Method for predicting specific energy and specific capacity of battery material based on RNN (Recurrent Neural Network) and Materials Proiect

Abstract

The invention provides a battery material specific energy and specific capacity prediction method based on RNN and Materials Project, and the method comprises the steps: 1, collecting the related data of a battery material from a Materials Project database, including the chemical components, crystal structure data and physical property data of the material; the physical property data comprises specific energy and specific capacity; data preprocessing is carried out to determine data for training an RNN model, and the data comprises crystal structure data, specific energy and specific capacity; 2, RNN model structure design: designing the number of neurons of an input layer according to the characteristics of crystal structure data; designing the number of hidden units of the hidden layer according to the complexity of the wafer structure; the specific energy and the specific capacity predicted value of the battery material are designed as the output of the output layer; step 3, RNN model training and optimization; and 4, predicting the specific energy and specific capacity of the battery material. According to the method, in the aspect of predicting the specific energy and the specific capacity of the battery material, effective combination of the RNN model and the material project database is realized.

Description

Technical Field

[0001] The present invention belongs to the field of battery material performance prediction, and in particular relates to a method for predicting the specific energy and specific capacity of battery materials based on RNN and Materials Project. Background Art

[0002] As the global demand for clean energy and sustainable development continues to grow, battery technology, as a key energy storage solution, plays a vital role in various fields. Among them, the performance of battery materials directly determines key indicators such as battery energy density, cycle life, and safety. Specific energy and specific capacity, as important parameters for measuring battery material performance, are of great significance for the development of high-performance batteries. Traditional methods for determining the specific energy and specific capacity of battery materials usually rely on experimental testing, which is not only time-consuming and labor-intensive, but also costly. In addition, due to the complexity and diversity of battery materials, it is difficult to comprehensively and quickly explore the impact of various material combinations and structures on performance through experimental methods.

[0003] In this context, using computational methods to predict the specific energy and capacity of battery materials has become an extremely attractive solution. In recent years, with the rapid development of materials science and computer technology, a large number of material databases, such as the Materials Project database, have emerged, providing a rich data resource for battery materials research. At the same time, deep learning technology, particularly recurrent neural network (RNN) models, has achieved tremendous success in fields such as image recognition and natural language processing. Its powerful feature extraction and pattern recognition capabilities provide new ideas and methods for accurately predicting the specific energy and capacity of battery materials.

[0004] However, using RNN models to predict specific energy and capacity for battery materials in the Materials Project database currently faces several challenges. For example, how to effectively extract key features from complex material data, how to select appropriate network structures and hyperparameters to improve prediction accuracy, and how to ensure the model's generalization capabilities all require further research and resolution. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting the specific energy and specific capacity of battery materials based on RNN and Materials Project, which realizes the effective combination of RNN model and materialsproject database in predicting the specific energy and specific capacity of battery materials.

[0006] A method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project, comprising the following steps:

[0007] Step 1: Data collection and preprocessing:

[0008] Collect relevant data on battery materials from the Materials Project database, including: chemical composition, crystal structure data, and physical property data of the materials; physical property data include specific energy and specific capacity;

[0009] After data preprocessing, the data used to train the RNN model is determined, including crystal structure data, specific energy and specific capacity;

[0010] Step 2: RNN model structure design:

[0011] The RNN model consists of an input layer, a hidden layer, a fully connected layer, and an output layer connected in sequence;

[0012] Design the number of neurons in the input layer based on the characteristics of the crystal structure data;

[0013] Design the number of hidden units in the hidden layer according to the complexity of the wafer structure;

[0014] Design the specific energy and specific capacity prediction values ​​of battery materials as the output of the output layer;

[0015] Step 3: RNN model training and optimization;

[0016] Step 4: Predict the specific energy and specific capacity of battery materials.

[0017] Preferably, determining the data for training the RNN model in step 1 includes the following steps:

[0018] Taking into account voltage, current, temperature, time series characteristics, crystal structure data, specific energy and specific capacity;

[0019] Use filtering method, wrapping method or embedding method for feature selection.

