Performance prediction method and system for full-flow process of copper indium gallium selenium solar cell

Abstract

The invention provides a performance prediction method and system for a full-flow process of a copper-indium-gallium-selenium solar cell, and the method comprises the steps: obtaining experimental data covering a complete device structure and a preparation process, removing redundant features through data cleaning, standardization and feature correlation analysis, introducing an SHAP game theory method to analyze a model, and carrying out the optimization of the performance of the whole-flow process of the copper-indium-gallium-selenium solar cell. Quantifying the marginal contribution degree of the process parameters to the photovoltaic performance; and finally, determining a parameter convergence interval of the high-efficiency device by using a parallel coordinate technology, and outputting a performance prediction value. According to the method, the modeling problem of the copper-indium-gallium-selenium solar hybrid heterogeneous data is effectively solved, the prediction result is ensured to conform to the physical law of a device through physical constraint, the trade-off effect among complex process parameters is revealed, clear parameter guidance is provided for efficient battery preparation, and the trial and error cost is greatly reduced.

Description

Technical Field

[0001] This invention relates to the fields of photovoltaic device manufacturing and artificial intelligence application technology, and in particular to a method and system for performance prediction of the entire process of copper indium gallium selenide solar cells. Background Technology

[0002] Copper indium gallium selenide (CIGS) thin-film solar cells have become one of the most competitive thin-film photovoltaic technologies due to their high photoelectric conversion efficiency and good long-term stability. However, CIGS cells are typical complex multilayer heterojunction systems, and their fabrication process involves a long and delicate entire process chain, including substrate cleaning, back electrode sputtering, absorption layer growth (such as co-evaporation and post-sputtering selenization), buffer layer chemical bath deposition, high-resistivity window layer and conductive window layer preparation, metal grid electrode evaporation, and anti-reflection layer coating.

[0003] This entire process involves dozens of process parameters (e.g., deposition temperature, elemental composition ratios, material composition of each layer, layer thickness, doping process, etc.). These parameters exhibit strong nonlinear coupling relationships and complex physical trade-offs. For example, while increasing the gallium (Ga) content can increase the bandgap and improve the open-circuit voltage, it often leads to a decrease in short-circuit current. Traditional CIGS R&D mainly relies on a "trial and error" approach, requiring researchers to conduct massive orthogonal experiments to find the optimal parameter combination. This not only results in long R&D cycles and high reagent consumption but also makes it extremely difficult to locate the global optimum in the multidimensional parameter space.

[0004] Existing auxiliary research and development methods mainly include physical simulation software (such as SCAPS-1D, AMPS, etc.), but the prediction accuracy of such software is highly dependent on the input of microscopic physical parameters such as defect energy level density and interface recombination rate. These microscopic parameters are extremely difficult to measure accurately in actual experiments, resulting in simulation results often having large deviations from experimental preparations.

[0005] In recent years, machine learning technology has shown great potential in the field of materials science. However, existing research in the CIGS field has limitations: First, CIGS process data has a typical "hybrid heterogeneous" characteristic, that is, it simultaneously contains continuous "numerical features" (such as temperature: 550℃, thickness: 2.5μm) and discrete "categorical features" (such as substrate type: soda-lime glass / polyimide; buffer layer material: CdS / Zn(O,S)). Traditional algorithms (such as linear regression, support vector regression, and random forest) usually require one-hot encoding when processing high-dimensional categorical features, which leads to an extremely sparse feature matrix. This not only increases the computational burden but also severs the potential connections between categorical features, resulting in limited prediction accuracy. Second, most existing research is limited to parameter optimization at the single level of the absorption layer (such as focusing only on the CGI and GGI ratios), ignoring the synergistic impact of the overall device structure parameters, such as the interface layer, window layer, and back contact layer, on the final efficiency. Third, many high-precision models are considered "black boxes," unable to explain the specific physical mechanisms to process engineers, nor can they provide intuitive parameter optimization ranges, making it difficult to implement the models in actual production lines. Summary of the Invention

[0006] In view of the above, the main objective of this invention is to propose a performance prediction method and system for the entire process of copper indium gallium selenide (CIGS) solar cells, in order to solve the aforementioned technical problems.

