A method for predicting and preparing compound silicon heterojunction solar cell parameters based on machine learning
By accurately predicting the UV/O3 treatment time of compound silicon heterojunction solar cells using a self-supervised machine learning system, the problems of strong experimental dependence and high cost in existing technologies are solved, and efficient process parameter optimization and cell performance improvement are achieved.
Patent Information
- Application Number
- CN202610257302.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies rely on destructive experiments and human experience to determine the optimal UV/O3 treatment time for compound silicon heterojunction solar cells. This makes it difficult to accurately capture the nonlinear coupling between interfacial oxidation kinetics and carrier transport, resulting in high time and cost, and poor generalization ability of the model when there is a small amount of noisy data.
A self-supervised machine learning approach was adopted, which uses a multilayer perceptron model with variational autoencoder and multi-head self-attention mechanism, combined with adversarial sample generation, to construct a self-supervised learning system that accurately predicts UV/O3 processing time and guides physical experiments to prepare silicon oxide passivation layers.
It significantly improved the accuracy of process parameter prediction, shortened the R&D cycle, improved the passivation performance and stability of the battery, enhanced the photoelectric conversion efficiency, and reduced R&D costs.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of photovoltaic device performance optimization and artificial intelligence computing, and in particular to a method for predicting the fabrication parameters of compound silicon heterojunction solar cells based on self-supervised machine learning. Background Technology
[0002] Compound silicon heterojunction solar cells have great application prospects in the field of high-efficiency photovoltaics due to their excellent photoelectric conversion efficiency. In cell fabrication, growing an ultrathin silicon oxide passivation layer on the surface of intrinsic hydrogenated amorphous silicon through ultraviolet / ozone (UV / O3) treatment is a core process for reducing the interface defect state density and improving the open-circuit voltage. However, the UV / O3 treatment time is extremely critical: too short a time will result in insufficient silicon oxide layer thickness, failing to effectively passivate dangling bonds; too long a time will result in an excessively thick silicon oxide layer, hindering carrier tunneling transport and increasing series resistance. In traditional methods, obtaining the optimal UV / O3 treatment time usually relies on a large amount of destructive experimental trial and error and physical characterization verification. This is not only time-consuming and material-costing, but also difficult to accurately capture due to the strong nonlinear coupling between interface oxidation kinetics and carrier transport, which is difficult to capture accurately by human experience. In recent years, although machine learning has begun to be applied to photovoltaic prediction, existing models have poor generalization ability when faced with a small amount of noisy test data. Therefore, there is an urgent need for an intelligent closed-loop solution that can deeply explore the mapping relationship between battery electrical parameters and preparation parameters, has high robustness, and can directly guide physical experiments. Summary of the Invention
[0003] This application aims to provide a method for predicting and preparing parameters of compound silicon heterojunction solar cells based on machine learning, including the following steps:
[0004] (1) Obtain test data of compound silicon heterojunction solar cells, separate the input feature set containing key JV parameters of the cell, and the output feature set containing preparation and other parameters such as UV / ozone treatment time, and perform data preprocessing by standard deviation standardization;
[0005] (2) Construct a variational autoencoder, which maps the preprocessed input feature set data to the latent space through the encoder, and reconstructs the data through the decoder. Self-supervised learning is performed using reconstruction loss and KL divergence loss to extract the latent representation of the data.
[0006] (3) Calculate the gradient of the input features based on the loss function of the model used in step (2), add adversarial perturbation to the original input feature set to generate adversarial samples, and merge the adversarial samples with the original samples to form an enhanced training set;
[0007] (4) Construct a multi-layer perceptron model that integrates multi-head self-attention mechanism and L2 regularization, and use the enhanced training set described in step (3) to perform adversarial iterative training on the model;
[0008] (5) The new data is standardized and then input into the trained multilayer perceptron model to obtain the predicted value. The mapping relationship of the output features is used to perform inverse standardization and output the optimal UV / O3 processing time and other parameters.
