A method for analyzing the ratio of perovskite components based on a deep generative model

Abstract

The present invention discloses a method for analyzing the component ratio of perovskite based on a deep generative model, comprising the following steps: by introducing a deep generative model and regularized flow technology, the representation capability of the latent space samples is optimized, and the sample position and structure are dynamically optimized in combination with an adaptive latent space adjustment mechanism and a self-attention mechanism to adapt to the characteristics of different types of perovskite materials. By combining a variational autoencoder with a regularized flow, multiple perovskite samples are generated, and by using a multi-dimensional reverse inference technique, the component ratio information of each sample is derived from the latent space to achieve accurate analysis of the component ratio of the perovskite material, and by combining the structural characteristics and chemical composition information of the perovskite material, chemical characterization is performed in the latent space of the generated sample, further improving the accuracy of the analysis. The present invention can achieve efficient and accurate analysis of the component ratio of perovskite materials, and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to the field of material science and technology, and in particular to a method for analyzing the proportion of perovskite components based on a deep generation model. Background Art

[0002] Perovskite materials are widely used in solar cells, photodetectors, lasers and other fields due to their excellent optoelectronic properties. The composition ratio of perovskite materials directly affects their performance, so accurate analysis and prediction of their composition ratio is crucial for optimizing material design and improving performance. However, existing perovskite composition analysis methods have some significant defects, which limit their application in efficient preparation and performance optimization.

[0003] At present, traditional methods for analyzing the components of perovskite materials mostly rely on chemical analysis and experimental methods, such as X-ray diffraction analysis, electron microscope imaging, etc. Although these methods can provide certain physical or chemical characterizations, they have the following problems:

[0004] 1. High cost and low efficiency: Traditional experimental methods usually require expensive equipment and complex experimental processes, and each experiment takes a long time, which limits their application in high-throughput material screening.

[0005] 2. Lack of accuracy: Existing technologies are often based on inferences based on surface or overall physical properties, which makes it difficult to accurately reflect the structure and composition distribution of perovskite materials in the potential space, resulting in the inability to achieve accurate component ratio analysis.

[0006] 3. Lack of automation and real-time performance: Most existing methods rely on manual intervention and lack automation and real-time analysis capabilities, making it difficult to provide timely feedback during large-scale material screening or online production.

[0007] 4. Single analysis perspective: Existing analysis methods mostly focus on the study of chemical composition or macroscopic structure, and fail to fully utilize the potential spatial characteristics and multi-dimensional data of the material for comprehensive analysis, resulting in limited depth and breadth of the analysis results.

[0008] Therefore, how to provide a method for analyzing the proportion of perovskite components based on a deep generation model is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for analyzing the component ratio of perovskite based on a deep generative model. The present invention combines technologies such as variational autoencoders, regularized flows, and adaptive latent space adjustment mechanisms. By generating multiple perovskite samples and optimizing their latent spaces, the component ratio of each sample is accurately derived. This method can effectively solve the problems of high cost, low efficiency, low precision, and lack of automation in traditional analysis methods, and has the advantages of high efficiency, accuracy, automation, and real-time analysis.

[0010] A method for analyzing the proportion of perovskite components based on a deep generation model according to an embodiment of the present invention includes the following steps:

[0011] S1. Use variational autoencoders to build a latent space model of perovskite materials, optimize the model parameters by maximizing the variational lower bound, learn the structural characteristics of perovskite materials, and map them to a low-dimensional latent space.

[0012] S2. Based on the regularization flow, the latent space generated by the variational autoencoder is transformed in distribution, and the reversible transformation technology is used to adjust the expression ability of the latent space, which specifically includes the following steps:

[0013] Latent space samples generated by variational autoencoders via regularized flow Transform to obtain new latent space samples , construct a regularized flow model, the transformation is composed of the transformation function of the regularized flow model Completed, including are the trainable parameters of the model:

[0014] ;

[0015] During the transformation process, the regularized flow model minimizes the distribution of samples in the latent space Distribution of actual perovskite materials The difference between , by minimizing Divergence measures the similarity of two distributions:

[0016] ;

[0017] in, The latent space sample distribution generated by the regularized flow, is the actual sample distribution of perovskite materials, for Divergence;

