Solar cell performance prediction method and device based on physical information fusion model

Abstract

The application relates to a solar cell performance prediction method and device based on a physical information fusion model, and relates to the technical field of computer data processing and semiconductor characterization. The technical scheme provided by the application adopts a deep neural network architecture to perform feature extraction and fusion on multi-dimensional spectral information, the prediction accuracy is improved by more than 35%, the reliability is improved by more than 40%, and the prediction performance is significantly improved; the training of the model adopts an innovative physical information fusion loss function, which comprises a data-driven prediction loss term and a constraint loss term based on the physical law of a semiconductor device, the prediction task and the physical principle are deeply fused, the physical causal relationship between spectral characteristics and device performance can be understood by the prediction model, more accurate performance prediction and better physical interpretability are realized, the prediction result is normalized by adopting multiple physical constraints, the model prediction result conforms to the physical principle of the semiconductor device, and the physical rationality and engineering application value of the prediction result are improved.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and semiconductor characterization technology, and in particular to a method and apparatus for predicting the performance of solar cells based on a physical information fusion model. Background Technology

[0002] Solar cells, as an important development direction of next-generation photovoltaic technology, have attracted much attention due to their excellent photoelectric conversion performance and relatively low manufacturing cost. However, solar cell materials are extremely sensitive to the preparation environment and process parameters; even slight changes in conditions can lead to significant differences in device performance. This makes controlling the yield and consistency of device fabrication a key challenge in the industrialization process.

[0003] In traditional solar cell R&D and manufacturing processes, device performance evaluation can only be conducted after all fabrication steps are completed. A complete device fabrication process involves multiple complex steps, including substrate cleaning, electron transport layer deposition, active layer fabrication, hole transport layer coating, and top metal electrode evaporation. Each step requires strict control of environmental parameters such as temperature, humidity, and atmosphere composition, making the process extremely complex. This delayed performance feedback mechanism leads to serious problems: First, it makes it impossible to identify and remove defective samples in the early stages of the fabrication process, resulting in a significant waste of materials and time; second, the optimization cycle for new material formulations and process parameters is greatly prolonged, severely hindering technological innovation and industrialization progress.

[0004] Existing early prediction methods mainly rely on single optical characterization techniques, such as empirical judgments based solely on certain characteristic parameters of absorption or photoluminescence spectra. These methods have limitations: low prediction accuracy, typically large errors; a lack of physical theoretical foundation, resulting in poor reliability and interpretability of prediction results; and insufficient adaptability to complex situations involving multiple coupled factors. In recent years, machine learning methods have been widely applied in materials science, but traditional data-driven methods face challenges in predicting solar cell performance, including limited training samples, insufficient model generalization ability, and a lack of physical plausibility in the prediction results. Summary of the Invention

[0005] Based on the above-mentioned situation of the prior art, the purpose of this invention is to provide a method and apparatus for predicting the performance of solar cells based on a physical information fusion model. The aim is to use deep learning technology combined with the physical principles of semiconductor devices to achieve early and accurate prediction of solar cell performance, thereby accelerating the material research and development process and improving production efficiency.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for predicting the performance of solar cells based on a physical information fusion model is provided, comprising the following steps:

[0007] Step 1: Collect and preprocess the spectral data of the half-cell sample of the solar cell to be predicted. The spectral data includes the absorption spectrum data, upper interface photoluminescence spectrum data and lower interface photoluminescence spectrum data of the half-cell sample.

[0008] Step 2: Construct a physical information fusion prediction model, which includes a spectral feature extraction module, a physical-guided feature enhancement module, a multimodal feature fusion module, and a performance index prediction module connected in sequence;

[0009] The spectral feature extraction module extracts features from three types of spectral data through three neural network branches;

[0010] The physical-guided feature enhancement module calculates key physical feature parameters of spectral data to guide the feature extraction process.

[0011] The multimodal feature fusion module fuses the features extracted by the spectral feature extraction module and the physically guided feature enhancement module through an attention mechanism;

[0012] The performance metric prediction module maps the fused feature vectors to predicted performance metric values.

[0013] Step 3: Design the fusion loss function and train the model;

[0014] Step 4: Use the trained model to predict the performance indicators of the half-cell samples.

[0015] Furthermore, the spectral feature extraction module specifically comprises three parallel one-dimensional convolutional neural network branches, used to process absorption spectral data, upper interface photoluminescence spectral data, and lower interface photoluminescence spectral data, respectively. Each branch adopts the same network structure, including multiple one-dimensional convolutional layers. Except for the last layer, each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and downsampling is performed using max pooling. The last convolutional layer is directly connected to an adaptive pooling layer to obtain three spectral data feature vectors of the same dimension.

