Training, prediction method and device of battery property prediction model and battery

Abstract

The application provides a battery attribute prediction model training and prediction method and device and a battery, relates to the technical field of battery prediction model application, and the training method comprises the following steps: establishing an initial prediction model; a composite loss function is constructed based on a data-driven loss term and a physical constraint loss term, wherein the physical constraint loss term represents the deviation degree of the prediction output relative to the prior physical law between the battery material components and the battery performance; and a target prediction model is trained to minimize the composite loss function. The initial prediction model can adopt a hybrid neural network architecture, including a shared classical layer, a classical branch, a quantum branch and a fusion layer, to realize collaborative prediction of PCE and T80 through double-branch parallel feature extraction and post-fusion, and through a physical constraint training mechanism and a quantum-classical hybrid architecture, the problem of insufficient model generalization ability under a small sample is solved, and high-precision, physically interpretable battery multi-attribute prediction is realized.

Description

Technical Field

[0001] This invention relates to the technical field of battery prediction model applications, and in particular to a method, apparatus, and battery for training and predicting battery property prediction models. Background Technology

[0002] Battery technology is a key support for driving the energy transition. Utilizing machine learning models to mine structure-property relationships from experimental data to accelerate the screening of high-performance battery materials and device optimization has become a cutting-edge paradigm in the field of battery research and development.

[0003] In training battery attribute prediction models, existing techniques generally employ loss functions composed of data-driven loss terms. This loss function aims solely to minimize the deviation between the model's predicted values ​​and the measured label values, guiding the model to learn the mapping relationship from features to performance on the training data.

[0004] However, acquiring high-performance experimental data in the battery field is costly and time-consuming, and the actual usable samples are often limited. Under these conditions, prediction models may overfit to the noise in the training data, and the mappings they learn may violate established fundamental laws of electrochemistry and thermodynamics. This leads to a serious lack of generalization ability in the model, and may even output physically impossible predictions, greatly weakening the model's practical value and engineering credibility. Summary of the Invention

[0005] The purpose of this invention is to provide a training and prediction method, device, and battery for a battery property prediction model, so as to alleviate the technical problem that when the training samples of a battery performance prediction model are limited, its prediction results may violate known physical and chemical laws, resulting in insufficient generalization ability and prediction reliability of the model.

[0006] In a first aspect, the present invention provides a method for training a battery attribute prediction model, comprising: Establish an initial prediction model for predicting battery performance attributes; A composite loss function is constructed for training the initial prediction model based on a data-driven loss term and a physical constraint loss term. Based on the battery sample dataset, the initial prediction model is trained with the goal of minimizing the composite loss function to obtain the target prediction model.

[0007] In an optional implementation, a composite loss function for training the initial prediction model is constructed based on a data-driven loss term and a physical constraint loss term, including: Using prior physical laws as constraints, calculate the residual of the predicted value output by the initial prediction model based on the input sample relative to the prior physical laws; The physical constraint loss term is constructed from the residuals; The physical constraint loss term and the data-driven loss term are weighted and summed to form the composite loss function.

[0008] In an optional implementation, the a priori physical laws include: a first quantitative relationship between the proportion of halide ions in the battery material composition and the photoelectric conversion efficiency, and a second quantitative relationship between the types of metal ions in the battery material composition and the device lifetime. The physical constraint loss term includes a first physical constraint loss term and / or a second physical constraint loss term; The first physical constraint loss term characterizes the predicted photoelectric conversion efficiency value output by the prediction model for samples with different halide ion ratios. The second physical constraint loss term characterizes the lifetime prediction value output by the prediction model for samples of different metal ion types.

[0009] In an optional implementation, the method further includes: The physical constraint loss term is embedded as a soft constraint regularization term into the composite loss function, guiding the model parameters to be updated in a direction that satisfies prior physical laws during training.

[0010] In an optional implementation, the initial prediction model is a multi-task output model, including a first output unit for outputting the predicted value of photoelectric conversion efficiency and a second output unit for outputting the predicted value of device lifetime.

[0011] In an optional implementation, the initial prediction model is built based on a hybrid neural network, including: A shared classic layer is used to map the original input features to a first feature vector of uniform dimension; The classic branch, connected to the output of the shared classic layer, is used to perform classic feature extraction on the first feature vector to generate a second feature vector; The quantum branch, connected to the output of the shared classical layer, is used to encode the first feature vector into a quantum state and extract the qubits through a variable quantum circuit to generate a third feature vector. The fusion layer, connected to the outputs of the classical branch and the quantum branch, is used to concatenate the second feature vector and the third feature vector, and output the predicted value of the battery performance attribute based on the concatenated joint feature vector.

[0012] In an optional implementation, the quantum branch includes: An amplitude coding layer is used to map the first feature vector into a quantum state of multiple qubits through amplitude coding. A parameterized quantum circuit comprises multiple circuit layers consisting of alternating rotational and entangled modules, used to evolve the quantum state; The measurement layer is used to measure the evolved qubits and obtain the output characteristics of the quantum circuit. A quantum post-processing layer is used to map the output features of the quantum circuit to the third feature vector.

[0013] In an optional implementation, when training the initial prediction model with the objective of minimizing the composite loss function, the method includes: The parameters of the variable quantum circuit in the quantum branch are updated using a non-gradient optimization method, while the parameters of the shared classical layer, the classical branch, and the fusion layer are updated using a gradient optimization method.

[0014] Secondly, the present invention provides a prediction method for a battery attribute prediction model, comprising: Obtain feature data of the battery to be predicted; The feature data is input into the target prediction model trained according to any of the aforementioned embodiments to obtain the battery performance attribute prediction result.

[0015] Thirdly, the present invention provides a training apparatus for a battery property prediction model, comprising: Establish a module to build an initial prediction model for predicting battery performance attributes; The module constructs a composite loss function for training the initial prediction model, based on a data-driven loss term and a physical constraint loss term. The training module, based on the battery sample dataset, trains the initial prediction model with the goal of minimizing the composite loss function to obtain the target prediction model.

[0016] Fourthly, the present invention provides a prediction device for a battery property prediction model, comprising: The acquisition module acquires the feature data of the battery to be predicted; The prediction module inputs the feature data into the target prediction model trained according to any of the aforementioned embodiments to obtain the battery performance attribute prediction result.

[0017] Fifthly, the present invention provides a battery in which the prediction device described in the foregoing embodiments is used to predict battery performance attributes.

[0018] In a sixth aspect, the present invention provides a battery pack comprising the battery described in the foregoing embodiments.

[0019] In a seventh aspect, the present invention provides an electrical device comprising the battery or battery pack described in the foregoing embodiments.