[0007] This invention proposes a performance prediction method for the entire process of copper indium gallium selenide (CIGS) solar cells, the method comprising the following steps: Step 1: Obtain experimental data of copper indium gallium selenide (CIGS) solar cells, and perform data cleaning and standardization to obtain a standard dataset; Step 2: Based on the device physics principle of copper indium gallium selenide solar cells, calculate the physical derived feature set from the fabrication process parameters in the standard dataset, and fuse the physical derived features with the fabrication process parameters to construct the physical enhancement feature matrix. Step 3: Analyze the correlation between features in the physical enhancement feature matrix to remove redundant features with multicollinearity, and identify numerical and categorical features to construct a hybrid feature input matrix; divide the dataset based on the hybrid feature input matrix into training and test sets. Step 4: Construct a prediction model based on a classification boosting tree model; train the prediction model using the training set and optimize it using a loss function guided by physical constraints to obtain the trained prediction model. Step 5: Introduce a game theory-based interpretability analysis method to analyze the trained prediction model in order to obtain the marginal contribution of each process parameter to the photovoltaic performance index. Step 6: Screen key process parameters based on marginal contribution, and use parallel coordinate visualization technology to map the distribution trajectory of high photovoltaic performance samples on the key process parameter axis to identify the dense convergence region of the sample trajectory; determine the dense convergence region of the sample trajectory as the convergence interval of the optimal process parameters.

[0008] This invention also proposes a performance prediction system for the entire process of copper indium gallium selenide (CIGS) solar cells. The system employs the performance prediction method for the entire CIGS solar cell process described above, and includes: The data management module is used for: Experimental data of copper indium gallium selenide (CIGS) solar cells were obtained, and the data was cleaned and standardized to obtain a standard dataset. The data processing module is used for: Based on the device physics principle of copper indium gallium selenide solar cells, a physical derived feature set is calculated from the fabrication process parameters in the standard dataset. The physical derived features are then fused with the fabrication process parameters to construct a physical enhancement feature matrix. The correlation between features in the physical enhancement feature matrix is ​​analyzed to remove redundant features with multicollinearity, and numerical and categorical features are identified to construct a hybrid feature input matrix. The dataset based on the hybrid feature input matrix is ​​then divided into a training set and a test set. The model calculation module is used for: Construct a prediction model based on a classification boosting tree model; train the prediction model using the training set and optimize it using a loss function guided by physical constraints to obtain the trained prediction model. The intelligent analysis module is used for: A game theory-based interpretability analysis method is introduced to analyze the trained prediction model in order to obtain the marginal contribution of each process parameter to the photovoltaic performance index. Key process parameters are selected based on marginal contribution, and the distribution trajectory of high photovoltaic performance samples on the key process parameter axis is mapped using parallel coordinate visualization technology to identify the dense convergence region of the sample trajectory; the dense convergence region of the sample trajectory is determined as the convergence interval of the optimal process parameters.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The machine learning model constructed in this invention has more than 20 key parameters from the substrate to the anti-reflection layer as input. Compared with the study of a single absorption layer, it can better reflect the synergistic effect of multilayer heterojunctions and the coupling mechanism of the entire process of real devices. 2. This invention constructs a set of physically derived features (equivalent band gap, interface barrier height, optical loss coefficient, etc.) to explicitly embed the band theory, carrier transport theory and optical theory of CIGS solar cells into the machine learning model, realizing a leap from pure data-driven to physical knowledge-guided, enabling the model to not only fit the data, but also understand the physical mechanism. 3. The physical constraint-guided loss function proposed in this invention introduces physical mechanism constraints such as theoretical limit constraints, current-voltage-fill factor consistency constraints, and bandgap-open circuit voltage relationship constraints to ensure that the model prediction results do not violate the physical laws of the device. This avoids the high-precision but unreasonable predictions that may be generated by pure black box models, and significantly improves the reliability and engineering practicality of the model in extrapolation scenarios. 4. By introducing adaptive ordered target statistical technology, this invention perfectly solves the high-dimensional sparse matrix problem caused by the mixing of "material category" and "process value" in CIGS process. Compared with the traditional one-hot encoding method, it avoids dimensionality explosion and retains the correlation information between category features and target values, significantly improving prediction accuracy. 5. The "parameter convergence interval" determined by visualization through parallel coordinates provides experimenters with a clear process window, significantly reducing trial and error costs.