[0016] Furthermore, the physical guidance feature enhancement module is specifically as follows:

[0017] The Tauc analysis method is used to extract the band gap energy from the absorption spectrum; the Urbach energy is extracted from the absorption spectrum; and the peak energy, full width at half maximum (FWHM), integral intensity, and Stokes shift are extracted from the photoluminescence spectrum. These physical parameters are then converted into high-dimensional feature vectors by a physical feature encoder, which serve as physical guidance feature vectors.

[0018] Furthermore, the performance index prediction module includes multiple hidden layers, with the number of neurons decreasing layer by layer, ReLU activation function, and the number of neurons in the output layer corresponding to the performance index.

[0019] Furthermore, the fusion loss function employs an adaptive weight adjustment mechanism based on gradient balancing, as shown in the following formula:

[0020]

[0021] in The total number of loss items, For the first Each loss item, For adaptive weighting coefficients, The transformation function is optional; the loss term includes a data-driven prediction loss term and a physical constraint term; the data-driven prediction loss term calculates the difference between the predicted value and the actual experimental measurement value; the physical constraint term learns the physical relationship between spectral features and device performance parameters;

[0022] The adaptive weight coefficients are determined through one or more of the following methods: fixed weights, gradient balancing adjustment, adaptive adjustment based on validation set performance, dynamic adjustment based on physical consistency, and weighted adjustment based on uncertainty.

[0023] Furthermore, the predicted performance metrics include photoelectric conversion efficiency, open-circuit voltage, short-circuit current density, and fill factor, and the physical constraints include:

[0024] The physical constraint term for open-circuit voltage is used to ensure that the predicted open-circuit voltage value conforms to the theoretical relationship.

[0025] The physical constraint term for short-circuit current density is used to constrain the predicted short-circuit current density value from not exceeding the root theoretical upper limit.

[0026] The fill factor physical constraint term is used to constrain the predicted fill factor values ​​to conform to theoretical relationships.

[0027] The upper limit constraint term for photoelectric conversion efficiency is used to ensure that the predicted photoelectric conversion efficiency value does not exceed the theoretical efficiency upper limit.

[0028] The photoelectric conversion efficiency self-consistency constraint term is used to ensure that the predicted photoelectric conversion efficiency value is physically consistent with the predicted open-circuit voltage value, short-circuit current density value, and fill factor value.

[0029] Furthermore, the model training includes a pre-training phase, a physical constraint introduction phase, and a joint optimization phase;

[0030] During the pre-training phase, network parameters are initialized using partially labeled data through purely data-driven supervised learning.

[0031] In the physical constraint introduction stage, the weight of the physical constraint terms is gradually increased so that the model learns the physical relationship between spectral features and device performance parameters.

[0032] During the joint optimization phase, all training data is used, and the complete fusion loss function is employed. End-to-end optimization is performed using gradient balancing mechanisms.

[0033] According to another aspect of the present invention, a solar cell performance prediction device based on a physical information fusion model is provided, comprising a spectral data acquisition module, a data preprocessing module, and a performance prediction module.

[0034] The spectral data acquisition module is used to acquire the absorption spectrum data, upper interface photoluminescence spectrum data, and lower interface photoluminescence spectrum data of the half-cell sample of the solar cell to be predicted.

[0035] The data preprocessing module is used to preprocess spectral data.

[0036] The performance prediction module is used to input preprocessed spectral data into a trained physical information fusion prediction model and output performance index prediction data.

[0037] In summary, the technical solution provided by this invention employs a deep neural network architecture to extract and fuse features from multi-dimensional spectral information. Compared to traditional single-spectral analysis methods, it can more comprehensively characterize the photoelectric properties of devices, improving prediction accuracy by over 35% and reliability by over 40%, thus significantly enhancing prediction performance. The model training utilizes an innovative physical information fusion loss function, deeply integrating the prediction task with physical principles. This allows the prediction model to understand the physical causal relationship between spectral features and device performance, achieving more accurate performance predictions and better physical interpretability. Multiple physical constraints are used to standardize the prediction results, ensuring that the model's predictions conform to the physical principles of semiconductor devices, improving the physical rationality and engineering application value of the prediction results. This invention's method is not only applicable to perovskite solar cells but can also be extended to the performance prediction of various photovoltaic devices such as organic photovoltaic cells, dye-sensitized solar cells, and quantum dot solar cells. Attached Figure Description

[0038] Figure 1 This is a flowchart of a solar cell performance prediction method based on a physical information fusion model provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the half-cell sample structure provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the overall structure of the physical information fusion prediction model provided in this embodiment of the invention;

[0041] Figure 4This is a schematic diagram of the spectral feature extraction module provided in an embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram of the structure of the multimodal feature fusion module provided in an embodiment of the present invention;

[0043] Figure 6 This is a schematic diagram of the fusion loss function provided in an embodiment of the present invention;

[0044] Figure 7 This is a flowchart of the model training process provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0046] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components.

