High thermoelectric figure of merit material reverse design and system based on depth generation model

Abstract

The invention relates to the field of machine learning and new material prediction, in particular to a high-thermoelectric figure of merit material reverse design method and system based on a depth generative model.The method comprises the steps that a thermoelectric material data set is obtained, temperature division operation is executed on multiple target material data, and a low-temperature material data, b medium-temperature material data and c high-temperature material data are obtained, performing corner mark interpolation operation on the a low-temperature material data, the b medium-temperature material data and the c high-temperature material data to obtain Z expansion matrix groups, constructing a generative adversarial network by using a variational auto-encoder and a authenticity identification model, and performing adversarial training on the generative adversarial network to obtain a target adversarial network, and a plurality of target material chemical formulas are generated by using a preset target zt value, a preset target temperature, the target adversarial network and the condition identification model, and reverse design of the high-thermoelectric figure of merit material is completed. According to the method, the accuracy of predicting the high-thermoelectric figure of merit material by the depth generation model can be improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning and new material prediction, and in particular to a method, system, electronic device and computer-readable storage medium for reverse design of high thermoelectric merit materials based on a deep generative model. Background Art

[0002] Based on the Seebeck effect and the Peart effect, thermoelectric materials are materials that can directly convert thermal energy into electrical energy. Compared with traditional energy conversion methods, thermoelectric materials have the advantages of high efficiency, environmental protection, reliability, and low maintenance cost.

[0003] At present, the method for predicting thermoelectric materials with high thermoelectric merit is usually to use a thermoelectric material database to train a generative adversarial network and a variational autoencoder, and then use the trained generative adversarial network and variational autoencoder to predict thermoelectric materials with high thermoelectric merit.

[0004] Although the existing generative adversarial networks and variational autoencoders can predict high thermoelectric figure of merit materials, the amount of data in the thermoelectric material database is insufficient to support the training of the model, and the number of thermoelectric materials in different temperature ranges is different. When training the model, overfitting problems often occur. At the same time, the discriminant model in the generative adversarial network only judges the authenticity of the chemical formula of the generated thermoelectric material, and does not judge the thermoelectric figure of merit and optimal temperature of the generated thermoelectric material, resulting in the generated thermoelectric material may not reach the thermoelectric figure of merit and optimal temperature required by researchers. Therefore, how to expand the data in the thermoelectric material database and how to judge the thermoelectric figure of merit and optimal temperature of the generated thermoelectric material have become urgent problems to be solved. Summary of the invention

[0005] The present invention provides a method for reverse design of high thermoelectric merit materials based on a deep generative model and a computer-readable storage medium, the main purpose of which is to improve the accuracy of deep generative models in predicting high thermoelectric merit materials.

[0006] To achieve the above object, the present invention provides a method for reverse design of high thermoelectric merit materials based on a deep generative model, comprising:

[0007] Acquire a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature, and a range of values ​​of a superscript;

[0008] Performing a chemical formula screening operation on a plurality of thermoelectric material data to obtain a plurality of target material data, performing a temperature division operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data;

[0009] Performing a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data;

[0010] Perform chemical formula quantization operation on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups;

[0011] Using Z expansion matrix groups to train the pre-constructed first discrimination model to obtain a conditional discrimination model;

[0012] The pre-constructed second identification model is trained using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder;

[0013] A generative adversarial network is constructed using a variational autoencoder and a true-false identification model, and adversarial training is performed on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder;

[0014] By using the preset target zt value, preset target temperature, target adversarial network and conditional identification model, multiple target material chemical formulas are generated to complete the reverse design of high thermoelectric merit materials.

[0015] Optionally, performing a chemical formula screening operation on a plurality of thermoelectric material data to obtain a plurality of target material data includes:

[0016] The following operations are performed on each of the multiple thermoelectric material data:

[0017] Identify the number of elements in the material chemical formula in the thermoelectric material data, and compare the number of elements with a preset first number threshold and a preset second number threshold, wherein the first number threshold is less than the second number threshold;

[0018] If the number of elements is greater than or equal to the first number threshold and the number of elements is less than or equal to the second number threshold, recording the thermoelectric material data as target material data;

[0019] The target material data are aggregated to obtain a plurality of target material data.

[0020] Optionally, performing a temperature division operation on a plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data includes:

[0021] The following operations are performed on each of the multiple target material data:

[0022] Comparing the optimal temperature corresponding to the target material data with a preset first temperature and a preset second temperature, wherein the first temperature is less than the second temperature;

[0023] If the optimal temperature is less than or equal to the first temperature, the target material data corresponding to the optimal temperature is recorded as low-temperature material data;

[0024] If the optimal temperature is greater than the first temperature and less than the second temperature, the target material data corresponding to the optimal temperature is recorded as medium-temperature material data;

[0025] If the optimal temperature is greater than or equal to the second temperature, the target material data corresponding to the optimal temperature is recorded as high-temperature material data;

[0026] The low-temperature material data, the medium-temperature material data and the high-temperature material data are summarized respectively to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data.

[0027] Optionally, performing a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data includes:

[0028] The low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio are calculated according to the preset target data quantity. The calculation formula is as follows:

[0029]

[0030] Among them, μ a , μ b and μ c are low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio respectively, a0 is the target data quantity, Represents a round-up operation;

[0031] A low-temperature interpolation ratio is used to perform a subscript interpolation operation on a low-temperature material data to obtain A low-temperature interpolation data. B medium-temperature interpolation data are obtained based on the medium-temperature interpolation ratio and b medium-temperature material data. C high-temperature interpolation data are obtained based on the high-temperature interpolation ratio and c high-temperature material data.

[0032] Optionally, performing a subscript interpolation operation on a low-temperature material data using a low-temperature interpolation ratio to obtain A low-temperature interpolation data includes:

[0033] The following operations are performed on each of the a low-temperature material data:

[0034] The low-temperature sampling interval is calculated according to the low-temperature interpolation ratio and the subscript value range corresponding to the low-temperature material data. The calculation formula is as follows:

[0035]

[0036] Among them, d a is the low temperature sampling distance, f max and f min are the maximum and minimum values ​​in the subscript value range, μ a is the low temperature interpolation ratio;

[0037] Based on the low-temperature sampling interval, a uniform sampling operation is performed on the subscript value range corresponding to the low-temperature material data to obtain m subscript values. The following operation is performed on each of the m subscript values:

[0038] Identify the subscript variable in the material chemical formula corresponding to the low-temperature material data, and update the subscript variable in the material chemical formula to the subscript value to obtain an updated chemical formula;

[0039] Integrate the maximum zt value corresponding to the updated chemical formula and low-temperature material data into low-temperature interpolation data;

[0040] The low-temperature interpolation data are summarized to obtain A low-temperature interpolation data, where A=m×a.

[0041] Optionally, the chemical formula quantization operation is performed on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups, including:

[0042] The following operation is performed on each of the A low-temperature interpolation data:

[0043] The low temperature element matrix is ​​constructed using the updated chemical formula in the low temperature interpolation data, where the low temperature element matrix is ​​as follows:

[0044]

[0045] Among them, D ea is the low temperature element matrix, E1, E2 and E n Update the atomic numbers of the first, second, and nth elements in the chemical formula, e1, e2, and e n are the subscripts of the first, second, and nth elements in the updated chemical formula, respectively, and n is the number of elements corresponding to the updated chemical formula;

[0046] The low temperature condition matrix is ​​constructed using the maximum zt value corresponding to the low temperature interpolation data. The low temperature condition matrix is ​​as follows:

[0047]

[0048] Among them, D xa is the low temperature condition matrix, X ztais the maximum zt value corresponding to the low temperature interpolation data;

[0049] Combining the low temperature element matrix and the low temperature condition matrix to obtain a low temperature matrix group, summing up the low temperature matrix groups to obtain A low temperature matrix groups;

[0050] The following operation is performed on each of the B medium-temperature interpolation data:

[0051] Obtaining a medium-temperature element matrix based on updated chemical formulas in the medium-temperature interpolation data;

[0052] The medium temperature condition matrix is ​​constructed using the maximum zt value corresponding to the medium temperature interpolation data. The medium temperature condition matrix is ​​as follows:

[0053]

[0054] Among them, D xb is the medium temperature condition matrix, X ztb is the maximum zt value corresponding to the medium temperature interpolation data;

[0055] Combining the medium temperature element matrix and the medium temperature condition matrix to obtain a medium temperature matrix group, and summing up the medium temperature matrix groups to obtain B medium temperature matrix groups;

[0056] The following operations are performed on each of the C high temperature interpolation data:

[0057] Obtaining a high temperature element matrix based on updated chemical formulas in the high temperature interpolation data;

[0058] The high temperature condition matrix is ​​constructed using the maximum zt value corresponding to the high temperature interpolation data. The high temperature condition matrix is ​​as follows:

[0059]

[0060] Among them, D xc is the high temperature condition matrix, X ztc is the maximum zt value corresponding to the high temperature interpolation data;

[0061] Combining the high temperature element matrix and the high temperature condition matrix to obtain a high temperature matrix group, and summing up the high temperature matrix groups to obtain C high temperature matrix groups;

[0062] A low-temperature matrix groups, B medium-temperature matrix groups, and C high-temperature matrix groups are summarized to obtain Z expanded matrix groups, wherein the expanded matrix group is a low-temperature matrix group, a medium-temperature matrix group, or a high-temperature matrix group, and Z=A+B+C. The expanded matrix group includes: an expanded element matrix and an expanded condition matrix. The expanded element matrix is ​​a low-temperature element matrix, a medium-temperature element matrix, or a high-temperature element matrix. The expanded condition matrix is ​​a low-temperature condition matrix, a medium-temperature condition matrix, or a high-temperature condition matrix.

