A small sample high-precision prediction model construction method for hydrogen storage materials

Abstract

The application provides a small sample high-precision prediction model construction method for hydrogen storage materials, relates to the technical field of machine learning, and comprises the following steps: constructing an original material data set and selecting characteristic variables; performing data disturbance and characteristic variable enhancement on the original material data set based on a data enhancement method of data fluctuation caused by physical constraints and experimental errors, generating an enhanced material data set, and constructing a model training set based on the enhanced material data set; respectively performing automatic adjustment of hyperparameters on a random forest model and a gradient boosting model; integrating and fusing the random forest model and the gradient boosting model, performing hyperparameter optimization, and constructing an integrated prediction model; performing regularization and deviation removal processing on each sample in the model, and inhibiting the influence of skewed data in the training process through sample weighting or resampling. The model constructed by the application ensures accurate and effective capture of the nonlinear relationship between the composition and performance of the hydrogen storage material.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for constructing a small-sample, high-precision prediction model for hydrogen storage materials. Background Technology

[0002] Materials play an irreplaceable and central role in the development of modern high technology. With the ever-increasing demand for novel materials with superior performance, materials research and development has made significant progress in recent decades. The development of high-throughput experimental techniques and advanced computational methods has made the acquisition of large-scale materials data increasingly convenient, laying the foundation for data-intensive research. Against this backdrop, the crucial role of data in driving scientific discovery has been gradually recognized, and data-driven science is considered the "fourth paradigm" after traditional experimental science, theoretical science, and computational science. One of the key technologies rapidly developing in this paradigm shift is machine learning (ML)—an interdisciplinary technology integrating statistics, computer science, and pattern recognition. Unlike traditional material exploration methods that rely on repeated trials, machine learning can automatically uncover potential patterns from existing data, analyzing the high-dimensional correlations hidden behind the data to achieve rapid prediction and performance evaluation of new materials. Therefore, machine learning not only accelerates the material screening and optimization process but also provides a completely new way of thinking and technical approach for materials design.

[0003] In recent years, new energy materials have experienced rapid development, with hydrogen energy emerging as a zero-carbon energy source. Hydrogen storage materials serve as the core carrier for the large-scale application of hydrogen energy. However, their performance is affected by multiple factors such as material composition, microstructure, and preparation process. Furthermore, the long experimental preparation cycle and high testing costs result in limited data availability, and the data also suffers from high noise and dispersion. These issues lead to a general shortage of samples in machine learning applications for materials science. On such small datasets, traditional machine learning models are prone to severe overfitting. Current research often employs single-model algorithms, such as random forests, which are highly sensitive to systematic biases and noise in the data and are prone to high variance in small sample sizes. Gradient boosting models are also prone to overfitting in small sample conditions, requiring extensive parameter tuning and resulting in long model runtimes. Linear models alone cannot capture the nonlinear relationship between composition and performance in hydrogen storage materials. Furthermore, hydrogen storage materials exhibit diverse characteristics, but with small sample sizes, high-dimensional feature variables make it difficult to establish reliable composition-performance mapping relationships. Therefore, small sample sizes, high noise levels, and multivariate coupling are common challenges in hydrogen storage material research, significantly limiting the effectiveness of traditional machine learning methods. This further highlights the importance of developing small-sample intelligent modeling strategies suitable for hydrogen storage materials, which is crucial for achieving efficient material screening, composition optimization, and performance prediction.

[0004] Given the characteristics of experimental data for hydrogen storage materials, such as small sample size, high noise, high feature dimensionality, and complex variable coupling, directly using a single machine learning model for modeling can easily lead to overfitting, resulting in low R-values. 2 Low R-values ​​(typically in the range of 0.1-0.5) lead to instability in predictions and systematic biases. To address this, this invention expands the effective sample space by introducing a data augmentation method based on physical constraints, mitigating insufficient model training and sensitivity to experimental errors under small sample conditions. Furthermore, it integrates multiple prediction models with complementary characteristics to collaboratively correct the limitations of a single model, thereby significantly improving R-values ​​while maintaining physical plausibility. 2 A value greater than 0.8 is used to enhance the model's stability, generalization ability, and prediction accuracy.

[0005] Machine learning, with its relatively low computational cost and powerful predictive capabilities, has been widely applied in the design and development of new materials. However, in recent years, the limitations of machine learning in materials science have become increasingly apparent, the most significant issue being data scarcity. Building reliable and accurate machine learning models using limited data is a challenging task. Furthermore, due to the slow accumulation of materials data, the problem of small sample sizes will persist in materials science for a long time. Therefore, improving the accuracy and reliability of machine learning models under conditions of insufficient data remains a challenging task. Summary of the Invention

[0006] One of the objectives of this invention is to provide a method for constructing a high-precision prediction model for hydrogen storage materials with small samples, in order to solve the problem of low model prediction performance and poor generalization ability caused by the scarcity of hydrogen storage material data in the prior art, so as to ensure the accurate and effective capture of the nonlinear relationship between composition and performance in hydrogen storage materials.