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. Embodiments of the present invention provide a method for predicting the performance of solar cells based on a physical information fusion model. Figure 1 The flowchart of the solar cell performance prediction method according to an embodiment of the present invention is shown below. Figure 1 As shown, the method includes the following steps:

[0048] Step S101: Prepare a half-cell sample and collect multi-dimensional spectral data. In this embodiment of the invention, the half-cell sample is an intermediate structure of a solar cell without a top metal electrode deposited. Figure 2 The diagram shows a schematic representation of the structure of a half-cell sample according to an embodiment of the present invention. Figure 2 As shown, the hierarchical structure of the half-cell sample from bottom to top consists of a transparent conductive substrate, an electron transport layer, an active layer, and a hole transport layer. The transparent conductive substrate is a fluorine-doped tin oxide glass substrate, the electron transport layer is made of tin dioxide, the active layer is made of a mixed cation material, and the hole transport layer is made of an organic hole transport material. Compared to a complete device, the half-cell sample omits subsequent processing steps such as metal electrode deposition, but it already contains the main functional layers that determine the device's photoelectric performance.

[0049] Before acquiring spectral data, the raw spectral data needs to be preprocessed. Preprocessing steps include baseline correction, noise filtering, and outlier removal. Baseline correction is performed using the asymmetric least squares method, and the baseline is obtained by minimizing the following objective function:

[0050]

[0051] in The index represents the discrete wavelength point number in the spectral data. The baseline signal at the wavelength point The value at this point represents the baseline spectral value obtained through fitting. The original spectral signal, These are the weighting coefficients. For smoothing parameters, It is a second-order difference operator. Noise filtering uses a Savitzky-Golay filter, and the window width is adaptively adjusted according to the spectral signal-to-noise ratio.

[0052] For each half-cell sample, three types of spectral data need to be collected as model input. Absorption spectral data were acquired using a UV-Vis-NIR spectrophotometer, and the total transmission spectrum of each half-cell sample was measured under controlled conditions. and total reflectance spectrum According to the law of conservation of energy, through calculation The absorption spectrum data of the sample were obtained. Indicates the wavelength of light. , , They represent wavelengths respectively. The absorption spectrum measures the absorptivity, transmittance, and reflectivity of the active layer. It directly reflects the active layer's ability to capture sunlight and is closely related to the battery's short-circuit current density.

[0053] The photoluminescence spectra of the upper and lower interfaces were acquired using a micro-confocal Raman spectroscopy system. In the lower interface photoluminescence spectroscopy measurement, the excitation source was incident perpendicularly from one side of the transparent conductive substrate, exciting the active layer to produce photoluminescence. The photoluminescence signals were collected from the same side to obtain the lower interface photoluminescence spectra. In the photoluminescence spectroscopy measurement of the upper interface, the excitation source was incident perpendicularly from one side of the hole transport layer, and the remaining measurement conditions were kept consistent with those of the lower interface to obtain the photoluminescence spectral data of the upper interface. .in and They represent wavelengths respectively. The photoluminescence intensities at the lower and upper interfaces are measured. The photoluminescence spectra at the upper and lower interfaces reflect the carrier transport efficiency and interfacial recombination characteristics, and are related to the open-circuit voltage and fill factor of the battery.

[0054] Through the above measurement procedure, three sets of one-dimensional spectral data were obtained for each half-cell sample: absorption spectrum. , lower interface photoluminescence spectrum and the photoluminescence spectrum of the upper interface These three types of spectral data can comprehensively reflect the optical absorption characteristics, carrier transport characteristics, and interface recombination characteristics of the device, providing a rich information foundation for subsequent performance prediction.

[0055] Step S102: Construct a physical information fusion prediction model. Figure 3 The diagram shows the overall structure of the physical information fusion prediction model according to an embodiment of the present invention, as shown below. Figure 3 As shown, the physical information fusion prediction model includes a spectral feature extraction module (1D-CNN), a physical-guided feature enhancement module, a multimodal feature fusion module (Attention), and a performance index prediction module (MLP) connected in sequence.

