A method and system for predicting the photocurrent of a perovskite / crystalline silicon tandem cell

Abstract

The application discloses a perovskite / crystalline silicon two-end laminated cell photocurrent prediction method and system, relates to the battery photocurrent calculation technical field, and aims to solve the problem that the perovskite / crystalline silicon two-end laminated cell is inefficient when simulated in a three-dimensional optical model; the method comprises the following steps: taking the perovskite material, band gap and wavelength of the laminated cell as input, and taking the complex refractive index as output to construct a first proxy model; obtaining a first result according to a three-dimensional wave optical model, inputting the first result into a database as sample data; preprocessing the sample data, combining the three-dimensional wave optical model and the artificial neural network trained by combining Bayesian optimization and Bayesian normalization to generate a second proxy model; inputting the full wavelength into the second proxy model to output a predicted wavelength, and then calculating the total photocurrent of the laminated cell; the application establishes an efficient multiple proxy model in a deep learning manner, and can be applied in multiple scenes without a high-configuration computer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery photocurrent calculation, in particular to a perovskite / crystalline silicon two-end stacked battery photocurrent prediction method and system. BACKGROUND

[0002] The perovskite / crystalline silicon two-end stacked battery integrates a perovskite thin film through a crystalline silicon substrate to form a series structure, and uses current matching and voltage superposition to improve efficiency. Currently, the perovskite / crystalline silicon two-end stacked battery has the problem of slow efficiency in three-dimensional optical model simulation. Specifically, it relates to: 1. The problem that the complex refractive index of the perovskite thin film cannot be accurately and quickly changed with its band gap; 2. The problem that full wavelength scanning is required when calculating the total photocurrent.

[0003] The perovskite / crystalline silicon two-end stacked battery depends on the current matching of the sub-battery, which needs to be realized by adjusting the material, band gap and thickness of the perovskite layer. Among them, the band gap adjustment can be completed by changing the chemical ratio of the perovskite, but at the same time it will change the complex refractive index (including real part and imaginary part), and then affect the light absorption rate. Since there is no direct mathematical relationship between the band gap, material and complex refractive index, and the complex refractive index changes nonlinearly with wavelength, a full wavelength complex refractive index database needs to be established for each band gap in the optimization simulation, resulting in low calculation efficiency.

[0004] The photocurrent of a solar cell is obtained by integrating the light absorption of each wavelength, so full wavelength scanning is required during simulation. For example, the light absorption range of a perovskite / crystalline silicon stacked battery is usually 300-1200nm, if calculated at a step of 1nm, the three-dimensional wave optical model needs to run 901 times, and the calculation cost is extremely high.

[0005] For example, the Chinese patent with publication number CN107507928A only discloses a method for regulating the total current of the battery, and cannot solve the above-mentioned technical problems. SUMMARY

[0006] The present application solves the problem of slow efficiency of perovskite / crystalline silicon two-end stacked battery in three-dimensional optical model simulation, and proposes a perovskite / crystalline silicon two-end stacked battery photocurrent prediction method and system. Through the deep learning method, an efficient multiple proxy model is established, which can be applied in multiple scenarios without high configuration computer.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme: a perovskite / crystalline silicon two-end stacked battery photocurrent prediction method, comprising the following steps:

[0008] S1, taking the perovskite material of the stacked battery, the band gap and the wavelength as the input, and the complex refractive index as the output, to construct a first agent model; S2, obtaining a first result according to a three-dimensional wave optical model, and inputting the first result into a database as sample data; the first result includes the input and output of the three-dimensional wave optical model;

[0009] S3, pre-processing the sample data, and combining the three-dimensional wave optical model and the artificial neural network trained by the Bayesian optimization and the Bayesian normalization to generate a second agent model;

[0010] S4, inputting the full wavelength into the second agent model to output a predicted wavelength, and then calculating the total photocurrent of the stacked battery.