[0010] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of a performance prediction method for a copper indium gallium selenide (CIGS) solar cell process proposed in this invention. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0014] Please see Figure 1 This embodiment provides a method for performance prediction of the entire process of copper indium gallium selenide (CIGS) solar cells, the method including the following steps: Step 1: Obtain experimental data of copper indium gallium selenide (CIGS) solar cells, and perform data cleaning and standardization to obtain a standard dataset.

[0015] In step 1, experimental data of copper indium gallium selenide (CIGS) solar cells are acquired and cleaned and standardized to obtain a standard dataset. The experimental data includes complete device structure parameters from substrate to top electrode, fabrication process parameters, and corresponding photovoltaic performance test indicators. The experimental data exhibits hybrid heterogeneous characteristics, containing both continuous numerical features and discrete categorical features. The complete device structure parameters and fabrication process parameters include at least: substrate type, maximum substrate temperature, back electrode material type, back electrode thickness, absorber layer fabrication process type, absorber layer thickness, gallium-indium ratio, copper-gallium-indium ratio, alkali metal doping method, potassium cyanide etching treatment, buffer layer material type, buffer layer thickness, window layer material type, window layer thickness, transparent conductive oxide layer material type, transparent conductive oxide layer thickness, grid electrode material type, grid electrode thickness, antireflection layer configuration, effective solar cell area, and grid electrode shading area ratio. The photovoltaic performance test indicators include: photoelectric conversion efficiency, open-circuit voltage, short-circuit current density, and fill factor. Specifically, data cleaning includes: using box plots to identify and remove outliers from continuous variables; The standardization process specifically includes: using the Z-Score standardization method to process numerical features, and the following relationship exists in the corresponding process: ; in, This represents the standardized parameter value. Indicates the original parameter value. Indicates parameters The mean, Indicates parameters The standard deviation.

[0016] Step 2: Based on the device physics principle of copper indium gallium selenide solar cells, calculate the physical derived feature set from the fabrication process parameters in the standard dataset, and fuse the physical derived features with the fabrication process parameters to construct the physical enhancement feature matrix.

[0017] In step 2, based on the device physics principle of copper indium gallium selenide solar cells, a physical derived feature set is calculated from the fabrication process parameters in the standard dataset. The physical derived features are then fused with the fabrication process parameters to construct a physical enhancement feature matrix. The physical derived feature set includes: equivalent bandgap, interface barrier height, and optical loss coefficient. The physical derived feature set is used to characterize the internal physical mechanisms of the device. The process of calculating a physical feature set based on the device physics principles of copper indium gallium selenide (CIGS) solar cells from fabrication process parameters in a standard dataset includes the following sub-steps: Based on the gallium-indium ratio, the equivalent bandgap is calculated using empirical formulas, and the following relationship exists in the process: ; in, Indicates the equivalent band gap. Indicates the gallium-indium ratio; The absorption layer composition is determined based on the gallium-indium ratio and the copper-gallium-indium ratio. The bottom conduction band level is determined by querying a material database based on the buffer layer material type and the absorption layer composition. The conduction band offset is then calculated to obtain the interface barrier height. The following relationship exists in this process: ; in, This indicates the height of the interface barrier, which affects the interface recombination rate and carrier transport efficiency. The conduction band bottom level indicates the type of buffer layer material. This represents the lower conduction band energy level of the absorber layer composition; The optical loss coefficient is calculated based on the antireflection layer configuration and the ratio of the transparent conductive oxide layer thickness to the gate electrode shading area. The following relationship exists in the process: ; in, Indicates the optical loss coefficient. This indicates the reflectivity of the antireflective layer. This represents the absorption coefficient of the transparent conductive oxide layer. Indicates the thickness of the transparent conductive oxide layer. This indicates the ratio of the light-shielding area of ​​the gate electrode.

[0018] It should be noted that physically derived features can explicitly characterize the internal band structure, carrier transport, and optical properties of CIGS solar cells, enabling the model to learn parameter-performance mapping relationships that conform to physical laws.

[0019] Step 3: Analyze the correlation between features in the physical enhancement feature matrix to remove redundant features with multicollinearity, and identify numerical and categorical features to construct a hybrid feature input matrix; divide the dataset based on the hybrid feature input matrix into training and test sets.