[0007] This invention provides a method for constructing a small-sample, high-precision prediction model for hydrogen storage materials, comprising: Constructing the original material dataset and selecting feature variables; Data augmentation methods based on physical constraints and data fluctuations caused by experimental errors perturb the original material dataset and enhance feature variables to generate an enhanced material dataset. The enhanced material dataset is then combined with the original material dataset to construct the model training set. Automatic hyperparameter tuning was performed on the random forest model and the gradient boosting model respectively to obtain the optimal parameter combination under small sample conditions; By integrating and fusing the random forest model and the gradient boosting model and performing hyperparameter tuning, an integrated prediction model is constructed. Regularization and debiasing are performed on each sample in the model, and the influence of skewed data during the training process is suppressed by sample weighting or resampling.

[0008] Preferably, the data augmentation method includes generating virtual samples based on data perturbation and / or interpolation based on experimental errors.

[0009] Preferably, the hyperparameters are automatically adjusted using Bayesian optimization algorithms, tree-structured parameter search algorithms, or genetic algorithms.

[0010] Preferably, the integration of random forest models and gradient boosting models includes: A stacked fusion method is adopted to integrate the prediction results of the random forest model and the gradient boosting model according to a set weight.

[0011] Preferably, the stacking and fusion methods include: A preset first number of random forest models and a preset second number of gradient boosting models are used as base learners for the base learner layer; The third number of gradient boosting models is used as the meta-learners of the meta-learner layer. The base learner is used to process the samples to obtain the prediction output; the meta learner is used to combine the prediction outputs of the various base learners to obtain the comprehensive prediction output.

[0012] Preferably, the method for constructing a small-sample, high-precision prediction model for hydrogen storage materials further includes: The model is then optimized based on the residuals and subjected to secondary training to obtain the final prediction model.

[0013] Preferably, the optimization steps for optimizing the samples in the model training set based on the residuals are as follows: The predicted values ​​obtained from the first cross-validation of the model; The predicted values ​​are compared with the actual hydrogen storage performance data, the residuals are calculated, and the residual distribution is constructed. Set a threshold to label samples with high residuals; High residual samples with labels are removed from the model training set.

[0014] Preferably, the hyperparameters include one or more combinations of the following: parameters controlling the number of trees in the random forest model, parameters controlling the maximum depth of the trees in the random forest model, parameters controlling the minimum number of samples in the leaf nodes of the random forest model, parameters controlling the number of iterations of the gradient boosting model, parameters controlling the contribution of each tree to the final result, parameters controlling the complexity of each tree, parameters controlling the proportion of random samples used by each tree, and parameters controlling the proportion of features used by each tree.

[0015] Preferably, the integrated prediction model adopts a weighted stacking structure, wherein the weights are adaptively generated based on the cross-validation performance of each model.

[0016] Preferably, the method for constructing a small-sample, high-precision prediction model for hydrogen storage materials further includes: Based on a pre-configured data acquisition rule base, corresponding reference data is obtained from the big data platform; Based on the acquisition type code in the data acquisition rule base corresponding to the reference data, the corresponding preset data conversion rule base is retrieved. Based on the retrieved data transformation rule base, the reference data is transformed to obtain the transformed dataset; Add the transformed dataset to the model training set.

[0017] The present invention has the following beneficial effects: This approach demonstrates significant benefits in improving model accuracy. By combining data augmentation, model hyperparameter optimization, and ensemble learning, the accuracy of the model is effectively enhanced. By introducing data augmentation strategies based on experimental error fluctuations and physical constraints, the inaccuracy of the data is effectively avoided, ensuring that the augmented data still conforms to the physical laws of materials science. Strict control over the perturbation amplitude and reasonable feature ranges during the augmentation process ensures that the augmented dataset maintains data validity while increasing sample diversity. Compared to the original dataset, the augmented dataset significantly reduces the risk of model overfitting and improves model accuracy.

[0018] By combining two complementary models and optimizing their hyperparameters, each individual model can fully leverage its learning capabilities. Through ensemble integration, the ensemble model achieves complementarity, resulting in higher prediction accuracy than the individual models.

[0019] Hydrogen storage material data often exhibits significant non-uniform or densely distributed patterns. Directly training the model can easily lead to bias towards high-density regions while ignoring low-frequency samples. This invention employs a sample weighting and debiasing strategy to enable the model to comprehensively learn the distribution characteristics of material properties across different regions, reducing interference from skewed samples. This not only improves the model's ability to identify a small number of key samples but also significantly enhances its applicability in real-world material systems, making the prediction results more reliable and more meaningful for engineering applications.