[0011] In the technical solution, the first agent model is constructed by using the perovskite material, the band gap and the complex refractive index relationship, then the first agent model is brought into the three-dimensional wave optical model for sample training and verification, the second agent model is finally obtained by combining the artificial neural network trained by the Bayesian optimization and the Bayesian normalization, and the total photocurrent of the stacked battery is calculated according to the structure of the second agent model.

[0012] The application further provides that the input of the three-dimensional wave optical model is the perovskite material band gap, wavelength, film thickness and roughness parameters; and the output of the three-dimensional wave optical model is the surface reflectivity under single wavelength, the quantum efficiency of the top perovskite and the bottom crystalline silicon sub-cell.

[0013] In the technical solution, the three-dimensional wave optical model can provide a large amount of virtual data.

[0014] The application further provides that the construction process of the first agent model is that the perovskite material, the band gap, the wavelength and the complex refractive index are taken as the first sample data, the Gaussian regression model with the Bayesian optimization is used for training, then the first sample data is tested, and finally the first agent model with the material, the band gap and the wavelength as the input and the complex refractive index as the output is generated.

[0015] In the technical solution, the first agent model is constructed, the complex refractive index is output, and finally used in the training and optimization process of the subsequent second agent model.

[0016] The application further provides that before the step S3 generates the second agent model, the test data set reserved in the sample data pre-processing process is used to detect the model performance and whether there is an overfitting problem; if the model cannot accurately reflect the relationship between the input and the output in the test data, the Bayesian optimization needs to be performed again; otherwise, if the model can accurately predict the test data, the verified second agent model is obtained.

[0017] The second agent model is further verified through the test process.

[0018] The first agent model is trained and optimized based on a Gaussian regression model with Bayesian optimization, and the regression model parameters and hyperparameters of the perovskite material are recorded after optimization.

[0019] In step S3, the parameters to be optimized by Bayesian optimization include the number of artificial neural network hidden layers, the hidden layer activation function, the hidden layer size, the initial learning rate, the minimum gradient descent, and the momentum of gradient descent.

[0020] The total photocurrent of the stacked battery is calculated as follows:

[0021] The photocurrents of the perovskite and crystalline silicon cells are calculated according to the vector integral corresponding to the predicted wavelength, and the photocurrent of the stacked battery is the minimum value of the photocurrents of the perovskite and crystalline silicon cells.

[0022] In the technical solution, the corresponding photocurrent is calculated according to the second agent model.

[0023] The perovskite layer thickness and the height of the textured surface in the sample data are calibrated after the first result is input into the database, and the perovskite thickness and the height of the textured surface are fine-tuned by minimizing the error between the simulated quantum efficiency and the actual quantum efficiency.

[0024] In the technical solution, the input perovskite layer thickness and the height of the textured surface are obtained by a scanning electron microscope, and the actual film thickness at different positions under the scanning electron microscope is not uniform, which causes errors, so calibration is used to further ensure the accuracy of the data.

[0025] The pre-processing of the sample data includes dividing the sample data into training, validation, and test data sets for training and optimization of the second agent model.

[0026] A perovskite / crystalline silicon two-end stacked battery photocurrent prediction system is suitable for the perovskite / crystalline silicon two-end stacked battery photocurrent prediction method, and includes:

[0027] The first agent model construction module constructs the first agent model with the perovskite material, band gap, and wavelength as input and the complex refractive index as output.

[0028] The first result generation module outputs the first result in combination with the three-dimensional wave optical model.

[0029] a second agent model construction module for training a second agent model according to the first result and the three-dimensional wave optical model;

[0030] a total photocurrent calculation module for inputting the full wavelength into the second agent model to output a predicted wavelength, and integrating to obtain the total photocurrent of the stacked cell.

[0031] In the technical solution, through the cooperative operation of the first agent model construction module, the first result generation module, the second agent model construction module and the total photocurrent calculation module, the calculation and prediction of the photocurrent of the perovskite / crystalline silicon two-end stacked cell can be realized, and the efficient operation of the perovskite / crystalline silicon two-end stacked cell in the three-dimensional optical model simulation can be ensured.