[0020] In step 3, the correlation between the features in the physical enhancement feature matrix is ​​analyzed to eliminate redundant features with multicollinearity. This includes the following sub-steps: Based on the standard dataset, the Pearson correlation coefficient between any two process parameters is calculated, and the following relationship exists in the process: ; in, express and The Pearson correlation coefficient between them Represents the mathematical expectation. and All represent process parameters. express The mean, express The mean, express standard deviation express Standard deviation; like If the correlation exceeds the preset threshold, then it is determined that... and Strong collinearity exists, and it is eliminated. and Features with low correlation to photovoltaic performance test indicators are selected to eliminate feature redundancy.

[0021] For example, "gate electrode thickness" is usually highly correlated with "gate electrode material type" (the choice of material determines its thickness). In this case, "gate electrode thickness" can be removed, and "gate electrode material type" with more information can be retained, thereby constructing a simplified hybrid feature input matrix.

[0022] It should be noted that the dataset based on the mixed feature input matrix is ​​divided into a training set and a test set in an 8:2 ratio, which are used for model training and model evaluation, respectively.

[0023] Step 4: Construct a prediction model based on a classification boosting tree model; train the prediction model using the training set and optimize it using a loss function guided by physical constraints to obtain the trained prediction model; the loss function includes: prediction error term, physical mechanism constraint term and process constraint.

[0024] In step 4, a prediction model based on a classification boosting tree model is constructed; the prediction model is trained using the training set and optimized using a loss function guided by physical constraints to obtain the trained prediction model; wherein, the loss function includes: a prediction error term, a physical mechanism constraint term, and a process constraint term, and the expression of the loss function is: ; in, Represents the loss function. This represents the prediction error term. and All represent weighting coefficients. Represents physical mechanism constraints. Indicates process constraints; The expression for the prediction error term is as follows: ; in, Indicates the number of samples. Indicates the first Predicted photoelectric conversion efficiency for each sample This represents the actual photoelectric conversion efficiency; The physical mechanism constraints include theoretical limit constraints, current-voltage-fill factor consistency constraints, and bandgap-open-circuit voltage relationship constraints. The expressions for the physical mechanism constraints are as follows: ; in, Represents theoretical limit constraints. This indicates taking the maximum value. This indicates the predicted photoelectric conversion efficiency. This represents the upper limit of theoretical efficiency. This indicates a current-voltage-fill factor consistency constraint. This represents the short-circuit current density predicted by the model. This represents the open-circuit voltage predicted by the model. This represents the fill factor predicted by the model. Indicates the standard incident light power. This represents the constraint on the bandgap-open-circuit voltage relationship. Indicates the amount of electron charge. Indicates empirical voltage loss; The expression for the process constraint term is as follows: ; in, and All represent weighting coefficients. This indicates the constraint on the rationality of layer thickness. This indicates constraints on the rationality of the components; It should be noted that the four weighting coefficients , , and All were determined through cross-validation to balance the importance of data fitting and physical constraints; the loss function guided by the physical constraints ensures that the model prediction results are not only statistically accurate, but also physically reasonable, avoiding predictions that may violate physical laws that may be produced by a purely data-driven model.

[0025] Furthermore, when training the prediction model using the training set, an adaptive ordered target statistical method based on the Sigmoid function is used to process the categorical features in the mixed feature input matrix and convert the categorical features into numerical features. The following relationship exists in the corresponding process: ; in, Indicates the first The first sample The numerical encoded values ​​of each category feature after conversion Indicates the weighting coefficient. This indicates the permutation index of the current sample. This indicates an indicator function, if the first element in the permutation... The first sample The category features and the first The first sample If all features have the same category value, then it equals 1; otherwise, it equals 0. Indicates the first The sample at the th Category values ​​on each feature, Indicates the first The sample at the th Category values ​​on each feature, This indicates the index of the sample currently undergoing encoding computation. Indicates that in a random permutation, it is located in the sample. Previous historical sample index, Index representing features, Indicates the first The target value for each sample This indicates that the position in the random permutation is the th The number of samples with the same category value preceding each sample. This represents a small constant used to prevent the denominator from being zero. Represents the prior probability. Represents the natural constant. This represents the center offset parameter of the Sigmoid function, which is used to control the sample size threshold for weight switching; This represents the slope adjustment parameter of the Sigmoid function, used to control the smoothness of weight changes.