[0020] This invention employs universal data cleaning, enhancement, fusion, and debiasing techniques, applicable to data prediction tasks across multiple fields of hydrogen storage materials. The entire method relies solely on common machine learning frameworks, offering advantages such as low implementation difficulty, minimal resource requirements, and easy integration into existing materials research and development processes. While ensuring high prediction accuracy, this invention significantly lowers the model development threshold, enhances the application value of small-sample machine learning techniques in materials research and development, and has broad prospects for widespread adoption.

[0021] Compared to existing machine learning models that primarily rely on large-scale datasets and perform poorly on small-sample tasks, this invention introduces material physics constraints into data augmentation, addressing the core issues of sparsity and high noise in hydrogen storage material data. Compared to simple random perturbations, the data generated by this invention has higher physical reliability, avoiding the accumulation of invalid and noisy samples. A complete "closed-loop framework for small-sample learning of hydrogen storage materials" is proposed, overcoming the fragmented and unsystematic shortcomings of existing technologies.

[0022] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a method for constructing a small-sample, high-precision prediction model for hydrogen storage materials in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the steps of a data expansion method in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the model construction steps in an embodiment of the present invention; Figures 4 to 6 These are schematic diagrams illustrating the accuracy changes of a model constructed using the method of this invention. Figure 7 This is a schematic diagram illustrating the application of the model constructed using the method of this invention in predicting the hydrogen storage cycle performance of V-based hydrogen storage alloys. Figures 8 to 10 These are schematic diagrams illustrating the application of the model constructed based on the comparative examples proposed in this invention to predict the hydrogen storage cycle performance in V-based hydrogen storage alloys. Detailed Implementation

[0025] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] The following examples use the cycling capacity of hydrogen storage materials as the main prediction object. By comparing the model prediction results with the actual cycling capacity test results, the effectiveness of the method of the present invention in cycling capacity prediction is verified.