[0032] The present application can bring the following beneficial effects:

[0033] The present application relates to a photocurrent prediction method of a perovskite / crystalline silicon two-end stacked cell, which constructs a first agent model and a second agent model by means of deep learning, ensures the efficient operation of the perovskite / crystalline silicon two-end stacked cell in the three-dimensional optical model simulation, and can be applied in multiple scenes without a high-configuration computer. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a flowchart of the present application for constructing a first agent model of a photocurrent prediction method of a perovskite / crystalline silicon two-end stacked cell.

[0035] Figure 2 is a flowchart of the present application for generating a first result of a photocurrent prediction method of a perovskite / crystalline silicon two-end stacked cell.

[0036] Figure 3 is a flowchart of the present application for constructing a second agent model and calculating a total photocurrent of a photocurrent prediction method of a perovskite / crystalline silicon two-end stacked cell. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific embodiments described herein are only one of the best embodiments of the present application, which are used to explain the present application and do not limit the protection scope of the present application. All other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0038] The perovskite / crystalline silicon two-end stacked cell integrates a perovskite thin film through a crystalline silicon substrate to form a series structure, and uses current matching and voltage superposition to improve efficiency. In the prior art, the perovskite / crystalline silicon two-end stacked cell has the problem of slow efficiency in three-dimensional optical model simulation. The specific problems mainly include the following three points: 1. The complex refractive index of the perovskite thin film cannot be accurately and quickly changed with the band gap; 2. Full wavelength scanning is required when calculating the total photocurrent; 3. Relying on a high-configuration computer is not conducive to convenient use and digital twin application.

[0039] For the problem that the complex refractive index of the perovskite thin film cannot be accurately and quickly changed with the band gap, specifically: the perovskite / crystalline silicon two-end stacked cell depends on the current matching of the sub-cell, which needs to be realized by adjusting the material, band gap and thickness of the perovskite layer. Among them, the band gap adjustment can be completed by changing the chemical ratio of the perovskite, but at the same time it will change the complex refractive index (including real part and imaginary part), and then affect the light absorption rate. Since there is no direct mathematical relationship between the band gap, material and complex refractive index, and the complex refractive index changes nonlinearly with the wavelength, a full-wavelength complex refractive index database needs to be established for each band gap in the optimization simulation, resulting in low calculation efficiency.

[0040] For the problem that full wavelength scanning is required when calculating the total photocurrent, specifically, the photocurrent of a solar cell is obtained by integrating the light absorption of each wavelength, so full wavelength scanning is required during simulation. For example, the light absorption range of the perovskite / crystalline silicon stacked cell is usually 300-1200nm, if calculated with 1nm step, the three-dimensional wave optical model needs to run 901 times, the calculation cost is extremely high.

[0041] For the problem that relying on a high-configuration computer is not conducive to convenient use and digital twin application, specifically: to accurately simulate the co-tuft morphology of the perovskite / crystalline silicon stacked cell (micron-level pyramid structure on the silicon surface), a three-dimensional wave optical model and a finite element method need to be used. Since the silicon thickness is hundreds of microns, and the wave optics requires a grid size less than 1 / 5 wavelength (60-240nm), resulting in a large number of grids, a high-performance server must be used for calculation.

[0042] Embodiment 1

[0043] In view of the deficiencies in the prior art, the embodiment proposes a photocurrent prediction method for a perovskite / crystalline silicon two-end stacked cell, which refers to Figure 1 , Figure 2 and Figure 3 , which includes the following steps.

[0044] Step S1, a first proxy model is constructed, the first proxy model takes the perovskite material, band gap and wavelength of the perovskite / crystalline silicon two-end stacked cell as input, and the complex refractive index as output.

[0045] Reference Figure 1 The first proxy model is trained based on the Gaussian regression model with Bayesian optimization, and the first proxy model is finally generated by testing the first sample data. The input of the first proxy model is the material, the band gap and the wavelength, and the output is the complex refractive index.

