Method, model training method, system and device for predicting solid-state electrolyte performance

Abstract

This disclosure provides a method, model training method, system, and device for predicting the performance of solid electrolytes. The training method includes: acquiring a first dataset and a second dataset; training a first pre-defined model on the first dataset to obtain a first prediction model; using the first prediction model to perform initial pseudo-label prediction on the second dataset; selecting target pseudo-labels that meet the pseudo-label selection criteria from the initial pseudo-labels to construct an extended training set; training a second pre-defined model on the extended training set to obtain a second prediction model; obtaining the residual distribution of the second prediction model on the first and second datasets, updating the key feature subset and pseudo-label selection criteria, and iteratively training the first and second prediction models to obtain a solid electrolyte performance prediction model. This disclosure can significantly reduce performance prediction errors and decrease reliance on experimental / computational labels when labeled data is limited.

Description

Technical Field

[0001] This disclosure relates to the fields of energy science, materials science and artificial intelligence, and is specifically applicable to high-throughput screening and performance evaluation of solid electrolyte materials for all-solid-state batteries. In particular, it relates to a method for predicting the performance of solid electrolytes, a model training method, a system and a device. Background Technology

[0002] All-solid-state batteries are considered an important development direction for next-generation energy storage systems due to their advantages such as high energy density, strong safety, and good interface stability. As the core component of all-solid-state batteries, the solid electrolyte's ionic conductivity, activation energy, electrochemical window, and interface stability directly determine the overall performance of the battery.

[0003] In recent years, machine learning technology has been gradually applied to the field of materials performance prediction. However, existing methods are still mainly supervised learning, and their stable and reliable regression models highly depend on large-scale labeled data for training. However, research on solid electrolytes faces practical bottlenecks such as the difficulty and high cost of experimental testing, resulting in an extreme scarcity of high-quality labeled data. Consequently, traditional supervised learning methods that rely on massive labeled samples are difficult to directly apply to the prediction of solid electrolyte performance. Summary of the Invention

[0004] The technical problem to be solved by this disclosure is to overcome the shortcomings of existing supervised learning methods that rely on massive labeled samples and are difficult to apply directly to the prediction of solid electrolyte performance, and to provide a method for predicting solid electrolyte performance, a model training method, a system and a device.

[0005] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0006] A first aspect of this disclosure provides a training method for a solid electrolyte performance prediction model, the training method comprising:

[0007] Obtain a first dataset and a second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data.

[0008] Several key features are extracted from the first training sample to form a key feature subset, and the first preset model is trained based on the key feature subset to obtain the first prediction model;

[0009] The first prediction model is used to predict the second training samples in the second dataset to obtain the initial pseudo-label corresponding to each second training sample;

[0010] Target pseudo-labels that meet the pseudo-label selection criteria are selected from the initial pseudo-labels, and an expanded training set is constructed based on the first dataset, the target pseudo-labels and their corresponding second training samples;

[0011] The second preset model is trained based on the third training samples in the extended training set to obtain the second prediction model;

[0012] Obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, update the key feature subset and / or the pseudo-label screening conditions, and return to execute the step of training the first preset model based on the key feature subset until the residual distribution meets the preset error range or reaches the preset stopping condition, and use the second prediction model as the solid electrolyte performance prediction model.

[0013] Optionally, the step of extracting several key features from the first training samples to form a key feature subset, and training the first preset model based on the key feature subset to obtain the first prediction model includes:

[0014] A first global view feature is extracted from the first training sample, and the first global view feature includes at least composition features, structural features and process features;

[0015] Using sequential forward selection or regularized regression, several key features are selected from the first global view features to form a subset of the key features of the first local view features;

[0016] The first preset model is trained based on the key feature subset to obtain the first prediction model.

[0017] Optionally, the step of using the first prediction model to predict the second training samples in the second dataset includes:

[0018] Extract the second local view features from the second training sample and input the second local view features into the first prediction model for prediction;

[0019] And / or,

[0020] The step of training the second preset model based on the third training samples in the expanded training set includes:

[0021] The second global view features are extracted from the third training samples to train the second prediction model, and the parameters of the second prediction model are updated by minimizing the joint loss function.

[0022] The joint loss function includes supervision by labeled performance labels, supervision by target pseudo-labels, and logical consistency constraints between the first prediction model and the second prediction model.

[0023] Optionally, the step of filtering target pseudo-tags that meet the pseudo-tag filtering conditions from the initial pseudo-tags includes:

[0024] Obtain the prediction residual of the first prediction model on the second dataset, and filter the first pseudo-labels from the initial pseudo-labels whose prediction residuals are less than a first preset threshold; or, obtain the output distribution of the first prediction model on the second dataset, and in response to the output distribution following a Gaussian distribution, filter the first pseudo-labels from the initial pseudo-labels whose variance is less than a second preset threshold.

[0025] Based on the mean of pseudo-label predictions from multiple models, the prediction dispersion of the first pseudo-label is obtained, and a second pseudo-label with a prediction dispersion less than a third preset threshold is selected.

[0026] The pseudo-label prediction mean is obtained based on multiple different third prediction models. These different third prediction models have the same model type as the first prediction model and are trained based on different third local view features. The third local view features are subsets of the key feature subset.

[0027] Obtain the first proportion of each second pseudo-label in each performance prediction interval, and filter out the third pseudo-labels in the same performance prediction interval whose deviation between the first proportion and the second proportion is less than a fourth preset threshold, as the target pseudo-labels;

[0028] The second percentage is obtained by dividing the first dataset according to each of the performance prediction intervals.

[0029] Optionally, the step of updating the key feature subset and the pseudo-label filtering conditions includes:

[0030] Based on the feature importance index of the second prediction model, the contribution of each input feature to the prediction result of solid electrolyte performance is obtained;

[0031] The key feature subset is updated based on the contribution level;

[0032] And / or,

[0033] The preset stopping conditions include at least one of the following: the model validation error decreases by less than a fifth preset threshold between two iterations; the size of the expanded training set no longer changes; or the maximum number of iterations is reached.

[0034] A second aspect of this disclosure provides a method for predicting the performance of a solid electrolyte, the method comprising:

[0035] Acquire data of the solid electrolyte under test;

[0036] Extract the global view features of the solid electrolyte data to be tested;

[0037] The global view features to be tested are input into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested.

[0038] The solid electrolyte performance prediction model is obtained using the training method described in the first aspect of this disclosure.

[0039] A third aspect of this disclosure provides a training system for a solid electrolyte performance prediction model, the training system comprising:

[0040] The first acquisition module is used to acquire a first dataset and a second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data.

