A deep learning-based method for predicting material spin Hall conductance

Abstract

The present invention discloses a method for predicting material spin Hall conductivity based on deep learning, comprising the following steps: S1, collecting material energy band data considering the coupling effect of electron spin channels, and performing data expansion and data processing on the material energy band data to obtain a data set for spin Hall conductivity prediction; S2, dividing the data set for spin Hall conductivity prediction into a training set, a validation set and a test set; S3, constructing a deep neural network model; S4, training the deep neural network model through the training set to obtain a trained deep neural network model; S5, inputting the test set into the trained deep neural network model, and outputting the prediction result of material spin Hall conductivity. The data set used in the present invention has sufficient data volume, the obtained prediction result is relatively close to the real result, the prediction speed is relatively fast, and less computing resources are occupied.

Description

Technical Field

[0001] The present invention belongs to the technical field of material property prediction, and specifically relates to the design of a material spin Hall conductivity prediction method based on deep learning. Background Art

[0002] Among the many physical effects of materials, the Hall effect is an important electromagnetic effect. When a conductor material is subjected to a magnetic field perpendicular to the direction of the current passing through it, under the influence of the Lorentz force, the electrons in the conductor material will gather at both ends of the conductor material to form a magnetic field, so that the carriers passing behind will not be offset. The Hall effect has a wide range of applications in the manufacture of sensors. The spin Hall effect is a special and important Hall effect. By spontaneously inducing vertical directions in semiconductor materials, the conductivity of semiconductor materials can be greatly improved. Spin Hall conductivity can reduce the energy dissipation of materials to a very low level, and the devices produced have a huge driving force in the fields of computers and quantum computing.

[0003] In the field of materials science, high-throughput methods are usually used to calculate spin Hall conductance. This method is generally time-consuming, with a calculation cycle of about one week, and requires a large amount of computing resources. In recent years, with the development of deep learning technology, more and more scientists have applied it to the field of material property prediction. Based on previous experimental data, they analyze and predict the factors that affect the properties to be predicted and the laws of their changes. At present, material scientists have built deep learning neural network models to predict the band gap structure of materials, and have also used convolutional neural networks to predict the formation energy and superconducting transition temperature of compounds.

[0004] At present, there are few studies on the prediction of material spin Hall conductance using deep learning technology. According to relevant theoretical research in materials science, the energy band data of materials has a certain influence on spin Hall conductance. Therefore, a technology is needed to use the material band data to predict the spin Hall conductance, so as to shorten the time of obtaining the spin Hall conductance of the material. At the same time, it can also verify the influence of certain physical properties of the material on the spin Hall conductance, and provide a reference for the theoretical development of materials science. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the existing spin Hall conductance calculation method is time-consuming and requires large computing resources, and proposes a material spin Hall conductance prediction method based on deep learning.

[0006] The technical solution of the present invention is: a method for predicting material spin Hall conductivity based on deep learning, comprising the following steps:

[0007] S1. Collect material energy band data considering the electron spin channel coupling effect, and perform data expansion and data processing on the material energy band data to obtain a data set for spin Hall conductance prediction.

[0008] S2. Divide the data set for spin Hall conductance prediction into training set, validation set and test set.

[0009] S3. Build a deep neural network model.

[0010] S4. Train the deep neural network model using the training set to obtain a trained deep neural network model.

[0011] S5. Input the test set into the trained deep neural network model and output the predicted result of the material's spin Hall conductivity.

[0012] Furthermore, the specific method of data expansion in step S1 is: selecting five energy bands near the energy where the spin Hall conductance is located as one piece of energy band data, each piece of energy band data includes 306 points and the energy value corresponding to each point.

[0013] Furthermore, the specific method of data processing in step S1 is: when the energy band splitting degree of a certain point is greater than 0.1 eV, the energy value of the point is the sum of the lower energy value of the same point on the two energy bands of the same orbit and the weighted energy band splitting degree.

[0014] Furthermore, in step S2, the ratio of the training set, validation set, and test set is 7:2:1.