[0009] (6) Provide a sample of crystalline silicon solar cell substrate, clean it with hydrofluoric acid solution to remove the natural oxide layer on the surface, rinse it with deionized water and blow it dry;
[0010] (7). Place the cleaned crystalline silicon solar cell substrate sample in the reaction chamber of the UV / O3 equipment, and use the optimal UV / O3 treatment time predicted in step (5) as the equipment preparation time parameter. Turn on the switch to perform in-situ ultraviolet irradiation and ozone oxidation treatment on the surface of the substrate sample to grow an ultrathin silicon oxide passivation layer.
[0011] (8) Transfer the solar cell sample with the silicon oxide passivation layer deposited to the deposition equipment, and deposit the subsequent heterojunction carrier transport layer functional film and metal electrode in sequence to complete the preparation and packaging of the compound silicon heterojunction solar cell.
[0012] The preparation method further includes the following steps:
[0013] S1, Data Collection and Preprocessing: Acquire test data of compound silicon heterojunction solar cells, separate the input feature set and output feature set, and standardize the input feature set and output feature set using the standardized models scaler_x (for standardizing input features) and scaler_y (for standardizing output features);
[0014] S2, Variational Autoencoder (VAE) Feature Learning: The variational autoencoder maps the input data to the latent space through the encoder and reconstructs the data through the decoder. It uses a combination of reconstruction loss and KL divergence loss as the loss function for self-supervised learning to extract the latent representation of the data and expand the data sample.
[0015] S3, Adversarial Sample Generation: To improve the robustness of the model, the system adds adversarial perturbations to the original input features based on the gradient of the input features relative to the loss function in step S2, thereby generating adversarial samples. These adversarial samples are then merged with the original samples to form a larger and more informative augmented training set for subsequent deep learning training.
[0016] S4, Model Construction and Adversarial Training: Construct a multilayer perceptron (MLP) model that integrates multi-head self-attention mechanism, and use the enhanced training set described in step S3 to iteratively train the multilayer perceptron (MLP) model that integrates multi-head self-attention mechanism.
[0017] S5, Parameter Prediction and Denormalization: After training, a multilayer perceptron (MLP) with a multi-head self-attention mechanism is used to predict new data (new data corresponding to the feature set): First, the new data is normalized, then the normalized input feature values are predicted, and finally, the predicted values output by the model are denormalized using the normalized model scaler_y of the output feature values, thereby outputting the optimal UV / O3 (UV / O3) treatment time and the corresponding core electrical performance parameters;
[0018] S6, Composite passivation layer experimental preparation: A crystalline silicon solar cell sample is provided, with one side being an n-type crystalline silicon surface.
[0019] First, the crystalline silicon solar cell sample is placed in a plasma-enhanced chemical vapor deposition (RF-PECVD) reaction chamber to deposit an intrinsic hydrogenated amorphous silicon layer on the n-type crystalline silicon surface. Based on the optimal UV / O3 treatment time predicted in step S5, an ultrathin silicon oxide layer is grown on the surface of the intrinsic hydrogenated amorphous silicon layer on the back side using a UV / O3 device. A molybdenum oxide hole transport layer is then prepared on the silicon oxide surface using a thermal evaporation device to create a multilayer passivated contact structure.
[0020] In steps (1) and S1:
[0021] The input feature set includes eight key JV parameters of solar cells: implicit open-circuit voltage iV. OC , pseudo-fill factor pFF, open-circuit voltage V OC Fill factor FF, short-circuit current J SC Maximum power point current I max Maximum power point voltage V max Maximum power P max ;
[0022] The output feature set includes four parameters: preparation parameters and performance parameters: UV / O3 treatment time (preparation parameter), effective carrier lifetime τ. eff Surface recombination current density J0, series resistance R s .
[0023] In steps (2) and S2, the specific structure of the variational autoencoder is as follows:
[0024] Encoder: Contains two fully connected layers to output the mean and log-variance of the latent space, and generates latent variables through a sampling function;
[0025] Decoder: Contains two fully connected layers that receive the latent variables and map them back to the original data space to output reconstructed data;
[0026] The variational autoencoder was compiled and optimized using the Adam optimizer.
[0027] In steps (4) and S4, the MLP model that integrates the multi-head self-attention mechanism includes:
[0028] Hidden layers: include multiple fully connected layers with ReLU activation function and L2 regularization;
[0029] Self-attention layer: It contains two multi-head self-attention layers. The first self-attention layer has 8 attention heads and a key dimension (key_dim) of 32, and the second self-attention layer has 16 attention heads and a key dimension of 64.