[0041] The first training module is used to extract several key features from the first training samples to form a key feature subset, and to train the first preset model based on the key feature subset to obtain the first prediction model.

[0042] The pseudo-label generation module is used to predict the second training samples in the second dataset using the first prediction model to obtain the initial pseudo-label corresponding to each second training sample.

[0043] The filtering module is used to filter target pseudo-labels that meet the pseudo-label filtering conditions from the initial pseudo-labels, and to construct an extended training set based on the first dataset, the target pseudo-labels and their corresponding second training samples;

[0044] The second training module is used to train the second preset model based on the third training samples in the extended training set to obtain the second prediction model.

[0045] An iterative module is used to obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, the key feature subset and / or the pseudo-label screening conditions are updated, and the first training module is called back until the residual distribution meets the preset error range or reaches the preset stopping condition, and the second prediction model is used as the solid electrolyte performance prediction model.

[0046] A fourth aspect of this disclosure provides a system for predicting the performance of a solid electrolyte, the system comprising:

[0047] The data acquisition module is used to acquire data of the solid electrolyte under test.

[0048] The extraction module is used to extract the global view features of the solid electrolyte data to be tested;

[0049] The prediction module is used to input the global view features to be tested into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested.

[0050] The solid electrolyte performance prediction model is obtained using the training system described in the third aspect of this disclosure.

[0051] A fifth aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement the training method of the first aspect of this disclosure, or the prediction method of the second aspect of this disclosure.

[0052] A sixth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the training method described in the first aspect of this disclosure, or the prediction method described in the second aspect of this disclosure.

[0053] A seventh aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the training method described in the first aspect of this disclosure, or the prediction method described in the second aspect of this disclosure.

[0054] Based on common knowledge in the field, the above optional conditions can be combined arbitrarily to obtain the optional examples of this disclosure.

[0055] The positive advancements of this disclosure are as follows: By introducing a first prediction model trained on a subset of key features for pseudo-label generation and iterative optimization, statistical regularities in a large amount of unlabeled solid electrolyte data are effectively mined, significantly reducing performance prediction errors when labeled data is limited, and reducing reliance on experimental / computational labels. Simultaneously, by utilizing pseudo-label selection criteria specific to solid electrolytes, the iterative process of pseudo-labels follows material physics constraints, avoiding the introduction of unreasonable pseudo-labeled samples and improving the credibility of the second prediction model in crystal structure space. Furthermore, through an interpretable second prediction model, not only can the performance prediction results of solid electrolytes be output, but the contribution of each input feature to the performance prediction results can also be quantified, providing quantitative evidence for understanding ion migration mechanisms and optimizing material design. Attached Figure Description

[0056] Figure 1 This is a first flowchart of the training method for the solid electrolyte performance prediction model disclosed herein;

[0057] Figure 2 This is a second flowchart of the training method for the solid electrolyte performance prediction model disclosed herein;

[0058] Figure 3 This is a schematic diagram illustrating the structure of the feature system disclosed herein;

[0059] Figure 4 This is a schematic diagram of the dual-model collaborative iterative training framework disclosed herein;

[0060] Figure 5 This is a flowchart of the method for predicting the performance of solid electrolytes disclosed herein;

[0061] Figure 6 This is a schematic diagram of the training system module for the solid electrolyte performance prediction model disclosed herein;

[0062] Figure 7 This is a schematic diagram of the solid electrolyte performance prediction system module disclosed herein;

[0063] Figure 8 This is a schematic diagram of the structure of an electronic device disclosed herein. Detailed Implementation

[0064] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0065] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not be construed as an unnecessary limitation. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0066] The performance of solid electrolytes is highly dependent on the crystal structure characteristics of the material (such as multi-anion frameworks, migration channel dimensions, and coordination environment), chemical composition (such as cation / anion ratio and electronegativity differences), and preparation process (such as sintering temperature and defect concentration). Furthermore, a large number of solid electrolyte candidate materials lacking performance labels but possessing complete structural information exist in publicly available structural databases, literature, and high-throughput computing platforms, resulting in a data pattern of "few labeled samples and many unlabeled samples."

[0067] In existing technologies, the performance evaluation of solid electrolytes mainly relies on experimental synthesis and testing, large-scale first-principles calculations, etc., which have the drawbacks of long cycle time, high cost, and difficulty in covering the vast space of material structure combinations. In terms of machine learning applications, supervised learning models are mostly used to train models through a large number of labeled samples to achieve performance prediction. However, due to the difficulty and high cost of experimental testing of solid electrolytes (such as the complexity and time-consuming steps of high-temperature sintering, tableting, electrochemical impedance spectroscopy, etc.), the high cost of simulation calculation (such as the calculation of ion migration paths, activation energies, diffusion coefficients, etc., which rely on a large number of first-principles or molecular dynamics simulations), and the lack of uniformity of data sources (such as the difficulty of cross-dataset alignment due to differences in experimental conditions in literature), high-quality labeled data for solid electrolytes is extremely limited. As a result, traditional supervised learning methods that rely on massive labeled samples are difficult to directly apply to the performance prediction of solid electrolytes.

[0068] In view of this, the present disclosure provides a method for predicting the performance of solid electrolytes, a model training method, a system, and an apparatus to solve the above problems.

[0069] Example 1

[0070] In a specific embodiment of this disclosure, a training method for a solid electrolyte performance prediction model is provided, such as... Figure 1 As shown, the training method includes:

[0071] S11. Obtain the first dataset and the second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data.

[0072] S12. Extract several key features from the first training sample to form a key feature subset, and train the first preset model based on the key feature subset to obtain the first prediction model.

[0073] S13. Use the first prediction model to predict the second training samples in the second dataset to obtain the initial pseudo-labels corresponding to each second training sample;

[0074] S14. Select target pseudo-labels that meet the pseudo-label selection criteria from the initial pseudo-labels, and construct an expanded training set based on the first dataset, the target pseudo-labels and their corresponding second training samples.

[0075] S15. The second preset model is trained based on the third training sample in the extended training set to obtain the second prediction model;

[0076] S16. Obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, update the key feature subset and / or pseudo-label screening conditions, and return to the step of training the first preset model based on the key feature subset until the residual distribution meets the preset error range or reaches the preset stopping condition, and use the second prediction model as the solid electrolyte performance prediction model.

[0077] The preset stopping conditions include at least one of the following: the model validation error decreases by less than a fifth preset threshold between two iterations; the size of the expanded training set no longer changes; or the maximum number of iterations is reached.