[0015] Furthermore, the deep neural network model constructed in step S3 includes an energy band feature extraction module, a time series feature extraction module, a Transformer feature extraction module and a fully connected layer which are connected in sequence.

[0016] The energy band feature extraction module is used to receive input two-dimensional energy band data, extract the physical features of each energy band, and use them as input to the time series feature extraction module.

[0017] The time series feature extraction module is used to extract the time series features of each energy band and serve as the input of the Transformer feature extraction module.

[0018] The Transformer feature extraction module is used to extract inter-band features and serve as the input of the fully connected layer.

[0019] The fully connected layer is used to output the material spin Hall conductance prediction results.

[0020] Furthermore, the number of input channels of the energy band feature extraction module is 5, the sequence length of each channel is 306, it is designed based on the residual network structure, and one-dimensional convolution is used for feature extraction. Four types of residual modules are stacked. The number of the first residual modules is 3, and the number of output channels is 256. The number of the second residual modules is 4, and the number of output channels is 512. The number of the third residual modules is 6, and the number of output channels is 1024. The number of the fourth residual modules is 3, the number of output channels is 2048, and the output sequence length is 73.

[0021] Furthermore, the input sequence length of the time series feature extraction module is 73, the number of input channels is 2048, and the basic module is designed by using a bidirectional LSTM with 2 layers and a residual method. The time series feature extraction module is composed of three basic modules stacked together.

[0022] Furthermore, the Transformer feature extraction module adopts the channel attention mechanism and the Transformer encoder to design the basic module. Three basic modules are stacked to form the Transformer feature extraction module. The output sequence length is 73 and the number of output channels is 128.

[0023] Furthermore, the output size of the fully connected layer is 1000, and the ReLU activation function is used to predict the spin Hall conductance.

[0024] Further, step S4 includes the following sub-steps:

[0025] S41. Set the mini-batch of the deep neural network model to 16, the batch-size to 64, the initial learning rate to 0.005, the number of iterations to 500, and the parameters to be randomly initialized.

[0026] S42. Use mean square error (MSE) as the loss function.

[0027] S43, input the training set and the validation set into the deep neural network model, and when the loss function value of the validation set does not decrease within a preset range, stop the training to obtain a trained deep neural network model.

[0028] The beneficial effects of the present invention are:

[0029] (1) The data set used in the present invention for predicting spin Hall conductance contains sufficient data, the deviation between the actual predicted data and the ideal predicted data is small, and the prediction accuracy is high.

[0030] (2) When predicting the spin Hall conductivity of a new material, the present invention only needs to consider the energy band data of the spin-orbit coupling effect, and the prediction speed is fast and less computing resources are occupied. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Shown is a flow chart of a material spin Hall conductivity prediction method based on deep learning provided by an embodiment of the present invention.

[0032] Figure 2 Shown is a schematic diagram of the energy band splitting phenomenon in the material energy band data provided by an embodiment of the present invention.

[0033] Figure 3 Shown is a schematic diagram of the structure of an energy band feature extraction module provided in an embodiment of the present invention.

[0034] Figure 4 FIG. 4 is a schematic diagram of the structure of a timing feature extraction module provided in an embodiment of the present invention.

[0035] Figure 5 The figure is a schematic diagram of the structure of the Transformer feature extraction module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0037] The present invention provides a method for predicting material spin Hall conductivity based on deep learning. Figure 1 As shown, the following steps S1 to S5 are included:

[0038] S1. Collect material energy band data considering the electron spin channel coupling effect, and perform data expansion and data processing on the material energy band data to obtain a data set for spin Hall conductance prediction.

[0039] In the embodiment of the present invention, the specific method of data expansion is: five energy bands near the energy of the spin Hall conductance are selected as one piece of energy band data, and each piece of energy band data includes 306 points and the energy value corresponding to each point.

[0040] In the embodiment of the present invention, the energy band data is expanded, and the original 114 energy band data are expanded to 10019 data. The energy band matrix of each data is cut from the original height 306 and width 27 to height 306 and width 5. Each energy band data corresponds to a spin Hall conductance value.