[0030] Regularization layers: Layer normalization and dropout layers are connected after each self-attention layer;
[0031] Output layer: A fully connected layer with a linear activation function.
[0032] During the training process in steps (4) and S4, the training data is randomly shuffled at the beginning of each training epoch, and the validation loss is evaluated using the validation set at the end of each epoch.
[0033] In steps (5) and S5: the output feature values before denormalization, i.e., the code language of the output feature values (code in the model, computer language), and the output feature values after denormalization, i.e., the physical language of the output feature values (UV / O3 processing time, τ). eff J0, R s ).
[0034] In step S6, the specific process of preparing the test includes:
[0035] The crystalline silicon solar cell is a silicon-based heterojunction solar cell, wherein the silicon base is n-type crystalline silicon, and a multilayer passivated contact structure is prepared on its surface.
[0036] The surface of the n-type crystalline silicon is a pyramidal textured surface, and the overall thickness of the sample silicon wafer is 70~200 micrometers.
[0037] The intrinsic hydrogenated amorphous silicon layer was prepared using RF-PECVD and has a thickness of 4–12 nanometers.
[0038] The silicon oxide layer is prepared using an ultraviolet / ozone cleaning machine, with a processing time of 1 to 10 seconds.
[0039] The molybdenum oxide hole transport layer is prepared using a thermal evaporation deposition apparatus with a chamber pressure below 10. -4 Pa, with a molybdenum oxide thickness of 10 nanometers.
[0040] The silicon heterojunction solar cell structure further described above: The upper surface of the n-type monocrystalline silicon substrate, from top to bottom, consists of: an intrinsic hydrogenated amorphous silicon layer, an n... ＋ The lower surface of the n-type single-crystal silicon substrate consists of a nanocrystalline silicon oxide doped layer, a TCO transparent conductive oxide layer, and silver gate lines. The layers arranged downwards are an intrinsic hydrogenated amorphous silicon layer, a silicon oxide layer, a molybdenum oxide layer, and a silver electrode layer.
[0041] This invention involves performing UV / O3 treatment on the back side of an intrinsic hydrogenated amorphous silicon layer on the back of an n-type monocrystalline silicon cell to generate a silicon oxide layer. The fabrication method employs machine learning-driven techniques to effectively capture complex nonlinear relationships between parameters. This significantly improves the prediction accuracy of process parameters and drastically shortens the development cycle by precisely guiding UV / O3 oxidation experiments. Simultaneously, the multilayer passivation contact structure for crystalline silicon solar cells provided by this invention effectively suppresses atomic diffusion between the intrinsic hydrogenated amorphous silicon and the molybdenum oxide transport layer, while simultaneously improving the passivation performance and stability of the cell. This effectively enhances the open-circuit voltage and fill factor of the crystalline silicon solar cell, thereby improving its photoelectric conversion efficiency. Attached Figure Description
[0042] Figure 1 A structural diagram of a crystalline silicon solar cell provided for an embodiment of the present invention.
[0043] Figure 2 This is a structural diagram of a crystalline silicon solar cell sample wafer with one side being an n-type crystalline silicon surface, provided as an embodiment of the present invention.
[0044] Figure 3 This is a schematic diagram of a multilayer passivated contact structure for use in crystalline silicon solar cells, provided as an example of an embodiment of the present invention.
[0045] Figure 1 In the middle, the structure of the crystalline silicon solar sample wafer: 1-silver gate line, 2-TCO transparent conductive oxide, 3-n ＋ 4-Nanocrystalline silicon oxygen doped layer, 5-Intrinsic hydrogenated amorphous silicon layer, 6-n-type single crystal silicon substrate, 7-Intrinsic hydrogenated amorphous silicon layer, 8-Silicon oxide layer (treated with UV / O3 in this invention), 9-Molybdenum oxide layer, 10-Silver electrode layer, 11-Multilayer passivated contact structure provided by this invention.
