An intelligent inverse design method for solid electrolytes integrating graph neural network and confidence analysis

Abstract

The present invention relates to the intersection of material design and artificial intelligence, and is an intelligent inverse design method for solid-state electrolytes that integrates graph neural networks and confidence analysis. The present invention constructs a prediction framework that integrates multiple models, including support vector regression, gradient boosting regression, deep neural networks and graph neural networks, fully integrating the three types of parameter features of components, processes and structures, and efficiently learning the nonlinear mapping relationship between input variables and performance parameters such as resistivity and conductivity. In order to improve credibility, the present invention further introduces Bayesian neural networks and Monte Carlo methods, outputs the confidence interval corresponding to each set of prediction results, and realizes quantitative evaluation of the credibility of the predicted values. In the inverse design module, high-dimensional latent space parameters are generated based on variational autoencoders, and combined with strategies such as Bayesian optimization and genetic algorithms, intelligent recommendation of parameter combinations driven by target performance is realized. The present invention can effectively improve the design efficiency of solid-state electrolytes and the success rate of material discovery.

Description

Technical Field

[0001] The present invention relates to the intersection of material design and artificial intelligence, and in particular to a solid electrolyte intelligent inverse design method that integrates graph neural networks and confidence analysis. Background Art

[0002] Solid-state electrolytes are core materials for a new generation of high-energy-density, high-safety electrochemical energy storage systems, and have broad application prospects in solid-state lithium batteries, sodium batteries, and flexible energy storage devices. Their non-flammability, high interfacial stability, and wide electrochemical window make them more suitable for next-generation high-energy battery systems than traditional liquid electrolytes. Currently, mainstream solid-state electrolyte materials include oxides (such as LLZO), sulfides (such as LGPS), and polymer systems. Their performance is controlled by the coupling of multiple material parameters and process conditions, such as component ratio, sintering temperature, stacking pressure, particle size, and interface control. The highly nonlinear relationship between these parameters makes their design challenging.

[0003] Traditional material performance optimization relies primarily on manual experimental screening and first-principles calculations, which have the following limitations: First, experimental methods are long and costly, making it difficult to achieve rapid global optimization in high-dimensional variable spaces. Second, while numerical simulations (such as DFT and MD) can reveal microscopic mechanisms, their ability to model complex components and multi-scale coupled systems is limited, and they are sensitive to initial assumptions and boundaries, resulting in poor versatility and low engineering conversion efficiency. This is especially true for crystalline solid-state electrolytes with unique topological structures, making it difficult for traditional modeling methods to effectively extract the deep correlation between structural information and performance. Summary of the Invention

[0004] In response to the technical problems raised above, a method for intelligent inverse design of solid-state electrolytes that integrates graph neural networks and confidence analysis is provided. This invention addresses the challenges of performance prediction and optimization of solid-state electrolyte materials by constructing an intelligent system consisting of three core modules: multi-model integrated prediction, result confidence modeling, latent space generation, and goal-oriented optimization. This method implements a full-process data-driven design, from input parameter modeling to performance prediction and reverse parameter recommendation, and possesses high-precision prediction capabilities, interpretable analysis capabilities, and automated inverse design capabilities.

[0005] The technical means adopted in the present invention are as follows:

[0006] A solid electrolyte intelligent inverse design method integrating graph neural network and confidence analysis, including:

[0007] Construct an electrolyte material performance prediction model based on multi-model fusion; train the performance prediction model using standardized input variables; calculate the confidence interval of the material performance prediction model; and construct an intelligent inverse design system based on the latent space generation mechanism combined with a multi-objective optimization method to realize solid electrolyte material screening.

[0008] Furthermore, the performance prediction model includes support vector regression, gradient boosting regression tree, multi-layer perceptron neural network and graph neural network; the support vector regression uses kernel function to map nonlinear input features to high-dimensional space to construct a robust regression hyperplane; the gradient boosting regression tree constructs multiple weak regression trees based on the Boosting integration method, and improves the prediction accuracy through residual iteration; the multi-layer perceptron neural network constructs a multi-layer fully connected deep neural network to extract high-order nonlinear representation of input features; the graph neural network models the connection relationship between atoms as a graph structure for topological information, and extracts a comprehensive representation of the local environment and global structure of atoms through the graph convolution mechanism.