[0027] Example 1: This embodiment of the invention provides a method for constructing a small-sample, high-precision prediction model for hydrogen storage materials, such as... Figure 1 As shown, it includes: Step 1: Construct the original material dataset and select feature variables; A small sample dataset of the target material system is collected, and the raw data is input into the data processing module for cleaning, normalization, and outlier removal. Key features are extracted based on knowledge of hydrogen storage materials to construct the original dataset. This original dataset provides the raw input for subsequent enhancement, ensemble learning, and regularization debiasing. Step 2: Based on the data augmentation method caused by physical constraints and experimental errors, the original material dataset is perturbed and the feature variables are enhanced to generate an enhanced material dataset. The enhanced material dataset is then combined with the original material dataset to construct the model training set. To address the common challenges in hydrogen storage materials, such as small sample sizes, high experimental costs, long testing cycles, and significant data noise, this invention proposes a data augmentation strategy. The data augmentation methods include generating virtual samples based on experimental error perturbation and / or interpolation. This invention's strategy does not simply apply random perturbations to the original data. Instead, it constructs a perturbation model simulating experimental errors based on the experimental error characteristics of hydrogen storage materials and the inherent physical constraints of the alloy system. This ensures that the augmented samples possess physical rationality and engineering feasibility. Traditional data augmentation methods often rely on Gaussian noise, numerical translation, or random perturbations. While these methods can increase the sample size, they often ignore the actual physical laws of the material system, easily generating distorted samples that mislead model learning. In contrast, this invention introduces the physical constraints of the hydrogen storage material system into the data augmentation process. For example, the composition of multi-element alloys (such as the proportions of V, Ti, and Cr) must satisfy a normalization condition where the sum is 1, and the content of each element cannot be negative or exceed the achievable range. This treatment makes the augmented samples not only formally reasonable but also reflects the requirements of composition conservation and structural stability in the design of real hydrogen storage materials. The specific operations for dataset construction and augmentation involve obtaining / setting basic parameters for the input dataset, which contains several samples. Each sample contains several numerical features and component columns (e.g., V proportion, Ti proportion, Cr proportion) and a target column (e.g., cycle capacity). Perturbations are designed according to the experimental error distribution, adding random noise to the data to increase its diversity. Known physical constraints are satisfied for each augmented sample. After generating all augmented samples, a simple check is performed: all components must satisfy the constraints; samples with identical repetitions or obviously unreasonable (negative / extreme values) are excluded; the mean and variance of the augmented dataset and the original data on each feature are compared to ensure that augmentation does not significantly alter the distribution (only increasing local density and perturbation); the extended dataset is generated through three methods: noise perturbation, interpolation, and random sampling based on physical constraints, effectively expanding the number of samples while ensuring physical rationality; the noise perturbation amplitude in data augmentation is set to 1%–5% of the feature standard deviation to ensure the physical rationality of the augmented samples. Step 3: Perform automatic hyperparameter tuning on the random forest model and the gradient boosting model respectively to obtain the optimal parameter combination under small sample conditions; An automated hyperparameter search algorithm (including Bayesian optimization, genetic algorithm, or grid search) is used to optimize the parameters of the random forest model and the gradient boosting model respectively, so that each model obtains the optimal hyperparameter combination under the condition of small sample data, thereby improving the model learning efficiency and prediction accuracy. The hyperparameters include: parameters that control the number of trees in the random forest model, parameters that control the maximum depth of the trees in the random forest model, parameters that control the minimum number of samples in the leaf nodes of the random forest model, parameters that control the number of iterations in the gradient boosting model, parameters that control the contribution of each tree to the final result, parameters that control the complexity of each tree, parameters that control the proportion of random samples used by each tree, and parameters that control the proportion of features used by each tree, or one or more of these parameters in combination. A single model is used to train several base learners to acquire different structural and nonlinear laws in the feature space of hydrogen storage materials, and feature importance is extracted for subsequent ensemble learning and material interpretability analysis. In the single model, this invention selects Random Forest and XGBoost. Random Forest (Random Forest Regressor) has strong robustness and can handle noise and small sample sizes in hydrogen storage material data; it can capture the highly nonlinear relationship between material features and cycling capacity; and it can output feature importance, providing material interpretability for understanding the relationship between composition (V, Ti, Cr, etc.) and performance. XGBoost (XGBRegressor) uses a gradient boosting framework, which converges quickly and is suitable for structured material data; it has a built-in regularization mechanism to suppress overfitting caused by data sparsity; and it excels at handling complex interactions between features, making it an effective tool for learning the proportional combination effect of hydrogen storage alloys. For feature selection, after removing the target column and non-numerical columns, important material parameters related to hydrogen storage performance, such as the proportion of multiple components and structural features, are retained. To ensure stable and reliable evaluation results even with small sample experimental data of hydrogen storage materials, K-fold cross-validation (such as 5-fold) or leave-one-out cross-validation (LOOCV) are chosen. Leave-one-out cross-validation works by setting aside one sample from the dataset as the test set for each N samples, using the remaining N-1 samples to train the model N times, ensuring each sample is tested exactly once. The errors from the N tests are then averaged. Stacking constructs out-of-fold predictions using RepeatedKFold cross-validation (5-fold × 5 repetitions), ensuring each training sample is used exactly once for training the meta-learner, avoiding information leakage. The meta-model learns the generalization ability of the base model rather than its memorization ability, and the 25 repetitions significantly reduce the impact of randomness. Step 4: Integrate the random forest model and the gradient boosting model and perform hyperparameter tuning to build an integrated prediction model; Integrating the random forest model and the gradient boosting model involves a stacked fusion approach, where the prediction results of the random forest model and the gradient boosting model are integrated according to set weights. This stacked fusion approach includes: using a preset first number of random forest models and a preset second number of gradient boosting models as base learners in the base learner layer; and using a preset third number of gradient boosting models as meta learners in the meta learner layer. The base learners process the samples to obtain the prediction output; the meta learners combine the prediction outputs of the various base learners to obtain a comprehensive prediction output. An example of an integrated model constructed using one random forest model and one gradient boosting model as base learners, and one random forest model as a meta learner, effectively addresses the systematic bias of a single model by constructing a two-layer learning architecture. The first layer (base learner layer): Random forest as base learner one: adept at capturing local nonlinear relationships and naturally robust to noise; XGBoost as base learner two: adept at learning global gradient optimization paths and capable of finely fitting residuals. The second layer (meta-learner) uses XGBoost as the meta-learner to learn how to optimally combine the predictions of the two base models, automatically identify the reliability of each base model in different composition regions, and adaptively correct the systematic biases of the base models. Its core idea is to use the prediction outputs of multiple base models as new features input to a "meta-learner," enabling it to learn how to adaptively integrate the prediction capabilities of each base model in different sample regions. Compared to simple weighting or averaging methods, stacking has the following advantages: it can learn the bias differences of different models and adaptively weight them in hydrogen storage material data scenarios with small sample sizes and uneven noise; the meta-model can further correct the systematic errors of the base models, especially the anomalous data commonly found in hydrogen storage materials; through multi-model collaboration, it can comprehensively capture the multi-scale nonlinearity of composition effects, structural changes, and material properties. Therefore, stacking is particularly suitable for hydrogen storage material performance prediction tasks; that is, the integrated prediction model adopts a weighted stacking structure, where the weights are adaptively generated based on the cross-validation performance of each model. Further hyperparameter tuning was performed on the ensemble model to optimize its performance under skewed distribution and high-dimensional feature conditions, constructing the final high-precision prediction model. Given that experimental measurements of hydrogen storage materials are often affected by factors such as sample differences and environmental conditions, some samples may have abnormally high residuals. To improve the robustness and reliability of the final prediction model, further performance tuning was performed on the ensemble model. Then, the performance of the ensemble model was further improved and corrected, including: identifying and removing outlier samples, regularizing the meta-model, and tuning hyperparameters. Commonly tuned parameters for random forests in hyperparameter tuning include: `n_estimators` (number of trees) (200-1000), which specifies the number of decision trees in the random forest; more trees result in a more stable model and less prone to overfitting; however, too many trees increase computation time. `max_depth` (maximum tree depth) (6-20 or None) controls the maximum depth of a single decision tree; a greater depth results in a more complex tree and a greater likelihood of overfitting; setting it to None indicates that nodes will split indefinitely until the leaf nodes are pure. `min_samples_leaf` (minimum number of samples per leaf node) (1-5) limits the minimum number of samples a leaf node must contain; it controls the "smoothness" of the tree and prevents fitting noise caused by excessively small leaf nodes. Commonly tuned parameters in XGBoost include `n_estimators` (number of iterations, number of trees) which controls the number of boosting iterations; more trees result in stronger fitting ability but may lead to overfitting. `learning_rate` (0.01–0.2) controls the contribution of each tree to the final result (scaling factor); a smaller learning rate is more stable but requires more trees. `max_depth` (tree depth) (3–10) controls the complexity of each tree; Boosting is more prone to overfitting than random forests, so the depth should not be too high. `subsample` controls the proportion of random samples used by each tree; it reduces overfitting and improves generalization ability. `colsample_bytree` controls the proportion of features used by each tree; similar to `subsample`, it is also a source of randomness and reduces overfitting. `reg_lambda` / `reg_alpha` (L2 regularization / L1 regularization) are used to suppress overfitting. Record the metrics for each model: MSE (mean squared error), MAE (mean absolute error), R² (coefficient of determination), and Pearson correlation.