[0056] Figure 4 The diagram shows a schematic representation of the spectral feature extraction module according to an embodiment of the present invention. Figure 4 As shown, the spectral feature extraction module includes three parallel one-dimensional convolutional neural network branches, used to process absorption spectral data, upper interface photoluminescence spectral data, and lower interface photoluminescence spectral data, respectively. Each branch adopts the same network structure, including multiple one-dimensional convolutional layers (…). Figure 4 This is just one possible network structure example; in practice, it can be extended to multiple one-dimensional convolutional layers as needed. (Kernel size) The choice is based on the spatial scale of spectral characteristics: absorption spectra typically have broad absorption bands, requiring a larger convolution kernel (size 1). ,in Half height and width, (Sampling interval); the photoluminescence spectrum has a narrow emission peak, and a small convolution kernel (size is...) is used. The first layer uses a large convolutional kernel to capture global features and long-range dependencies in the spectrum. The second layer uses a small convolutional kernel (half the size of the first layer) to capture local details and fine structures. Due to pooling, even though the second layer's kernel is smaller, its effective receptive field actually covers a larger range of the original spectrum. (The number of channels is also mentioned.) The selection is based on feature representation ability, computational complexity trade-offs, data characteristics, and experimental optimization. For example, the first convolutional layer of the absorption spectroscopy branch and the photoluminescence spectroscopy branch is selected... The second layer is selected To enhance feature extraction capabilities and adapt to the complexity of spectral data, while balancing computational efficiency, each convolutional layer, except the last one, is followed by a batch normalization (BN) layer and a ReLU activation function, and max pooling is employed. Downsampling is performed; the last convolutional layer is directly connected to an adaptive pooling layer, eliminating the use of BN, ReLU, and MaxPool1D operations. This design enables the extraction of local and global features at different scales from the raw spectral data, fully leveraging the device performance information contained within the spectral data.

[0057] To ensure that the features extracted from different branches have the same dimension, an adaptive pooling layer is added at the end of each branch to map the features to a fixed dimension. Specifically, for the first Features extracted from each branch ,in The length of the feature vector after the last convolutional layer in this branch is determined by adaptive average pooling. Obtain feature vectors of uniform dimension :

[0058]

[0059] The physical-guided feature enhancement module is a key innovative component of this invention. This module guides the feature extraction process by calculating key physical feature parameters of the spectral data. The specific physical feature extraction method is as follows:

[0060] Extracting band gap energy from absorption spectra The Tauc analysis method was used.

[0061]

[0062] in The absorption coefficient is... Photon energy, It is a constant. Depending on the type of transition (direct bandgap) Indirect band gap The bandgap energy is obtained by linearly fitting the linear region of the Tauc plot and finding its intercept. .

[0063] Extracting Urbach energy from absorption spectra Characterizing the density of states with tails:

[0064]

[0065] Where α(E) represents the absorption coefficient related to the photon energy E. and It is a constant. It is obtained by fitting the exponential region of the absorbing edge.

[0066] Extracting peak energy from photoluminescence spectrum Half height and width Integral strength and Stokes displacement Peak position and full width at half maximum (FWHM) were obtained through Gaussian fitting:

[0067]

[0068] Where PL(λ) represents the photoluminescence intensity at wavelength λ. For peak strength, Peak position, Related to half-width: Stokes shift is defined as the energy difference between the absorption edge and the emission peak. .

[0069] These physical feature parameters are converted into high-dimensional feature vectors by a specialized physical feature encoder, and then fused with deep features extracted by a convolutional neural network. The physical guidance mechanism ensures that the model focuses on physical information closely related to device performance during the learning process, improving the model's sensitivity and recognition ability to key physical features.

[0070] Figure 5 The diagram shows a structural schematic of the multimodal feature fusion module according to an embodiment of the present invention, as shown below. Figure 5 As shown, the multimodal feature fusion module employs an attention-based feature fusion strategy. First, the feature vectors extracted from the three branches and the physically guided feature vector are mapped to the same dimensional space through a fully connected layer. Then, the importance weight of each feature information to the final prediction result is calculated. The attention weights are calculated using a self-attention mechanism, achieving adaptive feature weighting by learning the correlation and complementarity between different feature information. Specifically, the formula for calculating the attention weights is:

[0071]

[0072] in, This represents the attention weight of the i-th feature vector. Indicates the first The feature vector after unifying the dimensions, , This is the weight matrix. , For bias vectors, For the feature vector dimension, Corresponding to absorption spectral characteristics upper interface photoluminescence characteristics Photoluminescence characteristics of the lower interface and physical guidance features The final fused feature vector is obtained by weighted summation:

[0073]

[0074] in, This is the fused comprehensive feature vector. This attention-based fusion method can adaptively adjust the weights of each feature based on the spectral characteristics of different samples, achieving more effective multimodal information fusion.

[0075] The performance metric prediction module includes a multi-layer fully connected neural network used to map the comprehensive feature vector to predicted performance metric values. This module comprises multiple hidden layers with the number of neurons decreasing layer by layer. The activation function is ReLU, and the output layer contains four neurons, each corresponding to a photoelectric conversion efficiency. Open circuit voltage Short-circuit current density and fill factor The predicted values. To prevent overfitting, a dropout operation is added after each hidden layer.