[0009] Furthermore, the input variables include the chemical composition ratio, crystal structure code, sintering temperature, stacking pressure and particle size of the electrolyte material; the performance prediction model is trained using the standardized input variables, and the multiple models fused in the performance prediction model are independently trained in the standardized feature space, and the prediction values ​​of multiple models are weighted averaged and fused:

[0010]

[0011] in, represents the performance prediction value after fusion, Indicates the number of fused sub-models, Indicates the Model pair samples The predicted value of To integrate the weights, they are dynamically adjusted according to the cross-validation error.

[0012] Furthermore, the performance prediction model adopts a standardized training process for model learning. The multiple models fused in the performance prediction model are independently trained on a normalized data set, and cross-validation is used to evaluate performance, and the fusion weight is determined based on the validation error. The data set is a combination of high-throughput experiments and simulation results, and the sample set contains the input feature vector and the corresponding target performance value , to minimize the mean square error To optimize the goal:

[0013]

[0014] in, represents the total number of training samples, is the model prediction function, are the parameters to be optimized. After the training is completed, the fusion module integrates the output results of each sub-model to form a performance prediction result.

[0015] Furthermore, calculating the confidence interval of the material performance prediction specifically includes: calculating the confidence interval of the performance prediction value corresponding to each set of input parameters based on the performance prediction result output by the performance prediction model; giving a certain input parameter combination according to the prediction result output by the performance prediction model Performance prediction value under ,in Represents a set of model parameters and evaluates the confidence bounds of the model predictions:

[0016]

[0017] in, represents the uncertainty width.

[0018] Furthermore, the latent space generation mechanism specifically includes: constructing a parameter generator based on a variational autoencoder, an encoder High-dimensional input Mapping to latent variables Space; Encoder Reconstruct feasible input combinations from the latent space; the latent space follows a normal distribution , the optimization objective function becomes the lower bound:

[0019]

[0020] in, Represents the encoder output distribution The expectation of the latent variable z on represents the Kullback-Leibler divergence between the encoding distribution and the prior distribution, represents the prior distribution of the latent variable.

[0021] Furthermore, the multi-objective optimization method is used to evaluate and screen the sample performance, and the multi-objective optimization method includes: Bayesian optimization, genetic algorithm and reinforcement learning optimization; an optimization feedback mechanism is constructed including candidate screening and confidence assessment, optimization round feedback and parameter reconstruction mechanism; the candidate screening and confidence assessment are used to output the confidence interval CI for each set of generated parameters in the prediction model, and if the predicted fluctuation Δ>5%, the next round of data enhancement is entered; the optimization round feedback dynamically tracks the difference between the optimal solution and the target of each round of optimization, and if it fails to converge for several consecutive rounds, it falls back to the previous solution set; the parameter reconstruction mechanism performs local perturbation or retraining in the latent space to expand the exploration area when the performance of continuously generated samples does not meet the standards.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] This paper proposes an intelligent inverse design method for solid-state electrolytes that integrates graph neural networks and confidence analysis. It constructs a fusion model of SVR, GBR, DNN, and graph neural networks, fully exploring the complex coupling relationships between material components, process parameters, and crystal structure characteristics. Compared to single models, this integrated architecture demonstrates superior multivariate modeling accuracy, generalization, and data adaptability. It is applicable to different types of solid-state electrolyte systems and improves the consistency and reliability of prediction results.

[0024] This paper presents an intelligent inverse design method for solid-state electrolytes that integrates graph neural networks and confidence analysis. By incorporating Bayesian neural networks and the Monte Carlo dropout method, it provides confidence intervals alongside performance predictions, effectively quantifying the uncertainty level of the model output. This mechanism helps researchers assess prediction risks, eliminate unstable solutions, improve the success rate of material screening and experimental verification, and enhance the practical application value of the model.