[0020] Preferably, step 2 further includes the step of setting hyperparameters, wherein the hyperparameters include hidden layer size, learning rate, sequence length, and batch size.

[0021] Preferably, the step of optimizing the hidden layer size specifically includes: setting an initial range, and then selecting the hidden layer size with the best performance by evaluating the performance of the model under different hidden layer sizes on the validation set.

[0022] The "hidden layer size" refers to the number of hidden units (neurons) in the hidden layer. Hidden units are responsible for feature extraction and transformation of input data.

[0023] The steps for optimizing the hidden layer size are as follows:

[0024] 1. Determine the initial scope:

[0025] Determine based on experience: Refer to similar battery material performance prediction tasks or successful practices of similar deep learning models. For example, if hidden layer sizes in the range of [64, 512] have achieved good results in similar tasks in the past, this range can be set as the initial range.

[0026] Determine based on data characteristics: Analyze the dimensionality and complexity of the input data. If the battery material data is high-dimensional and the relationships between features are complex, the lower limit of the initial range can be appropriately increased.

[0027] 2. Build and train the model:

[0028] Select a range of hidden layer size values ​​within the initial range, such as 64, 128, 256, 512, etc., at equal intervals.

[0029] For each selected hidden layer size, a corresponding neural network model is constructed. The other structures of the model (such as the number of layers and activation functions) remain the same, and only the number of neurons in the hidden layer is changed.

[0030] Use the training set data to train each model, and record key indicators during the training process, such as loss value, training time, etc.

[0031] 3. Evaluate on the validation set:

[0032] After training, the validation set is used to evaluate each model and calculate performance indicators such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R 2 ) etc. These indicators can measure the accuracy of the model in predicting battery material performance.

[0033] Record the performance index value of the model corresponding to each hidden layer size on the validation set.

[0034] 4. Choose the optimal hidden layer size:

[0035] Compare the performance indicators of models with different hidden layer sizes on the validation set and select the one with the best performance (such as the smallest MSE, R 2 The hidden layer size corresponding to the model with the largest value is taken as the final value.

[0036] Preferably, the method for optimizing the sequence length specifically includes: determining the sequence length according to the change in battery material performance. Specifically, it includes the following steps:

[0037] 1. Collect data and analyze performance change patterns:

[0038] Collect performance data for battery materials at different time points or at different charge and discharge cycle times, including voltage, current, temperature, specific energy, and specific capacity. Plot performance indicators over time or cycle number to observe performance fluctuations, trends, and periodicity. For example, analyze whether specific capacity shows a gradual decline with increasing cycle number, and the phased characteristics of this decline.

[0039] 2. Determine the candidate value of sequence length:

[0040] Determined based on the physical process: Based on the physical process of battery charging and discharging, if a complete charge and discharge cycle significantly affects performance, the time step or number of cycles corresponding to a cycle can be used as a candidate value for the basic sequence length. For example, if a complete charge and discharge process includes 100 time points, 100 can be used as one of the candidate values.

[0041] Determine based on the change cycle: If the performance data shows obvious periodic changes, the cycle length can be used as a candidate value. For example, if the battery performance shows a similar change pattern every 50 cycles, 50 can be used as a candidate sequence length.

[0042] Set candidate values ​​of different scales: In addition to the values ​​determined based on physics and cycles, set candidate sequence length values ​​of different scales, such as 20, 50, 100, 200, etc., to cover different time spans or cycle number ranges.

[0043] 3. Build the model and train it:

[0044] For each candidate sequence length, the battery material performance data is segmented or processed according to the length to form a training sample.

[0045] Build a neural network model (the model structure and hyperparameters remain consistent except for the sequence length), train the model using the processed training samples, and record the indicators during the training process.

[0046] 4. Evaluation and selection on the validation set:

[0047] Use the validation set to evaluate the performance of the model under different sequence lengths and calculate relevant indicators.

[0048] The model performance corresponding to each candidate sequence length is compared, and the sequence length that optimizes the model performance (such as minimizing the prediction error) is selected as the final sequence length.

[0049] Preferably, in step 3, the loss function selected is mean square error, and the optimization algorithm is stochastic gradient descent algorithm.