[0026] It should be noted that the data in the fabrication process of CIGS solar cells exhibits a significant "long-tailed distribution" characteristic. For example, the sample size of some mainstream substrates (such as soda-lime glass) is huge, while the sample size of some novel flexible substrates or specific doping processes is extremely small (sparse category). Traditional CatBoost ordered target statistics typically use fixed smoothing parameters, which has limitations when dealing with mixed CIGS data with vastly different sample sizes. For categories with sufficient sample size, fixing the smoothing parameter may introduce unnecessary prior bias. For categories with very small sample sizes, a fixed smoothing parameter may not be sufficient to suppress noise.

[0027] This invention creatively introduces an adaptive weighting mechanism based on the Sigmoid function, providing targeted improvements to the aforementioned problems: Weighting coefficient The physical meaning of is that it represents the model's confidence in the "current category statistical mean".

[0028] parameter Its function is to define the "flip threshold" for confidence levels. When the number of samples for a certain process category... < When the Sigmoid function outputs a value of 0, the model primarily relies on global prior probabilities. This avoids random errors caused by small samples (avoiding target leakage); when > When the Sigmoid function outputs a value of 1, the model mainly relies on the statistical mean of the category itself to accurately capture the specific impact of the process.

[0029] parameter Its function is to control the sensitivity to changes in confidence level. The smaller the value, the steeper the weight switching, enabling a rapid phase transition from "trusting priors" to "trusting data".

[0030] Step 5: Introduce a game theory-based interpretability analysis method to analyze the trained prediction model in order to obtain the marginal contribution of each process parameter to the photovoltaic performance index.

[0031] In step 5, a game theory-based interpretability analysis method is introduced to analyze the trained prediction model in order to obtain the marginal contribution of each process parameter to the photovoltaic performance index. This includes the following sub-steps: An interpretability analysis method based on game theory is introduced to calculate the SHAP values ​​of each process parameter, and the following relationship exists in the corresponding process: ; in, Representation of features SHAP value, Represents a set The middle does not contain features a subset of To represent factorial, Represents the set of all input features. Indicates in subset Add features based on The predicted value of the post-target prediction model. Indicates using only subsets The features in the target prediction model are the predicted values; The absolute value of the SHAP value of each process parameter is taken as the marginal contribution of each process parameter to the photovoltaic performance index.

[0032] It should be noted that, through SHAP analysis, this invention reveals key physical mechanisms, such as: the application of antireflection layer (ARC) is the dominant factor in improving photoelectric conversion efficiency; the GGI ratio has a significant positive contribution to open-circuit voltage, but an excessively high GGI will lead to a decrease in short-circuit current density, revealing the trade-off effect in bandgap modulation.

[0033] Step 6: Screen key process parameters based on marginal contribution, and use parallel coordinate visualization technology to map the distribution trajectory of high photovoltaic performance samples on the key process parameter axis to identify the dense convergence region of the sample trajectory; determine the dense convergence region of the sample trajectory as the convergence interval of the optimal process parameters.

[0034] In step 6, key process parameters are screened based on marginal contribution, and the distribution trajectory of high photovoltaic performance samples on the key process parameter axis is mapped using parallel coordinate visualization technology to identify dense convergence regions of the sample trajectories. These dense convergence regions are then determined as the optimal process parameter convergence intervals. This process includes the following sub-steps: Based on the marginal contribution of each process parameter to photovoltaic performance indicators, the process parameters are sorted in descending order, and process parameters with a marginal contribution greater than a preset threshold are selected as key process parameters. By using parallel coordinate visualization technology, key process parameters are mapped to a two-dimensional plane to obtain the key process parameter axes. Set a target optimal threshold for photoelectric conversion efficiency; Extract a set of samples whose performance prediction values ​​are higher than the target preferred threshold, and identify the dense convergence region of the sample set on the key process parameter axis, and determine the dense convergence region as the optimal process parameter convergence interval.