[0063] Optionally, the method of training the pre-constructed first identification model using Z expansion matrix groups to obtain a conditional identification model includes:

[0064] Get the training ratio and verification ratio, where the sum of the training ratio and the verification ratio is 1;

[0065] The training quantity is calculated based on the training ratio. The calculation formula is as follows:

[0066] z1=Z×k1

[0067] Among them, z1 is the number of training, k1 is the training ratio;

[0068] The verification quantity is obtained based on the verification ratio, and z1 training matrix groups and z2 verification matrix groups are obtained based on the training quantity, the verification quantity and the Z expansion matrix groups, where z2 is the verification quantity;

[0069] Randomly select x1 training matrix groups from z1 training matrix groups, where x1<z1;

[0070] Using x1 training matrix groups to train the first identification model to obtain an updated identification model;

[0071] For each of the z2 verification matrix groups, perform the following operations:

[0072] The expanded element matrix in the verification matrix group is input into the update identification model to obtain the condition identification matrix. The condition difference degree is calculated according to the condition identification matrix and the expanded condition matrix in the verification matrix group. The calculation formula is as follows:

[0073]

[0074] Among them, δ MSE is the conditional difference, P1 and P2 are the elements of the first row and first column and the elements of the first row and second column in the extended condition matrix in the verification matrix group, Q1 and Q2 are the elements of the first row and first column and the elements of the first row and second column in the conditional discrimination matrix, tanh is the hyperbolic tangent function;

[0075] Summarize the conditional differences to obtain multiple conditional differences, and calculate the difference mean according to the multiple conditional differences, wherein the difference mean is the average value of the multiple conditional differences;

[0076] Compare the difference mean with a preset difference threshold, if the difference mean is greater than the difference threshold, update the identification model as the first identification model, and return to the step of randomly selecting x1 training matrix groups from the n1 training matrix groups until the difference mean is greater than or equal to the difference threshold;

[0077] If the difference mean is less than or equal to the difference threshold, the updated identification model is used as the conditional identification model.

[0078] Optionally, the method of using a pre-constructed variational autoencoder and Z expansion matrix groups to train a pre-constructed second identification model to obtain a true and false identification model includes:

[0079] The following operations are performed on each of the Z expanded matrix groups:

[0080] Mapping the augmented element matrix in the augmented matrix group to a potential distribution using an encoder in a variational autoencoder;

[0081] Based on the preset number of samples, random sampling is performed in the potential distribution to obtain y potential variables, where y is the number of samples;

[0082] Use the decoder in the variational autoencoder to decode y latent variables and obtain y reconstruction matrices;

[0083] Sum up y reconstruction matrices to obtain Y reconstruction matrices, where Y = Z × y;

[0084] Extracting Z extended element matrices from the Z extended matrix groups, and randomly extracting Z reconstruction matrices from the Y reconstruction matrices;

[0085] The following operation is performed on each of the Z expanded element matrices:

[0086] The expanded element matrix and the pre-constructed true judgment matrix are combined into a true judgment matrix group, where the true judgment matrix is: [1];

[0087] Summarize the true judgment matrix groups to obtain Z true judgment matrix groups;

[0088] The following operations are performed on each of the Z reconstruction matrices:

[0089] The reconstructed matrix and the pre-constructed false judgment matrix are combined into a false judgment matrix group, where the false judgment matrix is: [0];

[0090] Summarize the false judgment matrix groups to obtain Z false judgment matrix groups;

[0091] Summarize Z true-judgment matrix groups and Z false-judgment matrix groups to obtain N true-false matrix groups, where a true-false matrix group is a true-judgment matrix group or a false-judgment matrix group, and N=2×Z;

[0092] Based on the training ratio, the verification ratio and the N true and false matrix groups, n1 true and false training matrix groups and n2 true and false verification matrix groups are obtained, wherein N=n1+n2;

[0093] Acquire a judgment identification model based on n1 true and false training matrix groups and a second identification model;

[0094] The following operations are performed on each of the n2 true and false verification matrix groups:

[0095] Input the expanded element matrix or the reconstructed matrix in the true and false verification matrix group into the updated discrimination model to obtain the true and false identification matrix, obtain the discrimination value based on the true and false identification matrix, obtain the original value based on the true judgment matrix or the false judgment matrix in the true and false verification matrix group, and calculate the discrimination difference value according to the discrimination value and the original value, wherein the discrimination value is the value of the first row and first column element in the true and false identification matrix, the original value is the value of the first row and first column element in the true judgment matrix or the false judgment matrix, and the discrimination difference value is the absolute difference between the discrimination value and the original value;

[0096] Summarize the discriminant difference values ​​to obtain n2 discriminant difference values, and calculate the discriminant mean value based on the n2 discriminant difference values, wherein the discriminant mean value is the average value of the n2 discriminant mean values;

[0097] Compare the discrimination mean value with the preset discrimination threshold value, if the discrimination mean value is greater than the discrimination threshold value, update the discrimination model as the second discrimination model, and return to the step of obtaining the discrimination model based on the n1 true and false training matrix groups and the second discrimination model until the discrimination mean value is greater than or equal to the discrimination threshold value;

[0098] If the discrimination mean value is less than or equal to the discrimination threshold, the updated discrimination model is used as the true and false discrimination model.

[0099] Optionally, the method of generating multiple target material chemical formulas using a preset target zt value, a preset target temperature, a target adversarial network and a conditional identification model includes:

[0100] Determine a maximum reference zt value and a minimum reference zt value based on the target zt value and the preset reference zt value, wherein the maximum reference zt value is the sum of the target zt value and the reference zt value, and the minimum reference zt value is the absolute difference between the target zt value and the reference zt value;

[0101] Determine a reference zt range based on a maximum reference zt value and a minimum reference zt value, wherein the maximum value of the reference zt value range is the maximum reference zt value, and the minimum value of the reference zt value range is the minimum reference zt value;

[0102] Compare the target temperature with the first temperature and the second temperature. If the target temperature is less than or equal to the first temperature, the temperature value is recorded as 1, and multiple reference low temperature data are selected from the A low temperature interpolation data, wherein the maximum zt value corresponding to the reference low temperature data is within the reference zt value range;

[0103] If the target temperature is greater than the first temperature and less than the second temperature, the temperature value is recorded as 2, and a plurality of reference medium temperature data are obtained based on B medium temperature interpolation data;

[0104] If the target temperature is greater than or equal to the second temperature, the temperature value is recorded as 3, and a plurality of reference high temperature data are obtained based on C high temperature interpolation data;

[0105] The following operation is performed on each of the plurality of reference low temperature data, the plurality of reference medium temperature data or the plurality of reference high temperature data:

[0106] Obtaining a reference element matrix based on reference low temperature data, reference medium temperature data, or reference high temperature data;

[0107] Aggregating the reference element matrices to obtain multiple reference element matrices, and inputting the multiple reference element matrices into a target encoder of a target variational autoencoder in a target adversarial network to obtain multiple reference potential distributions;

[0108] For each of the multiple reference latent distributions, perform the following operations:

[0109] Obtain y reference latent variables based on the number of samples and the reference latent distribution, and use the target decoder of the target variational autoencoder in the target adversarial network to decode the y reference latent variables to obtain y reference matrices;

[0110] The reference matrices are aggregated to obtain multiple reference matrices, and the following operations are performed on each of the multiple reference matrices:

[0111] Input the reference matrix into the condition identification model to obtain the reference condition matrix, and identify the reconstructed temperature value and the reconstructed zt value of the reference condition matrix, wherein the reconstructed zt value is the value of the element in the first row and first column of the reference condition matrix, and the reconstructed temperature value is the value of the element in the first row and second column of the reference condition matrix;

[0112] The temperature difference is calculated according to the reconstructed temperature value and the temperature value corresponding to the reference matrix, and the reconstructed zt difference is calculated according to the reconstructed zt value and the target zt value, wherein the temperature difference is the absolute difference between the reconstructed temperature value and the reference matrix, and the reconstructed zt difference is the absolute difference between the reconstructed zt value and the target zt value;

[0113] Determine whether the temperature difference is less than a preset temperature comparison threshold, and determine whether the reconstructed zt difference is less than a preset zt comparison threshold;

[0114] If the temperature difference is less than the temperature comparison threshold and the reconstructed zt difference is less than the zt comparison threshold, the reference matrix is ​​recorded as the final matrix, and the chemical formula of the target material is restored according to the final matrix;

[0115] The target material chemical formulas are summarized to obtain multiple target material chemical formulas.

[0116] To achieve the above object, the present invention also provides a reverse design system for high thermoelectric merit materials based on a deep generative model, comprising:

[0117] A material data partitioning module is used to obtain a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature and a range of angle values, a chemical formula screening operation is performed on the plurality of thermoelectric material data to obtain a plurality of target material data, and a temperature partitioning operation is performed on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data;

[0118] A material data expansion module is used to perform a subscript interpolation operation on a low-temperature material data, b medium-temperature material data and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data and C high-temperature interpolation data, and perform a chemical formula quantization operation on the A low-temperature interpolation data, the B medium-temperature interpolation data and the C high-temperature interpolation data to obtain Z expansion matrix groups;

[0119] A generative model training module is used to train a pre-constructed first identification model using Z expansion matrix groups to obtain a conditional identification model, and to train a pre-constructed second identification model using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder, and to construct a generative adversarial network using the variational autoencoder and the true and false identification model, and to perform adversarial training on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder;

[0120] Thermoelectric material generation module is used to generate multiple target material chemical formulas using preset target zt value, preset target temperature, target adversarial network and condition identification model to complete the reverse design of high thermoelectric merit materials.

[0121] In order to solve the above problem, the present invention further provides an electronic device, the electronic device comprising:

[0122] a memory storing at least one instruction; and

[0123] A processor executes instructions stored in the memory to implement the above-mentioned high thermoelectric figure of merit material reverse design method based on deep generative model.

[0124] In order to solve the above problems, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored, and the at least one instruction is executed by a processor in an electronic device to implement the above-mentioned high thermoelectric figure of merit material reverse design method based on deep generative model.