[0018] The regularization flow adopts a reversible transformation structure so that each generated latent space sample By inverse transformation function Recovering back to the original latent space sample :

[0019] ;

[0020] By introducing a hierarchical structure, the transformation process of the regularized flow adopts a multi-layer transformation function , where each layer transformation function Gradually optimize the expressive power of latent space samples, and complete the adjustment and optimization of latent space samples through multi-layer transformation;

[0021] In a multi-layer transformation structure, the parameters of the transformation function Automatically adjusted through an optimization process to minimize the latent space distribution of generated samples The actual distribution of perovskite materials The difference between them is calculated by adjusting the parameters of each layer’s transformation function through the back propagation algorithm to achieve the optimal potential space expression ability.

[0022] S3. Introduce an adaptive latent space adjustment mechanism, combined with the self-attention mechanism, dynamically optimize the representation of the latent space according to the characteristics of the perovskite material, adjust the position and structure of the latent space, and adapt to the changes of different types of perovskite materials;

[0023] S4. Construct a deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples, and derive the component ratio information of each sample from the latent space through multi-dimensional reverse inference technology;

[0024] S5. Based on the generated perovskite samples, unsupervised learning techniques are applied to perform data enhancement and expand the training set;

[0025] S6. Combining the structural characteristics and chemical composition information of the perovskite material, chemical characterization is performed in the latent space of the generated samples to derive the component ratio of each generated sample;

[0026] S7. Through the error minimization optimization algorithm, the error between the generated samples and the experimental data is minimized, and the parameters of the deep generative model are dynamically adjusted.

[0027] Optionally, the hierarchical structure in S2 specifically includes:

[0028] By introducing a hierarchical structure, the transformation process of the regularized flow adopts The transformation function of the layer is gradually optimized , where each layer transformation function Acting on latent space samples , mapping it to a new latent space sample ,in, is the dimension of the latent space samples, It is The output of the layer, It is the final latent space sample. Through multi-layer transformation, the feature representation ability of the latent space sample is gradually optimized, and the representation ability and accuracy of the sample after each layer of transformation are optimized.

[0029] In the transformation process of each layer, the transformation function Using a reversible transformation structure, the generated latent space samples Can be transformed by inverse function Recovering back to the original latent space sample :

[0030] ;

[0031] Transformation function of each layer Has trainable parameters , and adjust the trainable parameters through an optimization process to minimize the final latent space sample Distribution Distribution of actual perovskite materials The difference between:

[0032] L l = D KL [ p z (L) | p data ] ;

[0033] in, for Divergence, which is used to measure the difference between two distributions and optimize the parameters of each layer’s transformation function .

[0034] Optionally, the S3 specifically includes:

[0035] S31, introduce adaptive latent space adjustment mechanism, combined with self-attention mechanism, in each layer of latent space samples In the dynamic optimization sample The position and structure of the latent space samples are defined to adapt to the characteristics of different types of perovskite materials, calculate the similarity between the latent space samples, and define the latent space samples. and Similarity between ;

[0036] S32, the self-attention mechanism is through the latent space samples and Similarity between Perform calculations to generate a correlation weight matrix , which represents the latent space samples With other samples The correlation between:

[0037] ;

[0038] in, is the weight matrix of the query vector, is the dimension of the latent space;

[0039] S33. In the self-attention mechanism, according to the calculated relevance weight matrix , a sample of the latent space Will be weighted averaged to update the representation of the sample in the latent space, the latent space sample The update rule is:

[0040] ;

[0041] in, is the trainable matrix used for weighting, The latent space samples are updated by the self-attention mechanism. Through the weighted averaging process, the representation ability of the latent space samples is gradually optimized, and the position of the samples in the latent space is adjusted.