[0078] Specifically, step S11 obtains a first dataset and a second dataset. The first dataset includes solid electrolyte samples with input features and performance labels. The performance labels include at least one or more of ionic conductivity, activation energy, and electrochemical stability to meet the multi-dimensional performance prediction requirements of solid electrolytes, and there is a high degree of coupling between the performance dimensions. The second dataset includes solid electrolyte candidate samples with input features but without performance labels.

[0079] The performance labels in the first dataset can be obtained from experimental test data of solid electrolyte samples, simulation data obtained from first-principles calculations, molecular dynamics or ion migration path calculations, or performance data extracted from public literature by large models and manually verified. The solid electrolyte candidate samples in the second dataset can be from crystal structures in materials databases, or derived structures generated from known structures through element substitution, defect construction or superstructure design, or candidate structures automatically generated through structure generation models or high-throughput computing processes.

[0080] Of course, in order to improve data quality to ensure model performance and improve training efficiency, the first and second datasets can be cleaned, missing values ​​handled and normalized based on crystal structure files, stoichiometric composition, doping information, preparation process parameters and electrochemical test conditions. This specific implementation method does not limit this.

[0081] By unifying the management of a small amount of labeled solid electrolyte performance data obtained from existing experimental tests, electrochemical simulations, and literature extraction with a large amount of unlabeled solid electrolyte data from materials databases, high-throughput calculations, or structure generation processes, a solid electrolyte sample library consisting of a first dataset and a second dataset is formed.

[0082] Step S12 extracts several key features from each first training sample to form a key feature subset, and trains the first preset model based on the key feature subset and its corresponding performance label to obtain the first prediction model.

[0083] The first preset model can adopt a general interpretable regression model, such as a linear model, sparse regression, ensemble decision tree, neural network and other mainstream models. The core mapping relationship of the model is represented as f1(Z), as shown in formula (1):

[0084] (1)

[0085] in, This represents the first preset model. Represents the j-th feature or its basis function mapping. The parameter representing the feature of the j-th local view. Represents a parameter vector. Let P represent the feature matrix of the local view, where P represents the number of key features and N represents the number of samples.

[0086] It should be noted that the parameter forms of different models are adapted to the corresponding types. For example, the linear model is... The tree model uses the parameters of the splitting nodes, and the neural network uses the weight matrix. This is a general parameter symbol.

[0087] After training and obtaining the first prediction model, pseudo-label prediction is performed on the unlabeled solid electrolyte data in the second dataset in step S13 to obtain the initial pseudo-label corresponding to each unlabeled solid electrolyte data. Performance labels are then assigned to the unlabeled second training samples using the first prediction model to expand the training set size.

[0088] Next, in step S14, high-quality pseudo-labels are selected from the initial pseudo-labels as target pseudo-labels, and pseudo-label samples are formed with the second training samples corresponding to the target pseudo-labels to obtain a candidate sample set. The first dataset and the candidate sample set are then merged to form an extended training set.

[0089] Step S15 uses the third training samples in the extended training set to train the second preset model to obtain the second prediction model. The third training samples include the first training samples and pseudo-label samples.

[0090] Step S16 obtains the residual distribution of the second prediction model on the first and second datasets, updates the key feature subset and / or pseudo-label selection criteria to update the first prediction model, forming a consistency constraint and mutual promotion mechanism between the first and second prediction models. When the residual distribution meets the preset error range, or the model validation error decreases less than the fifth preset threshold between two iterations, or the size of the expanded training set no longer changes, or the maximum number of iterations is reached, a converged second prediction model is obtained. This second prediction model is the solid electrolyte performance prediction model and can be used for solid electrolyte performance prediction.

[0091] For example, based on the residual distribution of the second prediction model in the first and second datasets, and using the root mean square error (RMAE) as the error evaluation index of the residual distribution, the preset error range is determined to be [0, 0.1]. Then, the RMSE is compared with the preset error range. If RMSE ≤ 0.10, the training of the second prediction model is considered complete. If RMSE > 0.10, key features in the key feature subset are added, reduced, or replaced, and the pseudo-label selection conditions are adjusted. Then, the first preset model is retrained based on the updated key feature subset, and the target pseudo-label is selected using the adjusted pseudo-label selection conditions. A new extended training set is obtained to train the second preset model. After multiple iterations, the converged second prediction model is obtained.

[0092] like Figure 2 As shown, by acquiring a first dataset with labeled performance tags and a second dataset without labels, features are constructed and preprocessed on the first and second datasets. Based on the constructed features, a first prediction model is trained to generate initial pseudo-labels for the unlabeled data. The initial pseudo-labels are then evaluated and screened to select high-quality target pseudo-label samples to form an extended training set with the first dataset. The second prediction model is then trained based on the extended training set and iteratively optimized to finally obtain a converged solid electrolyte performance prediction model.

[0093] This specific implementation effectively mines statistical patterns in a large number of unlabeled solid electrolyte structures through pseudo-label generation and iterative optimization. It significantly reduces performance prediction errors with limited labeled data, decreases reliance on experimental / computational labels, and effectively improves data utilization efficiency. Compared to traditional experimental and simulation methods, it can shorten the material screening cycle by more than 60% and reduce R&D costs by more than 50%, demonstrating significant time and economic cost advantages. Without altering the fundamental principles of existing semi-supervised algorithms, it fully utilizes the statistical patterns of a large number of unlabeled samples, further enhancing its cost advantage.

[0094] In one specific embodiment, step S12 includes:

[0095] S121. Extract the first global view features from the first training samples. The first global view features include at least component features, structural features, and process features.

[0096] S122. Using sequential forward selection or regularized regression, select several key features from the first global view features to form a key feature subset of the first local view features;

[0097] S123. Train the first preset model based on the key feature subset to obtain the first prediction model.

[0098] Specifically, component features, structural features, and process features are extracted from the first training sample, such as... Figure 3 As shown, a feature vector is constructed that includes compositional, structural, and process characteristics as the first global view feature. Compositional characteristics can include one or more statistical features based on composition, such as element type, mole fraction, average ionic radius, and electronegativity difference. Structural characteristics can include one or more geometric features based on crystal structure, such as lattice constant, volume, coordination environment, and geometric indices of ion migration channels. Process characteristics can include one or more features based on process and testing conditions, such as sintering temperature, atmosphere, pressure, current density, testing temperature, and frequency. By designing a first global view feature that includes structural, compositional, and process characteristics for solid-state electrolytes, the problem of pseudo-label iteration updates deviating from the physical laws of materials can be effectively avoided.