[0028] Step 5: Perform regularization and debiasing on each sample in the model, and suppress the influence of skewed data during the training process by weighting or resampling the samples.

[0029] Based on the sample distribution, regularization constraints are applied to each sample, and biased samples are processed by sample weighting or resampling. By adjusting sample weights, increasing the proportion of low-frequency samples, or eliminating the training influence of extreme samples, the model effectively suppresses the interference of skewed data during training, improving the model's generalization performance and stability. The regularization process uses L1 / L2 regularization combined with sample reweighting to improve the model's stability against small sample noise. The bias removal strategy is combined with the SMOTE resampling method to improve the representativeness of samples in sparse regions.

[0030] Example 2: To further improve the stability and generalization ability of the constructed model's predictions, a method for constructing a high-precision prediction model for hydrogen storage materials using small samples also includes: optimizing the samples in the model training set based on residuals and then performing secondary training to obtain the final prediction model. The optimization steps for optimizing the samples in the model training set based on residuals are as follows: using the predicted values ​​of the model obtained from the first cross-validation; comparing the predicted values ​​with actual hydrogen storage performance data, calculating the residuals and constructing a residual distribution; setting a threshold to label high-residual samples; and removing the labeled high-residual samples from the model training set.

[0031] This embodiment further optimizes the final model by removing outlier samples. The specific process for identifying and removing outlier samples is as follows: The predicted values ​​of the stacking model are obtained using the first LOOCV (or K-fold) algorithm. These predicted values ​​are compared with actual hydrogen storage performance data to calculate the residuals and construct a residual distribution. A threshold is set to mark samples with high residuals. These samples may correspond to experimental measurement errors, data entry errors, etc. These marked samples are removed from the dataset. The stacking training process is then fully executed again (including re-LOOCV), thereby obtaining a more stable and generalizable hydrogen storage material prediction model. This residual-based secondary training strategy significantly reduces the impact of experimental noise on the model, making the final model closer to the actual structure-performance relationship of hydrogen storage materials. Example 3: To further expand the data used for model training, reference data is introduced. Based on the reference data, the model is transformed and expanded. This example provides a method for constructing a small-sample, high-precision prediction model for hydrogen storage materials, such as... Figure 2 As shown, it also includes: obtaining corresponding reference data from the big data platform based on a pre-configured data acquisition rule base; retrieving the corresponding preset data transformation rule base based on the acquisition type code in the data acquisition rule base corresponding to the reference data; transforming the reference data based on the retrieved data transformation rule base to obtain a transformation dataset; and adding the transformation dataset to the model training set.