[0046] Figure 2 The crystalline silicon solar cell sample used in the experiment has the following structure: 1-silver grid lines, 2-TCO transparent conductive oxide, 3-n ＋ Nanocrystalline silicon oxygen-doped layer, 4-Intrinsic hydrogenated amorphous silicon layer, 5-n-type single crystal silicon substrate, 11-Crystal silicon interface with multilayer passivated contact structure prepared in this invention.
[0047] Figure 3 In this invention, the multilayer passivated contact structure is: 5-n-type single-crystal silicon substrate, 6-intrinsic hydrogenated amorphous silicon layer, 7-silicon oxide layer, 8-molybdenum oxide layer, 11-crystalline silicon interface for preparing the multilayer passivated contact structure in this invention.
[0048] Figure 4 This is a schematic diagram of the Pearson correlation coefficient matrix.
[0049] Figure 5 The graph shows the loss function of the algorithm model. Detailed Implementation
[0050] The technical solution of the present invention will be described in detail below with reference to specific embodiments, but the present invention is not limited to the following embodiments.
[0051] Example 1
[0052] 1. A parameter prediction method for compound silicon heterojunction solar cells based on self-supervised machine learning, the specific execution logic and code implementation steps of which are as follows:
[0053] 2. Data Loading and Preprocessing: The training, validation, and test sets, separated by spaces, are read using `pd.read_csv`. Eight JV parameters are extracted as input features x (implied open-circuit voltage iV). OC , pseudo-fill factor pFF, open-circuit voltage V OC Fill factor FF, short-circuit current J SC Maximum power point current I max Maximum power point voltage V max Maximum power P max ); Extract four parameters as output features y (UV / O3 processing time, effective carrier lifetime τ). eff Surface recombination current density J0, series resistance R s The Standard Scaler is used to standardize features and labels, eliminating dimensional differences, and the standardization rules of the training set are applied simultaneously to the validation and test sets.
[0054] 3. Variational Autoencoder (VAE) Feature Learning: To learn the latent representation of photovoltaic data, a VAE model maximizing the lower bound of evidence (ELBO) is constructed. The encoder outputs the mean and log-variance of the latent space through two fully connected layers, which are then used to generate the latent variable z via a sampling function. The decoder receives the variable z and reconstructs the original data. The loss function consists of the reconstruction loss (mean squared error) and the KL divergence loss, which measures the difference between the latent distribution and the standard normal distribution, and is compiled using the Adam optimizer.
[0055] 4. Adversarial Training Design: Write the `adversarial_training` function to calculate the gradient of the input `x` with respect to the model's loss function. Based on the gradient direction, add adversarial perturbations to the original input to generate adversarial samples. In each epoch of training, concatenate the adversarial samples with the original samples to form a mixed dataset, allowing the model to learn to resist measurement fluctuations in the input parameters.
[0056] 5. Constructing an MLP model with multi-head attention: The front end uses multiple fully connected layers with ReLU activation and L2 regularization (penalty coefficient 0.1). This is followed by a two-stage multi-head self-attention module: the first layer uses 8 attention heads (key_dim=32); the second layer uses 16 attention heads (key_dim=64). Each layer is followed by Layer Normalization and Dropout layers to prevent overfitting. The final layer uses a linear activation function and directly outputs the predicted UV / O3 processing time and τ. eff J0, R s .
[0057] 6. Model Training, Prediction, and Destandardization: In train_with_adversarial (adversarial training, the same training model as above), the data is shuffled at the beginning of each epoch and the validation set loss is evaluated at the end. After the model converges, new battery design data to be optimized is input, and after the model's inference output, the output features are used to standardize the model scaler_y for destandardization to obtain absolute physical quantities, especially the optimal process guidance parameters: the optimal UV / O3 processing time.