[0041] In the embodiment of the present invention, the specific method of data processing is: when the energy band splitting degree of a certain point is greater than 0.1eV, the energy value of the point is the sum of the lower energy value of the same point on the two energy bands of the same orbit and the weighted energy band splitting degree.

[0042] In the embodiment of the present invention, for the energy band data with the energy band splitting phenomenon, such as Figure 2 As shown in the figure, NoSoc is the energy band data without considering orbital coupling, and Soc is the energy band data considering orbital coupling, with splitting phenomenon. According to the different degrees of energy band splitting, it is divided into 10 intervals, namely 0.1~1.0eV, interval length is 0.1, and greater than 1.0eV. The influence of different splitting degrees on the self-selected Hall conductance is statistically analyzed, and different weights are assigned to the splitting degree in each interval. The weight coefficient of each interval is shown in Table 1. The energy value that produces the splitting phenomenon is the sum of the lower energy value at the same point on the two energy bands of the same orbit and the weighted splitting degree.

[0043] Table 1

[0044]

[0045] S2. Divide the data set for spin Hall conductance prediction into training set, validation set and test set.

[0046] In the embodiment of the present invention, the ratio of the training set, the validation set and the test set is 7:2:1. The training set contains 7013 data, the validation set contains 2004 data, and the test set contains 1002 data.

[0047] S3. Build a deep neural network model.

[0048] In an embodiment of the present invention, the deep neural network model includes an energy band feature extraction module, a time series feature extraction module, a Transformer feature extraction module and a fully connected layer which are connected in sequence.

[0049] The energy band feature extraction module is used to receive input two-dimensional energy band data, extract the physical features of each energy band, and use them as input to the time series feature extraction module.

[0050] like Figure 3 As shown, in the embodiment of the present invention, the number of input channels of the energy band feature extraction module is 5, the sequence length of each channel is 306, and one-dimensional convolution is used for feature extraction. After 3 layers of convolution extraction, the number of channels becomes 128 and the sequence length becomes 163. The energy band feature extraction module is designed based on the residual network structure, and 4 types of residual modules are stacked. The number of the first type of residual modules is 3, and the number of output channels is 256. The number of the second type of residual modules is 4, and the number of output channels is 512. The number of the third type of residual modules is 6, and the number of output channels is 1024. The number of the fourth type of residual modules is 3, and the number of output channels is 2048, followed by an average pooling layer, and the output sequence length is 73.

[0051] The time series feature extraction module is used to extract the time series features of each energy band and serve as the input of the Transformer feature extraction module.

[0052] like Figure 4 As shown, in the embodiment of the present invention, the input sequence length of the time series feature extraction module is 73, the number of input channels is 2048, and the basic module is designed by adopting a bidirectional LSTM with 2 layers and a residual method. The time series feature extraction module is composed of three basic modules stacked together. After the LSTM performs time series feature extraction, a dropout layer is connected.

[0053] The Transformer feature extraction module is used to extract inter-band features and serve as the input of the fully connected layer.

[0054] like Figure 5 As shown, in an embodiment of the present invention, the Transformer feature extraction module adopts the channel attention mechanism and the Transformer encoder design basic module, and three basic modules are stacked to form a Transformer feature extraction module. The output sequence length is 73 and the number of output channels is 128.

[0055] The fully connected layer is used to output the material spin Hall conductance prediction results.

[0056] In the embodiment of the present invention, the output size of the fully connected layer is 1000, and the ReLU activation function is used to predict the spin Hall conductance.

[0057] S4. Train the deep neural network model using the training set to obtain a trained deep neural network model.

[0058] Step S4 includes the following sub-steps S41 to S43:

[0059] S41. Set the mini-batch of the deep neural network model to 16, the batch-size to 64, the initial learning rate to 0.005, the number of iterations to 500, and the parameters to be randomly initialized.

[0060] S42. Use mean square error (MSE) as the loss function.

[0061] S43, input the training set and the validation set into the deep neural network model, and when the loss function value of the validation set does not decrease within a preset range, stop the training to obtain a trained deep neural network model.

