representation interval A random number on .

[0021] As a further optimization solution of the present invention, the XRD space group identification model constructed based on the improved echo state network model in S2 is specifically:

[0022] Step 2.1, the basic echo state network (ESN) consists of an input layer, a reservoir, and an output layer;

[0023] remember The sample input sequence at time instant is , at this time the state sequence of the reserve pool is , the output sequence , then the ESN training process is:

[0024] ;

[0025] in, is a nonlinear reservoir activation function, is the activation function of the output layer, is the bias vector at output;

[0026] Step 2.2: Construct the RNN layer and convert the output sequence As the input of RNN, corresponding to the hidden state h1, h2…, ht, the output y1, y2…, yt, the operation process can be expressed as:

[0027] ;

[0028] Among them, U, W, V, b, c are all parameters, and f represents the activation function;

[0029] Step 2.3: Connect the first convolutional layer and the second convolutional layer in parallel after the RNN layer, and then connect the first pooling layer and the second pooling layer after the two convolutional layers respectively. Then merge the outputs of the two pooling layers, flatten them, and then connect another convolutional layer to output the result.

[0030] The beneficial effects of the present invention are:

[0031] 1. The present invention overcomes the shortcomings of existing crystal structure prediction technologies and can quickly and automatically obtain fairly accurate space groups of crystal materials on a large scale without any human intervention.

[0032] 2. The prediction method of the present invention does not involve experiments or use chemicals in the entire process, does not produce chemical pollution, conforms to the concept of green environmental protection, is easy to implement, and is suitable for promotion and application.

[0033] 3. This invention overcomes the problem of scarcity of experimental data in materials science and develops a new technology for predicting crystal structure using small samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the present invention.

[0035] Figure 2 Schematic diagram of the ESN infrastructure

[0036] Figure 3 This is the result of predicting the crystal structure using the XRD pattern in Example 1. DETAILED DESCRIPTION

[0037] The present application is described in further detail below in conjunction with the accompanying drawings. It is necessary to point out that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technicians in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0038] Example 1

[0039] In this embodiment, see Figure 1 A method for predicting crystal structure based on improved echo state network and XRD pattern comprises the following steps:

[0040] 1) Data collection:

[0041] 88 real crystal XRD data were collected from published articles.

[0042] 2) Data feature enhancement in the first step

[0043] In step 1), the first step of data enhancement is performed on the collected data. First, a horizontal shift scaling factor a and a diffraction peak scaling scaling factor b are set for the XRD spectral data. Then, reasonable step sizes c and d are set to horizontally shift and vertically scale the diffraction peaks while ensuring that the spectral data is not distorted, so as to expand the sample data set.

[0044] 3) Data processing

[0045] In step 2), the expanded XRD spectrum data is represented as a set of time series vectors, and then each data is labeled with a unique identifier as the original data set and divided into an 80% training set and a 20% test set.

[0046] 4) Second step of data feature enhancement

[0047] In step 3), the original data set is oversampled using the smote algorithm to further enhance data features and control data balance issues to form a new data set.

[0048] The SMOT algorithm is a synthetic minority oversampling technique. It is an improvement on the random oversampling algorithm. Since random oversampling adopts a strategy of simply duplicating samples to increase minority class samples, it is easy to cause the problem of model overfitting, that is, the information learned by the model is too specific and not general enough. The basic idea of ​​the SMOT algorithm is to analyze minority class samples and artificially synthesize new samples based on them and add them to the dataset. The algorithm process is as follows:

[0049] Input: imbalanced dataset D, number of neighbors K, sampling rate N.

[0050] Output: balanced dataset D new .

[0051] a) Filter out all minority class samples in the original data set D to form the minority class sample set D min , the remaining samples constitute the majority class sample set D maj .

[0052] b) For Minority class samples , calculate it to The distance of all samples in Neighbors, from Take any sample from the nearest neighbors and The distance between The calculation formula is shown as follows:

[0053] ;

[0054] Where: Indicates 2 Data samples of dimension attributes.

[0055] c) Calculate samples and The difference between the attribute values ​​of each corresponding attribute, multiply the difference by the interval A random number within, plus the sample Each corresponding attribute value can generate a new minority class sample , the specific calculation formula is shown in formula (2):

[0056] ;

[0057] Where: rand Representation interval A random number on .

[0058] 5) Training

[0059] In step 4), a crystal XRD data space group identification model is constructed based on an improved echo state network (ESN) model. The constructed XRD space group identification model is trained according to a training set and a test set to obtain a trained space group identification model. The ESN model is as follows:

[0060] S1. The basic echo state network (ESN) consists of three parts: input layer, reservoir and output layer. The connection matrix W between the input layer and the reservoir is in The connection matrix W of the neurons in the reserve pool is randomly generated during the network initialization phase and is no longer updated during the training process. out It is the readout connection weight matrix of the forward connection reserve pool and the output layer, and is also the only part that needs to be trained and updated. Its basic structure is as follows Figure 1 As shown. The time input sequence is , at this time the state sequence of the reservoir (i.e., the output of the reservoir neuron) is , and the output sequence , then the update during ESN training is:

[0061] ;

[0062] in, is a nonlinear reservoir activation function, often set to function, is the activation function of the output layer, usually set to the identity function. Is the bias vector at output. In simple terms, the ESN training process is to initialize and After that, update the reserve pool status .

