Spinel solar cell material multi-target performance prediction method and device based on multi-task gradient elevator, and storage medium

Abstract

The invention relates to a spinel solar cell material multi-target performance prediction method and device based on a multi-task gradient elevator and a storage medium. The multi-target performance prediction method comprises the steps that S1, spinel material data are extracted from a database and processed; s2, constructing a comprehensive feature set and screening key features; s3, MTGBM is selected for core model training; s4, dynamic weight distribution and other optimization models; and S5, predicting the multi-target performance of the spinel material by using the optimized model. Finally, multi-target performance prediction of the band gap value, the band gap type and the thermodynamic stability of the spinel material is achieved. Compared with the prior art, the method has the advantages that the influence trend of each characteristic of the spinel material on the multi-target performance can be provided, and meanwhile, a unified model capable of synchronously predicting the band gap value, the band gap type and the thermodynamic stability can be quickly constructed, so that the calculation cost of a traditional trial and error experiment is reduced.

Description

Technical Field

[0001] The present invention relates to the intersection of materials informatics and machine learning, and in particular to a multi-objective performance prediction method, device, and storage medium for spinel solar cell materials based on a multi-task gradient boosting machine. Background Art

[0002] In the research and development of spinel solar cell materials, traditional trial-and-error experiments and density functional theory calculations are time-consuming and costly, limiting the speed of screening and designing high-performance materials.

[0003] In the prior art, for example, reference patent CN119626395A provides a neural network-based screening method for multi-metal oxide positive electrode materials for lithium-ion batteries. Although the deep neural network prediction model effectively captures the complex relationship between material characteristics and performance, it focuses on lithium-ion battery materials and does not fully consider the unified prediction of multi-objective performance and the model reliability in small sample scenarios.

[0004] In addition, reference patent CN118783396A discloses a low-voltage distributed photovoltaic power prediction method. Although a personalized model training and fusion strategy is proposed in the training and optimization of the prediction model, it is applied to photovoltaic power prediction rather than multi-objective performance prediction of spinel solar cell materials, and fails to propose a specific solution for multi-task learning.

[0005] These reference patents fail to address key challenges such as feature screening, model generalization capability, and small sample prediction accuracy faced in multi-objective performance prediction of spinel solar cell materials.

[0006] Spinel materials (AB2X4) are considered potential candidate materials for solar cells due to their tunable composition and excellent optoelectronic properties. However, their research and development faces challenges with data scarcity and the complexity of multi-objective optimization: existing databases (such as the Materials Project) contain only 325 spinel compounds, resulting in insufficient data; traditional single-objective prediction models cannot account for the coordinated optimization of band gap value, band gap type (direct / indirect), and thermodynamic stability; and density functional theory (DFT)-based calculations are computationally expensive, making them difficult to use for large-scale screening. Existing machine learning methods (such as XGBoost and neural networks) are limited in data-scarce scenarios. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a multi-objective performance prediction method, equipment, and storage medium for spinel solar cell materials based on a multi-task gradient boosting machine. While providing the influence trend of various characteristics of spinel materials on multi-objective performance, it can also quickly construct a unified model that can simultaneously predict band gap values, band gap types and thermodynamic stability, so as to reduce the computational cost of traditional trial and error experiments.

[0008] On the other hand, the present invention also establishes a quantification system for the characteristic contribution of spinel materials. Through intelligent feature screening and sample weight compensation strategies, it ensures that the model training process is more in line with the physical mechanism of the material, significantly improving the prediction accuracy and generalization ability in small sample scenarios.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A first aspect of the present invention provides a multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine, comprising the following steps:

[0011] S1. Extract the band gap value, band gap type, and Ehull data of spinel materials from the Materials Project database, discretize the continuous values into binary classification labels, and expand the band gap value range to 0.9-2.0 eV to balance the data distribution;

[0012] S2. Based on the data processed by S1, a comprehensive feature set including basic descriptors, calculated descriptors, and composite descriptors is constructed, and key features are retained through pairwise correlation screening and stepwise forward selection iteration;

[0013] S3: Based on the feature set filtered by S2, MTGBM is selected as the core model for training;

[0014] S4: For the MTGBM model selected in S3, a dynamic weight allocation strategy is used to resolve multi-task gradient conflicts. Sample-level category compensation and the Optuna framework are combined to optimize hyperparameters to generate an optimized multi-task prediction model.