[0035] This embodiment also provides a performance prediction system for the entire process of copper indium gallium selenide (CIGS) solar cells. The system applies the performance prediction method for the entire CIGS solar cell process described above, and includes: The data management module is used for: Experimental data of copper indium gallium selenide (CIGS) solar cells were obtained, and the data was cleaned and standardized to obtain a standard dataset. The data processing module is used for: Based on the device physics principle of copper indium gallium selenide solar cells, a physical derived feature set is calculated from the fabrication process parameters in the standard dataset. The physical derived features are then fused with the fabrication process parameters to construct a physical enhancement feature matrix. The correlation between features in the physical enhancement feature matrix is ​​analyzed to remove redundant features with multicollinearity, and numerical and categorical features are identified to construct a hybrid feature input matrix. The dataset based on the hybrid feature input matrix is ​​then divided into a training set and a test set. The model calculation module is used for: Construct a prediction model based on a classification boosting tree model; train the prediction model using the training set and optimize it using a loss function guided by physical constraints to obtain the trained prediction model. The intelligent analysis module is used for: A game theory-based interpretability analysis method is introduced to analyze the trained prediction model in order to obtain the marginal contribution of each process parameter to the photovoltaic performance index. Key process parameters are selected based on marginal contribution, and the distribution trajectory of high photovoltaic performance samples on the key process parameter axis is mapped using parallel coordinate visualization technology to identify the dense convergence region of the sample trajectory; the dense convergence region of the sample trajectory is determined as the convergence interval of the optimal process parameters.

[0036] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0037] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0038] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0039] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for performance prediction of a copper indium gallium selenide (CIGS) solar cell process, characterized in that, The method comprises the following steps: Step 1, obtaining experimental data of copper-indium-gallium-selenium solar cells, and performing data cleaning and standardization processing to obtain a standard data set; Step 2, based on the device physical principle of copper-indium-gallium-selenium solar cells, the physical derived feature set is calculated from the preparation process parameters in the standard data set, the physical derived features and the preparation process parameters are fused to obtain a physical enhanced feature matrix; Step 3, analyze the correlation between each feature in the physical enhanced feature matrix to eliminate redundant features with multiple collinearity, and identify numerical features and categorical features to construct a mixed feature input matrix; Divide the data set based on the mixed feature input matrix into a training set and a test set; Step 4, constructing a prediction model based on a classification boosting tree model; training the prediction model using the training set, and using a loss function guided by physical constraints for optimization to obtain a trained prediction model; Step 5, introducing an explainability analysis method based on game theory to analyze the trained prediction model to obtain the marginal contribution of each process parameter to the photovoltaic performance index; Step 6, screening key process parameters based on marginal contribution, and using parallel coordinate visualization technology to map the distribution trajectory of high photovoltaic performance samples on the key process parameter axis to identify the dense convergence area of the sample trajectory; Determine the dense convergence area of the sample trajectory as the optimal process parameter convergence interval.

2. The method for performance prediction of a full process of a copper indium gallium selenide solar cell according to claim 1, characterized in that, In the step 1, the experimental data includes complete device structure parameters from the substrate to the top electrode, preparation process parameters and corresponding photovoltaic performance test indicators; the experimental data has mixed heterogeneous characteristics, including continuous numerical features and discrete categorical features; wherein, the complete device structure parameters and preparation process parameters at least include: substrate type, maximum substrate temperature, back electrode material type, back electrode thickness, absorption layer preparation process type, absorption layer thickness, gallium-indium ratio, copper-gallium-indium ratio, alkali metal doping treatment method, potassium cyanide etching treatment, buffer layer material type, buffer layer thickness, window layer material type, window layer thickness, transparent conductive oxide layer material type, transparent conductive oxide layer thickness, gate electrode material type, gate electrode thickness, anti-reflection layer setting, solar cell effective area, and gate electrode shading area ratio; the photovoltaic performance test indicators include: photoelectric conversion efficiency, open circuit voltage, short circuit current density, and fill factor. 3.The performance prediction method of a whole process of a copper-indium-gallium-selenium solar cell according to claim 2, characterized in that, In the step 1, the experimental data of copper-indium-gallium-selenium solar cells is obtained, and data cleaning and standardization processing is performed to obtain a standard data set, wherein the data cleaning processing specifically includes: using the box plot rule to identify and eliminate outliers in continuous variables; Wherein, the standardization processing specifically includes: using the Z-Score standardization method to process numerical features, and there is the following relationship in the corresponding process: ； wherein denotes the standardized parameter value, denotes the original parameter value, denotes the parameter of the mean, denotes the standard deviation of the parameter .