[0062] S5. Input the test set into the trained deep neural network model and output the predicted result of the material's spin Hall conductivity.

[0063] In the embodiment of the present invention, the mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient r are used. 2As evaluation indicators, various evaluation indicators of the material spin Hall conductance prediction results obtained by inputting the test set into the trained deep neural network model are shown in Table 2. The Soc-SHC in Table 2 is the spin Hall conductance value corresponding to the energy band data considering the orbital coupling effect. It can be seen that the prediction result of the method of the present invention has a small error with the actual result.

[0064] Table 2

[0065]

[0066] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A material spin Hall conductance prediction method based on deep learning, characterized in that: The following steps are involved: S1. Collecting material energy band data considering the electron spin channel coupling effect, and performing data expansion and data processing on the material energy band data to obtain a data set for spin Hall conductance prediction; S2, dividing the data set for spin Hall conductance prediction into training set, validation set and test set; S3, build a deep neural network model; S4, training the deep neural network model through the training set to obtain a trained deep neural network model; S5. Input the test set into the trained deep neural network model, and output the predicted result of the material spin Hall conductance; The specific method of data processing in step S1 is: when the energy band splitting degree of a certain point is greater than 0.1 eV, the energy value of the point is the sum of the lower energy value of the same point on the two energy bands of the same orbit and the weighted energy band splitting degree; The deep neural network model constructed in step S3 includes an energy band feature extraction module, a time series feature extraction module, a Transformer feature extraction module and a fully connected layer connected in sequence; The energy band feature extraction module is used to receive the input two-dimensional energy band data, extract the physical features of each energy band, and use them as input to the time series feature extraction module; The time series feature extraction module is used to extract the time series features of each energy band and use them as input of the Transformer feature extraction module; The Transformer feature extraction module is used to extract inter-band features and serve as the input of the fully connected layer; The fully connected layer is used to output the material spin Hall conductance prediction result; The energy band feature extraction module has 5 input channels, and the length of each channel sequence is 306. It is designed based on the residual network structure, and uses one-dimensional convolution for feature extraction. Four types of residual modules are stacked. The number of the first type of residual modules is 3, and the number of output channels is 256. The number of the second type of residual modules is 4, and the number of output channels is 512. The number of the third type of residual modules is 6, and the number of output channels is 1024. The number of the fourth type of residual modules is 3, and the number of output channels is 2048, and the output sequence length is 73.

2. The material spin Hall conductivity prediction method according to claim 1, characterized in that: The specific method of data expansion in step S1 is: selecting five energy bands near the energy of the spin Hall conductance as one piece of energy band data, each piece of energy band data includes 306 points and the energy value corresponding to each point.

3. The material spin Hall conductance prediction method according to claim 1, characterized in that: The ratio of the training set, the validation set and the test set in step S2 is 7:2:

1.

4. The material spin Hall conductance prediction method according to claim 1, characterized in that: The input sequence length of the time series feature extraction module is 73, the number of input channels is 2048, and the basic module is designed by using a bidirectional LSTM with 2 layers and a residual method. The time series feature extraction module is composed of three basic modules stacked together.

5. The material spin Hall conductance prediction method according to claim 1, characterized in that: The Transformer feature extraction module adopts the channel attention mechanism and the Transformer encoder to design the basic module. Three basic modules are stacked to form the Transformer feature extraction module. The output sequence length is 73 and the number of output channels is 128.

6. The material spin Hall conductance prediction method according to claim 1, characterized in that: The output size of the fully connected layer is 1000, and the ReLU activation function is used to predict the spin Hall conductance.

7. The material spin Hall conductance prediction method according to claim 1, characterized in that: The step S4 comprises the following sub-steps: S41, set the mini-batch of the deep neural network model to 16, the batch-size to 64, the initial learning rate to 0.005, the number of iterations to 500, and the parameters to be randomly initialized; S42, using mean square error MSE as loss function; S43, input the training set and the validation set into the deep neural network model, and when the loss function value of the validation set does not decrease within a preset range, stop the training to obtain a trained deep neural network model.

Images