[0015] S5: Use the optimized MTGBM model to simultaneously output multi-objective performance prediction results including the band gap value, band gap type, and thermodynamic stability of spinel materials.

[0016] Furthermore, S1 specifically includes the following steps:

[0017] Data extraction: Extract the raw data of spinel materials from the Materials Project database and screen out the three key attributes: band gap value, band gap type, and Ehull;

[0018] Discretization: Based on the extracted data, the continuous values that meet the band gap value range of 1.1-1.5eV, direct band gap, and Ehull ≤ 0.025eV / atom are discretized into binary classification labels to form a preliminary classification data set;

[0019] Range expansion and balancing: In response to the uneven distribution of band gap value samples in the preliminary classification data set, the band gap value range is further expanded to 0.9-2.0eV, and the data distribution is balanced by expanding the effective sample coverage range.

[0020] Furthermore, S2 specifically includes the following steps:

[0021] Comprehensive feature construction: Based on the data processed by S1, basic descriptors, calculated descriptors, and composite descriptors are extracted based on the spinel chemical formula AB2X4 to form a multidimensional comprehensive feature set;

[0022] Iterative feature screening: For the multidimensional comprehensive feature set, redundant features are first removed through pairwise correlation screening, and then stepwise forward selection is used to iteratively retain 6 key features using AUC as the indicator to achieve feature dimensionality reduction and enhance physical interpretability.

[0023] Furthermore, the six key features include: maximum heat of fusion, tolerance factor, LUMO energy, average number of valence electrons, mean space group number, and maximum periodic difference of elements;

[0024] The composite descriptor is generated by the Matminer tool, which covers the electron cloud overlap energy and valence electron distribution statistics between elements.

[0025] Furthermore, S4 specifically includes the following steps:

[0026] Dynamic weight allocation: To address the multi-task gradient conflicts in the MTGBM model, dynamic weight allocation is performed for the band gap prediction task;

[0027] Sample-level category compensation: Based on the model after dynamic weight allocation, the positive sample weights are set to address the scarcity of positive samples with band gap values, and the cross-entropy loss function is modified;

[0028] Hyperparameter collaborative optimization: After completing the weight and compensation strategy adjustments, the Optuna framework is used to jointly optimize the regularization coefficient, the number of leaf nodes, and the feature sampling ratio. The optimal parameter combination is screened through Pareto frontier analysis to generate the final optimized model.

[0029] Furthermore, the dynamic weight allocation process specifically includes the following steps:

[0030] Based on the criticality of the bandgap prediction task, an initial static weight coefficient of 1.2 is assigned to it. The bandgap prediction task is prioritized in the multi-task gradient fusion process, so that the bandgap prediction task will not be suppressed by the gradient updates of other tasks due to sample scarcity.

[0031] During model training, the weight of each task is dynamically adjusted by a scaling factor of α = 1.2 based on the AUC performance of each task on the validation set. The task weight is positively correlated with its AUC value. The higher the AUC, the greater the weight allocation ratio, thereby adaptively balancing the gradient update direction of multiple tasks.

[0032] Furthermore, the sample-level category compensation specifically includes the following steps:

[0033] Positive sample weight calculation: According to the ratio of positive samples to negative samples with band gap values, the positive sample weight coefficient is calculated based on the total sample size and the number of positive samples;

[0034] Loss function modification: The calculated positive sample weight coefficient is embedded in the cross-entropy loss function. By adjusting the category weight distribution during loss calculation, the model pays more attention to the classification accuracy of positive samples during training, thereby improving the recall rate of positive samples.