[0020] This invention provides a training and prediction method, apparatus, and battery for a battery property prediction model. It introduces a novel training paradigm that embeds prior physical laws governing the relationship between battery material composition and performance as soft-constraint regularization terms into model training optimization. The training method establishes an initial prediction model and constructs a composite loss function consisting of a data-driven loss term and a physical constraint loss term. The physical constraint loss term quantifies the degree to which the model's predicted output deviates from prior physical laws in the domain. Model training is then performed with the goal of minimizing this composite loss function. This training mechanism guides the model to converge in a direction consistent with physical laws while fitting the distribution of sample data. This significantly improves the model's generalization ability and prediction reliability under limited sample size conditions, solving the key technical problem that pure data-driven methods may result in predictions that violate fundamental physical principles.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a training method for a battery attribute prediction model provided in an embodiment of the present invention; Figure 2 This diagram illustrates the performance comparison between the classic model and the prediction model on the PCE prediction task. Figure 3 This diagram illustrates the performance comparison between the classic model and the prediction model on the T80 prediction task. Figure 4 A flowchart of a prediction method for a battery attribute prediction model provided in an embodiment of the present invention; Figure 5 A schematic diagram of a quantum branch parameterized quantum circuit structure for a battery property prediction model provided in an embodiment of the present invention; Figure 6A schematic diagram of the functional modules of a training device for a battery property prediction model provided in an embodiment of the present invention; Figure 7 A schematic diagram of the functional modules of a prediction device for a battery property prediction model provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the battery structure provided by the present invention; Figure 9 This is a schematic diagram of the battery pack provided by the present invention; Figure 10 This is a schematic diagram of the electrical device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Currently, the development of batteries such as perovskite solar cells faces challenges due to their vast compositional space and the multidimensional coupling of influencing factors. Traditional experimental trial-and-error methods are time-consuming and costly. Existing machine learning prediction schemes either only predict a single attribute or use purely traditional classical models. When processing high-dimensional, small-sample, and strongly nonlinear perovskite data, these schemes suffer from insufficient feature extraction capabilities, fragmentation of physical relationships between attributes, and the possibility that prediction results may violate basic physical laws. This results in poor model generalization ability and insufficient prediction reliability, failing to meet the industrial-grade requirement for accurate prediction of both Power Conversion Efficiency (PCE) and T80, a key lifetime indicator for measuring device stability.

[0027] Based on this, the present invention provides a battery attribute prediction model training, prediction method, device and battery, which, through the synergistic effect of physical constraint training mechanism and quantum and classical hybrid neural network architecture, significantly reduces the number of trainable parameters of the model while improving prediction accuracy, and realizes multi-attribute collaborative prediction of batteries that takes into account prediction accuracy, generalization ability and physical interpretability.

[0028] To facilitate understanding of this embodiment, a battery attribute prediction method disclosed in this embodiment of the invention will first be described in detail.

[0029] Figure 1 This is a flowchart illustrating a training method for a battery property prediction model provided in an embodiment of the present invention.

[0030] Reference Figure 1 The method includes: Step S102: Establish an initial prediction model for predicting battery performance attributes.

[0031] Here, a machine learning model framework is constructed based on the multidimensional feature data of perovskite solar cells, which can receive the above feature inputs and output predicted values ​​of battery performance attributes. This provides a model foundation for subsequent physical constraint training, enabling the model to have the basic ability to receive multidimensional feature data and generate predicted values ​​of performance attributes.

[0032] The photoelectric conversion efficiency (PCE) prediction task has 124 feature dimensions, while the long-term stability lifetime (T80) prediction task has 72 feature dimensions, covering all dimensions of information including microcrystalline structure parameters, physicochemical parameters, preparation environment parameters, process temperature parameters, and solvent engineering parameters. The stability model additionally incorporates test environment parameters based on the ISOS protocol. The initial prediction model adopts a hybrid neural network architecture, integrating the high-dimensional feature extraction capabilities of quantum neural networks with the nonlinear modeling capabilities of classical deep learning.

[0033] like Figure 2 As shown, two True vs Predicted scatter plots compare the performance of the 2(a) classical model and the 2(b) quantum model on the PCE prediction task. The scatter plots represent the prediction results, and the dashed line is the ideal fitting line. The classical model has an RMSE of 0.014 and an R² of 0.852, while the quantum model has an RMSE of 0.012 and an R² of 0.888, and the scatter plots are closer to the ideal line. In the T80 prediction task, the hybrid model also shows better prediction performance and significantly improved generalization ability on small sample datasets.

[0034] Step S104: Based on the data-driven loss term and the physical constraint loss term, construct a composite loss function for training the initial prediction model.

[0035] Among them, the physical constraint loss term characterizes the degree of deviation of the prediction output of the initial prediction model from the prior physical laws between the battery material composition and battery performance.

[0036] Here, the training optimization objective is extended from the traditional single data-driven loss to a dual-constraint approach. The data-driven loss term calculates the difference between the model's predicted value and the true label (e.g., mean squared error, MSE); the physical constraint loss term quantifies prior physical laws in the field of battery materials (e.g., the directional constraint of the effect of halide ion ratio changes on photoelectric conversion efficiency, the trend of the influence of specific metal ions on device stability, etc.) into a differentiable residual function, which is then embedded and backpropagated through partial derivative calculation. The composite loss function can be expressed as:

[0037] in, For data-driven loss terms, For physical constraint loss terms, These are the weighting coefficients for physical constraints, used to adjust the influence of physical constraints during training.

[0038] Among them, a priori physical laws refer to the quantitative relationships or trends between battery material components and battery performance that have been experimentally or theoretically verified in the field of battery materials. For example, in the field of perovskite materials, the regulation of band gap by changes in the proportion of halogen ions, and the influence of specific metal ions on the stability of perovskite structures.

[0039] This application organically embeds domain physics knowledge into the model training process, forcing the model to respect physical laws while fitting data, effectively avoiding physically unreasonable prediction results produced by purely data-driven models under small sample conditions.

[0040] Step S106: Based on the battery sample dataset, the initial prediction model is trained with the goal of minimizing the composite loss function to obtain the target prediction model.

[0041] Here, a dataset containing full-dimensional feature label pairs of multiple battery samples is used. The feature data is input into the initial prediction model to perform forward propagation to obtain predicted values. The composite loss function value is calculated, and the model parameters are updated through backpropagation and an optimizer (such as the Adam optimizer). This process is iteratively executed until the loss function converges or the preset number of training rounds is reached. Through end-to-end iterative training, the model parameters are continuously optimized under the dual guidance of data fitting and physical constraints, ultimately resulting in a target prediction model that combines data fitting accuracy and physical interpretability.