[0050] Compared with the prior art, the advantages of the present invention are:

[0051] 1. Excellent time series data processing capabilities

[0052] During the battery's charge and discharge process, specific energy and specific capacity exhibit complex dynamic trends over time and over the number of cycles. Through its unique recurrent structure, the RNN model processes input data at each time step and passes information from the previous time step to the current time step, effectively "memorizing" and leveraging past information to influence current predictions. This characteristic enables the RNN model to accurately capture subtle changes in specific energy and specific capacity over time when processing battery material performance data, providing a solid foundation for accurate predictions. For example, during the long-term charge and discharge cycles of a battery, the RNN model can reasonably infer changes in specific energy and specific capacity in subsequent cycles based on performance trends in previous cycles, which is crucial for evaluating battery life and stability.

[0053] 2. Effectively capture long-term dependencies

[0054] The performance changes of battery materials often exhibit long-term regularities. For example, the degradation process of a battery can be a gradual and long-term phenomenon. The hidden state mechanism in the RNN model enables it to link information from the distant past with current predictions across multiple time steps. This ability to capture long-term dependencies enables the RNN model to learn the evolution of battery performance throughout its entire life cycle, rather than being limited to short-term changes. Taking lithium-ion batteries as an example, as the number of charge and discharge cycles increases, the structure and performance of the battery materials gradually change. Based on the data characteristics of early cycles, the RNN model can predict the degree of performance degradation that may occur after a large number of future cycles, providing forward-looking guidance for the long-term use and maintenance of the battery.

[0055] 3. Accurately model nonlinear relationships

[0056] The specific energy and specific capacity of battery materials are influenced by a combination of factors, and the relationships between these factors are often nonlinear. The RNN model, through its powerful nonlinear activation functions and complex network structure, can automatically learn and fit these nonlinear relationships. Whether it's the complex interactions between factors such as chemical composition, crystal structure, and physical properties and specific energy and capacity, the RNN model can accurately model them by adjusting the weight matrix and bias terms. For example, changes in the content of trace elements in certain battery materials can have a nonlinear impact on specific energy. The RNN model can capture this subtle relationship and accurately reflect it in the prediction results, providing predictions that are more consistent with reality.

[0057] 4. Excellent adaptability and generalization ability

[0058] During training, the RNN model, by learning from a large amount of data on different battery materials, can summarize general patterns in battery performance. When faced with new, unseen battery material data, the model can leverage the patterns it has learned to make effective predictions. This adaptability and generalization capability gives the RNN model broad application potential in battery material research and development. Whether it's preliminary performance evaluation of new battery materials or consistency testing of material performance across production batches, the RNN model can quickly provide valuable predictions, saving researchers significant experimental time and costs. For example, when exploring new cathode materials, even if some of the material's properties differ from previously studied materials, the RNN model, based on its generalization capabilities, can still make reasonable predictions about its specific energy and capacity, providing guidance for further research.

[0059] 5. High prediction accuracy and computational efficiency

[0060] Furthermore, in practical applications, RNN models, through learning and optimizing large amounts of data, can achieve high accuracy in predicting specific energy and capacity. Furthermore, their computational efficiency is relatively high, enabling them to process large amounts of data and produce predictions within a reasonable timeframe. This makes RNN models not only valuable in laboratory research but also in meeting the demands of real-time monitoring and quality control in industrial production environments. For example, on a battery production line, RNN models can rapidly analyze raw material characteristic data, predict the performance of finished batteries, and promptly identify potential quality issues, ensuring stable and consistent product quality, thereby improving production efficiency and product competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flowchart of the battery material specific energy and specific capacity prediction method based on RNN and Materials Project;

[0062] Figure 2 This is the processing flow chart of the RNN neural network model;

[0063] Figure 3 is a flowchart including steps 1 to 3;

[0064] Figure 4 Schematic diagram of the expansion of the recurrent neural network (RNN). DETAILED DESCRIPTION

[0065] The following is a more detailed description of the battery material specific energy and specific capacity prediction method based on RNN and Materials Project of the present invention, which is combined with a schematic diagram. The preferred embodiment of the present invention is shown, and it should be understood that those skilled in the art can modify the invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being generally known to those skilled in the art and not as a limitation of the present invention.