[0125] The present invention is to solve the problem described in the background technology. The present invention obtains a thermoelectric material data set, wherein the thermoelectric material data set includes: multiple thermoelectric material data, and the thermoelectric material data includes: material chemical formula, maximum zt value, optimal temperature and superscript value range, performs a chemical formula screening operation on the multiple thermoelectric material data to obtain multiple target material data, performs a temperature division operation on the multiple target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data, performs a superscript interpolation operation on a low-temperature material data, b medium-temperature material data and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data and C high-temperature interpolation data. It can be seen that the embodiment of the present invention performs a plurality of thermoelectric materials by screening the multiple thermoelectric materials. The data is subjected to a chemical formula screening operation to screen out thermoelectric material data corresponding to the chemical formula of materials with too many or too few elements, thereby preventing the chemical formula of materials with too many or too few individual elements from affecting the training of subsequent models, and then performing a temperature division operation and a superscript interpolation operation on multiple target material data, so that the data amounts of low-temperature interpolation data, medium-temperature interpolation data, and high-temperature interpolation data are basically consistent, thereby preventing the subsequent trained conditional identification model from over-fitting the medium-temperature zone with more data amount, while ignoring the low-temperature zone and the high-temperature zone. A chemical formula quantization operation is performed on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups. It can be seen that the embodiment of the present invention divides A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data into Z expansion matrix groups. The temperature interpolation data and C high-temperature interpolation data are converted into a matrix form, which is convenient for the subsequent training of the first identification model and the second identification model. The pre-constructed first identification model is trained using Z expansion matrix groups to obtain a conditional identification model. It can be seen that the embodiment of the present invention trains a conditional identification model to facilitate the subsequent auxiliary target adversarial network to distinguish the generated target material chemical formula, thereby improving the accuracy of the deep generation model in predicting high thermoelectric figure of merit materials. The pre-constructed variational autoencoder and Z expansion matrix groups are used to train the pre-constructed second identification model to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder. It can be seen that the embodiment of the present invention trains a true and false identification model through Z expansion matrix groups, thereby improving the true and false identification. The model is accurate in discriminating the chemical formula of the material generated by the generative model, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. A generative adversarial network is constructed using a variational autoencoder and a true and false identification model, and the generative adversarial network is adversarially trained to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder. It can be seen that the embodiment of the present invention combines the variational autoencoder and the generative adversarial network to train a target adversarial network, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. A plurality of target material chemical formulas are generated using a preset target zt value, a preset target temperature, a target adversarial network and a conditional identification model.The reverse design of high thermoelectric merit material is completed. It can be seen that the embodiment of the present invention uses the conditional identification model to assist the target adversarial network to generate the target material chemical formula that meets the preset target zt value and the preset target temperature, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. Therefore, the present invention can improve the accuracy of the deep generative model in predicting high thermoelectric merit materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] Figure 1 A schematic diagram of a process for reverse design of high thermoelectric merit material based on a deep generative model provided in one embodiment of the present invention;

[0127] Figure 2 A functional module diagram of a reverse design system for high thermoelectric merit-of-merit materials based on a deep generative model provided by an embodiment of the present invention;

[0128] Figure 3 A schematic diagram of the structure of an electronic device for implementing the method for reverse design of high thermoelectric figure of merit materials based on a deep generative model provided in one embodiment of the present invention.

[0129] Description of reference numerals:

[0130] 1. Electronic device; 10. Processor; 11. Memory; 12. Bus.

[0131] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0132] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0133] The embodiment of the present application provides a method for reverse design of high thermoelectric merit materials based on a deep generative model. The execution subject of the method for reverse design of high thermoelectric merit materials based on a deep generative model includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided in the embodiment of the present application. In other words, the method for reverse design of high thermoelectric merit materials based on a deep generative model can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0134] Reference Figure 1 FIG. 1 is a flow chart of a method for reverse design of high thermoelectric merit-value materials based on a deep generative model according to an embodiment of the present invention. In this embodiment, the method for reverse design of high thermoelectric merit-value materials based on a deep generative model includes:

[0135] S1. Acquire a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature, and a superscript value range.

[0136] In the embodiment of the present invention, the thermoelectric material data set is obtained from the MRL Thermoelectric Database, which is a public database developed by the Materials Research Laboratory (MRL) of the University of Illinois at Urbana-Champaign (UIUC). The database focuses on the data collection and organization of thermoelectric materials, providing researchers with a systematic resource for evaluating and developing efficient thermoelectric materials.

[0137] In an embodiment of the present invention, the deep generative model refers to a deep learning model that combines a variational autoencoder and a generative adversarial network.

[0138] It should be explained that the material chemical formula refers to the chemical formula of the thermoelectric material. Since the thermoelectric figure of merit of the thermoelectric material changes with temperature, the maximum zt value refers to the maximum thermoelectric figure of merit that the thermoelectric material can achieve when the temperature is optional, and the optimal temperature is the temperature at which the thermoelectric material reaches the maximum zt value.

[0139] It should be understood that most of the thermoelectric material data in the MRL thermoelectric material database include the material chemical formula, maximum zt value, optimal temperature and subscript value range of the thermoelectric material. In order to ensure the uniformity of data in subsequent training models, when obtaining thermoelectric material data from the MRL thermoelectric material database, the embodiment of the present invention only selects the thermoelectric material data whose subscripts of the element symbols in the material chemical formula contain unknown numbers and whose value ranges of the unknown numbers are clearly marked in the MRL thermoelectric material database. For example, the material chemical formula of a thermoelectric material is ZnNiSn 1-x Sb x , and ZnNiSn is given in the MRL thermoelectric material database 1-x Sb x The value range of x is [0.1, 0.2], then the thermoelectric material data corresponding to the thermoelectric material is the thermoelectric material data required by the embodiment of the present invention, and the value range of x is the subscript value range.

[0140] S2. Perform a chemical formula screening operation on multiple thermoelectric material data to obtain multiple target material data, and perform a temperature division operation on the multiple target material data to obtain a low-temperature material data, b medium-temperature material data, and c high-temperature material data.

[0141] In detail, the chemical formula screening operation is performed on a plurality of thermoelectric material data to obtain a plurality of target material data, including:

[0142] The following operations are performed on each of the multiple thermoelectric material data:

[0143] Identify the number of elements in the material chemical formula in the thermoelectric material data, and compare the number of elements with a preset first number threshold and a preset second number threshold, wherein the first number threshold is less than the second number threshold;

[0144] If the number of elements is greater than or equal to the first number threshold and the number of elements is less than or equal to the second number threshold, recording the thermoelectric material data as target material data;

[0145] The target material data are aggregated to obtain a plurality of target material data.

[0146] It should be explained that the number of elements refers to the number of types of element symbols in the material chemical formula. For example, the material chemical formula is ZnNiSn 1-x Sb x Since the material chemical formula contains four element symbols: Zn, Ni, Sn and Sb, the number of elements is 4. Optionally, the first number threshold is set to 3 and the second number threshold is set to 5.

[0147] It should be understood that the embodiment of the present invention screens out thermoelectric material data corresponding to material chemical formulas with too many or too few elements, thereby preventing material chemical formulas with too many or too few individual elements from affecting subsequent model training.

[0148] In detail, performing a temperature division operation on a plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data includes:

[0149] The following operations are performed on each of the multiple target material data:

[0150] Comparing the optimal temperature corresponding to the target material data with a preset first temperature and a preset second temperature, wherein the first temperature is less than the second temperature;

[0151] If the optimal temperature is less than or equal to the first temperature, the target material data corresponding to the optimal temperature is recorded as low-temperature material data;

[0152] If the optimal temperature is greater than the first temperature and less than the second temperature, the target material data corresponding to the optimal temperature is recorded as medium-temperature material data;

[0153] If the optimal temperature is greater than or equal to the second temperature, the target material data corresponding to the optimal temperature is recorded as high-temperature material data;

[0154] The low-temperature material data, the medium-temperature material data and the high-temperature material data are summarized respectively to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data.

[0155] Optionally, the first temperature is set to 500 degrees Celsius and the second temperature is set to 1000 degrees Celsius.

[0156] It should be understood that the optimal temperatures of most thermoelectric material data are in the medium temperature zone (500℃-1000℃), while there are fewer thermoelectric material data corresponding to the low temperature zone (below 500℃) and the high temperature zone (above 1000℃). Therefore, the embodiment of the present invention divides the target material data into low-temperature material data, medium-temperature material data and high-temperature material data, and subsequently performs subscript interpolation operations on the low-temperature material data, medium-temperature material data and high-temperature material data, so that the data volume of the low-temperature interpolation data, the medium-temperature interpolation data and the high-temperature interpolation data is basically the same, thereby preventing the conditional identification model trained subsequently from over-fitting the medium temperature zone with more data volume and ignoring the low temperature zone and the high temperature zone.

[0157] It should be explained that a is the total number of low-temperature material data, b is the total number of medium-temperature material data, and c is the total number of high-temperature material data.

[0158] S3. Perform a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data.

[0159] In detail, the subscript interpolation operation is performed on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data, including:

[0160] The low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio are calculated according to the preset target data quantity. The calculation formula is as follows:

[0161]

[0162] Among them, μ a , μ b and μ c are low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio respectively, a0 is the target data quantity, Represents a round-up operation;

[0163] A low-temperature interpolation ratio is used to perform a subscript interpolation operation on a low-temperature material data to obtain A low-temperature interpolation data. B medium-temperature interpolation data are obtained based on the medium-temperature interpolation ratio and b medium-temperature material data. C high-temperature interpolation data are obtained based on the high-temperature interpolation ratio and c high-temperature material data.

[0164] It should be explained that the target data quantity is related to the quantity of data required for training, and the target data quantity is greater than the total quantity of low-temperature material data, medium-temperature material data and high-temperature material data. Optionally, the target data quantity is set to 2000.

[0165] It should be understood that the method of obtaining B medium-temperature interpolation data based on the medium-temperature interpolation ratio and b medium-temperature material data and the method of obtaining C high-temperature interpolation data based on the high-temperature interpolation ratio and c high-temperature material data are the same as the method of obtaining A low-temperature interpolation data using the low-temperature interpolation ratio and a low-temperature material data, and will not be repeated here.