[0042] In a preferred embodiment for practical application, firstly, a battery sample dataset is collected. Feature data includes microcrystalline structure parameters, physicochemical parameters, preparation environment parameters, process temperature parameters, and solvent engineering parameters. Tag data consists of measured power conversion efficiency (PCE) and long-term stability lifetime (T80). The data is then cleaned and normalized. Secondly, an initial prediction model is established, which is a hybrid neural network architecture. Then, a composite loss function is constructed, calculating the data-driven loss (e.g., MSE) between the model's predicted values ​​and the true labels, and the residual deviation of the predicted values ​​from prior physical laws as a physical constraint loss term. These two are then weighted and summed using a weighting coefficient λ. Furthermore, the specific form of the prior physical laws is determined, such as the quantitative relationship trend between the halide ion ratio and PCE. Finally, with the goal of minimizing the composite loss function, the Adam optimizer is used to iteratively train the model until convergence, yielding the target prediction model.

[0043] In some embodiments, step S104 may construct a composite loss function through the following steps: Step 2.1: Using prior physical laws as constraints, calculate the residual of the predicted value output by the initial prediction model based on the input sample relative to the prior physical laws.

[0044] Here, for each sample in the training batch, the prior physical laws are expressed as a mathematical function relationship between the input features and the expected output. Actual output of the calculation model The residual between the expected value of this physical law and For example, regarding the halide ion ratio-PCE relationship, according to physical laws, as the bromide ion ratio increases, the band gap widens and the PCE should exhibit a specific trend. If the model's predicted value deviates from this trend, non-zero residuals will occur. Quantifying qualitative physical laws into calculable residual indices provides a mathematical basis for constructing physical constraint loss terms.

[0045] Step 2.2: Construct the physical constraint loss term from the residuals.

[0046] Here, the residuals calculated in step 2.1 are mathematically transformed to construct the loss term. This can be expressed as the sum of squares or the mean square error of the residuals:

[0047] in, It is a physical constraint loss term. It is the first Performance prediction value for each sample : No. The full-dimensional input feature vector of a perovskite sample (including the ratio of A / B / X site ions, tolerance factor, process parameters, test environment parameters, etc.). It is a theoretical performance mapping function derived from prior physical laws; the model applies to the... The performance prediction value output for each sample For the sample size, For the first The prediction residuals of each sample.

[0048] It should be noted that this application transforms the degree of violation of physical laws into an optimizable, continuously differentiable loss function, enabling physical constraints to participate in gradient-based model parameter optimization.

[0049] Step 2.3: Weight the physical constraint loss term and the data-driven loss term to form a composite loss function.

[0050] Here, we introduce hyperparameters. As physical constraint weight coefficients, construct a composite loss function. . The larger the value, the stronger the influence of physical constraints on training; It degenerates into pure data-driven training. Adjustable weighting coefficients allow for a flexible balance between data fitting accuracy and physical consistency.

[0051] For example, the construction of this composite loss function includes: First, the selected prior physical laws are expressed as mathematical functions. For example, the quantitative trend relationship between the halide ion ratio and PCE is analyzed. Then, in each batch of training, the model's predicted value for the current batch of samples is calculated, and the residual between the predicted value and the expected value of the physical law is calculated. Next, a physical constraint loss term is constructed in the form of the mean squared error of the residual. Furthermore, a physical constraint weight coefficient λ is set, and the physical constraint loss term and the MSE data-driven loss term are weighted and summed to form the final composite loss function used for backpropagation. Finally, the model parameters are updated by the optimizer based on the gradient of the composite loss function.

[0052] Based on the foregoing embodiments, as an optional embodiment, the prior physical laws include a first quantitative relationship between the proportion of halide ions in the battery material composition and the photoelectric conversion efficiency, and a second quantitative relationship between the types of metal ions in the battery material composition and the device lifetime; the physical constraint loss term includes a first physical constraint loss term and / or a second physical constraint loss term; The first physical constraint loss term characterizes the degree to which the predicted photoelectric conversion efficiency output by the prediction model for samples with different halide ion ratios violates the first quantitative relationship.

[0053] For example, halogen ions (such as bromide ions, Br) in perovskite materials - With iodide ions I - The influence of bromide ion ratio changes on photoelectric conversion efficiency is used as a physical constraint. This constraint can be expressed as follows: when the bromide ion ratio increases within a specific range, the perovskite bandgap increases, the short-circuit current density decreases but the open-circuit voltage increases, and the PCE shows a trend of first increasing and then decreasing, indicating the existence of an optimal halogen ratio range. If the model's predicted PCE values ​​for samples with different halogen ratios violate this trend (e.g., PCE monotonically increases with increasing bromide ratio), then a physical constraint loss occurs.

[0054] Specifically, the physical constraint residual can be expressed as ,in For the halogen ratio-PCE prior function and These represent the ratios of bromide ions to iodide ions. It is the predicted photoelectric conversion efficiency output by the hybrid neural network for the i-th perovskite sample. It is a quantitative trend function of halogen ion ratio-PCE that has been experimentally verified in the field of perovskite solar cells.

[0055] This aims to leverage the core component-performance relationship in the battery field to guide model learning, ensuring that the prediction results are physically reasonable.

[0056] Based on this, the second physical constraint loss term characterizes the degree to which the lifetime prediction values ​​output by the prediction model for samples of different metal ion types violate quantitative relationships.

[0057] Here, the types of metal ions at the B site (such as lead ions Pb²) will be discussed. + With tin ions Sn² + The influence of metal ions on the long-term stability of devices is used as a physical constraint. Tin-based perovskites are known to be more easily oxidized than lead-based perovskites, and their stability is generally poorer. If the model's predicted T80 value for tin-based devices is abnormally high (violating the trend of stability degradation), a loss of physical constraint occurs. By constraining through the metal ion-stability relationship, the physical reliability of the model's lifetime predictions for different component systems is improved.

[0058] In application, the following example can be used: First, if a halide ion ratio-PCE constraint is used, the bromide and iodide ion ratios of each sample are extracted during training. Then, a halide ratio-PCE trend function is established based on the physical laws of perovskite, and the residual deviation of the model's PCE prediction value from this trend is calculated. Next, the residual is included in the physical constraint loss term. If a metal ion type-lifetime constraint is used, a stability ranking rule is established based on the metal ion type at site B of the sample, and the degree to which the T80 prediction value violates this ranking rule is calculated. Furthermore, the two physical constraints mentioned above can be enabled simultaneously or separately in the composite loss function as needed.

[0059] In some embodiments, the prediction model training method of this application further includes: Step 3.1: Embed the physical constraint loss term as a soft constraint regularization term into the composite loss function to guide the model parameters to be updated in a direction that satisfies the prior physical laws during the training process.

[0060] Here, the physical constraint loss term is added to the total loss function in the form of a regularization term, through weighting coefficients. The degree of hardness or softness in control constraints. Unlike hard constraints, which require the predicted output to strictly satisfy physical equations, soft constraints allow the model to strike a balance between data fit and physical consistency. When the data itself contains noise or measurement errors, the model is not forced to fit physical laws at the expense of its ability to fit the real data. The soft constraint mechanism ensures that physical laws guide the model training while avoiding over-constraint caused by data noise, enabling the model to achieve an optimal balance between physical consistency and data fitting accuracy.