[0046] In the technical solution, the first proxy model is constructed to output the complex refractive index, which is finally used in the training and optimization process of the subsequent second proxy model.

[0047] More specifically, the first proxy model is trained and optimized based on the Gaussian regression model with Bayesian optimization. After optimization, the regression model parameters and corresponding hyperparameters of the perovskite material are recorded.

[0048] The step S1 mainly includes a data acquisition process and a machine learning process.

[0049] For the data acquisition process, in the preparation stage, different band gap perovskite thin films are formed by changing the chemical ratio in the perovskite, and then the thin film band gap is confirmed and the material complex refractive index under full wavelength is measured by using a spectroscopic ellipsometer. Finally, an initial database composed of perovskite materials, band gaps, wavelengths and complex refractive indices is formed. Under a single material, the big data of different band gaps, wavelengths and complex refractive indices will be divided into training, validation and test data sets. In this embodiment, 100,000 samples (one band gap + one wavelength + one complex refractive index as one sample) are collected for the training of the proxy model.

[0050] For the machine learning process, it mainly includes sample data preprocessing, construction of Gaussian regression model with Bayesian optimization training process and model testing process.

[0051] For the sample data preprocessing process, more specifically, in this embodiment, the range of the band gap is usually 1 to 2eV, and the wide band gap perovskite is 1.6 to 1.8eV; the range of the wavelength is 300 to 1200nm; the range of the real part of the complex refractive index is generally 1.5 to 3, and the range of the imaginary part is generally 0 to 1.5. The above parameter ranges have no large span and do not have a difference of more than 10 times between the upper and lower limits, so normalization processing is not required. In order to train an accurate model more quickly and efficiently, the experimental data with the smallest band gap are used as the benchmark. This set of data will include the smallest band gap, full wavelength and corresponding complex refractive index, which will help the machine learning to quickly establish a data framework.

[0052] For the construction of the Gaussian regression model training process with Bayesian optimization. The selected hyperparameters are the basis function, the coefficients of the basis function, the noise variance, the kernel function and its corresponding hyperparameters. After the optimization program is completed, the regression model of the perovskite material and its final hyperparameters will be recorded, preparing for the future material database. For example: MAPbI material, FAPbI material and CsFAPbIBr material each train a model, match the material and model hyperparameters, and then add an artificial classification algorithm on top of the proxy regression model.

[0053] For the model testing process, specifically, use the test data set reserved during the data collection process to detect the model performance and whether there is an overfitting problem. If the model fails to accurately reflect the relationship between the input and output in the test data (the prediction result and the true response error are too large), more high-quality data needs to be retrained or a more complex algorithm needs to be used; in the present embodiment, the current regression model is accurate enough. Conversely, if the model can accurately predict the test data, we will get the verified proxy model.

[0054] Through the above steps, the first proxy model can be finally obtained, which can accurately output the thin film complex refractive index of the perovskite material at each wavelength by inputting the band gap and the considered wavelength.

[0055] After completing the process of step S1, refer to Figure 2 , perform step S2, combine the three-dimensional wave optical model to obtain the first result, and input the first result into the database as sample data for the subsequent second proxy model; wherein the first result includes the input and output of the three-dimensional wave optical model.

[0056] The input of the three-dimensional wave optical model is the band gap of the perovskite material, the wavelength, the film thickness and the roughness parameter; the output of the three-dimensional wave optical model is the surface reflectivity at a single wavelength, the quantum efficiency of the top perovskite and the bottom crystalline silicon subcell.

[0057] In the present technical solution, the three-dimensional wave optical model can provide a large amount of virtual data.