[0025] This paper proposes an intelligent inverse design method for solid-state electrolytes that integrates graph neural networks and confidence analysis. It uses a variational autoencoder to construct a latent space generative model and integrates Bayesian optimization with a genetic algorithm for performance-driven inverse search, enabling automatic recommendation of optimal design parameters. This strategy balances search accuracy with sample diversity, improving the efficiency of parameter combination exploration while enabling rapid model adaptation to new data. It exhibits excellent scalability and engineering practicality.

[0026] Based on the above reasons, the present invention can be widely promoted in the intersection of material design and artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0028] Figure 1 This is a flow chart of the solid electrolyte intelligent inverse design method that integrates graph neural network and confidence analysis in the present invention.

[0029] Figure 2 This is a diagram of the electrolyte material performance prediction model architecture in the present invention.

[0030] Figure 3 This is a framework diagram of the intelligent reverse design system in the present invention.

[0031] Figure 4 This is a predicted resistivity distribution diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0033] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0035] Unless otherwise specified, the relative arrangement of the parts and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be clear that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to actual proportional relationships. The technology, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be considered as a part of the specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of the exemplary embodiments can have different values. It should be noted that similar numbers and letters represent similar items in the following drawings, and therefore, once an item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0036] like Figure 1 As shown, the present invention provides a solid electrolyte intelligent inverse design method that integrates graph neural network and confidence analysis, including: constructing an electrolyte material performance prediction model based on multi-model fusion; in specific implementation, as a preferred embodiment of the present invention, Figure 2 As shown, the performance prediction models include support vector regression (SVR), gradient boosted regression tree (GBR), multilayer perceptron (MLP) neural network, and graph neural network (GNN). Support vector regression uses kernel functions to map nonlinear input features to a high-dimensional space, constructing a robust regression hyperplane; it is suitable for small and medium-sized samples and strong nonlinear scenarios. Gradient boosted regression tree constructs multiple weak regression trees based on the Boosting ensemble method, improving prediction accuracy through residual iteration; it is suitable for piecewise fitting problems of complex structures. Multilayer perceptron neural network constructs a multi-layer fully connected deep neural network to extract high-order nonlinear representations of input features. Graph neural network models the connectivity between atoms as a graph structure based on topological information. Through the graph convolution mechanism, it extracts a comprehensive representation of the local atomic environment and global structure, effectively improving the prediction's sensitivity to structure.

[0037] The performance prediction model is trained using standardized input variables; in specific implementation, as a preferred embodiment of the present invention, the input variables include the chemical component ratio, crystal structure encoding, sintering temperature, stacking pressure and particle size of the electrolyte material; in implementation, the crystal structure information is input through a graph structure to adapt to the processing method of the graph neural network. The performance prediction model is trained using standardized input variables, supports automatic loading of training data through a pipeline interface, and performs model initialization, training and parameter optimization. The multiple models fused in the performance prediction model are all trained independently in the standardized feature space, and the prediction values ​​of multiple models are weighted averaged and fused:

[0038]

[0039] in, represents the performance prediction value after fusion, Indicates the number of fused sub-models, Indicates the Model pair samples The predicted value of To integrate the weights, they are dynamically adjusted according to the cross-validation error.

[0040] Through model integration, the system can simultaneously utilize the prediction advantages of different models in different parameter regions, and comprehensively improve the accuracy, stability and generalization ability of performance prediction. It is particularly suitable for solid electrolyte systems involving complex coupling relationships between crystal structure and multiple variables. In specific implementation, as a preferred embodiment of the present invention, in order to ensure that each model has good generalization ability and stability, the performance prediction model adopts a standardized training process for model learning. The multiple models fused in the performance prediction model are independently trained on a normalized data set, and 5-fold cross-validation is used to evaluate performance, and the fusion weight is determined based on the validation error;

[0041] The dataset is a combination of high-throughput experiments and simulation results. The sample set contains the input feature vector and the corresponding target performance value , to minimize the mean square error To optimize the goal:

[0042]

[0043] in, represents the total number of training samples, is the model prediction function, For the parameters to be optimized, the DNN model uses the Adam optimizer, while the GNN model uses the graph convolution update rule. Dropout is also added during training to prevent overfitting. After training is complete, the fusion module integrates the output of each sub-model to form a performance prediction result.