[0066] like Figures 1 to 4 , a battery material specific energy and specific capacity prediction method based on RNN and Materials Project, including:

[0067] Step 1: Data collection and preprocessing.

[0068] Step 1A: Data collection.

[0069] In order to accurately obtain battery material data from the materials project database, it is first necessary to understand the database's interface specifications and data organization structure.

[0070] Then use the official interface provided by the database or the specially developed data download tool and follow the predetermined process.

[0071] Specifically, you need to visit the official website of the Materials Project database or use its API interface to ensure that the data is complete and accurate.

[0072] Relevant data on battery materials are collected from the Materials Project database, including but not limited to the chemical composition of the materials (element composition and their proportions), crystal structure data (such as unit cell parameters, symmetry, atomic coordinates, etc.), and physical property data (such as specific energy, specific capacity, etc.).

[0073] The battery material data collected from the Materials Project database is mainly used in the following aspects in the entire battery material specific energy and specific capacity prediction process:

[0074] Data preprocessing stage

[0075] Feature selection and extraction: The collected chemical composition and crystal structure data can be used for feature selection through filtering, wrapping, or embedding methods to screen out key features that significantly affect the battery's specific energy and capacity. For example, the unit cell parameters and atomic coordinates in the crystal structure data can be analyzed to extract effective information that reflects the material properties.

[0076] Data conversion and normalization: Physical property data (such as specific energy and specific capacity raw values) need to be normalized and converted to meet the subsequent model input requirements and improve model training effect and stability.

[0077] Model training phase

[0078] Constructing input-output pairs: The processed data is constructed into appropriate input-output pairs as samples for RNN model training. For example, multiple characteristic data of battery materials at different time points (such as voltage, current, crystal structure characteristics, etc.) are organized in sequence as input, and the corresponding specific energy and specific capacity are used as output.

[0079] Guided model learning: The model learns the patterns and relationships in this data and adjusts internal parameters (weights and biases) to grasp the mapping relationship between battery material characteristics and performance, thereby predicting the battery's specific energy and specific capacity.

[0080] Model evaluation and optimization phase

[0081] Validation set evaluation: Use a portion of the collected data as a validation set to evaluate the model's performance under different hyperparameter settings (such as hidden layer size, learning rate, etc.) and at different training stages. For example, calculate indicators such as mean square error and accuracy to determine whether the model has overfitting or underfitting problems.

[0082] Adjust the optimization direction: Based on the evaluation results of the validation set and referring to the actual relationship between material properties and performance in the collected data, make targeted adjustments to the model structure (such as increasing the number of hidden layer neurons) or training parameters (such as adjusting the learning rate) to optimize model performance.

[0083] When downloading crystal structure data of battery materials, ensure that complete and accurate unit cell parameters are obtained, including lattice constants (the lengths in the three directions of a, b, and c, accurate to the nanometer level or even higher, to accurately describe the geometric shape of the crystal), unit cell angles (the three angles α, β, and y, accurate to several decimal places to ensure the accuracy of the crystal symmetry information), and the coordinates of the atoms in the unit cell (recorded in the form of fractional coordinates, accurate to the appropriate decimal places, reflecting the relative positions of the atoms in the unit cell).

[0084] Downloading crystal structure data of battery materials includes the following steps:

[0085] First, you need to register on the Materials Project official website to obtain an API key (API_key).

[0086] Install the mp_api library and run pip install mp_api in the command line.

[0087] Call the Materials Project API through the MPRester class, use the search method to filter out battery material data containing lithium, and extract key information and store it in a pandas data frame. This can be further saved as a file for subsequent analysis and processing, such as use in battery material specific energy and specific capacity prediction models.

[0088] For physical property data, such as specific energy and specific capacity, it is necessary not only to obtain their numerical values, but also to record in detail the units of the data (such as specific energy in Wh / kg or other standard units, specific capacity in mAh / g, etc.) and relevant information such as measurement conditions.