[0099] Let X be the feature matrix of the complete view of the first dataset, containing all features such as composition, structure, and process. L The true performance label vector is y L Feature selection methods such as sequential forward selection or regularized regression are used to select features from X. L Key features are selected from the data to form the feature matrix Z of the local view. L (Z) L For X L The key feature subset of the subset (of which), the local view feature matrix Z LThe dimension P < M, where M represents the number of complete features. Sequential forward selection starts from an empty set, and each time the feature that maximizes the performance improvement of the first prediction model is added to the key feature subset, until the stopping condition is met.

[0100] Z based on the first dataset L With y L The model parameters of the first preset model are trained by minimizing the loss to obtain the first prediction model. The basic loss form is shown in formula (2):

[0101] (2)

[0102] in, Indicates the weight of the monitoring loss. Represents the L2 norm, This represents the model prediction value of the first prediction model.

[0103] Of course, you can also add corresponding regularization or penalty terms according to the model type to reduce the risk of overfitting. For example, for linear models, you can add... Examples include L1 regularization, complexity penalty terms for tree models, and weight decay terms for neural networks.

[0104] It should be noted that the first preset model f1 (·) can be flexibly selected from general models such as linear models, ensemble decision trees, and neural networks, as long as the fitting stability of the key feature subset is guaranteed.

[0105] This specific implementation uses a stable mapping relationship learned from a small amount of labeled data to assign reliable pseudo-label references to unlabeled samples. It constructs a multi-level feature system for solid electrolytes, incorporating compositional, structural, and process characteristics. Through elemental statistical features based on composition, geometric and topological features based on crystal structure, and process features reflecting preparation and testing conditions, it achieves physical constraints on structural type and conductivity distribution. This explicitly encodes key factors affecting ion migration channels, conductivity, and interface stability during pseudo-label generation and iterative optimization. During iterative updates, it ensures that pseudo-labels remain within a reasonable range to conform to material physics laws, avoiding the introduction of unreasonable pseudo-label samples. This improves the credibility of the first prediction model in crystal structure space, solves the problem of poor adaptability in existing semi-supervised techniques, and effectively enhances the generalization ability of the first prediction model.

[0106] In one specific implementation, step S13 includes: extracting the second local view features of the second training sample, and inputting the second local view features into the first prediction model for prediction.

[0107] Specifically, local view features corresponding to the key feature subset are extracted from the second training samples, and the initial pseudo-labels corresponding to the second training samples are predicted using the trained first prediction model.

[0108] Let Z be the local view feature matrix of the second dataset. U Initial pseudo-labels are generated using the first prediction model f1(·) after training. The mathematical expression of the initial pseudo-labels is as follows: ,in, This represents the initial pseudo-label vector of the second training sample in the second dataset.

[0109] This specific implementation introduces a first prediction model based on a subset of key features, uses a small amount of labeled data to learn a relatively stable and interpretable mapping relationship, and then generates pseudo-labels for unlabeled samples. This significantly reduces performance prediction error and decreases reliance on experimental / computational labels when labeled data is limited.

[0110] In one specific embodiment, step S14 includes:

[0111] S141. Obtain the prediction residual of the first prediction model on the second dataset, and filter the first pseudo-labels from the initial pseudo-labels whose prediction residuals are less than the first preset threshold; or, obtain the output distribution of the first prediction model on the second dataset, and in response to the output distribution following a Gaussian distribution, filter the first pseudo-labels from the initial pseudo-labels whose variance is less than the second preset threshold.

[0112] S142. Based on the mean of the pseudo-label prediction of multiple models, obtain the prediction dispersion of the first pseudo-label, and filter the second pseudo-label whose prediction dispersion is less than the third preset threshold.

[0113] The pseudo-label prediction mean is obtained based on multiple different third prediction models. These different third prediction models have the same model type as the first prediction model and are trained based on different third local view features. The third local view features are subsets of the key feature subsets.

[0114] S143. Obtain the first proportion of each second pseudo-label in each performance prediction interval, and filter out the third pseudo-labels whose deviation between the first proportion and the second proportion in the same performance prediction interval is less than the fourth preset threshold, and use them as target pseudo-labels.

[0115] The second percentage is based on the first dataset divided into different performance prediction intervals.

[0116] Specifically, to avoid the negative impact of mislabeling, the confidence level of the initial pseudo-labels is first evaluated based on metrics such as the prediction variance and residuals of the first prediction model, in order to select the first pseudo-labels. For example, suppose the prediction residuals of the first prediction model on the second dataset are... Then filter the predicted residuals Initial pseudo-labels that are less than the first preset threshold are used as the first pseudo-labels with high confidence.

[0117] Confidence intervals can also be defined based on the probability distribution of the output of the first prediction model, if the output distribution of the first prediction model follows a Gaussian distribution. Then select Initial pseudo-labels that are less than the second preset threshold are used as the first pseudo-labels with high confidence.

[0118] Next, a multi-model consistency evaluation is performed on the first pseudo-label. For example, multiple third preset models that are completely consistent with the model type of the first prediction model are trained based on different initialization parameters or different third local view features to obtain n third prediction models. k=1,2…n, each third prediction model is used to predict the pseudo-label of the second training sample corresponding to the first pseudo-label, and the pseudo-label predicted by each third prediction model is obtained. The consistency of the pseudo-labels predicted by each third prediction model is calculated by formula (3):

[0119] (3)

[0120] in, Let represent the mean of pseudo-label predictions from multiple third-party prediction models, where n represents the number of third-party prediction models. Let represent the k-th third prediction model, and Consis represent the degree of prediction dispersion of the same second training sample under multiple third prediction models.

[0121] Among them, the different third local view features are constructed based on the key feature subsets. They are subsets of the key feature subsets, rather than completely independent feature sets.

[0122] The first pseudo-label with a Consis less than the third preset threshold is selected as the second pseudo-label with high consistency.

[0123] Finally, by statistically analyzing the first proportion of the second pseudo-label in each performance prediction interval and comparing it with the second proportion determined based on the first dataset, high-quality third pseudo-labels with a deviation from the second proportion less than a fourth preset threshold are selected as target pseudo-labels, forming a candidate sample set. ,in, This represents the local view features of the second training sample corresponding to the target pseudo-label. This indicates a target pseudo-label.