[0058] The three-dimensional wave optical model uses wave optics and a finite element solver to simulate the light contact perovskite top cell under steady state in the unit structure of the stacked battery, to propagate on the uneven rough surface, and finally to calculate the quantum efficiency and photocurrent by the light field distribution after the crystalline silicon bottom cell, combined with the full wavelength incident spectrum, to assist in completing the current matching and structure design of the perovskite / crystalline silicon two-end stacked battery. Through the first proxy model and the three-dimensional wave optical model of the perovskite / crystalline silicon two-end stacked battery, a multi-proxy model based on machine learning can be finally trained, which is close to the actual perovskite material and accurately reflects the light field structure of the three-dimensional wave optical model.

[0059] After step S2, step S3 is performed, where the sample data obtained in step S2 is preprocessed, and then trained and optimized using a three-dimensional wave optical model and an artificial neural network with Bayesian optimization and Bayesian positive lateralization, finally generating the second surrogate model.

[0060] Among them, the parameters optimized by Bayesian optimization include the number of hidden layers in the artificial neural network; the activation function of the hidden layer; the size of the hidden layer; the initial learning rate; the minimum gradient descent; and the momentum of gradient descent.

[0061] Before generating the second surrogate model, test data is used to check the model performance and whether there is an overfitting problem. If the model cannot accurately reflect the relationship between the input and output in the test data (the error between the prediction result and the actual response is too large), it needs to be optimized again. If the model can accurately predict the test data, the model does not need to be optimized again, and thus the second surrogate model is obtained.

[0062] In addition, after inputting the first result into the database, the perovskite layer thickness and texture height in the sample data are calibrated. The calibration process is to fine-tune the perovskite thickness and texture height by minimizing the error between the quantum efficiency obtained from the simulation and the actual quantum efficiency.

[0063] In this technical solution, the input perovskite layer thickness and texture height are obtained by scanning electron microscopy. Under scanning electron microscopy, the actual film thickness is uneven at different locations, which may cause errors. Therefore, the above calibration is used to further ensure the accuracy of the data.

[0064] The construction of the second agent model mainly includes data collection and machine learning processes.

[0065] For the data acquisition process, refer to Figure 3 Specifically, it provides a large amount of virtual data through a three-dimensional wave optical model. The inputs include wavelength, perovskite material and its band gap, film thickness and textured surface geometry parameters; the outputs are the surface reflectivity at a single wavelength and the quantum efficiency of the top perovskite and bottom crystalline silicon sub-cells.

[0066] It should be noted that the data collection process for constructing the second agent model incorporated elements such as... Figure 1 The first surrogate model shown rapidly outputs the complex refractive index of the material through wavelength and band gap.

[0067] In addition, the data collection process includes a virtual data supplement process, which specifically requires calibration with experimental data (measured by a quantum efficiency tester). The parameters that need to be calibrated are the perovskite layer thickness and the height of the textured (pyramid) surface. The reason for choosing these two parameters is that the actual film thickness at different positions under a scanning electron microscope (SEM) is not uniform, which will cause some errors. The calibration method is to minimize the error between the simulated quantum efficiency and the actual quantum efficiency, and to fine-tune the perovskite thickness and the textured height.

[0068] In this embodiment, the number of calibrations is determined by the actual training data. In the case of using a large training data set of one hundred thousand at a wavelength of 300-1200 nm, at least 10 textured morphologies should be selected, and at most one calibration should be performed per 100 nm of perovskite thickness and per 100 nm of wavelength, resulting in about one thousand experimental calibrations (since the quantum efficiency tester can test all wavelengths at once, the actual number of experiments is about 100 times). After completing the calibration, the virtual data filling is completed by simulating the interpolation.

[0069] For the second agent model construction machine learning, it includes sample data preprocessing, establishing a training model of artificial neural network with Bayesian optimization and Bayesian regularization, and model testing process.