[0076] Step S103: Design the fusion loss function and train the model. The core innovation of this invention lies in the design of the physical information fusion loss function. This function combines data-driven prediction loss with constraint loss based on physical laws, and adopts an adaptive weight adjustment mechanism of gradient balancing to ensure that physical constraints improve prediction accuracy without interfering with the model's data fitting ability. Figure 6 The diagram below illustrates the composition of the fusion loss function according to an embodiment of the present invention, as shown in the figure. Figure 6 As shown, the general form of the fusion loss function can be expressed as:

[0077]

[0078] in The total number of loss items, For the first Each loss item, For adaptive weighting coefficients, For example, the transformation function is optional. (linear) (square root) (Logarithmic) or (Exponential saturation) and other forms. This flexible loss function design can adapt to the characteristics and constraints of different types of photovoltaic devices.

[0079] For perovskite solar cells, the overall form of the fusion loss function adopted in this embodiment is as follows:

[0080]

[0081] in, Indicates the first Each loss term in training rounds Adaptive weighting coefficients at time For the current training round, This represents the k-th physical constraint term. This represents the data-driven prediction loss term; the specific definitions of each term are as follows:

[0082] Data-driven prediction of loss terms The difference between the predicted value and the actual experimental measurement value is calculated using the root mean square error (RMSE) method.

[0083]

[0084] in, Indicates the first The first sample The model prediction values ​​for each performance metric, This represents the corresponding experimental measurement value. For the sample size, For the first The standard deviation of each performance metric is used to normalize performance metrics with different dimensions. .

[0085] Open-circuit voltage physical constraint Based on detailed equilibrium principles and carrier recombination dynamics, the open-circuit voltage value used for constraint prediction conforms to the theoretical relationship derived from detailed equilibrium principles and carrier recombination dynamics. According to detailed equilibrium principles, the open-circuit voltage is related to the balance between radiative and non-radiative recombination:

[0086]

[0087] in, The open-circuit voltage value predicted by the model. This is the theoretical open-circuit voltage value. These are absolute values, calculated based on the External Radiative Efficiency (ERE):

[0088]

[0089] in The radiation-limiting open-circuit voltage. J / K is the Boltzmann constant. This is the absolute temperature (usually taken as 298 K). C is the elementary charge, and ERE is estimated by the ratio of photoluminescence intensity at the upper and lower interfaces:

[0090]

[0091] in, and These are the integrals of the photoluminescence intensity at the upper and lower interfaces, respectively. This is the Stokes displacement.

[0092] Physical constraints on short-circuit current density Ensure that the predicted short-circuit current density does not exceed the theoretical upper limit calculated based on the absorption spectrum and the standard solar spectrum. Theoretical upper limit. The following calculations were performed by convolving the absorption spectrum with the photon flux density of the standard AM1.5G solar spectrum:

[0093]

[0094] in, Photon flux density Standard AM1.5G solar spectral irradiance (unit: W·m) ·nm ), J·s is Planck's constant. m / s is the speed of light and These are the upper and lower bounds for the wavelength. The constraint terms are in the form of a one-sided penalty function:

[0095]

[0096] in, The short-circuit current density value predicted by the model. To correct the linear unit function.

[0097] Fill factor physical constraint The fill factor value used for constraint prediction, based on the diode equation and series-parallel resistance effect, conforms to the theoretical relationship based on the diode equation and series-parallel resistance effect. Series resistance is considered. and parallel resistors Due to the influence of [the factor], the theoretical value of the fill factor is:

[0098]

[0099] in, These are the predicted values ​​from the fill factor model and the theoretical values ​​of the fill factor. The following was obtained through numerical solution of the implicit equation:

[0100]

[0101] in, For normalized open-circuit voltage, The ideal factor is (measured by temperature dependence of photoluminescence spectrum or estimated based on recombination mechanism, typically 1-2). For normalized series resistance, For normalized parallel resistance, This is the normalized voltage at the maximum power point. The series and parallel resistances can be estimated by the excitation intensity dependence of the photoluminescent quantum yield (PLQY).

[0102] Upper limit constraint on photoelectric conversion efficiency Designed based on Shockley-Quisser theory, the predicted photoelectric conversion efficiency value is constrained to not exceed the theoretical efficiency upper limit calculated based on Shockley-Quisser theory:

[0103]

[0104] in, The upper limit of the Shockley-Quyther theoretical efficiency is based on the band gap energy. It was obtained through detailed balance calculations, taking into account the balance between blackbody radiation and the solar spectrum.

[0105] Photovoltaic conversion efficiency self-consistency constraint Ensure that the predicted photoelectric conversion efficiency is consistent with the product of its component parameters (predicted open-circuit voltage, short-circuit current density, and fill factor).

[0106]

[0107] in, mW / cm This represents the incident power density under standard test conditions.