[0035] Furthermore, S5 specifically includes the following steps:

[0036] The spinel material data that has undergone preprocessing and feature screening is fed into the optimized MTGBM model in S4 to activate its multi-task prediction architecture.

[0037] Based on the dynamic weight allocation mechanism and sample compensation strategy embedded in the model, the three performance indicators of band gap value classification, band gap type, and thermodynamic stability are calculated simultaneously, and finally a unified multi-objective prediction result is output.

[0038] The second aspect of the present invention provides an electronic device comprising a memory and a processor, wherein the processor is used to execute a program in the memory, thereby implementing the multi-objective performance prediction method of spinel solar cell materials based on a multi-task gradient boosting machine as described above.

[0039] The third aspect of the present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, is used to execute the multi-objective performance prediction method for spinel solar cell materials based on the multi-task gradient boosting machine as described above.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention realizes efficient collaborative prediction of the band gap value, band gap type and thermodynamic stability of spinel materials through a multi-task dynamic fusion mechanism, and synchronously outputs multi-objective performance results in a single model inference, significantly reducing the computational redundancy required for step-by-step modeling in traditional methods. At the same time, based on a dynamic weight allocation strategy, it balances the gradient conflicts between tasks, making the prediction results more in line with the physical correlation between material properties. For example, direct band gap materials are more likely to meet a specific band gap range, thereby providing priority guidance for experimental synthesis and significantly reducing the cost of trial and error.

[0042] Furthermore, the present invention addresses the challenges of data scarcity and sample imbalance by adopting a physically interpretable key feature screening system and a two-level compensation strategy to enhance model reliability from two levels: feature contribution quantification and loss function optimization. This ensures that feature selection conforms to the intrinsic mechanisms of crystal structure and electronic energy bands, and improves prediction accuracy and generalization capabilities in small sample scenarios by dynamically adjusting sample weights and task priorities, providing an efficient and reliable intelligent tool for the rational design of solar cell materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a method in an application example of the present invention;

[0044] Figure 2 This is a performance comparison chart of the multi-objective model in the application example of the present invention on the spinel dataset;

[0045] Figure 3 This is a trend chart of the number of features screened by SFS and model performance in an application example of the present invention;

[0046] Figure 4 This is a heat map of the correlation between the features after feature screening and the three target performances in the application example of the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, circuit structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0048] Example 1

[0049] In this embodiment, a new multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine solves the problems of data scarcity and multi-task optimization of spinel materials, and achieves high-precision and high-efficiency performance prediction.

[0050] The novel multi-objective performance prediction method of spinel solar cell materials based on the multi-task gradient boosting machine in this embodiment includes the following steps:

[0051] Step 1: Data collection and preprocessing. Extract spinel related data from the Materials Project (MP) database, focusing on the band gap value (E g ), band gap type (“is_gap_direct”), and energy above the convex hull (E hull ) These three key attributes. Given that the data in the initial dataset that meets the expected characteristics are scarce and unevenly distributed, the continuous values are discretized into two categories (True(1) or False(0)) for binary classification, and E is expanded. g The selection range is 0.9-2.0eV to alleviate the sample imbalance problem and provide a high-quality data foundation for subsequent feature engineering and model training.

[0052] Step 2: Feature Engineering. Using the pre-processed data obtained in step 1, a feature extractor was developed to calculate a comprehensive descriptor set consisting of 62 basic descriptors, 10 calculated descriptors, and 274 integrated descriptors to better characterize material properties from chemical formulas. Through pairwise correlation screening (threshold 0.90) and stepwise forward selection (SFS), six key features were finally retained, including the maximum heat of melting (h max ), tolerance factor (τ), LUMO energy (l e ).