[0042] Optionally, the S4 specifically includes:

[0043] S41. A deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples , optimize the generative ability of the model by maximizing the variational lower bound:

[0044] ;

[0045] in, Representing the latent space The variational posterior distribution of , is the conditional distribution of the generative model, is the prior distribution of the latent space, for Divergence, Express expectations, represents the loss function in the variational autoencoder;

[0046] S42. Latent space samples generated by variational autoencoders , optimized by regularized flow, through a series of reversible transformations Gradually adjust the latent space samples so that the generated latent space samples is mapped to a new representation , and this transformation can make the distribution of generated samples closer to the actual perovskite material distribution. The optimization goal of each layer of transformation is to minimize the distribution of potential space samples. The distribution of real data Between Divergence:

[0047] ;

[0048] S43, based on the optimized latent space sample , using multi-dimensional reverse inference technology to derive the component ratio information from the latent space sample of each sample ,in, Indicates The component ratio of the latent space samples, is a sample in the latent space Middle The ratio of the perovskite components, is the number of perovskite components, a regression model is constructed, and the latent space samples are transformed into The proportion of components mapped to the latent space samples , regression loss function Defined as:

[0049] ;

[0050] in, is obtained by reverse inference from the latent space samples The predicted component ratio obtained is is the total number of samples in the latent space;

[0051] S44. Latent space samples generated by variational autoencoders and regularized flows , and the component ratio derived by multi-dimensional reverse inference technology , achieving accurate analysis of the component proportions of perovskite materials.

[0052] Optionally, the S6 specifically includes:

[0053] S61. Construct a chemical characterization network based on the structural characteristics and chemical composition information of perovskite materials By extracting the crystal structure information, elemental composition and chemical bonding characteristics of perovskite materials, a multi-dimensional representation is established, and the latent space sample of each generated sample is sampled through the chemical characterization network. Perform chemical characterization and derive component ratio information of latent space samples;

[0054] S62. Generate latent space samples in the model Enter the Chemical Characterization Network , through the processing of the chemical characterization network, a latent space sample with chemical characteristics is obtained Based on the actual chemical properties of perovskite materials, the potential space samples Adjust it to more closely match the chemical distribution of actual perovskite samples;

[0055] S63. Through regression model , the optimized latent space samples The proportion of components mapped to the latent space samples ,in c i =[ c i1 , c i2 ,…, c im ] Indicates The component ratio of the latent space samples, and For the The first sample in the latent space The regression model is calculated by minimizing the regression loss function To achieve the prediction of sample component ratio:

[0056] ;

[0057] in, is the sample from the optimized latent space through the regression model The predicted component ratios, is the number of samples in the latent space;

[0058] S64. Latent space samples by combining the outputs of the generative model , Latent space sample after chemical characterization network optimization , and the prediction of the regression model, complete the derivation of the component ratio of the perovskite sample.

[0059] The beneficial effects of the present invention are:

[0060] (1) The present invention achieves accurate analysis of the component ratio of perovskite materials by combining deep generative models, variational autoencoders, regularized flows, adaptive latent space adjustment mechanisms, and multi-dimensional reverse inference technology. This method can automatically generate perovskite samples, optimize latent space samples, and accurately derive the component ratio of each sample through reverse inference, overcoming the shortcomings of traditional methods in terms of high cost, low efficiency, and low precision. Through these technical means, the present invention can efficiently and accurately analyze the component ratio of perovskite materials, improving the degree of automation and accuracy of the analysis.

[0061] (2) The present invention provides global optimization capability for perovskite material component ratio analysis through regularization flow and adaptive latent space adjustment mechanism, combined with latent space samples generated by variational autoencoders. By gradually optimizing the representation capability of latent space samples at multiple layers, the generated samples can better adapt to the characteristics of different types of perovskite materials, further improving the accuracy of sample generation and component ratio derivation.

[0062] (3) The present invention provides a highly automated and accurate method for analyzing the ratio of perovskite components, which reduces the reliance on manual intervention and improves the efficiency and real-time performance of the analysis process. At the same time, the system can flexibly adapt to the characteristics of different materials, has strong versatility and adaptability, meets the analysis needs of a variety of perovskite materials, and has high application prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0064] Figure 1 This is a general framework diagram of a perovskite component ratio analysis method based on a deep generative model proposed in the present invention. DETAILED DESCRIPTION

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0066] refer to Figure 1 , a method for analyzing the ratio of perovskite components based on a deep generative model, comprising the following steps:

[0067] S1. Use variational autoencoders to build a latent space model of perovskite materials, optimize the model parameters by maximizing the variational lower bound, learn the structural characteristics of perovskite materials, and map them to a low-dimensional latent space.