[0058] 7. In this stage, the prediction results from the above machine learning (the optimal UV / O3 treatment time is 4s) are directly applied to physical preparation to form an ultrathin oxide layer passivation contact structure:
[0059] 8. Pretreatment of crystalline silicon substrate: The n-type crystalline silicon wafer is cleaned with hydrofluoric acid solution to remove the natural oxide layer on the surface, rinsed with deionized water and dried;
[0060] 9. UV / O3 In-situ Oxidation (Core Experimental Step): The cleaned silicon wafer is placed in the reaction chamber of the UV / O3 equipment. The "optimal UV / O3 treatment time" predicted by machine learning is input into the equipment's control program. Under the combined action of ultraviolet light excitation and ozone, a precise and controllable oxidation reaction occurs on the silicon wafer surface, growing a dense silicon oxide tunneling layer with uniform thickness (approximately 1.0~2.0 nm). The precise introduction of the predicted time ensures that the oxide layer is sufficient to passivate the dangling bonds on the silicon surface without blocking carrier tunneling.
[0061] 10. Subsequent thin film deposition: On the silicon wafer with the silicon oxide layer prepared, subsequent heterojunction functional thin films (such as carrier transport layer thin films) and front and rear metal electrodes are deposited sequentially using appropriate thin film deposition techniques (such as PECVD, thermal evaporation, etc.) to finally complete the fabrication and packaging of compound silicon heterojunction solar cells.
[0062] A comparative experiment was conducted between the embodiments of this application (a crystalline silicon solar cell using intrinsic hydrogenated amorphous silicon / silicon oxide / molybdenum oxide as the back contact) and a comparative example. The comparative example was a crystalline silicon solar cell using intrinsic hydrogenated amorphous silicon / molybdenum oxide as the back contact (equivalent to having no silicon oxide layer), while the remaining film structure was the same as that of the embodiments of this application. Standard solar intensity calibration was performed using a solar simulator, and the calibration was conducted on a 16 cm² area. 2 The device in the embodiment underwent a long-term JV test, with the starting voltage set to 1V, the cutoff voltage to 0V, the starting current to -1A, and the cutoff current to 1A. The test results are shown in the table below.
[0063] Devices Short-circuit current (mA) Open circuit voltage (mV) Fill factor (%) Photoelectric conversion efficiency (%) Examples of this application 38.78 732.3 82.48 23.42 Comparative Example 38.84 743.5 84.68 24.45
[0064] Beneficial effects
[0065] In summary, the method for predicting and preparing parameters of compound silicon heterojunction solar cells based on machine learning described in this application embodiment has at least the following advantages:
[0066] 1. Enhanced Anti-interference and Robustness: This invention dynamically generates adversarial samples by calculating the gradient of the relative loss of the input data, and then trains the model by mixing these adversarial samples with the original data. This step forces the model to learn more fundamental physical laws rather than data noise, greatly improving the model's robustness to noisy data in actual production lines.
[0067] 2. Precise Capture of Deep Features: This invention innovatively embeds two levels of multi-head self-attention layers (8-head / 32-dimensional and 16-head / 64-dimensional) into the MLP. The attention mechanism enables the model to dynamically focus on the feature weights most critical to a specific output (such as UV / O3 time) among the eight parameters, breaking the limitation of fixed weights in traditional MLPs and improving prediction accuracy.
[0068] 3. Effective overfit prevention: By combining L2 regularization (penalty coefficient 0.1), layer normalization, and dropout mechanism, the model maintains strong generalization ability even on small sample photovoltaic experimental datasets.
[0069] 4. Significantly reduced R&D costs: Researchers only need to provide conventional solar simulator test data (curve extraction parameters) to directly predict the underlying process optimization direction (UV / O3 time) and passivation quality, which greatly shortens the process iteration cycle of high-efficiency batteries.
[0070] 5. In the embodiments of this application, an intrinsic hydrogenated amorphous silicon / silicon oxide / molybdenum oxide multilayer passivation contact structure is prepared on the surface of a solar cell sample with one side being n-type crystalline silicon. The intrinsic hydrogenated amorphous silicon layer effectively passivates the dangling bonds on the crystalline silicon surface, while the silicon oxide layer suppresses hydrogen atom leakage from the intrinsic hydrogenated amorphous silicon, maintaining a high hydrogen content and thus improving the chemical passivation effect of the passivation contact in the crystalline silicon solar cell. Simultaneously, the silicon oxide layer suppresses oxygen diffusion in the molybdenum oxide, maintaining a high oxygen content and high work function, which further enhances the passivation effect of the passivation contact field in the crystalline silicon solar cell. Therefore, the fill factor and open-circuit voltage of the crystalline silicon solar cell can be improved, thereby increasing the photoelectric conversion efficiency of the crystalline silicon solar cell.