[0089] Specifically, materials databases such as the Materials Project have collected extensive data on battery materials, including physical properties like specific energy and capacity. You can obtain the specific energy and capacity values ​​for the desired battery materials by visiting the database's official website, using its API, or following its prescribed data query and download methods.

[0090] For example, when downloading relevant data on lithium-ion battery positive electrode materials, in addition to obtaining the coordinate information of conventional lithium, cobalt, oxygen and other atoms, it is also necessary to pay attention to whether there are doping elements and their coordinates, and at the same time compare the energy data to clarify under what charge and discharge rates, temperature and other conditions the measurement was obtained.

[0091] Step 1B, pretreatment.

[0092] (1) Remove outliers.

[0093] For crystal structure data, carefully check whether the unit cell parameters conform to physical common sense and material properties. For example, for a material with a specific crystal structure, its lattice constant should be within a known reasonable range. If the lattice constant of a certain data point is found to deviate significantly from this range, further in-depth analysis of the cause is required. This may be due to measurement errors, data entry errors, or special circumstances in the sample itself (such as defective structures, etc.). At this point, we can refer to a large amount of statistical data on similar materials, or make judgments based on theoretical knowledge of crystallography. If it is determined to be an outlier and cannot be corrected in a reasonable way, it will be deleted from the data set.

[0094] For physical property data, such as specific energy and specific capacity, outlier identification is performed using statistical methods such as boxplots and the 3G principle. The boxplot method can intuitively display the data distribution and quickly locate the upper and lower quartiles and the range of possible outliers. The 3σ principle, based on the assumption of normal distribution of data, considers data outside the range of ±3 times the standard deviation of the mean as outliers.

[0095] Comprehensively check the integrity of the data to ensure that each sample contains the necessary information. For example, if the specific capacity data of a battery material is missing, first try to find supplementary information from other relevant records or associated data in the database. If direct supplementary data cannot be found, a reasonable estimate can be made based on the category of the material, the distribution of specific capacity data of materials with similar components, and the physical and chemical principles of the material. In addition, the consistency of the data must be checked, such as whether there are any contradictions in the data of the same material obtained from different sources or different measurement methods. If there are any contradictions, further verification and correction are required.

[0096] (2) Remove erroneous and duplicate data.

[0097] Check data integrity and consistency, and delete erroneous data and duplicate records.

[0098] For example, for chemical composition data, check the rationality of elements and the proportion range to ensure that the element combination conforms to chemical laws.

[0099] For crystal structure data, check the crystal symmetry and lattice parameter rationality to ensure the accuracy of the crystal structure data.

[0100] For physical property data, check the rationality of values ​​and consistency of units to ensure the accuracy and consistency of data.

[0101] (3) Feature extraction and selection, that is, determining the data used to train the RNN model.

[0102] According to the characteristics of battery materials, combined with domain knowledge and data correlation analysis, the most appropriate input features are selected.

[0103] Select features that have an important impact on the performance of battery materials, such as voltage, current, temperature, etc., while considering the impact of historical data on the performance of battery materials and taking time series features into consideration.

[0104] Time series characteristics refer to the characteristics and patterns of data in the time dimension, reflecting the dynamic relationship of data changes over time. Specifically, they include:

[0105] 1. Trend: The overall direction of data changes over time, such as the trend of battery material specific capacity gradually decreasing with increasing usage time.

[0106] 2. Periodicity / Seasonality: Patterns (if any) in the data that repeat at regular intervals, such as the regular fluctuations in battery performance under specific temperature cycles (such as seasonal changes).

[0107] 3. Volatility: The fluctuation of data in a short period of time, such as the instantaneous fluctuation of voltage and current during battery charging and discharging.

[0108] 4. Time dependence: The relationship between the current data point and historical data points. For example, the temperature of a battery at a certain moment is affected by the previous charge and discharge current and duration.

[0109] In battery material performance analysis, time series features encompass the changes in parameters such as voltage, current, temperature, and number of charge and discharge cycles at different time points, as well as the impact of these parameters over time on battery performance, such as specific energy and capacity. By incorporating time series features, the model can capture the dynamic changes and dependencies of data over time, better understanding the evolution of battery performance and thus improving the accuracy of future performance predictions.