4. The method according to claim 3, wherein, In the step 2, based on the device physical principle of copper-indium-gallium-selenium solar cells, the physical derived feature set is calculated from the preparation process parameters in the standard data set, the physical derived features and the preparation process parameters are fused to obtain a physical enhanced feature matrix; The physical derived feature set includes: equivalent band gap, interface barrier height, optical loss coefficient; The physical derived feature set is calculated from the preparation process parameters in the standard data set based on the physical principles of the copper-indium-gallium-selenium solar cell device, and specifically includes the following sub-steps: Based on the gallium-indium ratio, the equivalent band gap is calculated using an empirical formula, and the following relationship exists in the corresponding process: ； wherein, represents the equivalent bandgap, represents the gallium to indium ratio; Based on the gallium-indium ratio and the copper-gallium-indium ratio, the absorption layer composition is determined, the conduction band minimum is determined according to the buffer layer material type and the absorption layer composition, and the conduction band offset is calculated to obtain the interface barrier height, and the following relationship exists in the corresponding process: ； wherein, represents an interface barrier height, the interface barrier affects the interface recombination rate and the carrier transport efficiency; represents a conduction band minimum energy level of the buffer layer material type, represents a conduction band minimum energy level of the absorber layer component; Based on the anti-reflection layer setting, the thickness of the transparent conductive oxide layer and the ratio of the gate electrode light shielding area, the optical loss coefficient is calculated, and the following relationship exists in the corresponding process: ； wherein, represents an optical loss coefficient, represents a reflectance of the antireflection layer, represents an absorption coefficient of the transparent conductive oxide layer, represents a thickness of the transparent conductive oxide layer, represents a light shielding area ratio of the gate electrode. 5.The performance prediction method of a full-process technology of a copper-indium-gallium-selenium solar cell according to claim 4, characterized in that, In step 3, the correlation between each feature in the physical enhanced feature matrix is analyzed to eliminate redundant features with multiple collinearity, specifically including the following sub-steps: Based on the standard data set, the Pearson correlation coefficient between any two process parameters is calculated, and the following relationship exists in the corresponding process: ； wherein denotes the Pearson correlation coefficient between , denotes the mathematical expectation, and both denote a process parameter, denotes the mean value of , denotes the mean value of , denotes the standard deviation of , denotes the standard deviation of ; If greater than a preset correlation threshold, it is determined that there is strong collinearity between and and the feature with lower correlation with the photovoltaic performance test index among is eliminated to eliminate feature redundancy. 6.The performance prediction method of a full-process technology of a copper-indium-gallium-selenium solar cell according to claim 5, wherein, In step 4, a prediction model based on a classification boosting tree model is constructed; the prediction model is trained using the training set, and a loss function guided by physical constraints is used for optimization to obtain the trained prediction model; wherein the loss function includes: a prediction error term, a physical mechanism constraint term and a process constraint term, and the expression of the loss function is: ； wherein, represents a loss function, represents a prediction error term, and both represent a weight coefficient, represents a physical mechanism constraint term, represents a process constraint term; Wherein, the expression of the prediction error term is: ； wherein, represents the number of samples, represents the predicted photoelectric conversion efficiency of the th sample, represents the true photoelectric conversion efficiency; Wherein, the physical mechanism constraint term includes a theoretical limit constraint, a current-voltage-fill factor consistency constraint and a band gap-open circuit voltage relationship constraint, and the expression of the physical mechanism constraint term is: ； wherein, denotes a theoretical limit constraint, denotes a maximum, denotes a predicted photoelectric conversion efficiency, denotes a theoretical efficiency upper limit, denotes a current-voltage- fill factor consistency constraint, denotes a model predicted short circuit current density, denotes a model predicted open circuit voltage, denotes a model predicted fill factor, denotes a standard incident light power, denotes a bandgap-open circuit voltage relationship constraint, denotes an electronic charge amount, denotes an empirical voltage loss; Wherein, the expression of the process constraint term is: ； wherein and both represent a weight coefficient, represents a layer thickness rationality constraint, represents a component rationality constraint. 7.The performance prediction method of a full-process technology of a copper-indium-gallium-selenium solar cell according to claim 6, characterized in that, In step 4, when the prediction model is trained using the training set, the adaptive ordered target statistical method based on the Sigmoid function is used to process the category type features in the mixed feature input matrix, and the category type features are converted into numerical type features, and the following relationship exists in the corresponding process: ； wherein, represents the th category feature of the th sample after conversion, represents a weight coefficient, represents a permutation index of the current sample, represents an indicator function, represents the th sample on the th feature, represents the th sample on the th feature, represents a sample index currently undergoing encoding calculation, represents a historical sample index located before the sample in a random permutation, represents an index of a feature, represents a target value of the th sample, represents a sample statistical number located before the th sample in a random permutation and having the same category value, represents a small constant to prevent the denominator from being zero, represents a prior probability, represents a natural constant, represents a center offset parameter of a Sigmoid function, represents a slope adjustment parameter of a Sigmoid function. 8.The performance prediction method of a full-process technology of a copper-indium-gallium-selenium solar cell according to claim 7, characterized in that, In step 5, a game theory-based interpretability analysis method is introduced to analyze the trained prediction model to obtain the marginal contribution of each process parameter to the photovoltaic performance index, specifically including the following sub-steps: A game theory-based interpretability analysis method is introduced to calculate the SHAP value of each process parameter, and the following relationship exists in the corresponding process: ； wherein, represents a feature SHAP value, represents a subset of the set of features represents a factorial, represents a set of all input features, represents a prediction value of the target prediction model after adding the feature to the subset , represents a prediction value of the target prediction model using only the features in the subset . The absolute value of the SHAP value of each process parameter is taken as the marginal contribution of each process parameter to the photovoltaic performance index. 9.The performance prediction method of a full-process technology of a copper-indium-gallium-selenium solar cell according to claim 8, characterized in that, In step 6, based on the marginal contribution, the key process parameters are screened, and the parallel coordinate visualization technology is used to map the distribution trajectory of the high photovoltaic performance sample on the key process parameter axis to identify the densely converging area of the sample trajectory; the densely converging area is determined as the optimal process parameter convergence interval, specifically including the following sub-steps: Based on the size of the marginal contribution of each process parameter to the photovoltaic performance index, each process parameter is sorted in descending order, and the process parameter with a marginal contribution greater than a preset threshold is selected as a key process parameter. The key process parameters are mapped to a two-dimensional plane by using parallel coordinate visualization technology to obtain a key process parameter axis; A target optimal threshold of photoelectric conversion efficiency is set; A sample set whose performance prediction value is higher than the target optimal threshold is extracted, and a dense convergence area of the sample set on the key process parameter axis is identified, and the dense convergence area is determined as an optimal process parameter convergence interval.