[0061] For example, first, the initial value of the weight coefficient λ of the physical constraints in the total loss function is determined (e.g., λ = 0.1). Then, during training, the physical constraint loss term is directly added to the data-driven loss as λ multiplied by the residual mean squared error. Furthermore, the value of λ can be adjusted based on the performance on the validation set; if the physical consistency of the validation set is insufficient, λ is increased; if the fitting accuracy of the validation set data is insufficient, λ is decreased, achieving a dynamic balance. Finally, the model parameters after training satisfy both the data distribution characteristics and better follow the prior physical laws.

[0062] In some embodiments, the initial prediction model is a multi-task output model, including a first output unit for outputting the predicted value of photoelectric conversion efficiency and a second output unit for outputting the predicted value of device lifetime.

[0063] Here, two independent neurons are set in the model output layer, corresponding to the PCE prediction value and the T80 prediction value, respectively. The two output units share the underlying feature extraction network (including shared classical layers, quantum branches, and classical branches), and realize information interaction and collaborative learning between the two tasks of PCE and T80 through shared feature representations. The shared underlying network captures the general mapping relationship between battery material components, such as perovskite material composition, structure, and process, and performance attributes. The two output units learn the task-specific final mapping respectively.

[0064] The multi-task output architecture can capture the inherent physical coupling between PCE and T80 (such as the stability trade-off that high-efficiency components often face). Compared with independent models predicting separately, collaborative learning can improve the prediction consistency and overall accuracy of the two tasks.

[0065] In practical applications, as an example, it includes: First, the model's output layer is designed to contain two parallel fully connected neurons: one for mapping PCE predictions and the other for mapping T80 predictions. Then, during training, a data-driven loss term is used to calculate the model's total loss for both tasks, which can be expressed as... α is the weighting coefficient for the T80 loss. Furthermore, when calculating the composite loss function, different prior physical constraints are applied to PCE and T80 in the physical constraint loss term (e.g., halogen ratio-PCE constraint applies to the first output unit, and metal ion-lifetime constraint applies to the second output unit). Finally, joint optimization is used to achieve coordinated prediction of the two attributes.

[0066] Based on the foregoing embodiments, the initial prediction model is constructed using a hybrid neural network, including: A shared classic layer is used to map the original input features to a first feature vector of uniform dimension.

[0067] Here, the shared classical layer consists of one or more fully connected layers and a ReLU activation function. Specifically, the first fully connected layer maps the original input feature dimension to 256 dimensions, and after ReLU activation, the second fully connected layer maintains the 256-dimensional output. The output dimension of this shared layer must satisfy the amplitude encoding requirement of the subsequent quantum branch (256=2^35). 8 (corresponding to 8 qubits). It provides a unified dimension of feature representation, laying the foundation for parallel processing of classical and quantum branches, while enhancing feature representation capabilities through preliminary high-dimensional mapping.

[0068] The classical branch, connected to the output of the shared classical layer, is used to extract classical features from the first feature vector to generate the second feature vector. The shared layer output undergoes progressive dimensionality reduction and nonlinear transformation to extract classical feature representations suitable for regression prediction.

[0069] Here, the classic branch consists of multiple fully connected layers and a ReLU activation function. Specifically, the first fully connected layer maps the 256-dimensional features to 64 dimensions. After ReLU activation, the second fully connected layer maps the 64-dimensional features to 16 dimensions, which is then output as the second feature vector.

[0070] The quantum branch, connected to the output of the shared classical layer, encodes the first feature vector into a quantum state and extracts quantum features through a variable quantum circuit to generate the third feature vector. Utilizing the parallel computational properties of quantum superposition and entanglement, it extracts deep abstract features in high-dimensional Hilbert space that are difficult for classical models to capture.

[0071] like Figure 3 As shown, two True vs Predicted scatter plots compare the performance of the 3(a) classical model and the 3(b) quantum model on the T80 prediction task. The scatter plots represent the prediction results, and the dashed line is the ideal fitting line. Experimental data show that in the PCE prediction task, the number of trainable parameters of the quantum-classical hybrid model is 116,041, which is about 51% less than the 237,873 of the all-classical model. In the T80 prediction task, the number of parameters of the hybrid model is 86,857, which is about 58% less than the 208,689 of the all-classical model. This significantly reduces the computational complexity while improving the prediction accuracy.

[0072] Here, the quantum branch receives the 256-dimensional first eigenvector from the shared classical layer output. It first maps the 256-dimensional classical data into an 8-qubit quantum state using amplitude encoding (256=2^3). 8The quantum state is then evolved using a variable quantum circuit that alternates between rotation modules and progressive entanglement modules. Finally, the evolved qubits are measured to obtain quantum characteristic representations. The parameterized quantum circuit consists of three repeating circuit layers, each containing a rotation module and a progressive entanglement module. The rotation module applies three parameterized rotation gates (RX, RY, and RZ) to each qubit sequentially. The progressive entanglement module uses a CNOT gate design with a span increasing with the number of layers. The first layer of CNOT gates has a span of 1 (adjacent qubits are entangled), the second layer has a span of 2, and the third layer has a span of 3, gradually establishing bit correlations from local to global.

[0073] The quantum branch includes: an amplitude encoding layer, used to map the first feature vector into the quantum state of multiple qubits through amplitude encoding; a parameterized quantum circuit, containing multiple circuit layers composed of alternating rotation modules and entanglement modules, used to evolve the quantum state; a measurement layer, used to measure the evolved qubits to obtain the output features of the quantum circuit; and a quantum post-processing layer, used to map the output features of the quantum circuit into the third feature vector.

[0074] The fusion layer, connected to the outputs of the classical and quantum branches, is used to concatenate the second and third feature vectors and output predicted values ​​of battery performance attributes based on the concatenated joint feature vector. This concatenation and fusion mechanism achieves information complementarity between quantum and classical features. The joint feature vector simultaneously contains high-dimensional abstract information from quantum space and nonlinear mapping information from classical space, improving prediction accuracy.

[0075] Here, the fusion layer first concatenates the 16-dimensional second feature vector output from the classical branch with the 16-dimensional third feature vector output from the quantum branch to form a 32-dimensional joint feature vector. Then, through a fully connected layer (32-dimensional → 16-dimensional, ReLU activation), a dropout layer (dropout rate 0.1), and an output layer (16-dimensional → 1-dimensional or 2-dimensional), the final attribute prediction value is generated.

[0076] The classical branch and the quantum branch each include a post-processing fully connected layer, which is used to map the extracted features to the same dimension. The dimension alignment design ensures the consistency of the features of the two branches during fusion, providing structural guarantee for subsequent splicing and fusion.