[0044] In the task of predicting the performance of solid electrolytes, in order to improve the practicality and reliability of the model output, the present invention not only provides the predicted value of the material performance, but also outputs its uncertainty range synchronously, and constructs an information expression containing confidence. Calculate the confidence interval of the material performance prediction of the performance prediction model; in specific implementation, as a preferred embodiment of the present invention, the confidence interval of the material performance prediction specifically includes: based on the performance prediction results output by the performance prediction model, calculate the confidence interval of the performance prediction value corresponding to each set of input parameters; the prediction result output by the performance prediction model gives a certain input parameter combination Performance prediction value under ,in Represents a set of model parameters and evaluates the confidence bounds of the model predictions:

[0045]

[0046] in, Represents the uncertainty width. The introduction of confidence intervals helps users identify design parameters with high volatility risk, improves robustness during the inverse design process, and provides a decision-making reference for subsequent experimental resource allocation. This paper uses Monte Carlo Dropout and Bayesian neural networks as examples to calculate confidence intervals.

[0047] Based on the Monte Carlo Dropout confidence estimation method, the Monte Carlo Dropout mechanism is introduced in deep neural network and graph neural network models. Unlike conventional training, this method retains the Dropout activation state during the inference phase and performs multiple forward propagation operations to obtain a set of prediction value samples:

[0048]

[0049] Calculate the mean and standard deviation for this group of samples:

[0050]

[0051] According to the given confidence level , the corresponding confidence interval is defined as:

[0052]

[0053] in, is the standard quantile under a normal distribution (e.g., 1.96 corresponds to 95%). This method requires no additional model structure modification and is applicable to all deep models with dropout architecture. It effectively captures prediction fluctuations caused by incomplete training data and is the preferred lightweight uncertainty modeling approach for engineering deployments.

[0054] Based on Bayesian Neural Network (BNN), considering that the performance data of solid electrolytes may be scarce or unstable in certain parameter ranges, in order to further enhance the modeling ability of the prediction model for epistemic uncertainty, a Bayesian Neural Network (BNN) is introduced. Assuming the model parameters Obey the prior distribution and update it to the posterior distribution based on the historical data set For any input x, the performance prediction is given by the sampling average under the distribution of the model parameters:

[0055]

[0056] By sampling the posterior distribution multiple times (e.g. 100 times):

[0057]

[0058] Generate corresponding output , construct the confidence interval:

[0059]

[0060] This method is more robust to low sample and high uncertainty areas and is suitable for high-precision fault-tolerant scenarios such as material screening in the R&D stage. As shown in Table 1, the Monte Carlo Dropout or Bayesian neural network calculation Typical ranges are:

[0061] Table 1 Calculation based on Monte Carlo Dropout or Bayesian neural network Typical range

[0062]

[0063] In order to realize the reverse design function of automatically obtaining material parameter combinations based on target performance, the present invention constructs an intelligent reverse design system that integrates latent space generation and multi-objective optimization based on multi-model prediction and confidence analysis. Figure 3 As shown in Figure 1, the system can efficiently search for the optimal design combination within the parameter space based on set performance indicators (such as resistivity and conductivity), achieving a closed-loop optimization from "objective to material design." An intelligent inverse design system is constructed based on a latent space generation mechanism combined with multi-objective optimization methods to enable solid electrolyte material screening.