[0068] S2. Based on the regularization flow, the latent space generated by the variational autoencoder is transformed in distribution, and the reversible transformation technology is used to adjust the expression ability of the latent space, which specifically includes the following steps:

[0069] Latent space samples generated by variational autoencoders via regularized flow Transform to obtain new latent space samples , construct a regularized flow model, the transformation is composed of the transformation function of the regularized flow model Completed, including are the trainable parameters of the model:

[0070] ;

[0071] During the transformation process, the regularized flow model minimizes the distribution of samples in the latent space Distribution of actual perovskite materials The difference between , by minimizing Divergence measures the similarity of two distributions:

[0072] ;

[0073] in, The latent space sample distribution generated by the regularized flow, is the actual sample distribution of perovskite materials, for Divergence;

[0074] The regularization flow adopts a reversible transformation structure so that each generated latent space sample By inverse transformation function Recovering back to the original latent space sample :

[0075] ;

[0076] By introducing a hierarchical structure, the transformation process of the regularized flow adopts a multi-layer transformation function , where each layer transformation function Gradually optimize the expressive power of latent space samples, and complete the adjustment and optimization of latent space samples through multi-layer transformation;

[0077] In a multi-layer transformation structure, the parameters of the transformation function Automatically adjusted through an optimization process to minimize the latent space distribution of generated samples The actual distribution of perovskite materials The difference between them is calculated by adjusting the parameters of each layer’s transformation function through the back propagation algorithm to achieve the optimal potential space expression ability.

[0078] This implementation method uses regularized flow to optimize the latent space samples generated by the variational autoencoder to accurately align the latent space sample distribution with the actual perovskite material distribution. This method gradually optimizes the expressiveness of the latent space samples through reversible transformations and multi-layer transformation structures, improves the accuracy and recoverability of the generated samples, and thus provides reliable support for the precise analysis of the perovskite component ratio, improving the expressiveness and analysis accuracy of the model.

[0079] S3. Introduce an adaptive latent space adjustment mechanism, combined with the self-attention mechanism, dynamically optimize the representation of the latent space according to the characteristics of the perovskite material, adjust the position and structure of the latent space, and adapt to the changes of different types of perovskite materials;

[0080] S4. Construct a deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples, and derive the component ratio information of each sample from the latent space through multi-dimensional reverse inference technology;

[0081] S5. Based on the generated perovskite samples, unsupervised learning techniques are applied to perform data enhancement and expand the training set;

[0082] S6. Combining the structural characteristics and chemical composition information of the perovskite material, chemical characterization is performed in the latent space of the generated samples to derive the component ratio of each generated sample;

[0083] S7. Through the error minimization optimization algorithm, the error between the generated samples and the experimental data is minimized, and the parameters of the deep generative model are dynamically adjusted.

[0084] In this implementation, the hierarchical structure in S2 specifically includes:

[0085] By introducing a hierarchical structure, the transformation process of the regularized flow adopts The transformation function of the layer is gradually optimized , where each layer transformation function Acting on latent space samples , mapping it to a new latent space sample ,in, is the dimension of the latent space samples, It is The output of the layer, It is the final latent space sample. Through multi-layer transformation, the feature representation ability of the latent space sample is gradually optimized, and the representation ability and accuracy of the sample after each layer of transformation are optimized.

[0086] In the transformation process of each layer, the transformation function Using a reversible transformation structure, the generated latent space samples Can be transformed by inverse function Recovering back to the original latent space sample :

[0087] ;

[0088] Transformation function of each layer Has trainable parameters , and adjust the trainable parameters through an optimization process to minimize the final latent space sample Distribution Distribution of actual perovskite materials The difference between:

[0089] L l = D KL [ p z (L) | p data ] ;

[0090] in, for Divergence, which is used to measure the difference between two distributions and optimize the parameters of each layer’s transformation function .

[0091] This implementation method optimizes the expressiveness of the latent space samples layer by layer by introducing a multi-level structure and a reversible transformation function, thereby effectively improving the accuracy and reliability of the latent space samples. Through trainable transformation parameters, the accuracy of each layer of transformation is optimized, so that the generated latent space samples are highly aligned with the actual perovskite material distribution, improving the accuracy of component ratio analysis and optimization efficiency.