[0166] In detail, the method of performing a subscript interpolation operation on a low-temperature material data using a low-temperature interpolation ratio to obtain A low-temperature interpolation data includes:

[0167] The following operations are performed on each of the a low-temperature material data:

[0168] The low-temperature sampling interval is calculated according to the low-temperature interpolation ratio and the subscript value range corresponding to the low-temperature material data. The calculation formula is as follows:

[0169]

[0170] Among them, d a is the low temperature sampling distance, f max and f min are the maximum and minimum values ​​in the subscript value range, μ a is the low temperature interpolation ratio;

[0171] Based on the low-temperature sampling interval, a uniform sampling operation is performed on the subscript value range corresponding to the low-temperature material data to obtain m subscript values. The following operation is performed on each of the m subscript values:

[0172] Identify the subscript variable in the material chemical formula corresponding to the low-temperature material data, and update the subscript variable in the material chemical formula to the subscript value to obtain an updated chemical formula;

[0173] Integrate the maximum zt value corresponding to the updated chemical formula and low-temperature material data into low-temperature interpolation data;

[0174] The low-temperature interpolation data are summarized to obtain A low-temperature interpolation data, where A=m×a.

[0175] For example, if the low temperature interpolation ratio is 6 and the subscript value range is [0.1, 0.2], the calculated low temperature sampling interval is 0.02, and the subscript value range is uniformly sampled to obtain 6 subscript values ​​of 0.1, 0.12, 0.14, 0.16, 0.18, and 0.2. If the material chemical formula is ZnNiSn 1-x Sb x , then the subscript variable is ZnNiSn 1-x Sb x After updating x to a subscript value of 0.1, the updated chemical formula is ZnNiSn 0.9 Sb 0.1 .

[0176] S4. Perform chemical formula quantization operation on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups.

[0177] In detail, the chemical formula quantization operation is performed on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups, including:

[0178] The following operation is performed on each of the A low-temperature interpolation data:

[0179] The low temperature element matrix is ​​constructed using the updated chemical formula in the low temperature interpolation data, where the low temperature element matrix is ​​as follows:

[0180]

[0181] Among them, D ea is the low temperature element matrix, E1, E2 and E n Update the atomic numbers of the first, second, and nth elements in the chemical formula, e1, e2, and e n are the subscripts of the first, second, and nth elements in the updated chemical formula, respectively, and n is the number of elements corresponding to the updated chemical formula;

[0182] The low temperature condition matrix is ​​constructed using the maximum zt value corresponding to the low temperature interpolation data. The low temperature condition matrix is ​​as follows:

[0183]

[0184] Among them, D xa is the low temperature condition matrix, X zta is the maximum zt value corresponding to the low temperature interpolation data;

[0185] Combining the low temperature element matrix and the low temperature condition matrix to obtain a low temperature matrix group, summing up the low temperature matrix groups to obtain A low temperature matrix groups;

[0186] The following operation is performed on each of the B medium-temperature interpolation data:

[0187] Obtaining a medium-temperature element matrix based on updated chemical formulas in the medium-temperature interpolation data;

[0188] The medium temperature condition matrix is ​​constructed using the maximum zt value corresponding to the medium temperature interpolation data. The medium temperature condition matrix is ​​as follows:

[0189]

[0190] Among them, D xb is the medium temperature condition matrix, X ztb is the maximum zt value corresponding to the medium temperature interpolation data;

[0191] Combining the medium temperature element matrix and the medium temperature condition matrix to obtain a medium temperature matrix group, and summing up the medium temperature matrix groups to obtain B medium temperature matrix groups;

[0192] The following operations are performed on each of the C high temperature interpolation data:

[0193] Obtaining a high temperature element matrix based on updated chemical formulas in the high temperature interpolation data;

[0194] The high temperature condition matrix is ​​constructed using the maximum zt value corresponding to the high temperature interpolation data. The high temperature condition matrix is ​​as follows:

[0195]

[0196] Among them, D xc is the high temperature condition matrix, X ztc is the maximum zt value corresponding to the high temperature interpolation data;

[0197] Combining the high temperature element matrix and the high temperature condition matrix to obtain a high temperature matrix group, and summing up the high temperature matrix groups to obtain C high temperature matrix groups;

[0198] A low-temperature matrix groups, B medium-temperature matrix groups, and C high-temperature matrix groups are summarized to obtain Z expanded matrix groups, wherein the expanded matrix group is a low-temperature matrix group, a medium-temperature matrix group, or a high-temperature matrix group, and Z=A+B+C. The expanded matrix group includes: an expanded element matrix and an expanded condition matrix. The expanded element matrix is ​​a low-temperature element matrix, a medium-temperature element matrix, or a high-temperature element matrix. The expanded condition matrix is ​​a low-temperature condition matrix, a medium-temperature condition matrix, or a high-temperature condition matrix.

[0199] It should be understood that when an element in the updated chemical formula has no subscript, the subscript of the element is assumed to be 1.

[0200] For example, if the updated chemical formula in the low temperature interpolation data is ZnNiSn 0.9 Sb 0.1, since the atomic numbers of Zn, Ni, Sn and Sb are 30, 28, 50 and 51 respectively, the low temperature element matrix is:

[0201]

[0202] If the maximum zt value corresponding to the low temperature interpolation data is 1.3, the low temperature condition matrix is:

[0203]

[0204] The same is true for the medium temperature condition matrix and the high temperature condition matrix.

[0205] It can be understood that the method of obtaining the medium-temperature element matrix based on the updated chemical formula in the medium-temperature interpolation data and the method of obtaining the high-temperature element matrix based on the updated chemical formula in the high-temperature interpolation data are the same as the method of constructing the low-temperature element matrix using the updated chemical formula in the low-temperature interpolation data, and will not be repeated here.

[0206] S5. Use Z expansion matrix groups to train the pre-constructed first identification model to obtain a conditional identification model.

[0207] In detail, the method of using Z expansion matrix groups to train the pre-constructed first identification model to obtain a conditional identification model includes:

[0208] Get the training ratio and verification ratio, where the sum of the training ratio and the verification ratio is 1;

[0209] The training quantity is calculated based on the training ratio. The calculation formula is as follows:

[0210] z1=Z×k1

[0211] Among them, z1 is the number of training, k1 is the training ratio;

[0212] The verification quantity is obtained based on the verification ratio, and z1 training matrix groups and z2 verification matrix groups are obtained based on the training quantity, the verification quantity and the Z expansion matrix groups, where z2 is the verification quantity;

[0213] Randomly select x1 training matrix groups from z1 training matrix groups, where x1<z1;

[0214] Using x1 training matrix groups to train the first identification model to obtain an updated identification model;

[0215] For each of the z2 verification matrix groups, perform the following operations:

[0216] The expanded element matrix in the verification matrix group is input into the update identification model to obtain the condition identification matrix. The condition difference degree is calculated according to the condition identification matrix and the expanded condition matrix in the verification matrix group. The calculation formula is as follows:

[0217]

[0218] Among them, δ MSE is the conditional difference, P1 and P2 are the elements of the first row and first column and the elements of the first row and second column in the extended condition matrix in the verification matrix group, Q1 and Q2 are the elements of the first row and first column and the elements of the first row and second column in the conditional discrimination matrix, tanh is the hyperbolic tangent function;

[0219] Summarize the conditional differences to obtain multiple conditional differences, and calculate the difference mean according to the multiple conditional differences, wherein the difference mean is the average value of the multiple conditional differences;

[0220] Compare the difference mean with the preset difference threshold, if the difference mean is greater than the difference threshold, update the identification model as the first identification model, and return to the step of randomly selecting x1 training matrix groups from z1 training matrix groups until the difference mean is greater than or equal to the difference threshold;

[0221] If the difference mean is less than or equal to the difference threshold, the updated identification model is used as the conditional identification model.

[0222] It should be explained that the training ratio and the verification ratio are set by the trainer of the first identification model based on experience. Optionally, the training ratio is 70% and the verification ratio is 30%.

[0223] It should be understood that the method for obtaining the verification quantity based on the verification ratio is the same as the method for calculating the training quantity based on the training ratio, and will not be repeated here.

[0224] Exemplarily, the number of training is 700, the number of verification is 300, and there are 1000 expansion matrix groups in total. Then, 700 expansion matrix groups are randomly extracted from the 1000 expansion matrix groups as 700 training matrix groups, and the remaining 300 expansion matrix groups after the extraction of the 1000 expansion matrix groups are used as 300 verification matrix groups.

[0225] It should be understood that the first identification model is a fully connected neural network, and the use of x1 training matrix groups to train the first identification model means: the expanded element matrix of each training matrix group in the x1 training matrix groups is input into the first identification model in turn, and the first identification model will output an output matrix. If the output matrix is ​​not an expanded condition matrix in the training matrix group, the output matrix is ​​corrected using the expanded condition matrix, so that the first identification model gradually learns the mapping relationship between the expanded element matrix and the expanded condition matrix, and the technology of using x1 training matrix groups to train the first identification model is a prior art and will not be repeated here. The updated identification model is the first identification model after training with x1 training matrix groups.

[0226] It is understandable that after the extended element matrix is ​​input into the update identification model, the update identification model will automatically output a conditional identification matrix.

[0227] Exemplarily, the expanded condition matrix is The conditional discriminant matrix is Then P1 and P2 are 1.3 and 3 respectively, and Q1 and Q2 are 1.6 and 2 respectively.

[0228] It is understandable that the condition difference reflects the degree of difference between the condition discrimination matrix and the expanded condition matrix. The greater the condition difference, the greater the degree of difference between the condition discrimination matrix and the expanded condition matrix, indicating that the training effect of the updated discrimination model is worse, so it is necessary to return to the step of randomly selecting x1 training matrix groups from z1 training matrix groups to train the updated discrimination model again. Optionally, the difference threshold is set to 0.1.

[0229] S6. Use the pre-constructed variational autoencoder and Z expansion matrix groups to train the pre-constructed second identification model to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder.