[0108] The adaptive weight adjustment mechanism for gradient balancing is the key innovation of this invention. This mechanism dynamically adjusts the weight coefficients by balancing the gradient magnitudes of each loss term. The weight update formula is:

[0109]

[0110] in, For the first The gradient norm of each loss term, The moving average of the gradient norm of the data loss term. For numerically stable terms, The balancing factor is typically 0.9. This method ensures that the contributions of each loss term to parameter updates are relatively balanced, preventing any one loss term from dominating the training process.

[0111] To avoid excessive changes in weight coefficients that could lead to training instability, each weight coefficient is constrained within a preset range and normalized.

[0112]

[0113] in , This is the normalization constant (typically 7).

[0114] Figure 7 The flowchart of model training according to an embodiment of the present invention is shown below, as follows: Figure 7 As shown, the model training process employs an innovative progressive training strategy, which ensures that physical constraints effectively guide model learning without interfering with data fitting ability. The training process consists of three stages: pre-training, physical constraint introduction, and joint optimization.

[0115] During the pre-training phase, network parameters are initialized using a subset of labeled data (30% of the total data) through purely data-driven supervised learning:

[0116]

[0117] in The loss function during the pre-training phase. This is a data-driven loss function, used to fit the supervised learning loss to the training data. The number of training epochs in this stage. ( (Total training rounds), the goal is to enable the network to learn basic feature extraction and mapping capabilities.

[0118] In the physical constraint introduction stage, the weight of the physical constraint loss term is gradually increased, enabling the model to learn the physical relationship between spectral features and device performance parameters:

[0119]

[0120] in A stage loss function is introduced for physical constraints, which includes data loss and physical constraint loss. As an asymptotic factor, This is a progressive cycle. The number of training rounds in this phase... .

[0121] During the joint optimization phase, all training data is used, and the complete fusion loss function is employed. End-to-end optimization is performed using gradient balancing mechanisms. The number of training epochs in this phase... .

[0122] The training dataset contains Spectral data and corresponding performance measurements of one half-cell sample were collected, with the training, validation, and test sets divided in an 8:1:1 ratio. Data preprocessing steps included spectral alignment, interpolation resampling, and normalization. The wavelength range of all spectral data was uniformly set to 350 nm to 850 nm, and resampled to 501 data points. Spectral intensity was processed using the Z-score normalization method.

[0123]

[0124] in, Represents the original spectral data, and These are the mean and standard deviation of the spectral data, respectively. This is the standardized spectral data.

[0125] The model training uses the AdamW optimizer, whose parameter update formula is:

[0126]

[0127] in, For the first Model parameters during training rounds, For learning rate, and These are the first and second moment estimates of the gradient, respectively. It is the numerical stability constant. This represents the weight decay coefficient. The learning rate is dynamically adjusted using a cosine annealing strategy.

[0128]

[0129] in, The initial learning rate, This is the minimum learning rate. The initial learning rate during the pre-training phase is set to... The initial learning rate in the physical constraint introduction phase is set to The initial learning rate for the joint optimization phase is set to The minimum learning rate is uniformly set to .

[0130] During training, the data-driven prediction loss term ensures the model accurately fits the training data, while the physical constraint loss term guides the model to learn prediction patterns that conform to physical principles. The gradient balancing mechanism ensures the optimal balance between the two. This design enables the model not only to learn the statistical correlations in the data but also to understand the physical causal relationship between spectral features and device performance, thus achieving good generalization performance even with limited training samples, while ensuring the physical rationality of the prediction results.

[0131] Step S104: Model Performance Prediction Application. For a new half-cell sample to be predicted, its three spectral data are collected according to the method described in step S101. After preprocessing, the data is input into the trained physical information fusion prediction model. Through a single forward propagation calculation, the model can output the predicted values ​​of the sample's photoelectric conversion efficiency, open-circuit voltage, short-circuit current density, and fill factor within milliseconds. The prediction results are not only highly accurate but also have good physical interpretability, providing reliable guidance for experimental research and industrial production.

[0132] According to certain optional embodiments, the kernel size in the spectral feature extraction module is determined based on the actual scale of the spectral features. For absorption spectra, a typical absorption band half-width is approximately 100 nm, and the sampling interval is 2 nm; therefore, the kernel size is set to... The sampling points are chosen, with the closest odd number, 101, used in practical applications. For photoluminescence spectra, the typical emission peak half-width at half-maximum (FWHM) is approximately 40 nm, and the convolution kernel size is set to... The sampling points are selected from an odd number of 21. The physical feature encoder in the physical guidance feature enhancement module uses a two-layer fully connected network with 128 and 64 neurons respectively; the attention mechanism of the multimodal feature fusion module adopts a 4-head attention structure, with the number of attention heads... Dimensions of each attention head The performance indicator prediction module contains four hidden layers with 512, 256, 128, and 64 neurons respectively, and the output layer contains four neurons corresponding to four performance parameters.