[0064] In specific implementation, as a preferred embodiment of the present invention, the latent space generation mechanism specifically includes: In the design of solid electrolyte material combinations, the parameter space dimension is high, the coupling between variables is complex, and the traditional rule search efficiency is low. In order to improve the search intelligence, a parameter generator based on variational autoencoder is constructed. High-dimensional input Mapping to latent variables Space; Encoder Reconstruct feasible input combinations from the latent space; the latent space follows a normal distribution , the optimization objective function becomes the lower bound:

[0065]

[0066] in, Represents the encoder output distribution The expectation of the latent variable z on represents the Kullback-Leibler divergence between the encoding distribution and the prior distribution, Represents the prior distribution of the latent variable. Through this mechanism, the system can quickly generate feasible parameter candidates that satisfy historical physical laws, avoiding the problem of invalid combinations caused by brute-force search and improving the efficiency of reverse engineering generation. In specific implementation, as a preferred embodiment of the present invention, after generating parameter combinations, multi-objective optimization methods are used to evaluate and screen sample performance. Multi-objective optimization methods include Bayesian optimization, genetic algorithms, and reinforcement learning optimization.

[0067] Bayesian Optimization: Building a Surrogate Function , use an acquisition function (such as EI or UCB) to guide the sampling of new candidates:

[0068]

[0069] Genetic Algorithm (GA): Encodes candidate parameters as individual chromosomes, performs operations such as crossover, mutation, and fitness selection, and optimizes the target value generation by generation:

[0070]

[0071] Reinforcement learning optimization: Using the objective function as the reward signal, we learn the optimal strategy path from the input space to the parameter combination. The optimization objectives are as follows:

[0072]

[0073] in, Predict resistivity for the model, is the cost function, is the feasible domain.

[0074] To ensure the effectiveness of generated parameters and the stability of model search, an optimization feedback mechanism is constructed, including candidate screening and confidence assessment, optimization round feedback, and parameter reconstruction mechanism. Candidate screening and confidence assessment are used to output the confidence interval CI for each set of generated parameters in the prediction model. If the prediction fluctuation Δ>5%, the next round of data enhancement is entered. The optimization round feedback dynamically tracks the difference between the optimal solution and the target in each round of optimization. If convergence fails for several consecutive rounds, it falls back to the previous solution set. The parameter reconstruction mechanism performs local perturbations or retraining in the latent space to expand the exploration area when the performance of continuously generated samples does not meet the standards.

[0075] The system supports online learning: When new experimental data is input, it can selectively update the parameters of some sub-models within the prediction model, avoiding full retraining. This module is based on an incremental training strategy, maintaining original performance while improving adaptability to new data distributions, supporting the long-term evolution of the model and closed-loop optimization capabilities.

[0076] Example

[0077] like Figure 1 As shown, the present invention provides a solid electrolyte intelligent inverse design method that integrates graph neural networks and confidence analysis. To verify the effectiveness of the system described in the present invention, the following uses a specific example of oxide solid electrolyte performance optimization to demonstrate the complete process from target performance setting to parameter inverse design output:

[0078] (1) Problem definition and goal setting: The goal is to design a class of oxide solid electrolyte materials with Li, La, and Zr as the main elements, requiring that the resistivity at room temperature is not higher than , and meet the following process constraints:

[0079] Sintering temperature ≤ 1200°C; stacking pressure ≥ 150MPa; average particle size range: [1,10]μm.

[0080] (2) Input parameter range definition

[0081] The design parameters are shown in Table 2:

[0082] Table 2 Design parameters

[0083]

[0084] All parameters are normalized and input into the trained multi-model fusion system.

[0085] (3) Model prediction and uncertainty assessment

[0086] The system first performs weighted fusion on the SVR, GBR, DNN, GNN and other models constructed with historical sample data. , output its predicted resistivity And the confidence interval:

[0087]

[0088] in, Derived from MC-Dropout repeated prediction or BNN output variance estimation.

[0089] (4) Reverse design process

[0090] The system is optimized using the following combination:

[0091] Generative model: Use the VAE model to generate 1,000 sets of parameter candidates in the latent space;

[0092] Performance evaluation: The multi-model fusion system predicts resistivity and confidence intervals for each set of parameters;

[0093] Optimizer: Use Bayesian optimizer and EI acquisition function to select the optimal solution that meets the constraints from the candidates.