[0124] For example, the first dataset is divided into several intervals according to the prediction target (such as ionic conductivity, activation energy, electrochemical stability, structure type, etc.), and the second proportion of samples in each interval is calculated. When screening the second pseudo-label, based on the deviation between the first and second proportions of the second pseudo-label in each interval, if the deviation between the first and second proportions of a certain interval is greater than a fourth preset threshold (e.g., 10%), it indicates that the error between the second pseudo-label in that interval and the actual performance is large, and the second pseudo-label in that interval is discarded. If the deviation between the first and second proportions of a certain interval is less than or equal to the fourth preset threshold, it indicates that the second pseudo-label in that interval has high credibility, and the second pseudo-label in that interval is retained as the target pseudo-label.

[0125] Merge global view features and performance labels from the first dataset And the global view features and high-quality pseudo-labels of the second dataset after filtering. To form an extended training set ,in, This represents the expanded global view feature matrix. and Concatenate according to sample dimensions This represents the expanded performance label vector. and Concatenate according to sample dimensions express The corresponding global view features.

[0126] In one specific implementation, step S15 includes: extracting second global view features from the third training samples to train the second prediction model, and updating the parameters of the second prediction model by minimizing the joint loss function;

[0127] The joint loss function includes supervision by labeled performance labels, supervision by target pseudo-labels, and logical consistency constraints between the first and second prediction models.

[0128] Specifically, the second preset model can adopt a general interpretable global view regression model, such as a linear model, sparse regression, ensemble decision tree, neural network and other mainstream models. The core mapping relationship of the second preset model is represented as f2(X), as shown in formula (4):

[0129] (4)

[0130] in, This represents the second prediction model. The parameter representing the j-th global view feature, Represents a parameter vector. This represents the feature matrix of the global view. It serves as a general parameter symbol, adaptable to different model types.

[0131] Based on the expanded training set, the model parameters of the second preset model are trained by minimizing the joint loss function. The joint loss function integrates real label supervision and pseudo label supervision, and incorporates logical consistency constraints between models, as shown in formula (5):

[0132] (5)

[0133] in, , , Assign weights to each loss item. This represents the supervision error between the true performance label and the target pseudo-label. This represents the logical consistency constraint between the first prediction model and the second prediction model. Indicates the regularity strength. This represents the complete set of learnable parameters for the second prediction model. This represents the regularization penalty term of the second prediction model. Matching settings can be configured according to different models; for example, a linear model is... L1 regularization, ensemble decision trees as complexity penalties, and neural networks as weight decay are examples of such methods.

[0134] This specific implementation introduces a first prediction model based on a subset of key features and a second prediction model with interpretability to balance prediction accuracy and interpretability. It can not only output the target performance prediction value of solid electrolyte, but also quantify the contribution of each input feature to the performance, providing a quantitative basis for understanding the ion migration mechanism and optimizing material design, thus making up for the lack of interpretability of existing models.

[0135] In one specific implementation, step S16 includes: obtaining the contribution of each input feature to the prediction result of the solid electrolyte performance based on the feature importance index of the second prediction model; and updating the key feature subset according to the contribution.

[0136] Specifically, after training the second prediction model, step S16 selectively updates the first prediction model and the pseudo-label generation strategy based on the parameter update results of the second prediction model. A dual-view collaborative training framework is adopted. In each iteration, the first prediction model generates initial pseudo-labels for the unlabeled second training samples based on the key feature subset, and updates the second prediction model based on the expanded training set. At the same time, based on the residual distribution of the second prediction model on the first and second datasets, the key feature subset or model parameters of the first prediction model are adjusted to form a consistency constraint and mutual promotion mechanism between the first and second prediction models.

[0137] Dual-view collaborative training refers to a training framework that simultaneously utilizes an interpretable model based on a subset of key features (the first prediction model) and a complete model based on global view features (the second prediction model) for collaborative training and mutual optimization. This training framework and process can be implemented based on a semi-supervised learning framework as a technical means to achieve pseudo-label generation and iterative optimization, enabling information interaction and consistency constraints on the outputs of the first and second prediction models, and jointly optimizing model parameters during the iteration process. Figure 4 As shown, the interpretable and stable first prediction model trained based on a subset of key features and the second prediction model trained based on global view features, which learns the complete structural performance relationship, update their parameters through consistency constraints during the iterative optimization process.

[0138] In each iteration, the pseudo-label selection criteria are updated using the training results of the current second prediction model, target pseudo-labels are added or removed, and the key feature subsets or model parameters of the first prediction model are appropriately adjusted based on the residuals and feature contributions of the second prediction model on the expanded training set. For example, the contribution of each input feature to the solid electrolyte performance prediction results is characterized by the built-in feature importance indicators, such as the absolute value of the parameters of the linear model, the Gini coefficient of the tree model, and the gradient contribution value of the neural network, so as to realize the interpretability analysis of the second prediction model and update the key feature subsets according to the contribution. Then, the process of pseudo-label generation, pseudo-label selection and construction of the expanded training set is repeated to retrain the second prediction model. The process continues to iterate until the residual distribution meets the preset conditions or reaches the preset stopping conditions to obtain a converged second prediction model, which is the solid electrolyte performance prediction model.

[0139] This specific implementation method constructs a feature representation adapted to solid electrolytes, a pseudo-label quality control strategy, and a dual-view collaborative iterative training process of "interpretability model + full feature model". It not only takes into account the prediction accuracy and interpretability, realizes a collaborative iterative mechanism for materials science, but also enables high-precision prediction of key performance indicators of solid electrolytes, while retaining the interpretability of feature contribution, reducing the cost of materials research and development, and effectively improving the screening efficiency of new materials.

[0140] This embodiment does not limit the specific semi-supervised learning algorithm or model structure. Any pseudo-label generation mechanism, collaborative training framework and semi-supervised regression method can be accessed. With minor modifications, it can be extended to the performance prediction tasks of other electrochemical materials or functional materials, and has a wide range of applications.

[0141] In a specific example, the data preparation begins by constructing a solid electrolyte sample library containing solid electrolyte crystal structure files, chemical composition, doping information, preparation process parameters, and electrochemical testing conditions. Some samples include target performance tags such as room temperature ionic conductivity and activation energy, forming the first dataset (500 labeled samples); the remaining samples, containing only input features, constitute the second dataset (5000 unlabeled samples).

[0142] A sequential forward selection method was used to filter a subset of key features from all features, ultimately selecting 12 core features, including elemental mole fraction, average ionic radius, lattice constant, ion migration channel width, sintering temperature, and testing temperature. A sparse regression model was selected as the first preset model, trained using the true labels in the first dataset, with a regularization parameter λ=0.01 and 100 iterations, to obtain the first prediction model after training.

[0143] The first prediction model was applied to the second dataset, and the predicted room temperature ionic conductivity and activation energy of each unlabeled sample were output as initial pseudo-labels.