[0053] Step 3: Model Training and Selection. Based on the feature dataset obtained after feature engineering in Step 2, evaluate multiple machine learning models, including neural networks (NN), CatBoost, XGBoost, and multi-task gradient boosting machines (MTGBM). Model evaluation is performed using metrics such as accuracy, F1 score, Hamming loss, and area under the curve (AUC). The best-performing model, MTGBM, is selected.

[0054] Step 4: Model construction and optimization. Based on the MTGBM model selected in step 3, the model construction and optimization for the three properties of spinel are carried out. First, the gradient weighting of dynamic weight allocation (formula: G n =w1G1+…+w n G n , where G i is the multi-task gradient, w i is the multi-task gradient weight coefficient) to solve the multi-task gradient conflict and give the bandgap task a higher weight; secondly, in view of the scarcity of positive samples of bandgap data, a weight matrix (positive sample weight c + =N total / N pos ), perform sample-level category compensation; then use the Optuna framework to optimize the regularization coefficient (λ 11 / λ12 ), the number of leaf nodes (num_leaves) and other parameters are used to screen the optimal combination of Pareto frontiers and optimize the model's hyperparameters.

[0055] Step 5: Model performance evaluation and validation. Using 5-fold cross-validation, the AUC metrics showed Eg predictions of 0.86, bandgap type of 0.91, and Ehull of 0.86. Compared to single-task GBDT, the Eg positive sample recall rate increased from 0.28 to 0.62.

[0056] In specific implementation, the specific process of data collection and preprocessing in step 1 is as follows:

[0057] Step A1: Extract 325 spinel data from the MP database, focusing on the band gap value (E g ), band gap type (“is_gap_direct”), and energy above the convex hull (E hull )) These three properties are crucial for solar cell applications. Step A2: Since the initial data set meets all the desired characteristics (E g In the range of 1.1-1.5eV, direct band gap, E hull ≤0.025eV / atom) is scarce and the data distribution is uneven, so the continuous values are discretized into two categories (True(1) or False(0)) for binary classification. Step A3: To alleviate the sample imbalance problem, E g The selection range is expanded to 0.9-2.0eV.

[0058] In specific implementation, the specific steps of feature engineering in step 2 are:

[0059] Step B1: Comprehensive descriptor construction, extracting the following features based on the material chemical formula (AB2X4): Basic descriptors (62 dimensions): including the ionic radius, oxidation state, electronegativity, atomic mass, etc. of the elements at the A / B / X sites; Computational descriptors (10 dimensions): band gap polarizability, crystal tolerance factor (τ=3(RB+RX) / 2(RA+RX)), octahedron factor, etc.; Composite descriptors (274 dimensions): generated using the Matminer tool, covering electron cloud overlap energy between elements, valence electron distribution statistics, etc. Step B2: Feature screening and dimensionality reduction, correlation screening: calculating the pairwise squared correlation between features, deleting features with correlations higher than 0.90, and reducing the feature dimension from 346 to 88; Stepwise forward selection (SFS): using AUC as the evaluation metric, iteratively adding features that improve model performance the most, and ultimately retaining 6 key features: maximum heat of fusion (h max ), tolerance factor (τ), LUMO energy (l e ), average valence electron number (e mean ), mean value of space group number (s mean), the maximum value of the element period difference (p max ).

[0060] In specific implementation, the specific steps of model training and selection in step 3 are:

[0061] Step C1: Constructing a candidate model set. The following machine learning models were selected for comparison: Neural Network (NN): a 3-layer fully connected network with Reluctant Unit (ReLU) activation function; CatBoost: default parameters, supporting categorical feature processing; XGBoost: tree depth set to 6, learning rate 0.1; MTGBM: a multi-task gradient boosting machine with an initial learning rate of 0.05. Step C2: Model evaluation and selection. Evaluation metrics included accuracy, F1 score, Hamming loss, and Area Under Correlation (AUC). Comparison results showed that MTGBM achieved an AUC of 0.86 in the Eg task, significantly higher than XGBoost (0.79) and CatBoost (0.75). MTGBM was ultimately selected.