[0092] In this implementation, S3 specifically includes:

[0093] S31, introduce adaptive latent space adjustment mechanism, combined with self-attention mechanism, in each layer of latent space samples In the dynamic optimization sample The position and structure of the latent space samples are defined to adapt to the characteristics of different types of perovskite materials, calculate the similarity between the latent space samples, and define the latent space samples. and Similarity between ;

[0094] S32, the self-attention mechanism is through the latent space samples and Similarity between Perform calculations to generate a correlation weight matrix , which represents the latent space samples With other samples The correlation between:

[0095] ;

[0096] in, is the weight matrix of the query vector, is the dimension of the latent space;

[0097] S33. In the self-attention mechanism, according to the calculated relevance weight matrix , a sample of the latent space Will be weighted averaged to update the representation of the sample in the latent space, the latent space sample The update rule is:

[0098] ;

[0099] in, is the trainable matrix used for weighting, The latent space samples are updated by the self-attention mechanism. Through the weighted averaging process, the representation ability of the latent space samples is gradually optimized, and the position of the samples in the latent space is adjusted.

[0100] This implementation introduces an adaptive latent space adjustment mechanism and a self-attention mechanism to dynamically optimize the position and structure of perovskite material samples in the latent space, significantly improving the accuracy of similarity calculation between samples. Through the self-attention mechanism of the weighted matrix, the sample representation capability is optimized, adapting to the characteristics of different perovskite materials, thereby enhancing the accuracy and robustness of component ratio analysis.

[0101] In this implementation manner, the S4 specifically includes:

[0102] S41. A deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples , optimize the generative ability of the model by maximizing the variational lower bound:

[0103] ;

[0104] in, Representing the latent space The variational posterior distribution of , is the conditional distribution of the generative model, is the prior distribution of the latent space, for Divergence, Express expectations, represents the loss function in the variational autoencoder;

[0105] S42. Latent space samples generated by variational autoencoders , optimized by regularized flow, through a series of reversible transformations Gradually adjust the latent space samples so that the generated latent space samples is mapped to a new representation , and this transformation can make the distribution of generated samples closer to the actual perovskite material distribution. The optimization goal of each layer of transformation is to minimize the distribution of potential space samples. The distribution of real data Between Divergence:

[0106] ;

[0107] S43, based on the optimized latent space sample , using multi-dimensional reverse inference technology to derive the component ratio information from the latent space sample of each sample ,in, Indicates The component ratio of the latent space samples, is a sample in the latent space Middle The ratio of the perovskite components, is the number of perovskite components, a regression model is constructed, and the latent space samples are transformed into The proportion of components mapped to the latent space samples , regression loss function Defined as:

[0108] ;

[0109] in, is obtained by reverse inference from the latent space samples The predicted component ratio obtained is is the total number of samples in the latent space;

[0110] S44. Latent space samples generated by variational autoencoders and regularized flows , and the component ratio derived by multi-dimensional reverse inference technology , achieving accurate analysis of the component proportions of perovskite materials.

[0111] This implementation combines the variational autoencoder and the regularized flow deep generative model to optimize the latent space sample generation, and accurately derives the component ratio of the perovskite sample through multi-dimensional reverse inference technology. This method significantly improves the accuracy of the generated samples, optimizes the latent space samples, and ultimately achieves high-precision analysis of the component ratio of the perovskite material, improving the reliability and efficiency of the perovskite material analysis.

[0112] In this implementation manner, S6 specifically includes:

[0113] S61. Construct a chemical characterization network based on the structural characteristics and chemical composition information of perovskite materials By extracting the crystal structure information, elemental composition and chemical bonding characteristics of perovskite materials, a multi-dimensional representation is established, and the latent space sample of each generated sample is sampled through the chemical characterization network. Perform chemical characterization and derive component ratio information of latent space samples;

[0114] S62. Generate latent space samples in the model Enter the Chemical Characterization Network , through the processing of the chemical characterization network, a latent space sample with chemical characteristics is obtained Based on the actual chemical properties of perovskite materials, the potential space samples Adjust it to more closely match the chemical distribution of actual perovskite samples;

[0115] S63. Through regression model , the optimized latent space samples The proportion of components mapped to the latent space samples ,in c i =[ c i1 , c i2 ,…, c im ] Indicates The component ratio of the latent space samples, and For the The first sample in the latent space The regression model is calculated by minimizing the regression loss function To achieve the prediction of sample component ratio:

[0116] ;

[0117] in, is the sample from the optimized latent space through the regression model The predicted component ratios, is the number of samples in the latent space;

[0118] S64. Latent space samples by combining the outputs of the generative model , Latent space sample after chemical characterization network optimization , and the prediction of the regression model, complete the derivation of the component ratio of the perovskite sample.