[0070] In the step of constructing the second agent model, the preprocessing process of the sample data specifically includes: dividing the collected big data into training, validation and testing data sets for the training and optimization of the agent model based on artificial neural network. Here, due to the consideration of multiple wavelengths (300-1200 nm, 1 nm precision), a large training data set of one hundred thousand and validation and testing data sets of ten thousand each are used. In terms of data preprocessing, the specific reference range of the considered parameters is: wavelength 300-1200 nm, perovskite thickness 400-1500 nm, perovskite band gap maximum range (depending on the material) 1-2 eV, pyramid base width 2-5 um, pyramid height 1-3 um, pyramid top to bottom width ratio 0.5-0.9, the upper and lower limits of which do not exceed 10 times, so there is no need to take the logarithm value or normalize. Furthermore, when the units of wavelength and perovskite thickness are converted to um, the values of all parameters are within the range of 0-5, without the need for additional data preprocessing. The output parameters are the total reflectivity of the surface, the quantum efficiency of the perovskite layer, and the quantum efficiency of the crystalline silicon layer, all of which range from 0 to 1.

[0071] In the step of constructing the second proxy model, the establishing of the artificial neural network training model with Bayesian optimization and Bayesian regularization specifically comprises: establishing an artificial neural network training program with Bayesian optimization and Bayesian regularization. The parameters to be optimized by the Bayesian optimization include the number of hidden layers of the artificial neural network (ranging from 3 to 5 layers, without considering 1-2 layers because they are not enough to complete a highly nonlinear regression problem, and without considering more than 5 layers because the training time is too long); the hidden layer activation function, including hyperbolic tangent sigmoid, radial basis, double curved function (log-sigmoid) and linear function; the size of the hidden layer (ranging from 2 to 100 for each layer); the maximum number of iterations is constant at 1000 (this parameter is not optimized); the initial learning rate (ranging from 0.01 to 1); the minimum gradient descent (one of the termination conditions for a single neural network training, ranging from to ) and the momentum of the gradient descent (ranging from 0.5 to 0.98). In addition to the minimum gradient descent, the termination conditions for a single neural network training also include early stopping (a kind of overfitting avoidance method when the training data error decreases and the validation data error increases), and the termination conditions for the overall Bayesian optimization are set as the maximum number of iterations limit (1000 times), the maximum time limit (usually 1 week is enough) and the minimum gradient descent (usually set to ). The artificial neural network adopts Bayesian regularization with back-propagation, which can effectively prevent overfitting and learn the covariance of multiple output parameters.

[0072] In the step of constructing the second proxy model, the model testing process comprises: using the test data set to detect the model performance and whether there is an overfitting problem. If the model fails to accurately reflect the relationship between the input and the output in the test data (the prediction result has a large error compared with the true response), the Bayesian optimization program needs to be performed again. Otherwise, if the model can accurately predict the test data, the verified second proxy model will be obtained.

[0073] In step S4, the full wavelength is input into the second proxy model to output the predicted wavelength, and then the total photocurrent of the stacked cell is calculated. Specifically, the photocurrents of the perovskite and crystalline silicon cells are calculated according to the vector integral corresponding to the predicted wavelength, and the photocurrent of the stacked cell is the minimum value of the photocurrents of the perovskite and crystalline silicon cells.

[0074] In the technical solution, a multi-agent model based on a neural network is finally obtained (a second agent model trained by a three-dimensional wave optical model is added to the first agent model), the input of which is six parameters (wavelength, perovskite film thickness, perovskite band gap, and three geometric parameters of the textured surface), and the output is the reflectivity of the incident surface, the quantum efficiency of the perovskite layer, and the quantum efficiency of the crystalline silicon layer. In addition, this model can perform vector calculation for the full wavelength, effectively solving the problem of the need for full wavelength scanning in the previous model. That is, the wavelength is a vector of 300-1200 nm, and the remaining inputs are single values, and the output is also a vector corresponding to the wavelength of 300-1200 nm. Then, the photocurrent of the perovskite and crystalline silicon cells is calculated by integration, and the photocurrent of the stacked cell is the minimum value of the two.