[0110] Specifically, scientific methods such as filtering, wrapping and embedding can be used for feature selection to extract the most representative and relevant features and improve the predictive performance of the model.

[0111] In this embodiment, the data used for subsequent training of the RNN model is finally determined to include: crystal structure data and physical property data.

[0112] RNNs (recurrent neural networks) were originally designed to process sequential data, and their input typically requires a time or sequence dimension (e.g., a three-dimensional tensor of [batch size, time step, number of features]). By serializing data (such as time-varying features like voltage, current, and temperature, as well as associated data like crystal structure and physical properties), the model can capture the dependencies and dynamic changes in the data sequence, effectively learning how battery material performance evolves over time or sequence. Without serialization, RNNs struggle to leverage their cyclical mechanisms to exploit the time series characteristics of the data, failing to fully leverage the model's advantages in processing dynamic data. Therefore, this step meets the input requirements and application logic of the RNN model.

[0113] The process of selecting RNN input from six types of data: voltage, current, temperature, time series, crystal structure, and chemical composition is as follows:

[0114] 1. Combining feature selection methods with the characteristics of 6 types of data:

[0115] Filtering method: Calculate the correlation between various data and battery performance (specific energy, specific capacity). For example, voltage, current, and temperature change over time during the charge and discharge process and are dynamically related to battery performance. Highly correlated data can be screened out by calculating the correlation coefficient. Crystal structure data (such as unit cell parameters and atomic coordinates) affects the intrinsic properties of the material and is intrinsically related to battery performance. The degree of its impact on performance is evaluated to determine the selection criteria. Chemical composition (element composition and ratio) is a fundamental property of the material. Analyze its correlation with performance and retain the component characteristics that significantly affect performance.

[0116] Parcel method: With the goal of improving the RNN model's predictive performance (e.g., minimizing mean squared error), the team searches for an optimal subset of features consisting of some or all of the six data types. For example, they try different combinations (e.g., "voltage + current + key crystal structure parameters," "temperature + time series features + key chemical composition elements," etc.), evaluate the performance of the RNN model through training, and select the feature combination that optimizes the model's performance.

[0117] Embedding: During RNN model training, the model's own mechanisms (such as weight updates) are used to automatically determine feature importance. For example, during training, the weights corresponding to certain crystal structure features or voltage variation features in a time series may be significantly updated, indicating that they are important for model prediction and are therefore retained as input features.

[0118] 2. Determine the final input: After the above method, the final RNN input may include:

[0119] The filtered voltage, current and temperature time series data reflect the dynamic working status of the battery at different time points.

[0120] The key features in the crystal structure data (such as unit cell parameters and atomic coordinates that have a great impact on performance) reflect the impact of the inherent properties of the material on performance.

[0121] Characteristics such as the composition and proportion of elements in the chemical composition that have a significant impact on performance.

[0122] Taking the time series characteristics as a whole into consideration, the above data are organized in chronological order, so that RNN can capture the dependencies and changing patterns of data over time, and improve the prediction performance of battery specific energy and capacity.

[0123] In summary, by combining the feature selection method with the six types of data characteristics, the most representative and relevant features are selected as RNN input to optimize the model's predictive ability for battery performance.

[0124] (4) Serialize the data to make it suitable for the input requirements of the RNN model.

[0125] (41) Map the data to the [0.1] interval.

[0126] For crystal structure data, select the appropriate normalization method based on its physical meaning and data distribution characteristics.

[0127] For example, for unit cell parameters, you can compare and analyze them with the common value ranges of similar materials. Suppose the lattice constant a of a certain material has a value range of [3.5, 4.5] nm in most samples, and the a value of the current sample is 4.8 nm. Using the linear formula:

[0128] x new=(xx min )÷(x max -x min )

[0129] where x new is the new value after normalization;

[0130] x is the original value;

[0131] x min and x max are the minimum and maximum values ​​of the feature respectively, mapping x to the corresponding value in the [0.1] interval.