[0062] In specific implementation, the specific steps of model construction and optimization in step 4 are:

[0063] Step D1: Dynamic weight allocation, to resolve multi-task gradient conflicts, define the integrated gradient G c =w1G1+w2G2+w3G3, where the weight coefficients are allocated according to task importance (w1=2.0, w2=0.5, w3=0.5); the weights are adaptively adjusted according to the task AUC correlation:

[0064] Where α is the scaling factor (set to 1.2 in the experiment).

[0065] Step D2: Sample-level category compensation, for E g Positive sample scarcity (positive-negative sample ratio 1:6.56), set the category weight matrix: positive sample weight where N total is the total number of samples (325), that is, the number of all spinel materials in the dataset. N pos is the number of positive samples (43), that is, to meet the target performance (such as E g The number of materials in the range of 1.1-1.5eV. c+ is the positive sample weight, which is used to compensate for data imbalance. The negative sample weight c_=1; modified cross entropy loss function: where y i,j is the true label of the jth task of the i-th sample (0 or 1). pi,j is the probability (between 0 and 1) that the jth task of the i-th sample is positive.

[0066] Step D3: Hyperparameter optimization, using the Optuna framework to search for the optimal parameter combination, with the optimization goal of maximizing Eg , bandgap type, E hull The weighted sum of AUC; key optimization parameters: regularization coefficient (λ 11 ,λ 12 ), the number of leaf nodes (num_leaves∈[20,200]), the feature sampling ratio (feature_fraction∈[0.4,0.9]); through Pareto frontier analysis (see Figure 3 ), and selected the optimal parameter combination (λ11=0.001,num_leaves=168,feature_fraction=0.47).

[0067] In specific implementation, the specific steps for model performance evaluation and verification in step 5 are as follows:

[0068] Step E1: Cross-validation and indicator calculation, using 5-fold cross-validation, dividing the data into 80% training set and 20% test set; calculate the average AUC value: E g Prediction 0.86, bandgap type 0.91, E hull 0.86; compared with the single-task GBDT model, E g The recall rate of positive samples increased from 0.28 to 0.62.

[0069] Application Example 1

[0070] like Figure 1 As shown in the figure, the multi-objective performance prediction method of spinel solar cell materials based on the multi-task gradient boosting machine (MTGBM) in this application example has the following specific steps:

[0071] Step 1: Data collection and preprocessing:

[0072] Step A1: Extract data of 325 spinel compounds from Materials Project (MP) database, focusing on band gap values (E g ), band gap type (direct / indirect) and energy above the convex hull (E hull ) three key attributes.

[0073] Step A2: Due to the compliance with the target characteristics (E g =1.1-1.5eV, direct band gap, E hull Since the data of Eg≤0.025eV / atom are scarce and unevenly distributed, the continuous values are discretized into binary labels (True / False), and the Eg range is extended to 0.9-2.0eV to balance the samples.

[0074] Step A3: E g , bandgap type, E hull Data standardization was performed to construct a preprocessed data set containing 325 samples.

[0075] Step 2: Feature Engineering:

[0076] Step B1: Extract a 346-dimensional feature set based on the chemical formula (AB2X4), including: basic descriptors (62 dimensions): ionic radius, electronegativity, atomic mass, etc. of the elements at the A / B / X sites; calculation descriptors (10 dimensions): tolerance factor LUMO energy (l e ), octahedral factor, etc.; composite descriptor (274 dimensions): generate electron cloud overlap energy between elements, valence electron distribution statistics, etc. through Matminer.

[0077] Step B2: Through pairwise correlation screening (threshold 0.90) and stepwise forward selection (SFS), 6 key features (such as Figure 3 As shown in Figure 2, the performance of the model during SFS screening significantly improved with the increase in the number of features): Maximum heat of melting (h max ),τ,l e 、Average valence electron number( emean ), mean value of space group number (s mean ), the maximum value of the element period difference (p max ).