[0071] 6. In this application example, silicon oxide is prepared by the low-temperature process of the UV / ozone cleaning machine, which is much lower than the traditional thermal oxidation process, avoiding thermal damage to the hydrogenated amorphous silicon film. It is particularly suitable for high-efficiency battery structures that are sensitive to temperature, and the process is simple and easy to integrate into existing production lines.
[0072] 7. The working principles of the RF-PECVD, UV / ozone cleaning machine, and thermal evaporation equipment involved in this invention are well known to those skilled in the art and will not be described in detail here.
Claims
1. A compound silicon heterojunction solar cell, characterized in that, The upper surface of the n-type single crystal silicon substrate, from top to bottom, consists of: an intrinsic hydrogenated amorphous silicon layer, an n+ nanocrystalline silicon oxygen-doped layer, a TCO transparent conductive oxide layer, and silver gate lines. The lower surface of the n-type single crystal silicon substrate, from bottom to top, consists of: an intrinsic hydrogenated amorphous silicon layer, a silicon oxide layer, a molybdenum oxide layer, and a silver electrode layer.
2. The method for preparing a compound silicon heterojunction solar cell according to claim 1, characterized in that, This is a machine learning-based method for predicting and fabricating parameters of compound silicon heterojunction solar cells, including the following steps: (1) Obtain test data of compound silicon heterojunction solar cells, separate the input feature set containing key JV parameters of the cell, and the output feature set containing preparation and other parameters such as UV / O3 treatment time, and perform data preprocessing by standard deviation standardization; (2) Construct a variational autoencoder, which maps the preprocessed input feature set data to the latent space through the encoder, and reconstructs the data through the decoder. Self-supervised learning is performed using reconstruction loss and KL divergence loss to extract the latent representation of the data. (3) Calculate the gradient of the input features based on the loss function of the model used in step (2), add adversarial perturbation to the original input feature set to generate adversarial samples, and merge the adversarial samples with the original samples to form an enhanced training set; (4) Construct a multi-layer perceptron model that integrates multi-head self-attention mechanism and L2 regularization, and use the enhanced training set described in step (3) to perform adversarial iterative training on the model; (5) The new data is standardized and then input into the trained multilayer perceptron model to obtain the predicted value. The mapping relationship of the output features is used to perform inverse standardization and output the optimal UV / O3 processing time and other parameters. (6) Provide a sample of crystalline silicon solar cell substrate, clean it with hydrofluoric acid solution to remove the natural oxide layer on the surface, rinse it with deionized water and blow it dry; The cleaned crystalline silicon solar cell substrate sample was placed in the reaction chamber of the UV / O3 device. The optimal UV / O3 treatment time predicted in step (5) was used as the preparation time parameter of the device. The switch was turned on to perform in-situ ultraviolet irradiation and ozone oxidation treatment on the surface of the substrate sample to grow an ultrathin silicon oxide passivation layer. The solar cell sample with a silicon oxide passivation layer is transferred to the deposition equipment, and the subsequent heterojunction carrier transport layer functional film and metal electrode are deposited in sequence to complete the fabrication and packaging of the compound silicon heterojunction solar cell.