[0077] Here, the post-processing fully connected layer of the classical branch maps the 64-dimensional intermediate features to 16 dimensions; the quantum post-processing layer of the quantum branch maps the 8-dimensional features of the quantum circuit measurement output to 16 dimensions. Maintaining consistent output dimensions between the two branches ensures compatibility of the fusion layer splicing operation.

[0078] For example, first, a shared classical layer is constructed, consisting of two fully connected layers (input dimension → 256, 256 → 256), each followed by a ReLU activation function. Second, a classical branch is constructed, consisting of two fully connected layers (256 → 64, 64 → 16), each followed by a ReLU activation function. Then, a quantum branch is constructed, consisting of an amplitude encoding layer (mapping 256-dimensional features to 8-qubit quantum states), a parameterized quantum circuit (3 layers of alternating rotation-entanglement circuits), and a quantum post-processing layer (8 → 16 fully connected layer). Furthermore, a fusion layer is constructed, concatenating the 16-dimensional output of the classical branch with the 16-dimensional output of the quantum branch to form a 32-dimensional joint feature, which is then processed through a fully connected layer (32 → 16, ReLU activation), Dropout (dropout rate 0.1), and an output layer (16 → 1 or 2 dimensions) to generate predicted values. Finally, based on a PCE or T80 sample dataset, this hybrid model is trained with the objective of minimizing the composite loss function.

[0079] The structure of the quantum branch parameterized quantum circuit (PQC) of the hybrid neural network in this embodiment of the invention is as follows: Figure 5 As shown, this is the core module for extracting high-dimensional quantum features from battery data (such as perovskite solar cells).

[0080] Quantum bit lines are Figure 5 The horizontal lines marked 0-9 on the left represent the evolution path of the qubit, which is the fundamental resource for quantum computing in this invention. This embodiment uses an 8-qubit system (…). Figure 5 The diagram shows bits 0 to 7. Figure 5 This is a general schematic diagram that can be extended to 10 bits according to actual needs to correspond to a parametric rotation module design of 10 bits × 3 gates.

[0081] Parametric revolving doors (RX / RY / RZ) are Figure 5 Within each layer, three consecutive boxes (labeled RX, RY, RZ) are applied to each qubit, serving as Pauli-X, Pauli-Y, and Pauli-Z parameterized rotation gates. Their function is to perform a full-space transformation of the single-qubit quantum state on the Bloch sphere through rotation operations with three degrees of freedom, achieving a nonlinear mapping of the input features. Trainable parameters: For an 8-qubit system, each rotation module contains 24 trainable parameters (8 bits × 3 gates), totaling 72 across 3 layers; for a 10-qubit system, each layer contains 30 trainable parameters, totaling 90 across 3 layers.

[0082] Progressive Entanglement Module (CNOT Gate) is Figure 5The controlled-NOT (CNOT) gate, composed of black dots (control bits) and ellipses (target bits), is used to establish correlations between qubits and achieve quantum entanglement. This invention employs a span-progressive entanglement design, with the three layers of entanglement modules connected as follows: the first-layer CNOT gate has a span of 1 (the...). The first bit controls the... (bits), establishing local associations between adjacent bits; the second layer has a span of 2 (the first bit). The first bit controls the... (bits), establish mid-distance bit associations; the third layer spans 3 (the first bit). The first bit controls the... (bits), establishing long-distance bit associations. This design can effectively capture the long-range nonlinear coupling relationships between features in perovskite data, solving the problem that traditional classical models are difficult to model high-dimensional coupled features.

[0083] The measurement module is Figure 5 The rightmost symbol represents performing Pauli-Z basis measurements on the qubit, mapping the evolution of the quantum state to a classical expectation value, and outputting a quantum eigenvector for subsequent feature fusion with the classical branch.

[0084] In practical applications, this parameterized quantum circuit is a core component of the quantum branch, and its workflow is as follows: Data encoding: The 256-dimensional feature vector of the shared classical layer output is mapped to an initial quantum state of 8 qubits through amplitude encoding; Quantum evolution: The initial quantum state undergoes three layers of evolution, namely "rotation module + entanglement module", to complete the extraction of high-dimensional nonlinear features; Measurement output: The evolved quantum state is measured using the Pauli-Z basis to obtain an 8-dimensional quantum feature vector, which is then mapped to a 16-dimensional feature vector through a fully connected layer. This feature vector is then concatenated and fused with the output features of the classical branch, and finally fed into the regression layer to complete the prediction of PCE or T80.

[0085] The progressively entangled parameterized quantum circuit employed in this invention, compared to traditional fixed-structure quantum circuits, can more efficiently capture the high-dimensional, nonlinear, and strongly coupled characteristic relationships in battery data such as perovskite data, solving the problem of poor generalization ability of classical models when processing small samples and low-quality data. While improving prediction accuracy (increasing R² and reducing RMSE), it significantly reduces the number of model parameters, echoing the technical effect of the quantum hybrid model in this application, which reduces the number of trainable parameters compared to the fully classical model while maintaining superior performance. At the same time, the progressively entangled design effectively balances modeling capability and computational complexity, providing quantum-level technical support for achieving accurate prediction of PCE and T80.

[0086] Building upon this, the amplitude encoding layer in the quantum branch encodes N-dimensional classical eigenvectors. Mapped to The quantum state of a qubit ,in And need (Automatically normalized to the unity norm). In this scheme, the classic layer outputs a 256-dimensional feature vector, corresponding to... One quantum bit.

[0087] The amplitude-encoded quantum state is represented as ,in For the first One calculated ground state ( The 8-bit binary representation corresponds to the state of 8 qubits. High-dimensional classical data is encoded into quantum states using exponential compression, making full use of the superposition property of qubits to store information.

[0088] Here, the parameterized quantum circuit (PQC) in the quantum branch consists of three repeating circuit layers (n_layers=3). Within each layer, a rotation module is executed sequentially, followed by an entanglement module. The rotation module applies a parameterized rotation gate to each qubit, while the entanglement module establishes correlations between qubits using CNOT gates. The connection span of the CNOT gates increases with the circuit layer number: the span is 1 for the first layer, 2 for the second, and 3 for the third, gradually expanding the entanglement range and capturing long-range coupling relationships between features. The total number of circuit parameters is 72 trainable parameters (3 layers × 8 qubits × 3 rotation gates), or 90 for a 10-qubit system. Through this multi-layered structure of alternating rotation and entanglement, the circuit can perform omnidirectional quantum state evolution for each qubit on the Bloch sphere, while simultaneously establishing correlations through inter-qubit entanglement to gradually extract multi-scale deep features from the input data, from local to global.