[0230] In detail, the pre-constructed second identification model is trained using the pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, including:

[0231] The following operations are performed on each of the Z expanded matrix groups:

[0232] Mapping the augmented element matrix in the augmented matrix group to a potential distribution using an encoder in a variational autoencoder;

[0233] Based on the preset number of samples, random sampling is performed in the potential distribution to obtain y potential variables, where y is the number of samples;

[0234] Use the decoder in the variational autoencoder to decode y latent variables and obtain y reconstruction matrices;

[0235] Sum up y reconstruction matrices to obtain Y reconstruction matrices, where Y = Z × y;

[0236] Extracting Z extended element matrices from the Z extended matrix groups, and randomly extracting Z reconstruction matrices from the Y reconstruction matrices;

[0237] The following operation is performed on each of the Z expanded element matrices:

[0238] The expanded element matrix and the pre-constructed true judgment matrix are combined into a true judgment matrix group, where the true judgment matrix is: [1];

[0239] Summarize the true judgment matrix groups to obtain Z true judgment matrix groups;

[0240] The following operations are performed on each of the Z reconstruction matrices:

[0241] The reconstructed matrix and the pre-constructed false judgment matrix are combined into a false judgment matrix group, where the false judgment matrix is: [0];

[0242] Summarize the false judgment matrix groups to obtain Z false judgment matrix groups;

[0243] Summarize Z true-judgment matrix groups and Z false-judgment matrix groups to obtain N true-false matrix groups, where a true-false matrix group is a true-judgment matrix group or a false-judgment matrix group, and N=2×Z;

[0244] Based on the training ratio, the verification ratio and the N true and false matrix groups, n1 true and false training matrix groups and n2 true and false verification matrix groups are obtained, wherein N=n1+n2;

[0245] Acquire a judgment identification model based on n1 true and false training matrix groups and a second identification model;

[0246] The following operations are performed on each of the n2 true and false verification matrix groups:

[0247] Input the expanded element matrix or the reconstructed matrix in the true and false verification matrix group into the updated discrimination model to obtain the true and false identification matrix, obtain the discrimination value based on the true and false identification matrix, obtain the original value based on the true judgment matrix or the false judgment matrix in the true and false verification matrix group, and calculate the discrimination difference value according to the discrimination value and the original value, wherein the discrimination value is the value of the first row and first column element in the true and false identification matrix, the original value is the value of the first row and first column element in the true judgment matrix or the false judgment matrix, and the discrimination difference value is the absolute difference between the discrimination value and the original value;

[0248] Summarize the discriminant difference values ​​to obtain n2 discriminant difference values, and calculate the discriminant mean value based on the n2 discriminant difference values, wherein the discriminant mean value is the average value of the n2 discriminant mean values;

[0249] Compare the discrimination mean value with the preset discrimination threshold value, if the discrimination mean value is greater than the discrimination threshold value, update the discrimination model as the second discrimination model, and return to the step of obtaining the discrimination model based on the n1 true and false training matrix groups and the second discrimination model until the discrimination mean value is greater than or equal to the discrimination threshold value;

[0250] If the discrimination mean value is less than or equal to the discrimination threshold, the updated discrimination model is used as the true and false discrimination model.

[0251] It should be explained that the variational autoencoder is a generative model, and the variational autoencoder includes an encoder and a decoder. Both the encoder and the decoder are fully connected neural networks, and the main function of the encoder is to map the extended element matrix in the extended matrix group to a potential distribution, and the main function of the decoder is to map the latent variables in the latent distribution back to the form of a matrix. The latent distribution is a normal distribution, and the latent variables are values ​​randomly sampled from the latent distribution. Random sampling in the latent distribution means: randomly selecting one or more values ​​from the latent distribution, and the probability of random selection follows the probability density function of the latent distribution. For example, the randomly selected value is likely to be near the mean of the normal distribution corresponding to the latent distribution.

[0252] It should be understood that the use of the decoder in the variational autoencoder to decode y potential variables means: the y potential variables are sequentially input into the decoder in the variational autoencoder, and the decoder in the variational autoencoder automatically maps the y potential variables into a matrix form, and the reconstruction matrix is ​​the matrix generated by the decoder according to the potential variables. The technology of mapping the extended element matrix in the extended matrix group to a potential distribution using the encoder in the variational autoencoder and the technology of decoding y potential variables using the decoder in the variational autoencoder to obtain y reconstruction matrices are both prior art and will not be repeated here.

[0253] It is understandable that the method of obtaining n1 true and false training matrix groups and n2 true and false verification matrix groups based on the training ratio, verification ratio and N true and false matrix groups is the same as the method of obtaining z1 training matrix groups and z2 verification matrix groups based on the training ratio, verification ratio and Z expansion matrix groups, which will not be repeated here. The method of obtaining the judgment identification model based on n1 true and false training matrix groups and the second identification model is the same as the method of training the first identification model using z1 training matrix groups to obtain an updated identification model, which will not be repeated here.

[0254] Exemplarily, the expanded element matrix in the true and false verification matrix group After inputting into the updated discriminant model, the updated discriminant model will automatically output a true and false identification matrix. If the true and false identification matrix is ​​[0.9], and the true matrix in the true and false verification matrix group is [1], the discriminant value is 0.9, the original value is 1, and the discriminant difference value is 0.1. Optionally, the discrimination threshold is set to 0.1.

[0255] S7. Use the variational autoencoder and the true and false identification model to construct a generative adversarial network, and perform adversarial training on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder.

[0256] It should be explained that the generative adversarial network is a deep learning model composed of a generative model and a discriminative model, and in an embodiment of the present invention, a variational autoencoder is used as the generative model of the generative adversarial network, and a true and false identification model is used as the discriminative model of the generative adversarial network.

[0257] It can be understood that the process of adversarial training of the generative adversarial network is as follows: multiple extended element matrices are input into the generative model of the generative adversarial network, the encoder in the generative model will generate multiple potential distributions based on the multiple extended element matrices, and the decoder in the generative model will randomly sample multiple latent variables from the multiple potential distributions, and generate multiple reconstruction matrices based on the multiple latent variables, and then randomly sample multiple extended element matrices from the multiple extended element matrices, and input the multiple reconstruction matrices and the multiple extended element matrices into the discriminant model. The discriminant model determines which of the matrices input into the discriminant model are false data generated by the generative model and which are real data from the multiple extended element matrices, and outputs the discrimination result. If the discrimination result is not equal to the true result, the discriminant model will The discriminant model is automatically corrected based on the real results, so that it can gradually distinguish which are the false data generated by the generative model and which are the real data from multiple extended element matrices. At the same time, at this stage, the generative model will also correct the generative model according to the discrimination results output by the discriminant model, so that the false data generated by the generative model is closer and closer to the real data from multiple extended element matrices. Finally, through the mutual confrontation between the generative model and the discriminant model, the generation ability of the generative model and the discrimination ability of the discriminant model will become stronger and stronger. Finally, when the generative model and the discriminant model converge (that is, after the discriminant model and the generative model are corrected using the real results and the discrimination results, the discrimination ability and generation ability of the discriminant model and the generative model almost no longer change), the adversarial training of the generative adversarial network is completed. The target adversarial network is the generative adversarial network after the adversarial training is completed. And the technology for adversarial training of the generative adversarial network is a prior art, which will not be repeated here.

[0258] It should be explained that the target variational autoencoder is the variational autoencoder in the target adversarial network after the adversarial training is completed, and the target encoder and target decoder are respectively the encoder and decoder in the target adversarial network after the adversarial training is completed.

[0259] S8. Use the preset target zt value, preset target temperature, target adversarial network and conditional identification model to generate multiple target material chemical formulas to complete the reverse design of high thermoelectric merit materials.

[0260] In detail, the method of generating multiple target material chemical formulas using a preset target zt value, a preset target temperature, a target adversarial network and a conditional identification model includes:

[0261] Determine a maximum reference zt value and a minimum reference zt value based on the target zt value and the preset reference zt value, wherein the maximum reference zt value is the sum of the target zt value and the reference zt value, and the minimum reference zt value is the absolute difference between the target zt value and the reference zt value;

[0262] Determine a reference zt range based on a maximum reference zt value and a minimum reference zt value, wherein the maximum value of the reference zt value range is the maximum reference zt value, and the minimum value of the reference zt value range is the minimum reference zt value;

[0263] Compare the target temperature with the first temperature and the second temperature. If the target temperature is less than or equal to the first temperature, the temperature value is recorded as 1, and multiple reference low temperature data are selected from the A low temperature interpolation data, wherein the maximum zt value corresponding to the reference low temperature data is within the reference zt value range;

[0264] If the target temperature is greater than the first temperature and less than the second temperature, the temperature value is recorded as 2, and a plurality of reference medium temperature data are obtained based on B medium temperature interpolation data;

[0265] If the target temperature is greater than or equal to the second temperature, the temperature value is recorded as 3, and a plurality of reference high temperature data are obtained based on C high temperature interpolation data;

[0266] The following operation is performed on each of the plurality of reference low temperature data, the plurality of reference medium temperature data or the plurality of reference high temperature data:

[0267] Obtaining a reference element matrix based on reference low temperature data, reference medium temperature data, or reference high temperature data;

[0268] Aggregating the reference element matrices to obtain multiple reference element matrices, and inputting the multiple reference element matrices into a target encoder of a target variational autoencoder in a target adversarial network to obtain multiple reference potential distributions;

[0269] For each of the multiple reference latent distributions, perform the following operations:

[0270] Obtain y reference latent variables based on the number of samples and the reference latent distribution, and use the target decoder of the target variational autoencoder in the target adversarial network to decode the y reference latent variables to obtain y reference matrices;

[0271] The reference matrices are aggregated to obtain multiple reference matrices, and the following operations are performed on each of the multiple reference matrices:

[0272] Input the reference matrix into the condition identification model to obtain the reference condition matrix, and identify the reconstructed temperature value and the reconstructed zt value of the reference condition matrix, wherein the reconstructed zt value is the value of the element in the first row and first column of the reference condition matrix, and the reconstructed temperature value is the value of the element in the first row and second column of the reference condition matrix;

[0273] The temperature difference is calculated according to the reconstructed temperature value and the temperature value corresponding to the reference matrix, and the reconstructed zt difference is calculated according to the reconstructed zt value and the target zt value, wherein the temperature difference is the absolute difference between the reconstructed temperature value and the reference matrix, and the reconstructed zt difference is the absolute difference between the reconstructed zt value and the target zt value;

[0274] Determine whether the temperature difference is less than a preset temperature comparison threshold, and determine whether the reconstructed zt difference is less than a preset zt comparison threshold;

[0275] If the temperature difference is less than the temperature comparison threshold and the reconstructed zt difference is less than the zt comparison threshold, the reference matrix is ​​recorded as the final matrix, and the chemical formula of the target material is restored according to the final matrix;

[0276] The target material chemical formulas are summarized to obtain multiple target material chemical formulas.