[0063] S2, build the RNN layer, and add it to the output layer of S1 , …, As the input of RNN, corresponding to the hidden state , …, , output , …, , the operation process can be expressed as:

[0064] ;

[0065] Among them, U, W, V, b, and c are all parameters, and Represents the activation function, specifically the tanh function;

[0066] S3: Connect convolutional layer 1 and convolutional layer 2 in parallel after the RNN layer of S2, and then connect pooling layer 1 and pooling layer 2 after the two convolutional layers respectively. Then merge the outputs of pooling layer 1 and pooling layer 2, flatten them, and then connect another convolutional layer to output the classification result.

[0067] This example uses a method for predicting crystal structure based on an improved echo state network and XRD patterns to model experimental crystal XRD data from published articles. This model establishes a method that can rapidly and automatically obtain highly accurate space group predictions for crystal structures on a large scale. The final prediction accuracy for the space group was 98.7%.

[0068] The method of this embodiment overcomes the shortcomings of existing crystal structure acquisition technologies and can rapidly and automatically obtain a fairly accurate crystal space group on a large scale without any human intervention. It can also automatically analyze the XRD data of crystals.

[0069] Based on the existing technology, the present invention overcomes the shortcomings of low efficiency and improves the accuracy of obtaining crystal space groups. It proposes a method of predicting crystal structure using XRD patterns and based on an improved echo state network. Without any human intervention, a fairly accurate crystal space group can be obtained quickly and automatically on a large scale.

[0070] This invention overcomes the problem of scarcity of experimental data in materials science and develops a new technology for predicting crystal structure under small samples.

[0071] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A method for predicting crystal structure based on improved echo state network and XRD pattern, characterized in that: The steps include: S1. Collect XRD pattern data of various crystal structures from different sources as a sample data set, pre-process the sample data set and divide it into a training set and a test set; the XRD pattern data specifically includes diffraction angle and diffraction peak intensity data; S2. Constructing an XRD space group recognition model based on the improved echo state network model, and training the constructed XRD space group recognition model according to the training set and test set of S1 to obtain a trained XRD space group recognition model, wherein the improved echo state network model is composed of a basic echo state network and a RNN neural network; The XRD space group identification model based on the improved echo state network model is constructed as follows: Step 2.1, the basic echo state network (ESN) consists of an input layer, a reservoir, and an output layer; remember The sample input sequence at time instant is , at this time the state sequence of the reserve pool is , the output sequence , then the ESN training process is: ； in, is a nonlinear reservoir activation function, is the activation function of the output layer, is the bias vector at output, is the connection matrix between the output layer and the reservoir, is the connection matrix of neurons in the reserve pool, The readout connection weight matrix for the forward connection reservoir and the output layer; Step 2.2: Construct the RNN layer and convert the output sequence As input to RNN , corresponding to the hidden state , …, , output , …, , the operation process can be expressed as: ； in, is the output sequence of RNN, U, W, V, b, c are all parameters, and represents the nonlinear hidden layer activation function; Step 2.3: Connect the first convolutional layer and the second convolutional layer in parallel after the RNN layer, and then connect the first pooling layer and the second pooling layer after the two convolutional layers respectively. Then merge the outputs of the two pooling layers, flatten them, and then connect another convolutional layer to output the result; S3. For the crystal material to be predicted, input the XRD spectrum data of the crystal material to be predicted into the trained XRD space group recognition model, and output the crystal structure of the material to be predicted.

2. The method for predicting crystal structure based on improved echo state network and XRD pattern according to claim 1, characterized in that: The different sources in S1 are specifically experimental data, database data, and published paper data.

3. The method for predicting crystal structure based on improved echo state network and XRD pattern according to claim 1, characterized in that: The preprocessing in S1 is specifically as follows: Step 1.1, performing the first step of data enhancement processing on the sample data set: setting a horizontal shift scaling factor a and a diffraction peak scaling scaling factor b, setting step sizes c and d to horizontally shift and vertically scale the diffraction peaks while ensuring that the spectral data is not distorted, thereby expanding the sample data set; Step 1.2: Represent the expanded sample dataset as a set of time series vectors as the original dataset D. Perform the second step of enhancement processing on the original dataset using the pre-built smote algorithm. Step 1.2.1: Filter out all minority class samples in the original dataset D to form the minority class sample set D min , the remaining samples constitute the majority class sample set D maj ; Step 1.2.2, for Minority class samples , calculate it to The distance of all samples in Neighbor, from Take any sample from the nearest neighbors and The distance between The calculation formula is as follows: ； Where: Indicates 2 Data samples of dimension attributes; Step 1.2.

3. Calculate samples and The difference between the attribute values ​​of each corresponding attribute, multiply the difference by the interval A random number within, plus the sample Each corresponding attribute value can generate a new minority class sample , the specific calculation formula is as follows: ； Where: rand Representation interval A random number on .

Images