[0078] Step B3: Figure 3 The impact of feature number on model performance was revealed: During the key feature screening phase (1-3 features), the AUC rapidly increased from 0.78 to 0.86, demonstrating that the addition of τ and le significantly improved the model's representational capabilities. During the performance saturation phase (4-6 features), the AUC stabilized at 0.86-0.88, validating the rationale for selecting 6 features to avoid overfitting.

[0079] Step 3: Model training and selection

[0080] Step C1: Compare the neural network (NN), CatBoost, XGBoost, and MTGBM models. The evaluation metrics include AUC, F1 score, and Hamming loss.

[0081] Step C2: Figure 2 As shown in the figure, MTGBM achieves an accuracy of 0.84 in multi-task collaborative prediction, which is significantly better than XGBoost (0.79) and CatBoost (0.82), and obtains a higher F1 score and lower Hamming loss, verifying the effectiveness of its gradient sharing mechanism. Finally, MTGBM was selected as the core model.

[0082] Step 4: Model optimization:

[0083] Step D1: Dynamic weight allocation (Formula: G n=w1G1+…+w n G n ) Solve the multi-task gradient conflict and give E g The task has a higher weight (w1=1.2).

[0084] Step D2: Target E g Positive sample scarcity (positive-negative ratio 1:6.56), set category weight compensation (positive sample weight ).

[0085] Step D3: Optimize the regularization coefficient (λ using the Optuna framework 11 = 0.001), the number of leaf nodes (num_leaves = 168) and other parameters to screen the optimal combination of Pareto frontier (such as Figure 4 As shown in Figure 3, the SHAP value heatmap of feature and target performance verifies the physical rationality of parameter optimization).

[0086] Step D4: The correlation between the features and the target performance is quantified by the SHAP value heat map, such as Figure 4 Shown: Bandgap value (E g ): Tolerance factor (τ) and E g Strong positive correlation (SHAP = 0.68), LUMO energy (le) is negatively correlated (SHAP = -0.71), consistent with crystal field theory. Band gap type: space group number mean (s mean ) is positively correlated with the direct band gap (SHAP=0.41), and high symmetry crystals are more likely to form direct band gaps. Thermodynamic stability (E hull ): Maximum melting heat (h max ) is negatively correlated with stability (SHAP = -0.58). Materials with high melting heat have higher lattice energy and lower stability.

[0087] Step 5: Performance Evaluation and Verification:

[0088] Step E1: Using 5-fold cross validation, MTGBM is g , bandgap type, E hull The AUCs in prediction were 0.86, 0.91, and 0.86, respectively.

[0089] Step E2: E g The positive sample recall rate increased from 0.28 of single-task GBDT to 0.62, verifying the effectiveness of the dynamic weight and sample compensation strategy.

[0090] In summary, the present invention, by studying known spinel material experimental and computational data, can analyze the influence of various features on the multi-objective performance of materials while being able to quickly predict the optimal material composition that meets specific performance requirements, thereby greatly reducing the resources required for traditional trial-and-error experiments and density functional theory calculations, and significantly accelerating the research and development process of spinel solar cell materials. By constructing a multi-task dynamic fusion mechanism and an intelligent feature screening strategy, 6 key descriptors with clear physical meanings are extracted from 346-dimensional features, and combined with the Optuna hyperparameter optimization framework, high precision and strong generalization capabilities of the multi-objective prediction model are achieved. In addition, the established spinel material feature contribution quantification system and gradient sharing mechanism provide guidance for multi-task learning, ensure that the model optimization process closely fits the intrinsic properties of the material, and provide a data-driven innovative solution for the design and screening of high-performance spinel solar cells.

[0091] Example 2

[0092] This embodiment provides an electronic device for implementing multi-objective performance prediction of spinel solar cell materials. The electronic device includes a memory and a processor, wherein the processor is configured to execute a program in the memory.