[0032] The data acquisition rule base and data transformation rule base were both pre-constructed by professionals. They were mainly built by analyzing the data differences between the same materials for storing hydrogen or other gases, and quantified into performance ratios, which represent the ratio of the parameters for hydrogen storage performance to the parameters for storing other gases. Based on the experimental data of materials storing other gases, the hydrogen storage performance can be derived from the ratio, thereby further expanding the data used to train the model. When constructing the data acquisition rule base, the storage data items (including search keywords, data types, etc.) corresponding to various gases associated with hydrogen storage are encoded to obtain data acquisition type codes. Different data transformation rule bases are used for different gases (acquisition type codes). This embodiment obtains a transformed dataset by acquiring reference data from a big data platform. This transformed dataset is then used to further expand the training data for model training. During model training, the ratio of transformed data samples, augmented samples, and original data samples in the training data is controlled to avoid model distortion caused by excessive expansion of the training data. Furthermore, model training can be summarized as follows: Figure 3 The steps shown are as follows: data collection and feature selection, data augmentation, training a single model, integrating the single models to obtain an ensemble model, and then regularizing and debiasing the ensemble model to form a model that can be used for high-precision prediction of small samples of hydrogen storage materials.

[0033] As time progresses, the amount of data on big data platforms is increasing, necessitating data filtering to obtain data more favorable for model training. The filtering steps are as follows: Based on the source and last update date of the reference data, determine the quantization value through a pre-configured quantization library; determine the evaluation value based on the weight coefficients and quantization values ​​corresponding to the type codes in the data acquisition rule library; filter the reference data in descending order of evaluation values; the weight coefficients are determined through the following steps: configure initial weight coefficients; when the transformed dataset participates in model training and the number of times the model is used after training reaches a preset number, analyze the statistical data of the results of the transformed dataset participating in model use to obtain analysis data; determine the reference weight coefficients for each type code using the entropy weight method; replace the initial weight coefficients with the reference weight coefficients.

[0034] To address the need for model updates, reference data also needs to be reacquired. Reacquiring this data involves recalculating evaluation values ​​and then filtering the data from the big data platform, which is labor-intensive. A more efficient approach is to pre-construct a marker dataset for filtering. This involves comparing the data in the marker dataset with the data acquired from the big data platform, improving efficiency. Specifically: The number of samples in the marker dataset is determined based on a pre-configured data update ratio (the product of the data update ratio and the number of filtered reference data); reference data is then entered in ascending order of evaluation values; the data is then divided into multiple comparison groups based on its source; and the data acquired from the big data platform is compared with the first reference data in each comparison group. Data with an update date in the comparison group is selected based on the reference data. Before the data is processed, data in the comparison group is filtered out. Afterward, it is compared with the previous reference data; if identical data is found, it is saved as reference data. Data with the smallest evaluation value in the marker dataset is removed. This process continues until all sample data in the marker dataset is removed. The remaining data from the previous reference dataset is then combined with newly added data to obtain the updated reference data. To ensure accurate filtering, the previous reference data can be grouped (according to the data source) before constructing the marker dataset. Data in each group of the previous reference data that matches the sample data in the marker dataset is deleted, and the data with the smallest evaluation value in each group is used as the calibration data. If no corresponding group is found when comparing with the marker dataset, the calibration data can be used for comparison. After a successful comparison, a determination of whether the data is identical is made.

[0035] Since the data on the big data platform is uploaded by various data sources, the more realistic the data is when used for training, the more accurate the trained model will be. Therefore, this application evaluates the authenticity and reliability of the data in the following way to ensure the accuracy of the model. The specific evaluation method is as follows: Based on the confidence level of the data source, a first evaluation value is determined; the experimental conditions that generated the data are evaluated to determine a second evaluation value; approximate data is extracted and used to perform confirmatory evaluation on the data sample to determine a third evaluation value; when the sum of the first, second, and third evaluation values ​​is less than a preset threshold, the data sample is removed; wherein, the first evaluation value is obtained by querying a pre-configured first evaluation table based on the confidence level, and the higher the confidence level, the higher the first evaluation value; the second evaluation value is the weighted sum of the first quantization values ​​obtained by quantifying each condition parameter of the experimental conditions using a pre-configured first quantization database, wherein each condition parameter (equipment model, usage time) is... Different weights are assigned to data sources (such as temperature and humidity during the experiment); approximate data is extracted and used to perform confirmatory evaluation on the data samples to determine the third evaluation value. This includes: calculating the similarity between the data samples and the data entering the second step of the screening process, and extracting data with a similarity greater than or equal to a threshold as approximate data; when the number of approximate data is greater than or equal to a preset number (at least 3), calculating the average data based on the approximate data; calculating the deviation of each parameter of the data samples from the average data; determining the second quantification value according to the pre-configured deviation quantification table corresponding to each parameter, and then performing a weighted sum to obtain the third evaluation value, with different weights assigned to each second quantification value. The confidence level is configured after professional personnel evaluate the situation of each data source.