[0144] Calculate the prediction variance for each initial pseudo-label, set a variance threshold of 0.001, and select samples corresponding to the first pseudo-label with variances below the threshold. Simultaneously, combine this with a sample local density index to assist in assessing confidence. The local density index is the proportion of true label samples or high-confidence pseudo-labels.

[0145] By constructing sparse regression models with three different initial parameters, pseudo-labels are generated for the second dataset. Samples corresponding to the second pseudo-labels with a prediction deviation of less than 5% are selected from the three models. The consistency of the model prediction results under different key feature subsets is also considered.

[0146] Based on the conductivity distribution range in the first dataset (10 -6 ~10 -2 (S / cm), eliminating samples with conductivity exceeding the range of the second pseudo-label, ensuring that the samples corresponding to the third pseudo-label are reasonably supplemented in the composition space and structure space.

[0147] After screening, 1,800 high-quality pseudo-labeled samples were retained from 5,000 unlabeled samples to form a set of candidate samples.

[0148] The first dataset (500 sets) was merged with the selected candidate sample set (1800 sets) to form an expanded training set containing 2300 sets of samples.

[0149] A gradient boosting regression model was selected as the second preset model. All features (28 items) of the expanded training set were input, and the number of decision trees was set to 300, the learning rate to 0.05, and the maximum tree depth to 8. The model was then trained to obtain the second prediction model. This second model is able to learn a more complete and detailed structural relationship between the input features and the target performance.

[0150] Next, the first iteration was performed: the prediction residuals of the expanded training set were calculated using the second prediction model, the key feature subset of the first prediction model was adjusted based on the residual distribution, and the "coordination environment complexity" feature was added; the pseudo-label selection criteria were updated, the initial pseudo-labels were regenerated and selected, and 1950 high-quality samples were retained; the expanded training set was updated and the second prediction model was retrained, and the model validation error was reduced by 12% compared with the initial result.

[0151] Iterations 2-5: Repeat the above process. In each iteration, based on the residuals and feature contributions of the second prediction model on the expanded training set, adjust the key feature subsets or parameters of the first prediction model, update the pseudo-label screening threshold, and gradually optimize the second prediction model.

[0152] After the 5th iteration, the model validation error between the two iterations decreased by 1.8%, which is less than the preset threshold of 3%. At the same time, the size of the expanded training set tended to stabilize, and the iteration stopped, resulting in a converged second prediction model, which is the solid electrolyte performance prediction model.

[0153] This embodiment introduces a first prediction model trained on a subset of key features for pseudo-label generation and iterative optimization. This effectively uncovers statistical regularities in a large amount of unlabeled solid electrolyte data, significantly reducing performance prediction errors and decreasing reliance on experimental / computational labels when labeled data is limited. Simultaneously, by utilizing pseudo-label selection criteria specific to solid electrolytes, the iterative process of pseudo-labels follows material physics constraints, avoiding the introduction of unreasonable pseudo-labeled samples and improving the credibility of the second prediction model in crystal structure space. Furthermore, through an interpretable second prediction model, not only can the performance prediction results of solid electrolytes be output, but the contribution of each input feature to the performance prediction results can also be quantified, providing quantitative evidence for understanding ion migration mechanisms and optimizing material design.

[0154] Example 2

[0155] In a specific embodiment of this disclosure, a method for predicting the performance of solid electrolytes is provided, such as... Figure 5 As shown, the prediction method includes:

[0156] S21. Obtain the data of the solid electrolyte to be tested;

[0157] S22. Extract the global view features of the solid electrolyte data to be tested;

[0158] S23. Input the global view features to be tested into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested.

[0159] The solid electrolyte performance prediction model was obtained using the training method described in Example 1.

[0160] Specifically, a solid-state electrolyte performance prediction model is used to predict the performance of solid-state electrolytes under test. By using the crystal structure file of the solid-state electrolyte data, a global view of the test's composition, structure, and processing characteristics is generated. This global view is then input into the solid-state electrolyte performance prediction model to obtain corresponding predicted values ​​for room temperature conductivity, activation energy, or interface stability. This enables rapid evaluation of a large number of solid-state electrolyte materials without the need for complex experiments and costly simulations, providing a basis for material selection and design.

[0161] Following the previous example, 100 newly designed solid electrolyte candidate materials were selected, their crystal structure files were analyzed, and composition information and preparation process parameters (sintering temperature 800-1200℃, atmosphere is argon, pressure 5MPa, etc.) were collected to generate corresponding 28 feature vectors.

[0162] By inputting the feature vector into the solid electrolyte performance prediction model, the predicted values ​​of room temperature ionic conductivity and activation energy of each candidate material are obtained, enabling rapid evaluation of a large number of candidate materials without the need for complex experiments and high-cost simulations.

[0163] The top 10 candidate materials based on the predicted results were selected for experimental synthesis and performance testing. For 8 of these materials, the relative error between the measured ionic conductivity and the predicted value was less than 8%, and the relative error of the activation energy was less than 10%, validating the model's prediction accuracy. Furthermore, through the model interpretation and visualization module, "ion migration channel width" and "average ionic radius difference" were identified as the top two key features affecting ionic conductivity, providing a quantitative basis for subsequent material structure optimization.

[0164] Example 3

[0165] In one specific embodiment of this disclosure, a training system for a solid electrolyte performance prediction model is provided, such as... Figure 6 As shown, the training system includes:

[0166] The first acquisition module 101 is used to acquire a first dataset and a second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data. The first acquisition module 101 receives and stores the first dataset and the second dataset containing solid electrolyte samples to provide data support for subsequent feature construction and model training.

[0167] The first training module 102 is used to extract several key features from the first training samples to form a key feature subset, and to train the first preset model based on the key feature subset to obtain the first prediction model.

[0168] The pseudo-label generation module 103 is used to predict the second training samples in the second dataset using the first prediction model to obtain the initial pseudo-label corresponding to each second training sample; the pseudo-label generation module 103 trains the first prediction model based on the first dataset, and uses the model to generate initial pseudo-labels for the samples in the second dataset, and outputs them to the filtering module 104.

[0169] The filtering module 104 is used to filter target pseudo-labels that meet the pseudo-label filtering conditions from the initial pseudo-labels, and to construct an extended training set based on the first dataset, the target pseudo-labels and their corresponding second training samples; the filtering module 104 filters high-quality pseudo-label samples from the initial pseudo-labels according to the preset confidence standard and / or consistency standard, and outputs them to the second training module 105.