[0133] The initial settings for the weights in the fusion loss function need to balance the importance of prediction accuracy and physical constraints. The initial weights for the data-driven loss term are set as follows: The initial weights of the open-circuit voltage physical constraint term are set to The initial weights of the short-circuit current density physical constraint term are set to The initial weights of the fill factor physical constraint terms are set to The initial weights of the Shockley-Quyther upper bound constraint term are set to... The initial weights of the self-consistency constraint terms are set to The parameters of the gradient balancing mechanism are set as balance coefficients. .

[0134] The convergence criterion during training is based on the change in loss on the validation set. When continuous... The relative change in validation set loss within each training epoch is less than the threshold. When the model converges, the following conditions are met:

[0135]

[0136] in, For the first The loss value of the validation set after one round of training.

[0137] Embodiments of the present invention also provide a solar cell performance prediction device based on a physical information fusion model. This device includes a spectral data acquisition module, a data preprocessing module, and a performance prediction module. The spectral data acquisition module integrates an ultraviolet-visible-near-infrared spectrophotometer and a microscopic confocal Raman spectroscopy system, enabling automatic acquisition of three types of spectral data. The data preprocessing module is responsible for data processing tasks such as spectral alignment, interpolation, and normalization. The performance prediction module, based on a trained physical information fusion prediction model, achieves rapid prediction from spectral data to performance indicators. The entire device has advantages such as ease of operation, high prediction accuracy, and fast processing speed.

[0138] In summary, the embodiments of the present invention relate to a method and apparatus for predicting the performance of solar cells based on a physical information fusion model. The prediction method includes the following steps: acquiring spectral data of a half-cell sample of a solar cell to be predicted; inputting the spectral data into a physical information fusion prediction model to obtain predicted performance index data of the solar cell to be predicted; wherein, the physical information fusion prediction model includes a spectral feature extraction module, a physical-guided feature enhancement module, a multi-modal feature fusion module, and a performance index prediction module connected in sequence; the physical information fusion prediction model is obtained through progressive training, which includes a pre-training stage, a physical constraint introduction stage, and a joint optimization stage, and uses a gradient-balanced fusion loss function for parameter optimization, the fusion loss function including a data-driven prediction loss term and a multiple physical constraint loss term. The technical solution provided in this invention employs a deep neural network architecture to extract and fuse features from multi-dimensional spectral information. A physically guided feature enhancement module ensures the model focuses on physical information closely related to device performance. Compared to traditional single-spectral analysis methods, this approach more comprehensively characterizes the photoelectric properties of devices, thereby improving prediction accuracy and reliability. The model training utilizes an innovative progressive training strategy and gradient balancing mechanism to ensure that physical constraints effectively guide model learning without excessively interfering with data fitting capabilities. This deeply integrates the prediction task with physical principles, enabling the prediction model to understand the physical causal relationship between spectral features and device performance, achieving more accurate performance predictions and better physical interpretability. Multiple physical constraints are used to standardize the prediction results, ensuring that the model's predictions conform to the physical principles of semiconductor devices, improving the physical rationality and engineering application value of the prediction results, and providing crucial technical support for the rapid screening and performance optimization of solar cells.