[0036] like Figure 4 The figure shows a schematic diagram illustrating the changes in the accuracy of a data augmentation model based on physical constraints. Figure 4 The corresponding model, based on the original dataset, introduces physical constraints of the hydrogen storage material system to perform controlled perturbations on the original samples, generating enhanced samples. These physical constraints include alloy composition conservation constraints, reasonable range constraints for component values, and material structural stability constraints. Comparing the changes in model prediction accuracy before and after introducing physical constraints shows that, after data augmentation using physical constraints, the prediction accuracy gradually improves with the increase in the number of samples during model training, and the model's prediction results on test data are more stable. This demonstrates that data augmentation methods based on physical constraints can effectively avoid the adverse effects of physically meaningless samples on model training and are an important technical means to improve model accuracy.

[0037] like Figure 5 As shown, this is also a schematic diagram illustrating the changes in model accuracy based on the ensemble model. Figure 5The corresponding model employs multiple basic prediction models to learn the performance of hydrogen storage materials. Each basic prediction model fits the mapping relationship between material composition parameters and hydrogen storage performance parameters from different perspectives. Subsequently, the prediction results of multiple basic prediction models are integrated and fused to construct an integrated prediction model. By comparing the changes in prediction accuracy between the single model and the integrated model, it can be seen that the overall prediction accuracy of the integrated model is higher than that of any single prediction model. This demonstrates that integrating multiple prediction models can fully leverage the complementary advantages of different models and effectively improve the overall prediction performance of the model.

[0038] like Figure 6 As shown, this is also a schematic diagram of the changes in model accuracy based on algorithm tuning in an ensemble model. Figure 6 After constructing the ensemble prediction model, the corresponding model undergoes further optimization strategies, including sample weight adjustment, regularization constraints, and adaptive optimization of model parameters. These optimization measures reduce the interference of outlier and high-density samples on the model training process and improve the model's predictive consistency across different component regions. Figure 6 The changes in model accuracy shown indicate that the optimized ensemble model further improves its prediction accuracy, demonstrating that proper optimization of the ensemble model helps to fully leverage its predictive advantages under small sample conditions.

[0039] like Figure 7 The diagram illustrates the application of the proposed method in performance prediction of V-based hydrogen storage alloys. The small-sample, high-precision prediction model construction method described in this invention is applied to V-based hydrogen storage alloy systems to predict the performance of V-based hydrogen storage alloys with different composition ratios. The predicted performance parameters include hydrogen storage capacity and cycle stability. The prediction results can be used to guide the composition screening and optimization of V-based hydrogen storage alloys, thereby reducing the number of experimental trials and improving the efficiency of material development. This embodiment demonstrates that the method of this invention can achieve stable and reliable performance prediction in specific hydrogen storage material systems, and has good engineering application value.

[0040] The optimized integrated model showed a significantly improved overall goodness of fit, reaching approximately 0.9. Under this high goodness of fit, the predicted high-cycle-capacity stability alloy range was around V65Ti15Cr20. Using this composition, the alloy was arc-melted under a purified argon atmosphere (99.999% purity) and labeled V65. Before the experiment, the sample surface was polished to remove any possible oxides, and the blocky alloy sample was mechanically crushed into small particles. This treatment reduces oxide interference and increases the reaction surface area, thereby improving the efficiency of the activation reaction. The sample underwent 50 hydrogen absorption / desorption cycles. The final cycle stability of V65 reached 92.14%. This demonstrates that the results obtained under high goodness of fit can achieve high-precision prediction of alloy composition and properties.

[0041] Comparative Example 1: A method for constructing a high-precision prediction model for small samples of hydrogen storage materials, which is related to Figure 7 The difference between the corresponding models is that only a single model was used, while the rest of the preparation steps remained unchanged.

[0042] To verify the progressiveness of the embodiments of this application, this application... Figure 7 The corresponding model and the V-based alloy prepared in Comparative Example 1 were subjected to the following performance tests: The overall goodness of fit obtained from the single model test was around 0.3. Under low goodness of fit, the predicted alloy composition range with high cycle capacity stability was approximately V25Ti32Cr43. Using this composition, the alloy was prepared and arc-melted under a purified argon atmosphere (99.999% purity), and labeled as V25. Before the experiment, the sample surface was polished to remove any possible oxides, and the blocky alloy sample was mechanically crushed into small particles. This treatment reduces the interference of the oxide layer and increases the reaction surface area, thereby improving the efficiency of the activation reaction. The sample was subjected to 50 hydrogen absorption and desorption cycles. Figure 8 As shown, the final cycle stability of V25 is only 41.41%.

[0043] Comparative Example 2 A method for constructing a high-precision prediction model for small samples of hydrogen storage materials, which is related to Figure 7 The difference between the corresponding models is that data augmentation was used, while the rest of the preparation steps remained the same.