[0094] (5) Optimal parameter recommendation

[0095] Finally, several groups of recommended parameters are output, one of which is shown in Table 3:

[0096] Table 3 Recommended parameters

[0097]

[0098] The distribution diagram of the predicted resistivity is as follows Figure 4 shown.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A solid electrolyte intelligent inverse design method integrating graph neural network and confidence analysis, characterized in that: include: Construct an electrolyte material performance prediction model based on multi-model fusion; the performance prediction model includes support vector regression, gradient boosting regression tree, multi-layer perceptron neural network and graph neural network; the support vector regression uses kernel function to map nonlinear input features to high-dimensional space to construct a robust regression hyperplane; the gradient boosting regression tree constructs multiple weak regression trees based on the Boosting ensemble method, and improves prediction accuracy through residual iteration; the multi-layer perceptron neural network constructs a multi-layer fully connected deep neural network to extract high-order nonlinear representations of input features; the graph neural network models the connection relationship between atoms as a graph structure based on topological information, and extracts a comprehensive representation of the local environment and global structure of atoms through the graph convolution mechanism; The performance prediction model is trained using standardized input variables, wherein the input variables include chemical component ratios, crystal structure codes, sintering temperatures, stacking pressures, and particle sizes of the electrolyte material; The performance prediction model is trained using standardized input variables. The multiple models fused in the performance prediction model are trained independently in the standardized feature space, and the prediction values ​​of multiple models are weighted averaged and fused: in, represents the performance prediction value after fusion, Indicates the number of fused sub-models, Indicates the Model pair samples The predicted value of To integrate weights, dynamically adjust according to cross-validation error; Calculating the confidence interval of the material property prediction model; An intelligent inverse design system is constructed based on the latent space generation mechanism combined with a multi-objective optimization method to realize the screening of solid electrolyte materials; The latent space generation mechanism specifically includes: constructing a parameter generator based on a variational autoencoder, an encoder High-dimensional input Mapping to latent variables Space; Encoder Reconstruct feasible input combinations from the latent space; the latent space follows a normal distribution , the optimization objective function becomes the lower bound: in, Represents the encoder output distribution The expectation of the latent variable z is, represents the Kullback-Leibler divergence between the encoding distribution and the prior distribution, represents the prior distribution of the latent variable.

2. The solid electrolyte intelligent inverse design method integrating graph neural network and confidence analysis according to claim 1 is characterized in that: The performance prediction model adopts a standardized training process for model learning. The multiple models fused in the performance prediction model are independently trained on a normalized dataset, cross-validation is used to evaluate performance, and the fusion weight is determined based on the validation error. The dataset is a combination of high-throughput experiments and simulation results. The sample set contains the input feature vector Corresponding target performance value , to minimize the mean square error To optimize the goal: in, represents the total number of training samples, is the model prediction function, are the parameters to be optimized. After the training is completed, the fusion module integrates the output results of each sub-model to form a performance prediction result.

3. The solid electrolyte intelligent inverse design method integrating graph neural network and confidence analysis according to claim 1 is characterized in that: Calculating the confidence interval for the material property prediction specifically includes: Based on the performance prediction results output by the performance prediction model, the confidence interval of the performance prediction value corresponding to each set of input parameters is calculated; the prediction results output by the performance prediction model give a certain input parameter combination Performance prediction value under ,in Represents a set of model parameters and evaluates the confidence bounds of the model predictions: in, represents the uncertainty width.

4. The solid electrolyte intelligent inverse design method integrating graph neural network and confidence analysis according to claim 1 is characterized in that: The multi-objective optimization method is used to evaluate and screen sample performance, and the multi-objective optimization method includes: Bayesian optimization, genetic algorithm and reinforcement learning optimization; an optimization feedback mechanism is constructed including candidate screening and confidence assessment, optimization round feedback and parameter reconstruction mechanism; the candidate screening and confidence assessment are used to output a confidence interval CI for each set of generated parameters in the prediction model, and if the predicted fluctuation Δ>5%, the next round of data enhancement is entered; the optimization round feedback dynamically tracks the difference between the optimal solution and the target of each round of optimization, and if convergence fails for several consecutive rounds, it falls back to the previous solution set; the parameter reconstruction mechanism performs local perturbation or retraining in the latent space to expand the exploration area when the performance of continuously generated samples does not meet the standards.

Images