[0075] Based on the above-mentioned perovskite / crystalline silicon two-end stacked cell photocurrent prediction method, the embodiment further proposes a perovskite / crystalline silicon two-end stacked cell photocurrent prediction system, which mainly includes a first agent model construction module, a first result generation module, a second agent model construction module, and a total photocurrent calculation module. The first agent model construction module is connected with the first result generation module, the first result generation module is connected with the second agent model construction module, and the second agent model construction module is connected with the total photocurrent calculation module.

[0076] For the first agent model construction module, the first agent model can be constructed by taking the perovskite material, band gap, and wavelength of the stacked cell as the input and the complex refractive index as the output. The specific construction process can refer to Figure 1 and the process of step S1 in the above prediction method.

[0077] For the first result generation module, the first result can be output according to the three-dimensional wave optical model. The specific generation process of the first result can refer to the process of step S2 in the above prediction method.

[0078] For the second agent model construction module, the second agent model can be finally obtained by training and optimizing the first result and the three-dimensional wave optical model.

[0079] For the total photocurrent calculation module, the full wavelength can be input into the second model after training and optimization, and the predicted wavelength can be finally output. The photocurrent of the perovskite and crystalline silicon cells can be obtained by integration operation, and the photocurrent of the stacked cell is the minimum value of the two.

[0080] In the technical solution, through the cooperative operation of the first proxy model construction module, the first result generation module, the second proxy model construction module and the total photocurrent calculation module, the calculation and prediction of the photocurrent of the perovskite / crystalline silicon two-end stacked battery can be realized, and the efficient operation of the perovskite / crystalline silicon two-end stacked battery in three-dimensional optical model simulation is ensured.

[0081] 1. The embodiment solves the problem that the complex refractive index of the perovskite thin film cannot change accurately and quickly with the band gap: (1) Since the first proxy model based on Gaussian regression is trained, the complex refractive index of the thin film and the material band gap are mapped, that is, the complex refractive index of the full wavelength can be output by inputting the perovskite material, the band gap and the wavelength; (2) The time consumption of the embodiment is instant, and when the photoelectric simulation is performed, the total calculation time required will be greatly reduced, facilitating data collection and analysis; (3) After confirming the perovskite material, the first proxy model can complete the mapping of the band gap parameter to the full wavelength (300-1200nm) complex refractive index of 901*2=1802 parameters, greatly reducing the memory and disk usage during calculation.

[0082] 2. The embodiment solves the problem of full wavelength scanning when calculating the total photocurrent: (1) The second proxy model obtained finally can perform vectorization calculation of the full wavelength, solving the problem of full wavelength scanning, and simplifying 901 simulations to 1 vectorization calculation simulation; (2) Because the amount of training data collected is large and the parameter range is wide, the proxy model of the embodiment can cope with different topographies of the textured surface (from sharp triangle to wide pyramid), different perovskite band gaps and thicknesses, and is beneficial to global optimization.

[0083] 3. The embodiment solves the problem that relying on a high-configuration computer is not conducive to convenient use and digital twin application: (1) The multiple proxy models obtained finally only use six input parameters (wavelength, band gap, film thickness and three geometric parameters determining the textured surface structure) and three output parameters (reflectivity and quantum efficiency of two sub-cells), which are simple and convenient. On this basis, only six inputs can cope with the adjustment of perovskite material ratio, perovskite thin film preparation process and textured surface treatment process of crystalline silicon, instant calculation, and complementing actual laboratory preparation, which can be applied to digital twin; (2) Compared with the traditional simulation method, there is no need for equation and physical field construction and grid division steps, so there is no hardware requirement for computer memory, disk and CPU, and it does not rely on a high-configuration computer; (3) Because it is a proxy model based on machine learning, it is convenient for retraining of new data, and there is room for progress with the times.