[0132] For physical property data, such as specific energy and specific capacity, the maximum-minimum normalization method can be directly applied. However, before normalization, carefully examine the data for unusually large or small values, as these may affect the normalization effect. If so, further analyze their rationality, such as whether they are due to measurement errors or special experimental conditions. For reasonable but extreme values, consider preprocessing with methods such as logarithmic transformation before normalization to ensure that the data are of the same magnitude and distributed within the [0.1] interval, facilitating effective model learning and calculation.

[0133] (42) Determine the sequence length.

[0134] Determine the appropriate sequence length based on data characteristics and model requirements.

[0135] Available methods include truncation (if the sequence length exceeds the maximum allowed length, you can choose to truncate it from the beginning or end of the sequence) and padding (if the sequence length is less than the minimum allowed length, you can fill the sequence with a specific padding value (such as 0)).

[0136] The way to determine the sequence length is to comprehensively determine it based on the data characteristics (such as the physical meaning and distribution of unit cell parameters, atomic coordinates, etc. of crystal structure data, the changing pattern of physical property data, etc.) and the requirements of the RNN model for sequence input (it is necessary to process data with time or sequence dependencies, and the input is often in the form of a sequence with a fixed time step).

[0137] The requirement of RNN is that the input data is presented in the form of a sequence with a clear time step or sequence order, so as to capture the temporal dependencies and dynamic changes between data.

[0138] The characteristic of crystal structure data is that it contains precise unit cell parameters (lattice constant, unit cell angle, atomic coordinates, etc.), which are used to accurately describe the crystal geometry, symmetry and relative positions of atoms, and are crucial for reflecting the intrinsic properties of the material; the characteristics of physical property data (such as specific energy and specific capacity) are that they have clear units and measurement conditions, and their numerical changes are directly related to battery performance, and attention should be paid to the impact of data outliers on the overall rules.

[0139] (43) Construct a sequence dataset.

[0140] For text data, the text is segmented into words, characters, or subword units (such as BPE, WordPiece, etc.). For time series data, continuous time points need to be organized into sequences.

[0141] Text data: chemical composition data (such as element composition and proportion, element symbols, composition ratios, etc. described in text form), crystal structure data (such as unit cell parameters, atomic coordinates, crystal symmetry, and other descriptive content of the structure).

[0142] Time series data: time-dependent measurements of physical property data (such as voltage, current, temperature at different times, which change continuously in the time dimension to form a time series).

[0143] Chemical composition data and crystal structure data are static descriptions of the inherent properties of materials and do not rely on time series. The measured values ​​of physical properties that change over time (such as voltage, current, temperature, etc.) have continuous change characteristics in the time dimension and can constitute time series data, which is used to capture time-related laws.

[0144] (44) Create input and output pairs.

[0145] For sequence-to-sequence tasks (such as machine translation), both the input and output are sequences; the output is a specific numerical value (i.e., a numerical label) rather than a sequence form (such as the continuous sequence output in sequence-to-sequence tasks).

[0146] For sequence classification tasks such as sentiment analysis, the input is a sequence and the output is a class label.

[0147] (45) Before inputting data into RNN, it is necessary to ensure that the data is presented in the form of a sequence. That is, the sequence dataset is organized into a two-dimensional tensor (or matrix), where each row represents a sequence and each column represents a time step in the sequence.

[0148] When combined with a materials database:

[0149] A sequence is a collection of data arranged in chronological order, used to describe how material characteristics (such as physical properties and structural parameters) change over time. For example, in a materials database, the voltage, current, specific energy, specific capacity, and other data recorded over multiple charge and discharge cycles of a battery material form a sequence, demonstrating the dynamic evolution of the material's performance over time or over the course of an experiment.

[0150] A time step is each time point or time interval in a sequence, representing a specific moment or stage in the sequence. Each time step contains the relevant data characteristics of the material at that moment. For example, the crystal structure parameters (such as unit cell parameters) and physical properties (such as voltage) of a battery material are recorded once every hour. Each recording is a time step, and each time step corresponds to a specific set of material data, reflecting the state of the material at that moment.