[0089] Here, performing a projection measurement on the evolved qubit under the Pauli-Z basis, the expected value function of the complete forward output of the quantum branch can be expressed as:

[0090] In the formula: Input feature x, circuit parameters Corresponding global measurement output; The initial vacuum direct product state is 8 qubits. It is its dual state; For amplitude encoding, unitary operator, Its Hermitian conjugate; For the global evolution operator of parameterized quantum circuits, For all trainable rotation parameters, Its Hermitian conjugate; For Pauli-Z basis observation operators, It is the total number of qubits used in this parameterized quantum circuit.

[0091] Based on this, the single-bit subdivision measurement result can be obtained: the expected measurement value of the i-th qubit is ,in The Pauli-Z operator acts on the i-th qubit. The measurement results of the 8 qubits are concatenated to obtain an 8-dimensional expectation vector, which serves as the original output feature of the quantum circuit.

[0092] The expected measurement value of the i-th qubit is Quantum measurement collapses quantum state information into classically readable eigenvectors, thus completing the conversion from quantum information to classical information.

[0093] Here, the quantum post-processing layer in the quantum branch is a fully connected network that maps the 8-dimensional quantum features output from the measurement layer into a 16-dimensional third feature vector, which is then activated by ReLU and output. This 16-dimensional feature vector has the same dimension as the 16-dimensional second feature vector output from the classical branch, satisfying the splicing requirements of the subsequent fusion layer. By adjusting the dimension and representation space of the quantum features through post-processing mapping, it aligns with the classical features in terms of dimension, providing compatibility assurance for the fusion of features from both branches.

[0094] For example, quantum branching can be constructed in the following ways: First, an amplitude encoding layer is constructed. The 256-dimensional feature vector shared by the classical layer output is L2 normalized and then mapped to an 8-qubit quantum state through amplitude encoding. Second, a parameterized quantum circuit is constructed, containing a three-layer rotation-entanglement alternating structure: the rotation module applies RX, RY, and RZ parameterized rotation gates sequentially to each qubit (with a weight matrix of shape (3,8,3)), and the entanglement module adopts a cyclic entanglement structure, using CNOT gates to establish correlations between qubits. The numbering difference between qubits connected by CNOT gates in different circuit layers increases with the layer number (span of 1 for layer 1, 2 for layer 2, and 3 for layer 3). Then, in the measurement layer, the expected values ​​of the Pauli-Z operator are measured for each of the eight evolved qubits to obtain the 8-dimensional quantum circuit output features. Furthermore, the 8-dimensional features are mapped to 16 dimensions through a quantum post-processing fully connected layer, outputting a third feature vector. Finally, this third feature vector is concatenated with the 16-dimensional second feature vector of the classical branch in a fusion layer.

[0095] Based on this, the rotation module includes an RX parameterized rotation gate, an RY parameterized rotation gate, and an RZ parameterized rotation gate that act sequentially on each qubit.

[0096] Here, in each circuit layer of the parameterized quantum circuit, the rotation module sequentially applies three parameterized rotation gates to each qubit: RX(θ1): Rotation angle θ1 around the X-axis of the Bloch sphere, in matrix form. ; RY(θ2): Rotation angle θ2 around the Y-axis of the Bloch sphere, in matrix form. ; RZ(θ3): Rotation angle θ3 around the Z-axis of the Bloch sphere, in matrix form. .

[0097] The three parameters θ1, θ2, and θ3 are all trainable variables. During model training, the gradient is calculated using the parameter shift rule and updated by the optimizer. For an 8-qubit system, each rotating module layer contains 24 trainable parameters (8 bits × 3 gates), for a total of 72 trainable parameters across the three layers. For a 10-qubit system, each layer contains 30 trainable parameters, for a total of 90 trainable parameters across the three layers.

[0098] Among them, the parameter shifting rule is a gradient calculation method for quantum circuits. For the form... The quantum expectation function, whose gradient can be obtained through... The computation does not require backpropagation and is suitable for optimizing quantum circuit parameters.

[0099] The three rotation gates RX, RY, and RZ cover the rotational degrees of freedom around the X, Y, and Z axes on the Bloch sphere, respectively. Their combination enables state transformations of a single qubit at any position on the Bloch sphere, covering all possible quantum states of a single qubit. This ensures complete representation of quantum states while maintaining parameter efficiency.

[0100] In practical applications, firstly, in the rotation module of the parameterized quantum circuit, three parameterized rotation gates (RX, RY, and RZ) are sequentially applied to each qubit. For an 8-qubit system, each rotation module layer contains 24 trainable parameters. Secondly, during training, the gradient of each rotation parameter is calculated using the parameter shift rule: the expected value is measured after shifting the parameter by +π / 2 and -π / 2 respectively, and the gradient of the parameter is obtained by dividing the difference by 2. Then, all rotation parameters are updated on a classical computer using the Adam optimizer. Furthermore, the above process is repeated in three circuit layers, involving a total of 72 rotation parameters in the training. Finally, the trained rotation parameters enable the quantum circuit to adaptively explore the optimal single-qubit quantum state space, achieving efficient quantum feature extraction in conjunction with the entanglement module.

[0101] In some embodiments, when training the initial prediction model in step S106 with the objective of minimizing the composite loss function, the method further includes: Step 4.1: Update the parameters of the variable quantum circuit in the quantum branch using a non-gradient optimization method, and update the parameters of the shared classical layer, classical branch, and fusion layer using a gradient optimization method.

[0102] Here, a hybrid optimization strategy is employed during model training. For the rotation parameters of the variable quantum circuits in the quantum branch, the gradient is calculated using the parameter shifting rule and then optimized and updated (a non-traditional backpropagation optimization method). For the weights and bias parameters of the fully connected layers in the shared classical layer, classical branch, and fusion layer, the gradient is calculated using the standard backpropagation algorithm and updated by the Adam optimizer. This hybrid optimization strategy balances the characteristics of quantum parameters (requiring multiple executions of the quantum circuit to obtain the desired gradient) and classical parameters (which can efficiently calculate the gradient through automatic differentiation), achieving efficient joint training of the quantum-classical hybrid model.

[0103] Specifically, firstly, in each training batch, sample features are input into a shared classical layer for forward propagation to obtain a 256-dimensional first feature vector. Secondly, the first feature vector is fed into both the classical and quantum branches. In the quantum branch, a parameterized quantum circuit is executed using a quantum simulator to obtain the expected measurement value, which is then output as a quantum feature through a quantum post-processing layer. In the classical branch, classical features are output through forward propagation via a fully connected layer. Then, the features are concatenated in the fusion layer to generate predicted values, and the composite loss function is calculated. Furthermore, for the parameters of the classical branch, the shared classical layer, and the fusion layer, gradients are calculated using the automatic differentiation function of frameworks such as PyTorch and updated by the Adam optimizer. For the rotation parameters of the variable quantum circuit in the quantum branch, the gradient of each parameter is calculated using the parameter shifting rule and then updated by the Adam optimizer. Finally, the above process is repeated until training converges, yielding the target prediction model.