[0084] Finally, it should be noted that the above examples are merely used to illustrate the technical solutions of the present application but not to limit. Although the present application is explained in detail with reference to the examples, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting the photocurrent of a perovskite / crystalline silicon tandem cell, characterized in that, The method comprises the following steps: S1, constructing a first agent model with perovskite material, band gap and wavelength as input and complex refractive index as output; specifically, training a Gaussian regression model with Bayesian optimization, testing according to the first sample data, and generating a first agent model with material, band gap and wavelength as input and complex refractive index as output; S2, obtaining a first result according to a three-dimensional wave optical model, and inputting the first result into a database as sample data; the first result comprises input and output of the three-dimensional wave optical model, and the input comprises perovskite material band gap, wavelength, film thickness and surface roughness parameters; The output comprises surface reflectivity under single wavelength, quantum efficiency of top perovskite and bottom crystalline silicon sub-cell; S3, preprocessing the sample data, and training a second agent model by combining the three-dimensional wave optical model and artificial neural network with Bayesian optimization and Bayesian normalization; The construction of the second agent model comprises data collection and machine learning; the data collection provides a large amount of virtual data through the three-dimensional wave optical model; Machine learning comprises preprocessing of sample data, training of artificial neural network with Bayesian optimization and Bayesian normalization, and model testing process; 2. The method of claim 1, wherein the method is a method of predicting the photocurrent of a perovskite / silicon tandem cell, characterized in that, S4, inputting full wavelength into the second agent model to output predicted wavelength, and then calculating total photocurrent of the stacked cell.

3. The method of claim 1, wherein the method is a method of predicting the photocurrent of a perovskite / silicon tandem cell, characterized in that, Before the step S3 generates the second agent model, the test data set reserved in the sample data preprocessing process is used to detect the model performance and whether there is overfitting problem; if the model fails to accurately reflect the relationship between the input and output in the test data, Bayesian optimization needs to be performed again; otherwise, if the model can accurately predict the test data, the verified second agent model will be obtained.

4. The method of claim 1, wherein the method is a method of predicting the photocurrent of a perovskite / silicon tandem cell, characterized in that, The first agent model is trained and optimized based on the Gaussian regression model with Bayesian optimization, and the regression model parameters and hyperparameters of the perovskite material are recorded after optimization is completed.

5. The method of claim 1, wherein the method is a method of predicting the photocurrent of a perovskite / silicon tandem cell, characterized in that, In the step S3, the parameters to be optimized by Bayesian optimization include the number of hidden layers of artificial neural network, hidden layer activation function, hidden layer size, initial learning rate, minimum gradient descent and momentum of gradient descent. The total photocurrent of the stacked cell is calculated as follows:

6. The method of predicting photocurrent of a perovskite / silicon tandem cell according to claim 1, wherein The photocurrents of perovskite and crystalline silicon cells are calculated according to the vector integral corresponding to the predicted wavelength, and the photocurrent of the stacked cell is the minimum value of the photocurrents of perovskite and crystalline silicon cells.

7. The method of claim 1, wherein the method is a method of predicting the photocurrent of a perovskite / silicon tandem cell, characterized in that, After the first result is input into the database, the thickness of the perovskite layer and the height of the surface roughness in the sample data are calibrated; specifically, the thickness of the perovskite layer and the height of the surface roughness are fine-tuned by minimizing the error between the simulated quantum efficiency and the actual quantum efficiency.

8. A photogenerated current prediction system for a perovskite / silicon tandem cell, adapted to a method for predicting the photogenerated current of a perovskite / silicon tandem cell according to any one of claims 1 to 7, characterized in that, The preprocessing of the sample data comprises dividing the sample data set into training, verification and test data sets for training and optimization of the second agent model. The method comprises the following steps: A first agent model construction module constructs a first agent model with perovskite material, band gap and wavelength as input and complex refractive index as output; A first result generation module outputs a first result by combining a three-dimensional wave optical model; A second proxy model construction module is configured to train a second proxy model according to the first result and the three-dimensional wave optics model; A total photocurrent calculation module is configured to input the full wavelength into the second proxy model to output a predicted wavelength, and to calculate a total photocurrent of the stacked battery by integration.

Images