[0044] To verify the progressiveness of the embodiments of this application, this application... Figure 7 The corresponding model and the V-based alloy prepared in Comparative Example 2 were subjected to the following performance tests: The overall goodness of fit obtained by the single model plus data augmentation test is about 0.5. Under this goodness of fit, the predicted alloy composition range with high cycle capacity stability is about V40Ti30Cr40. Using the above composition, the alloy was arc-melted under a purified argon atmosphere (99.999% purity) and labeled as V40. Before the experiment, the sample surface was polished to remove any possible oxides, and the blocky alloy sample was mechanically crushed into small particles. This treatment reduces the interference of the oxide layer and increases the reaction surface area, thereby improving the efficiency of the activation reaction. The sample was subjected to 50 hydrogen absorption and desorption cycles. Figure 9 As shown, the final cycle stability of V40 is only 47.64%.

[0045] Comparative Example 3 A method for constructing a high-precision prediction model for small samples of hydrogen storage materials, which is related to Figure 7 The difference between the corresponding models is that data augmentation and ensemble models are used, while the rest of the preparation steps remain the same.

[0046] To verify the progressiveness of the embodiments of this application, this application... Figure 7 The corresponding model and the V-based alloy prepared in Comparative Example 3 were subjected to the following performance tests: The overall goodness of fit obtained by the integrated model and data augmentation test was approximately 0.7. Under this goodness of fit, the predicted alloy composition range with high cycle capacity stability was approximately V50Ti21Cr29. Using this composition, the alloy was prepared and arc-melted under a purified argon atmosphere (99.999% purity), and labeled V50. Before the experiment, the sample surface was polished to remove any possible oxides, and the blocky alloy sample was mechanically crushed into small particles. This treatment reduces the interference of the oxide layer and increases the reaction surface area, thereby improving the efficiency of the activation reaction. The sample was subjected to 50 hydrogen absorption and desorption cycles. Figure 10 As shown, the final V50 cycle stability is only 82.74%.

[0047] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for constructing a high-precision prediction model for small samples of hydrogen storage materials, characterized in that, include: Constructing the original material dataset and selecting feature variables; Data augmentation methods based on physical constraints and data fluctuations caused by experimental errors perturb the original material dataset and enhance feature variables to generate an enhanced material dataset. The enhanced material dataset is then combined with the original material dataset to construct the model training set. Automatic hyperparameter tuning was performed on the random forest model and the gradient boosting model respectively to obtain the optimal parameter combination under small sample conditions; By integrating and fusing the random forest model and the gradient boosting model and performing hyperparameter tuning, an integrated prediction model is constructed. Regularization and debiasing are performed on each sample in the model, and the influence of skewed data during the training process is suppressed by sample weighting or resampling.

2. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, Data augmentation methods include generating virtual samples based on data perturbation and / or interpolation methods based on experimental errors.

3. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, Automatic hyperparameter tuning is performed automatically using Bayesian optimization algorithms, tree-structured parameter search algorithms, or genetic algorithms.

4. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, Integrating and fusing random forest models and gradient boosting models includes: A stacked fusion method is adopted to integrate the prediction results of the random forest model and the gradient boosting model according to a set weight.

5. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 4, characterized in that, Stacking and fusion methods include: A preset first number of random forest models and a preset second number of gradient boosting models are used as base learners for the base learner layer; The third preset gradient boosting model is used as the meta-learner of the meta-learner layer; The base learner is used to process the samples to obtain the prediction output; the meta learner is used to combine the prediction outputs of the various base learners to obtain the comprehensive prediction output.

6. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, Also includes: The model is then optimized based on the residuals and subjected to secondary training to obtain the final prediction model.

7. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 6, characterized in that, The optimization steps for optimizing the samples in the model training set based on residuals are as follows: The predicted values ​​obtained from the first cross-validation of the model; The predicted values ​​are compared with the actual hydrogen storage performance data, the residuals are calculated, and the residual distribution is constructed. Set a threshold to label samples with high residuals; High residual samples with labels are removed from the model training set.

8. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, The hyperparameters include: parameters controlling the number of trees in the random forest model, parameters controlling the maximum depth of the trees in the random forest model, parameters controlling the minimum number of samples in the leaf nodes of the random forest model, parameters controlling the number of iterations of the gradient boosting model, parameters controlling the contribution of each tree to the final result, parameters controlling the complexity of each tree, parameters controlling the proportion of random samples used by each tree, and parameters controlling the proportion of features used by each tree.

9. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, The integrated prediction model adopts a weighted stacking structure, where the weights are adaptively generated based on the cross-validation performance of each model.

10. The method for constructing a small-sample, high-precision prediction model for hydrogen storage materials as described in claim 1, characterized in that, Also includes: Based on a pre-configured data acquisition rule base, corresponding reference data is obtained from the big data platform; Based on the acquisition type code in the data acquisition rule base corresponding to the reference data, the corresponding preset data conversion rule base is retrieved. Based on the retrieved data transformation rule base, the reference data is transformed to obtain the transformed dataset; Add the transformed dataset to the model training set.

Images