[0277] For example, Xiao Zhang is a researcher in the thermoelectric material research laboratory. Now Xiao Zhang hopes to predict a thermoelectric material with an optimal temperature of 600°C and a maximum zt value of 1.3, so 600°C is used as the target temperature and 1.3 is used as the target zt value. Optionally, the reference zt value is set to 0.1.

[0278] It should be understood that, since each of the A low-temperature interpolation data corresponds to a maximum zt value, the screening out of multiple reference low-temperature data from the A low-temperature interpolation data means: screening out multiple low-temperature interpolation data whose maximum zt value among the A low-temperature interpolation data is within the reference zt value range, and using the multiple low-temperature interpolation data as multiple reference low-temperature data.

[0279] It is understandable that the method of obtaining multiple reference medium-temperature data based on B medium-temperature interpolation data and the method of obtaining multiple reference high-temperature data based on C high-temperature interpolation data are the same as the method of obtaining multiple reference low-temperature data using A low-temperature interpolation data, and will not be repeated here. The method of obtaining a reference element matrix based on reference low-temperature data, reference medium-temperature data or reference high-temperature data is the same as the method of constructing a low-temperature element matrix using the updated chemical formula in the low-temperature interpolation data, and will not be repeated here.

[0280] It should be understood that the inputting of multiple reference element matrices into the target encoder of the target variational autoencoder in the target adversarial network to obtain multiple reference potential distributions means: inputting each of the multiple reference element matrices into the target encoder of the target variational autoencoder in the target adversarial network in turn, using the target encoder to map the reference element matrix to a reference potential distribution, and summarizing the reference potential distribution to obtain multiple reference potential distributions. And the method of mapping the reference element matrix to the reference potential distribution using the target encoder is the same as the method of mapping the extended element matrix in the extended matrix group to the potential distribution using the encoder in the variational autoencoder, which will not be repeated here.

[0281] It can be understood that the method of obtaining y reference latent variables based on the number of samplings and the reference latent distribution is the same as the method of obtaining y latent variables by randomly sampling the latent distribution based on a preset number of samplings. The method of decoding y reference latent variables using the target decoder of the target variational autoencoder in the target adversarial network to obtain y reference matrices is the same as the method of decoding y latent variables using the decoder in the variational autoencoder to obtain y reconstruction matrices. The method of inputting the reference matrix into the conditional identification model to obtain the reference conditional matrix is ​​the same as the method of inputting the expanded element matrix in the verification matrix group into the updated identification model to obtain the conditional identification matrix, and they will not be repeated here.

[0282] Optionally, the temperature comparison threshold is set to 50° C., and the zt comparison threshold is set to 0.1.

[0283] For example, if the final matrix is Since the element symbol corresponding to atomic number 12 is Mg, the element symbol corresponding to atomic number 51 is Sb, the element symbol corresponding to atomic number 83 is Bi, and the element symbol corresponding to atomic number 52 is Te, and 3.1 is the subscript of the element symbol Mg, 0.5 is the subscript of the element symbol Sb, 1.45 is the subscript of the element symbol Bi, and 0.05 is the subscript of the element symbol Te, the chemical formula of the target material restored according to the final matrix is ​​Mg 3.1 Sb 0.5 Bi 1.45 Te 0.05 , namely Mg3.1 Sb 0.5 Bi 1.45 Te 0.05 Thermoelectric materials predicted by the target adversarial network and conditional discrimination model required by researchers.

[0284] The present invention is to solve the problem described in the background technology. The present invention obtains a thermoelectric material data set, wherein the thermoelectric material data set includes: multiple thermoelectric material data, and the thermoelectric material data includes: material chemical formula, maximum zt value, optimal temperature and superscript value range, performs a chemical formula screening operation on the multiple thermoelectric material data to obtain multiple target material data, performs a temperature division operation on the multiple target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data, performs a superscript interpolation operation on a low-temperature material data, b medium-temperature material data and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data and C high-temperature interpolation data. It can be seen that the embodiment of the present invention performs a plurality of thermoelectric materials by screening the multiple thermoelectric materials. The data is subjected to a chemical formula screening operation to screen out thermoelectric material data corresponding to the chemical formula of materials with too many or too few elements, thereby preventing the chemical formula of materials with too many or too few individual elements from affecting the training of subsequent models, and then performing a temperature division operation and a superscript interpolation operation on multiple target material data, so that the data amounts of low-temperature interpolation data, medium-temperature interpolation data, and high-temperature interpolation data are basically consistent, thereby preventing the subsequent trained conditional identification model from over-fitting the medium-temperature zone with more data amount, while ignoring the low-temperature zone and the high-temperature zone. A chemical formula quantization operation is performed on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups. It can be seen that the embodiment of the present invention divides A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data into Z expansion matrix groups. The temperature interpolation data and C high-temperature interpolation data are converted into a matrix form, which is convenient for the subsequent training of the first identification model and the second identification model. The pre-constructed first identification model is trained using Z expansion matrix groups to obtain a conditional identification model. It can be seen that the embodiment of the present invention trains a conditional identification model to facilitate the subsequent auxiliary target adversarial network to distinguish the generated target material chemical formula, thereby improving the accuracy of the deep generation model in predicting high thermoelectric figure of merit materials. The pre-constructed variational autoencoder and Z expansion matrix groups are used to train the pre-constructed second identification model to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder. It can be seen that the embodiment of the present invention trains a true and false identification model through Z expansion matrix groups, thereby improving the true and false identification. The model is accurate in discriminating the chemical formula of the material generated by the generative model, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. A generative adversarial network is constructed using a variational autoencoder and a true and false identification model, and the generative adversarial network is adversarially trained to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder. It can be seen that the embodiment of the present invention combines the variational autoencoder and the generative adversarial network to train a target adversarial network, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. A plurality of target material chemical formulas are generated using a preset target zt value, a preset target temperature, a target adversarial network and a conditional identification model.The reverse design of high thermoelectric merit material is completed. It can be seen that the embodiment of the present invention uses the conditional identification model to assist the target adversarial network to generate the target material chemical formula that meets the preset target zt value and the preset target temperature, thereby improving the accuracy of the deep generative model in predicting high thermoelectric merit materials. Therefore, the present invention can improve the accuracy of the deep generative model in predicting high thermoelectric merit materials.

[0285] like Figure 2 , which is a functional module diagram of a high thermoelectric merit material reverse design system based on a deep generative model provided by an embodiment of the present invention.

[0286] The high thermoelectric merit material reverse design system 100 based on the deep generative model of the present invention can be installed in an electronic device. According to the functions implemented, the high thermoelectric merit material reverse design system 100 based on the deep generative model can include a material data partitioning module 101, a material data expansion module 102, a generative model training module 103 and a thermoelectric material generation module 104. The module of the present invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can complete fixed functions, which are stored in the memory of the electronic device.

[0287] The material data partitioning module 101 is used to obtain a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature and a range of angle values, perform a chemical formula screening operation on the plurality of thermoelectric material data to obtain a plurality of target material data, perform a temperature partitioning operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data;

[0288] The material data expansion module 102 is used to perform a subscript interpolation operation on a low-temperature material data, b medium-temperature material data and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data and C high-temperature interpolation data, and perform a chemical formula quantization operation on the A low-temperature interpolation data, the B medium-temperature interpolation data and the C high-temperature interpolation data to obtain Z expansion matrix groups;

[0289] The generative model training module 103 is used to train the pre-constructed first identification model using Z expansion matrix groups to obtain a conditional identification model, and to train the pre-constructed second identification model using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder, and to construct a generative adversarial network using the variational autoencoder and the true and false identification model, and to perform adversarial training on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder;

[0290] The thermoelectric material generation module 104 is used to generate multiple target material chemical formulas using a preset target zt value, a preset target temperature, a target adversarial network and a condition identification model to complete the reverse design of high thermoelectric merit materials.

[0291] In detail, each module in the high thermoelectric merit material reverse design system 100 based on the deep generative model in the embodiment of the present invention is used in the same manner as described above. Figure 1 The same technical means are used as the reverse design method of high thermoelectric merit materials based on deep generative model described in the specification and can produce the same technical effects, so they will not be repeated here.

[0292] like Figure 3 , is a schematic diagram of the structure of an electronic device for implementing a reverse design method for high thermoelectric figure of merit materials based on a deep generative model provided by an embodiment of the present invention.

[0293] The electronic device 1 may include a processor 10, a memory 11 and a bus 12, and may also include a computer program stored in the memory 11 and executable on the processor 10, such as a reverse design method program for high thermoelectric merit materials based on a deep generative model.

[0294] Among them, the memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (for example: SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device 1, such as a mobile hard disk of the electronic device 1. In other embodiments, the memory 11 can also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1. Further, the memory 11 also includes an internal storage unit of the electronic device 1 and an external storage device. The memory 11 can not only be used to store application software and various types of data installed in the electronic device 1, such as the code of the reverse design method program of high thermoelectric merit material based on the deep generative model, but also can be used to temporarily store data that has been output or is to be output.