[0119] This implementation method constructs a chemical characterization network and combines the structural characteristics and chemical composition information of perovskite materials to accurately derive the component ratio of each generated sample. This method optimizes the chemical properties of the latent space samples and combines them with the regression model for mapping, which greatly improves the accuracy and precision of the sample component ratio prediction and enhances the depth and effectiveness of the generative model for perovskite material property analysis.

[0120] Embodiment 1:

[0121] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the synthesis laboratory of perovskite materials. The laboratory synthesizes hundreds of perovskite materials with different compositions every month for research in the fields of new energy, photovoltaics, catalysis, etc. However, due to the complex composition of perovskite materials and the presence of diverse composition combinations, traditional chemical analysis methods such as X-ray diffraction (XRD) and energy dispersive spectroscopy (EDS) are often unable to efficiently and accurately derive the composition ratio of the material due to large differences between samples, long test time, and insufficient analysis accuracy.

[0122] In the application scenario of this laboratory, the synthesis process of perovskite samples involves changes in different chemical raw materials and conditions such as temperature and time, which generates a large amount of experimental data and samples. In order to quickly and accurately derive the component ratio of each sample, the traditional method is not only time-consuming, but also has large analysis errors. In order to improve analysis efficiency and reduce errors, the laboratory decided to introduce the perovskite component ratio analysis method based on the deep generative model of the present invention to achieve more accurate and real-time sample component analysis.

[0123] During the implementation process, a large number of perovskite samples were first generated using a deep generative model that combines a variational autoencoder (VAE) with a regularized flow. These samples were dynamically optimized in the latent space, combined with an adaptive latent space adjustment mechanism, so that the latent space samples of each sample can adapt to the characteristics of different perovskite materials. Through this method, the laboratory was able to generate a large number of perovskite samples that conform to the distribution of real materials and derive the specific component ratio information of each sample.

[0124] Next, the laboratory performed weighted optimization of the generated samples using the self-attention mechanism, further improving the similarity and accuracy between samples. Finally, the component ratio information of each sample was extracted from the latent space samples through multi-dimensional reverse inference technology, avoiding the high error problem in traditional analysis methods. According to this method, the laboratory can derive the accurate component ratio of each perovskite sample in real time when conducting synthesis experiments.

[0125] In order to verify the effectiveness of this technology, the laboratory compared the accuracy and efficiency of the method of the present invention with the traditional X-ray diffraction (XRD) analysis method in sample analysis. The following is the comparative data of the component ratio analysis of the perovskite sample before and after implementation:

[0126] Table 1 Comparison of component ratio analysis of perovskite samples

[0127]

[0128] As can be seen from Table 1, after the introduction of the deep generative model analysis method of the present invention, the component ratio analysis error of the sample was reduced from 6.5% of the traditional method to 1.2%, and the analysis accuracy was significantly improved. In addition, the sample analysis time was also greatly shortened, from 45 minutes of the traditional method to 15 minutes, which improved the experimental efficiency. More importantly, in terms of the number of samples processed, the method of the present invention can analyze 200 samples in one experiment, while the traditional method can only process up to 50 samples, which significantly improves the experimental efficiency of the laboratory. At the same time, the repeatability of the experiment has also been optimized, and the consistency of the analysis results has increased by 6 percentage points, from 92% to 98%.