[0170] The second training module 105 is used to train the second preset model based on the third training samples in the extended training set to obtain the second prediction model; the second training module 105 trains the second prediction model based on the extended training set composed of real labeled samples and high-quality pseudo-label samples.

[0171] The iteration module 106 is used to obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, the key feature subset and / or pseudo-label selection conditions are updated, and the first training module 102 is called back until the residual distribution meets the preset error range or reaches the preset stopping condition. The second prediction model is then used as the solid electrolyte performance prediction model. During the iteration process, the iteration module 106 updates the pseudo-label generation strategy and model parameters to obtain the converged second prediction model.

[0172] In one specific implementation, the first training module 102 is specifically used to extract a first global view feature from the first training sample. The first global view feature includes at least compositional features, structural features, and process features. Using sequential forward selection or regularized regression, several key features are selected from the first global view feature to form a key feature subset of the first local view feature. Based on the key feature subset, a first preset model is trained to obtain a first prediction model. The first training module 102 generates feature vectors by parsing and preprocessing the structural files, compositional information, and process parameters of the solid electrolyte sample.

[0173] In one specific implementation, the pseudo-label generation module 103 is specifically used to extract the second local view features of the second training sample and input the second local view features into the first prediction model for prediction.

[0174] In one specific implementation, the second training module 105 is specifically used to extract second global view features from the third training samples to train the second prediction model, and to update the parameters of the second prediction model by minimizing the joint loss function; wherein, the joint loss function includes labeled performance label supervision, target pseudo-label supervision, and logical consistency constraints between the first prediction model and the second prediction model.

[0175] In one specific implementation, the filtering module 104 is specifically used to obtain the prediction residual of the first prediction model on the second dataset, and filter the first pseudo-labels whose prediction residuals are less than a first preset threshold from the initial pseudo-labels; or, obtain the output distribution of the first prediction model on the second dataset, and in response to the output distribution following a Gaussian distribution, filter the first pseudo-labels whose variance is less than a second preset threshold from the initial pseudo-labels; based on the pseudo-label prediction mean of multiple models, obtain the prediction dispersion of the first pseudo-labels, and filter the second pseudo-labels whose prediction dispersion is less than a third preset threshold; wherein, the pseudo-label prediction mean is obtained based on multiple different third prediction models, the different third prediction models have the same model type as the first prediction model, and are trained based on different third local view features; the third local view features are a subset of the key feature subset; obtain the first proportion of each second pseudo-label in each performance prediction interval, and filter the third pseudo-labels whose deviation between the first proportion and the second proportion in the same performance prediction interval is less than a fourth preset threshold, as target pseudo-labels;

[0176] The second percentage is based on the first dataset divided into different performance prediction intervals.

[0177] In one specific implementation, the iteration module 106 is specifically used to obtain the contribution of each input feature to the performance prediction result of the solid electrolyte based on the feature importance index of the second prediction model; and to update the key feature subset according to the contribution. Based on the parameters or feature importance index of the second prediction model, the iteration module 106 ranks and visualizes the compositional and structural features of the solid electrolyte according to their importance, providing a quantitative basis for material design and mechanism analysis.

[0178] In one specific implementation, the preset stopping conditions include at least one of the following: the decrease in model validation error between two iterations is less than a fifth preset threshold, the size of the expanded training set no longer changes, and the maximum number of iterations is reached.

[0179] This system can be deployed on servers or scientific computing platforms, with each module working collaboratively to complete the entire process of data processing, model training, and performance prediction. At the system architecture level, a solid-state electrolyte performance prediction system, including the aforementioned modules, has been constructed. This system can be deployed on scientific servers or cloud platforms, supporting batch evaluation and online querying of solid-state electrolyte candidate materials.

[0180] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0181] This embodiment introduces a first prediction model trained on a subset of key features for pseudo-label generation and iterative optimization. This effectively uncovers statistical regularities in a large amount of unlabeled solid electrolyte data, significantly reducing performance prediction errors and decreasing reliance on experimental / computational labels when labeled data is limited. Simultaneously, by utilizing pseudo-label selection criteria specific to solid electrolytes, the iterative process of pseudo-labels follows material physics constraints, avoiding the introduction of unreasonable pseudo-labeled samples and improving the credibility of the second prediction model in crystal structure space. Furthermore, through an interpretable second prediction model, not only can the performance prediction results of solid electrolytes be output, but the contribution of each input feature to the performance prediction results can also be quantified, providing quantitative evidence for understanding ion migration mechanisms and optimizing material design.

[0182] Example 4

[0183] In one specific embodiment of this disclosure, a system for predicting the performance of solid-state electrolytes is provided, such as... Figure 7 As shown, the prediction system includes:

[0184] Data acquisition module 201 is used to acquire data of the solid electrolyte under test;

[0185] Extraction module 202 is used to extract the global view features of the solid electrolyte data to be tested;

[0186] Prediction module 203 is used to input the global view features to be tested into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested.

[0187] The solid electrolyte performance prediction model was obtained using the training system described in Example 3.

[0188] This prediction system can be deployed on research servers or cloud platforms, supporting batch evaluation and online querying of solid electrolyte candidate materials. Users can upload data such as crystal structure files, composition information, and preparation process parameters of solid electrolyte candidate materials. The system automatically completes feature construction, model calling, and performance prediction, outputting target performance prediction results and a visualization analysis report on feature importance, providing efficient and accurate decision support for materials researchers.

[0189] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0190] This embodiment introduces a first prediction model trained on a subset of key features for pseudo-label generation and iterative optimization. This effectively uncovers statistical regularities in a large amount of unlabeled solid electrolyte data, significantly reducing performance prediction errors and decreasing reliance on experimental / computational labels when labeled data is limited. Simultaneously, by utilizing pseudo-label selection criteria specific to solid electrolytes, the iterative process of pseudo-labels follows material physics constraints, avoiding the introduction of unreasonable pseudo-labeled samples and improving the credibility of the second prediction model in crystal structure space. Furthermore, through an interpretable second prediction model, not only can the performance prediction results of solid electrolytes be output, but the contribution of each input feature to the performance prediction results can also be quantified, providing quantitative evidence for understanding ion migration mechanisms and optimizing material design.

[0191] Example 5

[0192] Figure 8This is a schematic diagram of the structure of an electronic device according to an example embodiment of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the training method or prediction method described in any of the above embodiments. Figure 8 The electronic device 30 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0193] like Figure 8 As shown, the electronic device 30 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including memory 32 and processor 31).

[0194] Bus 33 includes a data bus, an address bus, and a control bus.

[0195] The memory 32 may include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and may further include read-only memory (ROM) 323.