[0295] The processor 10 may be composed of an integrated circuit in some embodiments, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and combinations of various control chips, etc. The processor 10 is the control core (Control Unit) of the electronic device, and uses various interfaces and lines to connect various components of the entire electronic device, and executes or executes programs or modules stored in the memory 11 (such as a reverse design method program for high thermoelectric merit materials based on a deep generative model, etc.), and calls data stored in the memory 11 to execute various functions of the electronic device 1 and process data.

[0296] The bus 12 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 12 may be divided into an address bus, a data bus, a control bus, etc. The bus 12 is configured to realize connection and communication between the memory 11 and at least one processor 10, etc.

[0297] Figure 3 Only an electronic device with components is shown, and those skilled in the art will understand that Figure 3 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

[0298] For example, although not shown, the electronic device 1 may also include a power source (such as a battery) for supplying power to each component. Preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that the power management device can realize functions such as charging management, discharging management, and power consumption management. The power source may also include any components such as one or more DC or AC power sources, recharging devices, power failure detection circuits, power converters or inverters, power status indicators, etc. The electronic device 1 may also include a variety of sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be repeated here.

[0299] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is generally used to establish a communication connection between the electronic device 1 and other electronic devices.

[0300] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), or a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode) touch device. The display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device 1 and to display a visual user interface.

[0301] The high thermoelectric merit material reverse design method program based on deep generative model stored in the memory 11 of the electronic device 1 is a combination of multiple instructions. When running in the processor 10, it can achieve:

[0302] Acquire a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature, and a range of values ​​of a superscript;

[0303] Performing a chemical formula screening operation on a plurality of thermoelectric material data to obtain a plurality of target material data, performing a temperature division operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data;

[0304] Performing a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data;

[0305] Perform chemical formula quantization operation on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups;

[0306] Using Z expansion matrix groups to train the pre-constructed first discrimination model to obtain a conditional discrimination model;

[0307] The pre-constructed second identification model is trained using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder;

[0308] A generative adversarial network is constructed using a variational autoencoder and a true-false identification model, and adversarial training is performed on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder;

[0309] By using the preset target zt value, preset target temperature, target adversarial network and conditional identification model, multiple target material chemical formulas are generated to complete the reverse design of high thermoelectric merit materials.

[0310] Specifically, the specific implementation method of the processor 10 for the above instructions can refer to Figures 1 to 3 The description of the relevant steps in the corresponding embodiments will not be repeated here.

[0311] Furthermore, if the module / unit integrated in the electronic device 1 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0312] The present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor of an electronic device, the computer program can implement:

[0313] Acquire a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature, and a range of values ​​of a superscript;

[0314] Performing a chemical formula screening operation on a plurality of thermoelectric material data to obtain a plurality of target material data, performing a temperature division operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data;

[0315] Performing a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data;

[0316] Perform chemical formula quantization operation on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups;

[0317] Using Z expansion matrix groups to train the pre-constructed first discrimination model to obtain a conditional discrimination model;

[0318] The pre-constructed second identification model is trained using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder;

[0319] A generative adversarial network is constructed using a variational autoencoder and a true-false identification model, and adversarial training is performed on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder;

[0320] By using the preset target zt value, preset target temperature, target adversarial network and conditional identification model, multiple target material chemical formulas are generated to complete the reverse design of high thermoelectric merit materials.

[0321] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, systems and methods can be implemented in other ways. For example, the system embodiments described above are only illustrative, and actual implementation may have other division methods.

[0322] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0323] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0324] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0325] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for reverse design of high thermoelectric merit materials based on deep generative model, characterized in that: The method comprises: Acquire a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature, and a range of values ​​of a superscript; Performing a chemical formula screening operation on a plurality of thermoelectric material data to obtain a plurality of target material data, performing a temperature division operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data; Performing a subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data; Perform chemical formula quantization operation on A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data to obtain Z expansion matrix groups; Using Z expansion matrix groups to train the pre-constructed first discrimination model to obtain a conditional discrimination model; The pre-constructed second identification model is trained using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder; A generative adversarial network is constructed using a variational autoencoder and a true-false identification model, and adversarial training is performed on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder; By using the preset target zt value, preset target temperature, target adversarial network and conditional identification model, multiple target material chemical formulas are generated to complete the reverse design of high thermoelectric merit materials.

2. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 1, characterized in that: The chemical formula screening operation is performed on the multiple thermoelectric material data to obtain multiple target material data, including: The following operations are performed on each of the multiple thermoelectric material data: Identify the number of elements in the material chemical formula in the thermoelectric material data, and compare the number of elements with a preset first number threshold and a preset second number threshold, wherein the first number threshold is less than the second number threshold; If the number of elements is greater than or equal to the first number threshold and the number of elements is less than or equal to the second number threshold, recording the thermoelectric material data as target material data; The target material data are aggregated to obtain a plurality of target material data.

3. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 1, characterized in that: The performing of temperature division operation on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data comprises: The following operations are performed on each of the multiple target material data: Comparing the optimal temperature corresponding to the target material data with a preset first temperature and a preset second temperature, wherein the first temperature is less than the second temperature; If the optimal temperature is less than or equal to the first temperature, the target material data corresponding to the optimal temperature is recorded as low-temperature material data; If the optimal temperature is greater than the first temperature and less than the second temperature, the target material data corresponding to the optimal temperature is recorded as medium-temperature material data; If the optimal temperature is greater than or equal to the second temperature, the target material data corresponding to the optimal temperature is recorded as high-temperature material data; The low-temperature material data, the medium-temperature material data and the high-temperature material data are summarized respectively to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data.

4. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 1, characterized in that: The step of performing the subscript interpolation operation on a low-temperature material data, b medium-temperature material data, and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data, and C high-temperature interpolation data includes: The low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio are calculated according to the preset target data quantity. The calculation formula is as follows: Among them, μ a , μ b and μ c are low temperature interpolation ratio, medium temperature interpolation ratio and high temperature interpolation ratio respectively, a0 is the target data quantity, Represents a round-up operation; A low-temperature interpolation ratio is used to perform a subscript interpolation operation on a low-temperature material data to obtain A low-temperature interpolation data. B medium-temperature interpolation data are obtained based on the medium-temperature interpolation ratio and b medium-temperature material data. C high-temperature interpolation data are obtained based on the high-temperature interpolation ratio and c high-temperature material data.

5. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 4, characterized in that: The method of performing a subscript interpolation operation on a low-temperature material data using a low-temperature interpolation ratio to obtain A low-temperature interpolation data includes: The following operations are performed on each of the a low-temperature material data: The low-temperature sampling interval is calculated according to the low-temperature interpolation ratio and the subscript value range corresponding to the low-temperature material data. The calculation formula is as follows: Among them, d a is the low temperature sampling distance, f max and f min are the maximum and minimum values ​​in the subscript value range, μ a is the low temperature interpolation ratio; Based on the low temperature sampling interval, a uniform sampling operation is performed on the subscript value range corresponding to the low temperature material data to obtain m subscript values. The following operation is performed on each of the m subscript values: Identify the subscript variable in the material chemical formula corresponding to the low-temperature material data, and update the subscript variable in the material chemical formula to the subscript value to obtain an updated chemical formula; Integrate the maximum zt value corresponding to the updated chemical formula and low-temperature material data into low-temperature interpolation data; The low-temperature interpolation data are summarized to obtain A low-temperature interpolation data, where A=m×a.

6. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 5, characterized in that: The chemical formula quantization operation is performed on the A low-temperature interpolation data, the B medium-temperature interpolation data, and the C high-temperature interpolation data to obtain Z expansion matrix groups, including: The following operation is performed on each of the A low-temperature interpolation data: The low temperature element matrix is ​​constructed using the updated chemical formula in the low temperature interpolation data, where the low temperature element matrix is ​​as follows: Among them, D ea is the low temperature element matrix, E1, E2 and E n Update the atomic numbers of the first, second, and nth elements in the chemical formula, e1, e2, and e n are the subscripts of the first, second, and nth elements in the updated chemical formula, respectively, and n is the number of elements corresponding to the updated chemical formula; The low temperature condition matrix is ​​constructed using the maximum zt value corresponding to the low temperature interpolation data. The low temperature condition matrix is ​​as follows: Among them, D xa is the low temperature condition matrix, X zta is the maximum zt value corresponding to the low temperature interpolation data; Combining the low temperature element matrix and the low temperature condition matrix to obtain a low temperature matrix group, and summing up the low temperature matrix groups to obtain A low temperature matrix groups; The following operation is performed on each of the B medium-temperature interpolation data: Obtaining a medium-temperature element matrix based on updated chemical formulas in the medium-temperature interpolation data; The medium temperature condition matrix is ​​constructed using the maximum zt value corresponding to the medium temperature interpolation data. The medium temperature condition matrix is ​​as follows: Among them, D xb is the medium temperature condition matrix, X ztb is the maximum zt value corresponding to the medium temperature interpolation data; Combining the medium temperature element matrix and the medium temperature condition matrix to obtain a medium temperature matrix group, and summing up the medium temperature matrix groups to obtain B medium temperature matrix groups; The following operations are performed on each of the C high temperature interpolation data: Obtaining a high temperature element matrix based on updated chemical formulas in the high temperature interpolation data; The high temperature condition matrix is ​​constructed using the maximum zt value corresponding to the high temperature interpolation data. The high temperature condition matrix is ​​as follows: Among them, D xc is the high temperature condition matrix, X ztc is the maximum zt value corresponding to the high temperature interpolation data; Combining the high temperature element matrix and the high temperature condition matrix to obtain a high temperature matrix group, and summing up the high temperature matrix groups to obtain C high temperature matrix groups; A low-temperature matrix groups, B medium-temperature matrix groups, and C high-temperature matrix groups are summarized to obtain Z expanded matrix groups, wherein the expanded matrix group is a low-temperature matrix group, a medium-temperature matrix group, or a high-temperature matrix group, and Z=A+B+C. The expanded matrix group includes: an expanded element matrix and an expanded condition matrix. The expanded element matrix is ​​a low-temperature element matrix, a medium-temperature element matrix, or a high-temperature element matrix. The expanded condition matrix is ​​a low-temperature condition matrix, a medium-temperature condition matrix, or a high-temperature condition matrix.

7. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 6, characterized in that: The method of using Z expansion matrix groups to train the pre-constructed first identification model to obtain a conditional identification model includes: Get the training ratio and verification ratio, where the sum of the training ratio and the verification ratio is 1; The training quantity is calculated based on the training ratio. The calculation formula is as follows: z1=Z×k1 Among them, z1 is the number of training, k1 is the training ratio; The verification quantity is obtained based on the verification ratio, and z1 training matrix groups and z2 verification matrix groups are obtained based on the training quantity, the verification quantity and the Z expansion matrix groups, where z2 is the verification quantity; Randomly select x1 training matrix groups from z1 training matrix groups, where x1<z1; Using x1 training matrix groups to train the first identification model to obtain an updated identification model; For each of the z2 verification matrix groups, perform the following operations: The expanded element matrix in the verification matrix group is input into the update identification model to obtain the condition identification matrix. The condition difference degree is calculated according to the condition identification matrix and the expanded condition matrix in the verification matrix group. The calculation formula is as follows: Among them, δ MSE is the conditional difference, P1 and P2 are the elements of the first row and first column and the elements of the first row and second column in the extended condition matrix in the verification matrix group, Q1 and Q2 are the elements of the first row and first column and the elements of the first row and second column in the conditional discrimination matrix, tanh is the hyperbolic tangent function; Summarize the conditional differences to obtain multiple conditional differences, and calculate the difference mean according to the multiple conditional differences, wherein the difference mean is the average value of the multiple conditional differences; Compare the difference mean with the preset difference threshold, if the difference mean is greater than the difference threshold, update the identification model as the first identification model, and return to the step of randomly selecting x1 training matrix groups from z1 training matrix groups until the difference mean is greater than or equal to the difference threshold; If the difference mean is less than or equal to the difference threshold, the updated identification model is used as the conditional identification model.

8. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 7, characterized in that: The method of using the pre-constructed variational autoencoder and the Z expansion matrix groups to train the pre-constructed second identification model to obtain a true and false identification model includes: The following operations are performed on each of the Z expanded matrix groups: Mapping the augmented element matrix in the augmented matrix group to a potential distribution using an encoder in a variational autoencoder; Based on the preset number of samples, random sampling is performed in the potential distribution to obtain y potential variables, where y is the number of samples; Use the decoder in the variational autoencoder to decode y latent variables and obtain y reconstruction matrices; Sum up y reconstruction matrices to obtain Y reconstruction matrices, where Y = Z × y; Extracting Z extended element matrices from the Z extended matrix groups, and randomly extracting Z reconstruction matrices from the Y reconstruction matrices; The following operation is performed on each of the Z expanded element matrices: The expanded element matrix and the pre-constructed true judgment matrix are combined into a true judgment matrix group, where the true judgment matrix is: [1]; Summarize the true judgment matrix groups to obtain Z true judgment matrix groups; The following operations are performed on each of the Z reconstruction matrices: The reconstructed matrix and the pre-constructed false judgment matrix are combined into a false judgment matrix group, where the false judgment matrix is: [0]; Summarize the false judgment matrix groups to obtain Z false judgment matrix groups; Summarize Z true-judgment matrix groups and Z false-judgment matrix groups to obtain N true-false matrix groups, where a true-false matrix group is a true-judgment matrix group or a false-judgment matrix group, and N=2×Z; Based on the training ratio, the verification ratio and the N true and false matrix groups, n1 true and false training matrix groups and n2 true and false verification matrix groups are obtained, wherein N=n1+n2; Acquire a judgment identification model based on n1 true and false training matrix groups and a second identification model; The following operations are performed on each of the n2 true and false verification matrix groups: Input the expanded element matrix or the reconstructed matrix in the true and false verification matrix group into the updated discrimination model to obtain the true and false identification matrix, obtain the discrimination value based on the true and false identification matrix, obtain the original value based on the true judgment matrix or the false judgment matrix in the true and false verification matrix group, and calculate the discrimination difference value according to the discrimination value and the original value, wherein the discrimination value is the value of the first row and first column element in the true and false identification matrix, the original value is the value of the first row and first column element in the true judgment matrix or the false judgment matrix, and the discrimination difference value is the absolute difference between the discrimination value and the original value; Summarize the discriminant difference values ​​to obtain n2 discriminant difference values, and calculate the discriminant mean value based on the n2 discriminant difference values, wherein the discriminant mean value is the average value of the n2 discriminant mean values; Compare the discrimination mean value with the preset discrimination threshold value, if the discrimination mean value is greater than the discrimination threshold value, update the discrimination model as the second discrimination model, and return to the step of obtaining the discrimination model based on the n1 true and false training matrix groups and the second discrimination model until the discrimination mean value is greater than or equal to the discrimination threshold value; If the discrimination mean value is less than or equal to the discrimination threshold, the updated discrimination model is used as the true and false discrimination model.

9. The method for reverse design of high thermoelectric merit materials based on deep generative model according to claim 8, characterized in that: The method of generating multiple target material chemical formulas by using a preset target zt value, a preset target temperature, a target adversarial network and a conditional identification model includes: Determine a maximum reference zt value and a minimum reference zt value based on the target zt value and the preset reference zt value, wherein the maximum reference zt value is the sum of the target zt value and the reference zt value, and the minimum reference zt value is the absolute difference between the target zt value and the reference zt value; Determine a reference zt range based on a maximum reference zt value and a minimum reference zt value, wherein the maximum value of the reference zt value range is the maximum reference zt value, and the minimum value of the reference zt value range is the minimum reference zt value; Compare the target temperature with the first temperature and the second temperature. If the target temperature is less than or equal to the first temperature, the temperature value is recorded as 1, and multiple reference low temperature data are selected from the A low temperature interpolation data, wherein the maximum zt value corresponding to the reference low temperature data is within the reference zt value range; If the target temperature is greater than the first temperature and less than the second temperature, the temperature value is recorded as 2, and a plurality of reference medium temperature data are obtained based on B medium temperature interpolation data; If the target temperature is greater than or equal to the second temperature, the temperature value is recorded as 3, and a plurality of reference high temperature data are obtained based on C high temperature interpolation data; The following operation is performed on each of the plurality of reference low temperature data, the plurality of reference medium temperature data or the plurality of reference high temperature data: Obtaining a reference element matrix based on reference low temperature data, reference medium temperature data, or reference high temperature data; Aggregating the reference element matrices to obtain multiple reference element matrices, and inputting the multiple reference element matrices into a target encoder of a target variational autoencoder in a target adversarial network to obtain multiple reference potential distributions; For each of the multiple reference latent distributions, perform the following operations: Obtain y reference latent variables based on the number of samples and the reference latent distribution, and use the target decoder of the target variational autoencoder in the target adversarial network to decode the y reference latent variables to obtain y reference matrices; The reference matrices are aggregated to obtain multiple reference matrices, and the following operations are performed on each of the multiple reference matrices: Input the reference matrix into the condition identification model to obtain the reference condition matrix, and identify the reconstructed temperature value and the reconstructed zt value of the reference condition matrix, wherein the reconstructed zt value is the value of the element in the first row and first column of the reference condition matrix, and the reconstructed temperature value is the value of the element in the first row and second column of the reference condition matrix; The temperature difference is calculated according to the reconstructed temperature value and the temperature value corresponding to the reference matrix, and the reconstructed zt difference is calculated according to the reconstructed zt value and the target zt value, wherein the temperature difference is the absolute difference between the reconstructed temperature value and the reference matrix, and the reconstructed zt difference is the absolute difference between the reconstructed zt value and the target zt value; Determine whether the temperature difference is less than a preset temperature comparison threshold, and determine whether the reconstructed zt difference is less than a preset zt comparison threshold; If the temperature difference is less than the temperature comparison threshold and the reconstructed zt difference is less than the zt comparison threshold, the reference matrix is ​​recorded as the final matrix, and the chemical formula of the target material is restored according to the final matrix; The target material chemical formulas are summarized to obtain multiple target material chemical formulas.

10. A reverse design system for high thermoelectric merit materials based on deep generative model, characterized in that: The system comprises: A material data partitioning module is used to obtain a thermoelectric material data set, wherein the thermoelectric material data set includes: a plurality of thermoelectric material data, and the thermoelectric material data includes: a material chemical formula, a maximum zt value, an optimal temperature and a range of angle values, a chemical formula screening operation is performed on the plurality of thermoelectric material data to obtain a plurality of target material data, and a temperature partitioning operation is performed on the plurality of target material data to obtain a low-temperature material data, b medium-temperature material data and c high-temperature material data; A material data expansion module is used to perform a subscript interpolation operation on a low-temperature material data, b medium-temperature material data and c high-temperature material data to obtain A low-temperature interpolation data, B medium-temperature interpolation data and C high-temperature interpolation data, and perform a chemical formula quantization operation on the A low-temperature interpolation data, the B medium-temperature interpolation data and the C high-temperature interpolation data to obtain Z expansion matrix groups; A generative model training module is used to train a pre-constructed first identification model using Z expansion matrix groups to obtain a conditional identification model, and to train a pre-constructed second identification model using a pre-constructed variational autoencoder and Z expansion matrix groups to obtain a true and false identification model, wherein the variational autoencoder includes: an encoder and a decoder, and to construct a generative adversarial network using the variational autoencoder and the true and false identification model, and to perform adversarial training on the generative adversarial network to obtain a target adversarial network, wherein the target adversarial network includes: a target variational autoencoder, and the target variational autoencoder includes: a target encoder and a target decoder; Thermoelectric material generation module is used to generate multiple target material chemical formulas using preset target zt value, preset target temperature, target adversarial network and condition identification model to complete the reverse design of high thermoelectric merit materials.

Images