[0093] The memory stores a program for multi-objective performance prediction of spinel solar cell materials. The program is based on a multi-task gradient boosting machine (MTGBM) model and can implement a complete process from data preprocessing to final performance prediction.

[0094] When the processor runs the program, it first extracts data such as the band gap value, band gap type, and Ehull number of spinel materials from the Materials Project database. It then discretizes this data, converting continuous values into binary classification labels, and expands the band gap range to 0.9-2.0 eV to balance the data distribution.

[0095] Then, a comprehensive feature set is constructed based on the processed data, including basic descriptors, calculated descriptors, and composite descriptors. Through pairwise correlation screening and stepwise forward selection iteration, key features are retained to form a streamlined and effective feature set.

[0096] Based on this feature set, we selected MTGBM as the core model for training. For the MTGBM model, we adopted a dynamic weight allocation strategy to resolve multi-task gradient conflicts. We also combined sample-level category compensation and the Optuna framework to optimize hyperparameters, generating an optimized multi-task prediction model.

[0097] Ultimately, the optimized MTGBM model simultaneously outputs multi-objective performance predictions for spinel materials, including bandgap value, bandgap type, and thermodynamic stability. This provides accurate performance evaluation and guidance for the research and development of spinel solar cell materials. The electronic device efficiently executes this prediction process, meeting the demand for material performance prediction in practical applications.

[0098] Example 3

[0099] This embodiment provides a storage medium containing computer-executable instructions. When executed by a computer processor, the storage medium is used to perform the multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine. The specific steps are as follows:

[0100] (1) The band gap values, band gap types, and Ehull data of spinel materials were extracted from the Materials Project database. The continuous values were discretized into binary classification labels and the band gap value range was expanded to 0.9-2.0 eV to balance the data distribution.

[0101] (2) Based on the processed data, a comprehensive feature set consisting of basic descriptors, calculated descriptors, and composite descriptors is constructed, and key features are retained through pairwise correlation screening and stepwise forward selection iteration.

[0102] (3) Based on the filtered feature set, MTGBM is selected as the core model for training.

[0103] (4) For the selected MTGBM model, a dynamic weight allocation strategy is adopted to resolve multi-task gradient conflicts, and the hyperparameters are optimized by combining sample-level category compensation and the Optuna framework to generate an optimized multi-task prediction model.

[0104] (5) The optimized MTGBM model is used to simultaneously output multi-objective performance prediction results including the band gap value, band gap type, and thermodynamic stability of spinel materials.

[0105] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention, without departing from the scope of the present invention, should be within the scope of protection of the present invention.

Claims

1. A multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine, characterized in that: The following steps are involved: S1. Extract the band gap value, band gap type, and Ehull data of spinel materials from the Materials Project database, discretize the continuous values into binary classification labels, and expand the band gap value range to 0.9-2.0 eV to balance the data distribution; S2. Based on the data processed by S1, a comprehensive feature set including basic descriptors, calculated descriptors, and composite descriptors is constructed, and key features are retained through pairwise correlation screening and stepwise forward selection iteration; S3: Based on the feature set filtered by S2, MTGBM is selected as the core model for training; S4: For the MTGBM model selected in S3, a dynamic weight allocation strategy is used to resolve multi-task gradient conflicts. Sample-level category compensation and the Optuna framework are combined to optimize hyperparameters to generate an optimized multi-task prediction model. S5: Use the optimized MTGBM model to simultaneously output multi-objective performance prediction results including the band gap value, band gap type, and thermodynamic stability of spinel materials.

2. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 1, characterized in that: S1 specifically includes the following steps: Data extraction: Extract the raw data of spinel materials from the Materials Project database and screen out the three key attributes: band gap value, band gap type, and Ehull; Discretization: Based on the extracted data, the continuous values that meet the band gap value range of 1.1-1.5eV, direct band gap, and Ehull ≤ 0.025eV / atom are discretized into binary classification labels to form a preliminary classification data set; Range expansion and balancing: In response to the uneven distribution of band gap value samples in the preliminary classification data set, the band gap value range is further expanded to 0.9-2.0eV, and the data distribution is balanced by expanding the effective sample coverage range.

3. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 2, characterized in that: In S2, specifically The following steps are involved: Comprehensive feature construction: Based on the data processed by S1, basic descriptors, calculated descriptors, and composite descriptors are extracted based on the spinel chemical formula AB2X4 to form a multidimensional comprehensive feature set; Iterative feature screening: For the multidimensional comprehensive feature set, redundant features are first removed through pairwise correlation screening, and then stepwise forward selection is used to iteratively retain 6 key features using AUC as the indicator to achieve feature dimensionality reduction and enhance physical interpretability.

4. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 3, characterized in that: The six key features include: maximum heat of fusion, tolerance factor, LUMO energy, average number of valence electrons, mean space group number, and maximum periodic difference of elements; The composite descriptor is generated by the Matminer tool, which covers the electron cloud overlap energy and valence electron distribution statistics between elements.

5. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 1, characterized in that: S4 specifically includes the following steps: Dynamic weight allocation: To address the multi-task gradient conflicts in the MTGBM model, dynamic weight allocation is performed for the band gap prediction task; Sample-level category compensation: Based on the model after dynamic weight allocation, the positive sample weights are set to address the scarcity of positive samples with band gap values, and the cross-entropy loss function is modified; Hyperparameter collaborative optimization: After completing the weight and compensation strategy adjustments, the Optuna framework is used to jointly optimize the regularization coefficient, the number of leaf nodes, and the feature sampling ratio. The optimal parameter combination is screened through Pareto frontier analysis to generate the final optimized model.

6. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 5, characterized in that: The dynamic weight allocation process specifically includes the following steps: Based on the criticality of the bandgap prediction task, an initial static weight coefficient of 1.2 is assigned to it. The bandgap prediction task is prioritized in the multi-task gradient fusion process, so that the bandgap prediction task will not be suppressed by the gradient updates of other tasks due to sample scarcity. During model training, the weight of each task is dynamically adjusted by a scaling factor of α = 1.2 based on the AUC performance of each task on the validation set. The task weight is positively correlated with its AUC value. The higher the AUC, the greater the weight allocation ratio, thereby adaptively balancing the gradient update direction of multiple tasks.

7. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 5, characterized in that: The sample-level category compensation specifically includes the following steps: Positive sample weight calculation: According to the ratio of positive samples to negative samples with band gap values, the positive sample weight coefficient is calculated based on the total sample size and the number of positive samples; Loss function modification: The calculated positive sample weight coefficient is embedded in the cross-entropy loss function. By adjusting the category weight distribution during loss calculation, the model pays more attention to the classification accuracy of positive samples during training, thereby improving the recall rate of positive samples.

8. The multi-objective performance prediction method for spinel solar cell materials based on a multi-task gradient boosting machine according to claim 1, characterized in that: S5 specifically includes the following steps: The spinel material data that has undergone preprocessing and feature screening is fed into the optimized MTGBM model in S4 to activate its multi-task prediction architecture. Based on the dynamic weight allocation mechanism and sample compensation strategy embedded in the model, the three performance indicators of band gap value classification, band gap type, and thermodynamic stability are calculated simultaneously, and finally a unified multi-objective prediction result is output.

9. An electronic device comprising a memory and a processor, characterized in that: The processor is used to execute the program in the memory to implement the multi-objective performance prediction method of spinel solar cell materials based on a multi-task gradient boosting machine as described in any one of claims 1 to 8.

10. A storage medium containing computer-executable instructions, characterized in that: The storage medium of the computer executable instructions is used to execute the multi-objective performance prediction method of spinel solar cell materials based on a multi-task gradient boosting machine as described in any one of claims 1 to 8 when executed by a computer processor.

Images