10. A performance prediction system for a full process of a copper indium gallium selenide solar cell, characterized by, The system adopts the performance prediction method of the full-process technology of the copper-indium-gallium-selenium solar cell according to any one of claims 1 to 9, and the system comprises: The data management module is configured to: Obtain experimental data of the copper-indium-gallium-selenium solar cell, and perform data cleaning and standardization processing to obtain a standard data set; The data processing module is configured to: Based on the device physics principle of the copper-indium-gallium-selenium solar cell, a physical derived feature set is calculated from the preparation process parameters in the standard data set, the physical derived features and the preparation process parameters are fused to obtain a physical enhanced feature matrix; The correlation between the features in the physical enhanced feature matrix is analyzed to eliminate redundant features with multiple collinearity, and the numerical features and the categorical features are identified to obtain a mixed feature input matrix; the data set based on the mixed feature input matrix is divided into a training set and a test set; The model calculation module is configured to: Construct a prediction model based on a classification boosting tree model, train the prediction model using the training set, and optimize the prediction model using a loss function guided by physical constraints to obtain a trained prediction model; The intelligent analysis module is configured to: Introduce an explainability analysis method based on game theory to analyze the trained prediction model to obtain the marginal contribution of each process parameter to the photovoltaic performance index; Based on the marginal contribution, key process parameters are screened, and the distribution trajectory of high photovoltaic performance samples on the key process parameter axis is mapped by using parallel coordinate visualization technology to identify a dense convergence area of the sample trajectory; and the dense convergence area of the sample trajectory is determined as an optimal process parameter convergence interval.

Images