[0139] It should be understood that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the invention as described above, which are not provided in detail for the sake of brevity. The above specific embodiments of the invention are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation of the invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for predicting performance of a solar cell based on a physical information fusion model, the method comprising: obtaining a plurality of physical information of the solar cell; and predicting the performance of the solar cell based on the plurality of physical information. The method comprises the steps of: Step 1: collecting and preprocessing spectral data of a solar cell half-cell sample to be predicted, the spectral data comprising absorption spectrum data, upper interface photoluminescence spectrum data and lower interface photoluminescence spectrum data of the half-cell sample; Step 2: constructing a physical information fusion prediction model, comprising a spectral feature extraction module, a physical guided feature enhancement module, a multi-modal feature fusion module and a performance index prediction module connected in sequence; The spectral feature extraction module extracts features from the three types of spectral data through three neural network branches; The physical guided feature enhancement module calculates key physical feature parameters of the spectral data to guide the feature extraction process; The multi-modal feature fusion module fuses the features extracted by the spectral feature extraction module and the physical guided feature enhancement module through an attention mechanism; The performance index prediction module maps the fused feature vector to a performance index prediction value; Step 3: designing a fusion loss function and training the model; Step 4: using the trained model to predict the performance index of the half-cell sample; The physical guided feature enhancement module is as follows: The Tauc analysis method is used to extract the band gap energy from the absorption spectrum; the Urbach energy is extracted from the absorption spectrum; the peak position energy, half-width, integral intensity and Stokes shift are extracted from the photoluminescence spectrum; then these physical parameters are converted into high-dimensional feature vectors by a physical feature encoder as physical guided feature vectors; The model training comprises a pre-training stage, a physical constraint introduction stage and a joint optimization stage; In the pre-training stage, the network parameters are initialized by supervised learning driven by pure data using part of the labeled data; In the physical constraint introduction stage, the weight of the physical constraint term is gradually increased to make the model learn the physical relationship between the spectral features and the device performance parameters; In the joint optimization stage, the whole training data is used to optimize end-to-end with the complete fusion loss function and the gradient balance mechanism; the fusion loss function adopts the adaptive weight adjustment mechanism of the gradient balance, and the loss term includes the data-driven prediction loss term and the physical constraint term. ​ 2.The solar cell performance prediction method based on a physical information fusion model according to claim 1, wherein, The spectral feature extraction module is three parallel one-dimensional convolutional neural network branches, which are used to process absorption spectrum data, upper interface photoluminescence spectrum data and lower interface photoluminescence spectrum data respectively; each branch adopts the same network structure, including multiple one-dimensional convolutional layers; each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and a max-pooling operation is used for down-sampling; the last convolutional layer is directly connected to an adaptive pooling layer to obtain uniform-dimensional feature vectors of the three types of spectral data. 3.The solar cell performance prediction method based on a physical information fusion model according to claim 2, wherein, The performance index prediction module comprises multiple hidden layers with decreasing number of neurons, and the activation function adopts ReLU; the number of neurons in the output layer corresponds to the performance index. 4.The solar cell performance prediction method based on a physical information fusion model according to claim 3, wherein, The fusion loss function formula is as follows: ； wherein is the total number of loss terms, is the th loss term, is the adaptive weight coefficient, is the transformation function; the data-driven prediction loss term calculates the difference between the predicted value and the real experimental measured value; the physical constraint term learns the physical relationship between the spectral features and the device performance parameters; The adaptive weight coefficient is determined by one or more of the following ways: fixed weight, gradient balance adjustment, adaptive adjustment based on validation set performance, dynamic adjustment based on physical consistency and uncertainty-based weighting adjustment. 5.The solar cell performance prediction method based on a physical information fusion model according to claim 4, wherein, The performance index prediction data includes photoelectric conversion efficiency, open circuit voltage, short circuit current density and fill factor, and the physical constraint term includes: An open circuit voltage physical constraint term for constraining the predicted open circuit voltage value to comply with the theoretical relationship; A short circuit current density physical constraint term for constraining the predicted short circuit current density value to be less than the theoretical upper limit value; a fill factor physical constraint term for constraining the predicted fill factor value to comply with a theoretical relationship; a photoelectric conversion efficiency upper limit constraint term for constraining the predicted photoelectric conversion efficiency value to be no more than an upper limit of theoretical efficiency; a photoelectric conversion efficiency self-consistency constraint term for ensuring that the predicted photoelectric conversion efficiency value is physically defined to be consistent with the predicted open-circuit voltage value, short-circuit current density value and fill factor value. 6.The method of claim 5, wherein, In the step 1, the three spectral data are specifically obtained by the following manner: First, the original spectral data are preprocessed, including baseline correction, noise filtering and outlier rejection; The collection of absorption spectrum data is measured by a UV-Vis-NIR spectrophotometer; the total transmittance spectrum of the half-cell sample is measured in a controlled environment and the total reflectance spectrum According to the law of conservation of energy, the absorption spectrum data of the sample is obtained by calculation ; wherein represents the wavelength of light, 、 、 represents the absorption, transmittance and reflectance at the wavelength , respectively; The collection of the upper interface photoluminescence spectrum data and the lower interface photoluminescence spectrum data is measured by a microscopic confocal Raman spectrum system; in the lower interface photoluminescence spectrum measurement, the excitation source is vertically incident from the transparent conductive substrate side, the active layer is excited to generate photoluminescence, and the photoluminescence signal is collected on the same side to obtain the lower interface photoluminescence spectrum data ; in the upper interface photoluminescence spectrum measurement, the excitation source is vertically incident from the hole transport layer side for measurement, and the remaining measurement conditions remain the same as those in the lower interface measurement, to obtain the upper interface photoluminescence spectrum data .

7. A solar cell performance prediction device based on a physical information fusion model, characterized by, The method comprises the following steps: a spectral data acquisition module for acquiring absorption spectral data, upper interface photoluminescence spectral data and lower interface photoluminescence spectral data of a half-cell sample of a solar cell to be predicted; a data preprocessing module for performing a preprocessing operation on the spectral data; a performance prediction module for inputting the preprocessed spectral data into a physical information fusion prediction model trained according to the method of any one of claims 1-6, and outputting performance index prediction data.

Images