[0196] The memory 32 may also include a program tool 325 (or utility) having a set (at least one) program module 324, such program module 324 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0197] The processor 31 executes various functional applications and data processing by running computer programs stored in the memory 32, such as the training or prediction methods provided in any of the above embodiments.

[0198] Electronic device 30 can also communicate with one or more external devices 34 (e.g., keyboard, pointing device, etc.). This communication can be performed through input / output (I / O) interface 35. Furthermore, electronic device 30 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network, such as the Internet) via network adapter 36. As shown, network adapter 36 communicates with other modules of electronic device 30 via bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with electronic device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0199] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0200] Example 6

[0201] This disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the training method or prediction method provided in any of the above embodiments.

[0202] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0203] Example 7

[0204] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements any of the training methods or prediction methods described above.

[0205] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0206] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A training method for a solid electrolyte performance prediction model, characterized in that, The training method includes: Obtain a first dataset and a second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data. Several key features are extracted from the first training sample to form a key feature subset, and the first preset model is trained based on the key feature subset to obtain the first prediction model; The first prediction model is used to predict the second training samples in the second dataset to obtain the initial pseudo-label corresponding to each second training sample; Target pseudo-labels that meet the pseudo-label selection criteria are selected from the initial pseudo-labels, and an expanded training set is constructed based on the first dataset, the target pseudo-labels and their corresponding second training samples; The second preset model is trained based on the third training samples in the extended training set to obtain the second prediction model; Obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, update the key feature subset and / or the pseudo-label screening conditions, and return to execute the step of training the first preset model based on the key feature subset until the residual distribution meets the preset error range or reaches the preset stopping condition, and use the second prediction model as the solid electrolyte performance prediction model.

2. The training method according to claim 1, characterized in that, The steps of extracting several key features from the first training samples to form a key feature subset, and training the first preset model based on the key feature subset to obtain the first prediction model include: A first global view feature is extracted from the first training sample, and the first global view feature includes at least composition features, structural features and process features; Using sequential forward selection or regularized regression, several key features are selected from the first global view features to form a subset of the key features of the first local view features; The first preset model is trained based on the key feature subset to obtain the first prediction model.

3. The training method according to claim 2, characterized in that, The step of using the first prediction model to predict the second training samples in the second dataset includes: Extract the second local view features from the second training sample and input the second local view features into the first prediction model for prediction; And / or, The step of training the second preset model based on the third training samples in the expanded training set includes: The second global view features are extracted from the third training samples to train the second prediction model, and the parameters of the second prediction model are updated by minimizing the joint loss function. The joint loss function includes supervision by labeled performance labels, supervision by target pseudo-labels, and logical consistency constraints between the first prediction model and the second prediction model.

4. The training method according to claim 2, characterized in that, The step of filtering target pseudo-tags that meet the pseudo-tag filtering criteria from the initial pseudo-tags includes: Obtain the prediction residual of the first prediction model on the second dataset, and filter the first pseudo-labels from the initial pseudo-labels whose prediction residuals are less than a first preset threshold; or, obtain the output distribution of the first prediction model on the second dataset, and in response to the output distribution following a Gaussian distribution, filter the first pseudo-labels from the initial pseudo-labels whose variance is less than a second preset threshold. Based on the mean of pseudo-label predictions from multiple models, the prediction dispersion of the first pseudo-label is obtained, and a second pseudo-label with a prediction dispersion less than a third preset threshold is selected. The pseudo-label prediction mean is obtained based on multiple different third prediction models. These different third prediction models have the same model type as the first prediction model and are trained based on different third local view features. The third local view features are subsets of the key feature subset. Obtain the first proportion of each second pseudo-label in each performance prediction interval, and filter out the third pseudo-labels in the same performance prediction interval whose deviation between the first proportion and the second proportion is less than a fourth preset threshold, as the target pseudo-labels; The second percentage is obtained by dividing the first dataset according to each of the performance prediction intervals.

5. The training method according to any one of claims 1 to 4, characterized in that, The step of updating the key feature subset and the pseudo-label filtering conditions includes: Based on the feature importance index of the second prediction model, the contribution of each input feature to the prediction result of solid electrolyte performance is obtained; The key feature subset is updated based on the contribution level; And / or, The preset stopping conditions include at least one of the following: the model validation error decreases by less than a fifth preset threshold between two iterations; the size of the expanded training set no longer changes; or the maximum number of iterations is reached.

6. A method for predicting the performance of solid electrolytes, characterized in that, The prediction method includes: Acquire data of the solid electrolyte under test; Extract the global view features of the solid electrolyte data to be tested; The global view features to be tested are input into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested. The solid electrolyte performance prediction model is obtained using the training method described in any one of claims 1 to 5.

7. A training system for a solid electrolyte performance prediction model, characterized in that, The training system includes: The first acquisition module is used to acquire a first dataset and a second dataset. The first dataset includes a first number of first training samples, and the second dataset includes a second number of second training samples. The first training samples include solid electrolyte data with performance labels, and the second training samples include unlabeled solid electrolyte data. The first training module is used to extract several key features from the first training samples to form a key feature subset, and to train the first preset model based on the key feature subset to obtain the first prediction model. The pseudo-label generation module is used to predict the second training samples in the second dataset using the first prediction model to obtain the initial pseudo-label corresponding to each second training sample. The filtering module is used to filter target pseudo-labels that meet the pseudo-label filtering conditions from the initial pseudo-labels, and to construct an extended training set based on the first dataset, the target pseudo-labels and their corresponding second training samples; The second training module is used to train the second preset model based on the third training samples in the extended training set to obtain the second prediction model. An iterative module is used to obtain the residual distribution of the second prediction model on the first dataset and the second dataset. In response to the residual distribution not meeting the preset error range, the key feature subset and / or the pseudo-label screening conditions are updated, and the first training module is called back until the residual distribution meets the preset error range or reaches the preset stopping condition, and the second prediction model is used as the solid electrolyte performance prediction model.

8. A system for predicting the performance of solid electrolytes, characterized in that, The prediction system includes: The data acquisition module is used to acquire data of the solid electrolyte under test. The extraction module is used to extract the global view features of the solid electrolyte data to be tested; The prediction module is used to input the global view features to be tested into the solid electrolyte performance prediction model to obtain the performance prediction results of the solid electrolyte data to be tested. The solid electrolyte performance prediction model is obtained using the training system described in claim 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the training method of any one of claims 1 to 5, or the prediction method of claim 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the training method as described in any one of claims 1 to 5, or the prediction method as described in claim 6.

Images