[0104] As an optional embodiment, the dropout layer in the fusion layer may also employ a dynamic dropout rate strategy during training, with the dynamic dropout rate decreasing linearly as the training rounds increase; The initial discard rate was between 0.1 and 0.3, and the final discard rate was 0.

[0105] Here, a higher dropout rate (e.g., 0.1-0.3) is used in the early stages of training to enhance regularization and prevent the model from overfitting to noisy data in the early stages. As the training epochs increase, the dropout rate is linearly reduced to 0, allowing the model to fully utilize all neuron connections and learn more refined joint feature representations in the later stages of training. The update formula for the dropout rate is as follows: ,in This is the initial discard rate. The total number of training epochs is represented by epoch, which is the current training epoch number. It represents the real-time activation / discard probability of the target operation when training progresses to the epoch.

[0106] The dynamic dropout rate strategy enhances regularization to prevent overfitting in the early stages of training and retains the full model capacity to improve fitting accuracy in the later stages of training. Compared with a fixed dropout rate, it can more effectively balance generalization ability and fitting accuracy in quantum-classical heterogeneous feature fusion scenarios, and is especially suitable for perovskite small sample datasets.

[0107] Based on the aforementioned embodiments, the a priori physical laws may also include a third quantitative relationship between the tolerance factor and the stability of the perovskite structure; the physical constraint loss term may also contain a third physical constraint loss term that characterizes the degree to which the lifetime prediction value output by the prediction model for samples with different tolerance factors violates the third quantitative relationship.

[0108] Here, in addition to the halide ion ratio-PCE constraint and the metal ion type-lifetime constraint, a priori physical laws governing the relationship between the tolerance factor (Goldschmidt tolerance factor) and the stability of battery-related structures such as perovskite crystals are further introduced. The tolerance factor is known. (in , , The effective ionic radii (A, B, and X sites, respectively) indicate that perovskite structures are stable within the range of 0.8–1.0; beyond this range, the structure is prone to distortion and instability. If the tolerance factor deviates from the ideal range while the model's T80 prediction remains abnormally high, a physical constraint loss occurs.

[0109] By introducing tolerance factor constraints, the intrinsic physical stability laws of perovskite crystal structures are embedded into the training, further constraining the physical rationality of T80 predictions and supplementing the structural stability dimensions not covered by halogen ratio and metal ion constraints.

[0110] In addition, the composite loss function in the aforementioned embodiments may also include a model complexity regularization term, which is used to constrain the number of trainable parameters of the initial prediction model; the model complexity regularization term is calculated based on the deviation between the current number of model parameters and a preset parameter threshold.

[0111] Here, a model complexity regularization term is added to the composite loss function. ,in This represents the number of trainable parameters for the model. Set the preset parameter threshold (e.g., 120,000). This is the complexity weighting coefficient. No penalty is incurred when the number of model parameters is below a threshold, and a loss proportional to the degree of exceeding the threshold is incurred when the number of parameters exceeds the threshold.

[0112] The complexity regularization term of this model guides the training process to favor parameter-efficient model architectures, which in turn creates a positive incentive with the fact that the quantum-classical hybrid model itself has fewer parameters, ensuring that the model remains lightweight while pursuing prediction accuracy.

[0113] In some embodiments, the present invention also provides a prediction method for a battery property prediction model.

[0114] like Figure 4 As shown, it includes the following steps: Step S202: Obtain the feature data of the battery to be predicted.

[0115] The cells to be predicted can include chemical cells and novel material cells such as new perovskite solar cells. Here, for the cell materials to be predicted, full-dimensional feature data is collected, including the A / B / X site ion ratio, tolerance factor, octahedral factor, electron transport layer (ETL) and hole transport layer (HTL) material and process parameters, deposition process parameters, annealing conditions, encapsulation state, and test environment parameters (temperature, humidity, light intensity, ISOS protocol type), etc. This data is then cleaned and normalized according to the same preprocessing procedure as the training data to generate standardized feature vectors to be predicted. Ensuring that the input data format is consistent with the training data guarantees the accuracy and consistency of the model inference.

[0116] Step S204: Input the feature data into the target prediction model trained according to the above method to obtain the battery performance attribute prediction result.

[0117] Here, a pre-trained target prediction model (a complete hybrid network containing a shared classical layer, classical branch, quantum branch, and fusion layer) is loaded. The model undergoes forward propagation, sequentially passing through high-dimensional mapping of the shared classical layer, parallel feature extraction via two branches, concatenation of the fusion layer, and a regression output layer. This ultimately generates predicted values ​​for photoelectric conversion efficiency and long-term stability lifetime. The end-to-end forward inference process requires no human intervention and can output predicted PCE and T80 values ​​for new materials within seconds, providing a rapid and quantitative basis for material selection and process optimization in perovskite solar cells.

[0118] In practical applications, firstly, full-dimensional feature data of the battery to be predicted is acquired, covering the ratio of A / B / X site ions, tolerance factor, ETL / HTL material type, deposition process parameters, annealing conditions, and test environment parameters. Secondly, following the same preprocessing procedure as the training data, the feature data is cleaned (handling missing and outlier values) and normalized (e.g., Min-Max normalization or Z-score standardization). Then, the target prediction model trained by the physical constraint training method in the aforementioned embodiment is loaded. This model contains a complete quantum-classical hybrid network structure. Furthermore, the feature vector to be predicted is input into the model and sequentially undergoes parallel processing through a shared classical layer (original dimension → 256 dimensional → 256 dimensional), a classical branch (256 dimensional → 64 dimensional → 16 dimensional) and a quantum branch (amplitude encoding → parameterized quantum circuit evolution → measurement → post-processing mapping to 16 dimensional), and a fusion layer (splicing 32 dimensional → 16 dimensional → 1 dimensional regression output). Finally, the model outputs predicted values ​​for the photoelectric conversion efficiency and long-term stability lifetime of the new material battery, providing researchers with quantitative reference for material selection and process optimization.

[0119] In some embodiments, such as Figure 6 As shown, an embodiment of the present invention provides a training apparatus for a battery property training model, comprising: Establish a module to build an initial prediction model for predicting battery performance attributes; The module constructs a composite loss function for training the initial prediction model, based on a data-driven loss term and a physical constraint loss term. The training module, based on the battery sample dataset, trains the initial prediction model with the goal of minimizing the composite loss function to obtain the target prediction model.

[0120] In some embodiments, such as Figure 7 As shown, an embodiment of the present invention provides a training device for a battery property prediction model, comprising: Establish a module to build an initial prediction model for predicting battery performance attributes; The module constructs a composite loss function for training the initial prediction model, based on a data-driven loss term and a physical constraint loss term. The training module, based on the battery sample dataset, trains the initial prediction model with the goal of minimizing the composite loss function to obtain the target prediction model.