[0129] In order to further verify the application effect of this technology in high-throughput analysis, the laboratory conducted a high-throughput perovskite sample analysis involving about 1,000 perovskite materials of different compositions. The following is a data comparison of the high-throughput experiment:

[0130] Table 2 Comparison of high-throughput perovskite sample analysis data

[0131]

[0132] As can be seen from Table 2, when performing high-throughput analysis, the number of samples processed by the method of the present invention is increased from 200 to 1000, and the total analysis time is shortened from 60 hours to 20 hours, which greatly improves the work efficiency of the laboratory. At the same time, the accuracy of sample component analysis is also significantly improved, and the analysis error is reduced from 5.8% to 1.1%. More importantly, the experimental error is greatly reduced from 4.3% to 0.5%, indicating that the method of the present invention has strong stability and reliability.

[0133] In summary, by introducing the perovskite component ratio analysis method based on the deep generative model of the present invention, the laboratory can achieve efficient and accurate analysis of the component ratio of perovskite materials, significantly improve the experimental efficiency, and maintain high precision and high stability in high-throughput analysis. The successful application of this technology not only solves the defects of traditional analysis methods in analysis accuracy and efficiency, but also provides more advanced tools and methods for the research and application of perovskite materials, which has important application value and promotion prospects.

[0134] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing the ratio of perovskite components based on a deep generative model, characterized in that: The following steps are involved: S1. Use variational autoencoders to build a latent space model of perovskite materials, optimize the model parameters by maximizing the variational lower bound, learn the structural characteristics of perovskite materials, and map them to a low-dimensional latent space. S2. Based on the regularization flow, the latent space generated by the variational autoencoder is transformed in distribution, and the reversible transformation technology is used to adjust the expression ability of the latent space, which specifically includes the following steps: Latent space samples generated by variational autoencoders via regularized flow Transform to obtain new latent space samples , construct a regularized flow model, the transformation is composed of the transformation function of the regularized flow model Completed, including are the trainable parameters of the model: ； During the transformation process, the regularized flow model minimizes the distribution of samples in the latent space Distribution of actual perovskite materials The difference between , by minimizing Divergence measures the similarity of two distributions: ； in, The latent space sample distribution generated by the regularized flow, is the actual sample distribution of perovskite materials, for Divergence; The regularization flow adopts a reversible transformation structure so that each generated latent space sample By inverse transformation function Recovering back to the original latent space sample : ； By introducing a hierarchical structure, the transformation process of the regularized flow adopts a multi-layer transformation function , where each layer transformation function Gradually optimize the expressive power of latent space samples, and complete the adjustment and optimization of latent space samples through multi-layer transformation; In a multi-layer transformation structure, the parameters of the transformation function Automatically adjusted through an optimization process to minimize the latent space distribution of generated samples The actual distribution of perovskite materials The difference between them is calculated by adjusting the parameters of each layer’s transformation function through the back propagation algorithm to achieve the optimal potential space expression capability. S3. Introduce an adaptive latent space adjustment mechanism, combined with the self-attention mechanism, dynamically optimize the representation of the latent space according to the characteristics of the perovskite material, adjust the position and structure of the latent space, and adapt to the changes of different types of perovskite materials; S4. Construct a deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples, and derive the component ratio information of each sample from the latent space through multi-dimensional reverse inference technology; S5. Based on the generated perovskite samples, unsupervised learning techniques are applied to perform data enhancement and expand the training set; S6. Combining the structural characteristics and chemical composition information of the perovskite material, chemical characterization is performed in the latent space of the generated samples to derive the component ratio of each generated sample; S7. Through the error minimization optimization algorithm, the error between the generated samples and the experimental data is minimized, and the parameters of the deep generative model are dynamically adjusted.

2. A method for analyzing the proportion of perovskite components based on a deep generative model according to claim 1, characterized in that: The hierarchical structure in S2 specifically includes: By introducing a hierarchical structure, the transformation process of the regularized flow adopts The transformation function of the layer is gradually optimized , where each layer transformation function Acting on latent space samples , mapping it to a new latent space sample ,in, is the dimension of the latent space samples, It is The output of the layer, It is the final latent space sample. Through multi-layer transformation, the feature representation ability of the latent space sample is gradually optimized, and the representation ability and accuracy of the sample after each layer of transformation are optimized. In the transformation process of each layer, the transformation function Using a reversible transformation structure, the generated latent space samples Can be transformed by inverse function Recovering back to the original latent space sample : ； Transformation function of each layer Has trainable parameters , and adjust the trainable parameters through an optimization process to minimize the final latent space sample Distribution Distribution of actual perovskite materials The difference between: ； in, for Divergence, which is used to measure the difference between two distributions and optimize the parameters of each layer’s transformation function .