3. The preparation method according to claim 2, characterized in that, The corresponding steps include: S1, Data Collection and Preprocessing: Acquire test data of compound silicon heterojunction solar cells, separate the input feature set and output feature set, and standardize the input feature set and output feature set using the standardized models scaler_x (for standardizing input features) and scaler_y (for standardizing output features); S2, Variational Autoencoder (VAE) Feature Learning: The variational autoencoder maps the input data to the latent space through the encoder and reconstructs the data through the decoder. It uses a combination of reconstruction loss and KL divergence loss as the loss function for self-supervised learning to extract the latent representation of the data and expand the data sample. S3, Adversarial Sample Generation: To improve the robustness of the model, the system adds adversarial perturbations to the original input features based on the gradient of the input features relative to the loss function in step S2, thereby generating adversarial samples. These adversarial samples are then merged with the original samples to form a larger and more informative augmented training set for subsequent deep learning training. S4, Model Construction and Adversarial Training: Construct a multilayer perceptron (MLP) model that integrates multi-head self-attention mechanism, and use the enhanced training set described in step S3 to iteratively train the multilayer perceptron (MLP) model that integrates multi-head self-attention mechanism. S5, Parameter Prediction and Denormalization: After training, a multilayer perceptron (MLP) with a multi-head self-attention mechanism is used to predict new data: First, the new data is normalized, then the normalized input feature values are predicted, and finally, the predicted values output by the model are denormalized using the normalized model scaler_y of the output feature values, thereby outputting the optimal UV / O3 (UV / O3) treatment time and the corresponding core electrical performance parameters. S6, Composite passivation layer experimental preparation: A crystalline silicon solar cell sample is provided, with one side being an n-type crystalline silicon surface. First, the crystalline silicon solar cell sample is placed in a plasma-enhanced chemical vapor deposition (RF-PECVD) reaction chamber to deposit an intrinsic hydrogenated amorphous silicon layer on the n-type crystalline silicon surface. Based on the optimal UV / O3 treatment time predicted in step S5, an ultrathin silicon oxide layer is grown on the surface of the intrinsic hydrogenated amorphous silicon layer on the back side using a UV / O3 device. A molybdenum oxide hole transport layer is then prepared on the silicon oxide surface using a thermal evaporation device to create a multilayer passivated contact structure.
4. The preparation method according to claim 2 or 3, characterized in that, In steps (1) and S1: The input feature set includes eight key JV parameters of solar cells: implicit open-circuit voltage iV. OC , pseudo-fill factor pFF, open-circuit voltage V OC Fill factor FF, short-circuit current J SC Maximum power point current I max Maximum power point voltage V max Maximum power P max ; The output feature set includes four parameters: preparation parameters and performance parameters: UV / O3 treatment time, effective carrier lifetime τ. eff Surface recombination current density J0, series resistance R s .
5. The preparation method according to claim 2 or 3, characterized in that, In steps (2) and S2, the specific structure of the variational autoencoder is as follows: Encoder: Contains two fully connected layers to output the mean and log-variance of the latent space, and generates latent variables through a sampling function; Decoder: Contains two fully connected layers that receive the latent variables and map them back to the original data space to output reconstructed data; The variational autoencoder was compiled and optimized using the Adam optimizer.
6. The preparation method according to claim 2 or 3, characterized in that, In steps (4) and S4, the MLP model that integrates the multi-head self-attention mechanism includes: Hidden layers: include multiple fully connected layers with ReLU activation function and L2 regularization; Self-attention layer: It contains two multi-head self-attention layers. The first self-attention layer has 8 attention heads and a key dimension (key_dim) of 32, and the second self-attention layer has 16 attention heads and a key dimension of 64. Regularization layers: Layer normalization and dropout layers are connected after each self-attention layer; Output layer: A fully connected layer with a linear activation function.
7. The preparation method according to claim 2 or 3, characterized in that, In steps (5) and S5: the output feature values before denormalization, i.e., the code language of the output feature values (code in the model, computer language), and the output feature values after denormalization, i.e., the physical language of the output feature values (UV / O3 processing time, τ). eff J0, R s ).
8. The preparation method according to claim 2 or 3, characterized in that, In step S6, the specific process of the test preparation includes: the silicon oxide layer is prepared using an ultraviolet / ozone cleaning machine, and the processing time is 1 to 10 seconds.
9. The preparation method according to claim 2 or 3, characterized in that, In step S6, the crystalline silicon solar cell is a silicon-based heterojunction solar cell, wherein the silicon-based material is n-type crystalline silicon, and a multilayer passivated contact structure is prepared on its surface. The surface of the n-type crystalline silicon is a pyramidal textured surface, and the overall thickness of the sample silicon wafer is 70~200 micrometers.
10. The intrinsic hydrogenated amorphous silicon layer was prepared using RF-PECVD and had a thickness of 4~12 nm; The molybdenum oxide hole transport layer is prepared using a thermal evaporation deposition apparatus with a chamber pressure below 10. -4 Pa, with a molybdenum oxide thickness of 10 nanometers.