[0121] In some embodiments, such as Figure 7 As shown, an embodiment of the present invention provides a prediction device for a battery property prediction model, comprising: The acquisition module acquires the feature data of the battery to be predicted; The prediction module inputs the feature data into the target prediction model trained according to the aforementioned method to obtain the battery performance attribute prediction result.

[0122] According to one aspect of the present invention, a battery is provided that uses the prediction device in the above embodiments to predict battery performance attributes.

[0123] This invention does not impose any particular limitation on the shape of the battery; it can be cylindrical, square, or any other arbitrary shape. For example, Figure 8 This is a square-structured battery as an example. It is understood that the arrangement of the terminals (or electrode terminals) is not limited to the same-side positive and negative terminals shown in the figure, and they can also be arranged at two opposite ends of the battery. This invention does not limit this arrangement.

[0124] In some embodiments, the battery includes a casing, within which electrode assemblies are encapsulated. The battery may contain one or more electrode assemblies, which can be selected by those skilled in the art according to specific practical needs.

[0125] In some implementations, batteries can be assembled into battery modules, and battery modules can be assembled into battery packs. The number of battery modules contained in a battery pack can be one or more, and the specific number can be selected by those skilled in the art according to the application and capacity of the battery pack.

[0126] Figure 9 An exemplary structure of a battery pack is shown. Figure 9 As shown, the battery pack includes a housing and multiple battery modules housed within the housing. The housing comprises a lower housing and an upper housing, with the upper housing covering the lower housing and together forming a closed space to accommodate the battery modules. The multiple battery modules can be arranged in any manner within the housing.

[0127] According to another aspect of the present invention, an electrical device is provided, comprising at least one of the above-described solid-state battery, battery module, or battery pack. The solid-state battery, battery module, or battery pack can be used as a power source for the electrical device or as an energy storage unit for the electrical device. The electrical device may include mobile devices (typically mobile phones, laptops, etc.), electric vehicles (typically pure electric vehicles, hybrid electric vehicles, plug-in hybrid electric vehicles, electric bicycles, electric scooters, electric golf carts, electric trucks, etc.), electric trains, ships, satellites, or energy storage systems, etc.

[0128] Figure 10 This is an example of an electrical device. The device could be a pure electric vehicle, a hybrid electric vehicle, or a plug-in hybrid electric vehicle. To meet the high power and high energy density requirements of the secondary battery for this device, a battery pack or battery module can be used.

[0129] It is understood that the specific operation methods of each functional module in this embodiment can be referred to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0130] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program. When the computer program code is executed, it can implement the method described in any of the above embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0131] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0132] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0133] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0134] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A training method for a battery property prediction model, characterized in that, include: Establish an initial prediction model for predicting battery performance attributes; A composite loss function is constructed for training the initial prediction model based on a data-driven loss term and a physical constraint loss term. Based on the battery sample dataset, the initial prediction model is trained with the goal of minimizing the composite loss function to obtain the target prediction model.

2. The method according to claim 1, characterized in that, A composite loss function for training the initial prediction model is constructed based on a data-driven loss term and a physical constraint loss term, including: Using prior physical laws as constraints, calculate the residual of the predicted value output by the initial prediction model based on the input sample relative to the prior physical laws; The physical constraint loss term is constructed from the residuals; The physical constraint loss term and the data-driven loss term are weighted and summed to form the composite loss function.

3. The method according to claim 2, characterized in that, The a priori physical laws include: a first quantitative relationship between the proportion of halide ions in the battery material composition and the photoelectric conversion efficiency, and a second quantitative relationship between the types of metal ions in the battery material composition and the device lifetime; The physical constraint loss term includes a first physical constraint loss term and / or a second physical constraint loss term; The first physical constraint loss term characterizes the predicted photoelectric conversion efficiency value output by the prediction model for samples with different halide ion ratios. The second physical constraint loss term characterizes the lifetime prediction value output by the prediction model for samples of different metal ion types.

4. The method according to claim 1, characterized in that, The method further includes: The physical constraint loss term is embedded as a soft constraint regularization term into the composite loss function, guiding the model parameters to be updated in a direction that satisfies prior physical laws during training.

5. The method according to claim 1, characterized in that, The initial prediction model is a multi-task output model, which includes a first output unit for outputting the predicted value of photoelectric conversion efficiency and a second output unit for outputting the predicted value of device lifetime.

6. The method according to any one of claims 1 to 5, characterized in that, The initial prediction model is built based on a hybrid neural network and includes: A shared classic layer is used to map the original input features to a first feature vector of uniform dimension; The classic branch, connected to the output of the shared classic layer, is used to perform classic feature extraction on the first feature vector to generate a second feature vector; The quantum branch, connected to the output of the shared classical layer, is used to encode the first feature vector into a quantum state and extract the qubits through a variable quantum circuit to generate a third feature vector. The fusion layer, connected to the outputs of the classical branch and the quantum branch, is used to concatenate the second feature vector and the third feature vector, and output the predicted value of the battery performance attribute based on the concatenated joint feature vector.

7. The method according to claim 6, characterized in that, The quantum branch includes: An amplitude coding layer is used to map the first feature vector into a quantum state of multiple qubits through amplitude coding. A parameterized quantum circuit comprises multiple circuit layers consisting of alternating rotational and entangled modules, used to evolve the quantum state; The measurement layer is used to measure the evolved qubits and obtain the output characteristics of the quantum circuit. A quantum post-processing layer is used to map the output features of the quantum circuit to the third feature vector.

8. The method according to claim 6, characterized in that, When training the initial prediction model with the objective of minimizing the composite loss function, the method includes: The parameters of the variable quantum circuit in the quantum branch are updated using a non-gradient optimization method, while the parameters of the shared classical layer, the classical branch, and the fusion layer are updated using a gradient optimization method.

9. A prediction method for a battery attribute prediction model, characterized in that, include: Obtain feature data of the battery to be predicted; The feature data is input into the target prediction model trained by the method according to any one of claims 1 to 8 to obtain the battery performance attribute prediction result.

10. A training device for a battery property prediction model, characterized in that, include: Establish a module to build an initial prediction model for predicting battery performance attributes; The module constructs a composite loss function for training the initial prediction model, based on a data-driven loss term and a physical constraint loss term. The training module, based on the battery sample dataset, trains the initial prediction model with the goal of minimizing the composite loss function to obtain the target prediction model.

11. A prediction device for a battery property prediction model, characterized in that, include: The acquisition module acquires the feature data of the battery to be predicted; The prediction module inputs the feature data into the target prediction model trained by the method according to any one of claims 1 to 8 to obtain the battery performance attribute prediction result.

12. A battery, characterized in that, The battery uses the prediction device described in claim 11 to predict battery performance attributes.

13. A battery pack, characterized in that, Includes the battery as described in claim 12.

14. An electrical appliance, characterized in that, Includes the battery of claim 12 or the battery pack of claim 13.

Images