3. The method for analyzing the ratio of perovskite components based on a deep generative model according to claim 1, characterized in that: The S3 specifically includes: S31, introduce adaptive latent space adjustment mechanism, combined with self-attention mechanism, in each layer of latent space samples In the dynamic optimization sample The position and structure of the latent space samples are defined to adapt to the characteristics of different types of perovskite materials, calculate the similarity between the latent space samples, and define the latent space samples. and The similarity between ; S32, the self-attention mechanism is through the latent space samples and The similarity between Perform calculations to generate a correlation weight matrix , which represents the latent space samples With other samples The correlation between: ； in, is the weight matrix of the query vector, is the dimension of the latent space; S33. In the self-attention mechanism, according to the calculated relevance weight matrix , a sample of the latent space Will be weighted averaged to update the representation of the sample in the latent space, the latent space sample The update rule is: ； in, is the trainable matrix used for weighting, The latent space samples are updated by the self-attention mechanism. Through the weighted averaging process, the representation ability of the latent space samples is gradually optimized, and the position of the samples in the latent space is adjusted.

4. The method for analyzing the ratio of perovskite components based on a deep generative model according to claim 1, characterized in that: The S4 specifically includes: S41. A deep generative model based on variational autoencoder combined with regularized flow to generate multiple perovskite samples , optimize the generative ability of the model by maximizing the variational lower bound: ； in, Represents a latent space sample The variational posterior distribution of , is the conditional distribution of the generative model, is the prior distribution of the latent space, for Divergence, Express expectations, represents the loss function in the variational autoencoder; S42. Latent space samples generated by variational autoencoders , optimized by regularized flow, through a series of reversible transformations Gradually adjust the latent space samples so that the generated latent space samples is mapped to a new representation , and this transformation can make the distribution of generated samples closer to the actual perovskite material distribution. The optimization goal of each layer of transformation is to minimize the distribution of potential space samples. The distribution of real data Between Divergence: ； S43, based on the optimized latent space sample , using multi-dimensional reverse inference technology to derive the component ratio information from the latent space sample of each sample ,in, Indicates The component ratio of the latent space samples, is a sample in the latent space Middle The ratio of the perovskite components, is the number of perovskite components, a regression model is constructed, and the latent space samples are transformed into The proportion of components mapped to the latent space samples , regression loss function Defined as: ； in, is obtained by reverse inference from the latent space samples The predicted component ratio obtained is is the total number of samples in the latent space; S44. Latent space samples generated by variational autoencoders and regularized flows , and the component ratio derived by multi-dimensional reverse inference technology , achieving accurate analysis of the component proportions of perovskite materials.

5. The method for analyzing the ratio of perovskite components based on a deep generative model according to claim 4, characterized in that: The S6 specifically includes: S61. Construct a chemical characterization network based on the structural characteristics and chemical composition information of perovskite materials By extracting the crystal structure information, elemental composition and chemical bonding characteristics of perovskite materials, a multi-dimensional representation is established, and the latent space sample of each generated sample is sampled through the chemical characterization network. Perform chemical characterization and derive component ratio information of latent space samples; S62. Generate latent space samples in the model Enter the Chemical Characterization Network , through the processing of the chemical characterization network, a latent space sample with chemical characteristics is obtained Based on the actual chemical properties of perovskite materials, the potential space samples Adjust it to more closely match the chemical distribution of actual perovskite samples; S63. Through regression model , the optimized latent space samples The proportion of components mapped to the latent space samples ,in Indicates The component ratio of the latent space samples, and For the The first sample in the latent space The regression model is calculated by minimizing the regression loss function To achieve the prediction of sample component ratio: ； in, is the sample from the optimized latent space through the regression model The predicted component ratios, is the number of samples in the latent space; S64. Latent space samples by combining the outputs of the generative model , Latent space sample after chemical characterization network optimization , and the prediction of the regression model, complete the derivation of the component ratio of the perovskite sample.

Images