[0151] A sequence is a whole arrangement of material data over time, and a time step is a record of material data at each specific time point in the sequence. The combination of the two can be used to analyze how material properties, structure, etc. change over time, providing support for material research and optimization.

[0152] In summary, the above steps include:

[0153] (41) Map the data to the interval [0,1] and normalize them to the same order of magnitude, laying the foundation for subsequent processing.

[0154] (42) After data normalization, the sequence length is determined based on the data characteristics and model requirements, and the time step or structural framework of the data is clarified.

[0155] (43) A sequence dataset is constructed based on a determined sequence length, and the data is organized into a structure that meets the requirements of the sequence model.

[0156] (44) Create input and output pairs based on the sequence dataset to clarify the sample format for model training.

[0157] (45) Finally, before inputting the data into the RNN, ensure that the data is presented in sequence form and complete the final arrangement before data input.

[0158] This process, from data scaling, sequence length determination, dataset construction, input-output pair creation to final data format confirmation, is closely linked and conforms to the logic of data processing and RNN model input preparation, which can effectively provide a suitable data format for model training.

[0159] Step 2: RNN model structure design.

[0160] Step 2A: Design the RNN model structure.

[0161] When designing the RNN recurrent neural network model structure, the design of the input layer should be closely integrated with the characteristics of the crystal structure data.

[0162] If the crystal structure data represents information such as atomic coordinates and unit cell parameters in vector form, the input layer needs to have corresponding reception and processing capabilities.

[0163] For example, if the atomic coordinates are represented by three-dimensional vectors (x, y, z coordinates) and the unit cell parameters are represented by several other numerical values, the number of neurons in the input layer should be determined based on the sum of the dimensions of these vectors.

[0164] Suppose that certain crystal structure data contains atomic coordinates (three-dimensional vectors, such as (x, y, z)) and unit cell parameters (lattice constants a, b, c, unit cell angles \alpha, \beta, \gamma, a total of 6 parameters). At this time, the crystal structure data of each sample can be represented as a 3+6=9-dimensional vector. When designing the input layer, the number of neurons is determined to be 9, and each neuron corresponds to one-dimensional data (such as the first neuron corresponds to the x-coordinate, the second corresponds to the y-coordinate, and so on). In this way, the input layer can directly receive and process information of each dimension of the crystal structure data, enabling the model to effectively learn the impact of atomic coordinates and unit cell parameters on battery performance. This reflects that the input layer design is closely integrated with the characteristics of the crystal structure data (dimensional composition), ensuring that the data is reasonably connected to the model for subsequent calculations.

[0165] The number of neurons in the input layer is determined by the number of features of the input data, that is, the number of neurons in the input layer is equal to the number of features of the data.

[0166] For example, when processing crystal structure data, if atomic coordinates are represented by three-dimensional vectors (x, y, z) (three features), and the unit cell parameters (lattice constants a, b, c, unit cell angles alpha, beta, and gamma) have a total of six features, then the total number of features is 3 + 6 = 9. In this case, the number of neurons in the input layer should be set to 9, with each neuron corresponding to a feature (e.g., the first neuron corresponds to the x-coordinate, the second to the y-coordinate, and so on). This allows the model to receive and process this feature information and learn the relationship between the crystal structure data and battery performance.

[0167] The hidden layer uses an appropriate number of hidden units, a choice that requires comprehensive consideration of multiple factors. Generally speaking, if the data is complex, such as if the crystal structure is complex and the physical properties have complex nonlinear relationships with multiple factors, the number of hidden units can be appropriately increased, such as 64 or 128 hidden units. The hidden layer's computational process strictly adheres to the basic principles of RNNs, processing the temporal dependencies in the input sequence through a recurrent structure. At each time step, the hidden unit receives the current time step data and the hidden state information from the previous time step from the input layer. After calculating the weight matrix and bias term and transforming it through a nonlinear activation function (such as tanh or ReLU), the hidden state for the current time step is obtained.

[0168] The fully connected layer integrates the feature information extracted by the hidden layer and maps it to the dimension of the output layer.

[0169] The output layer is designed to output the predicted values ​​of specific energy and specific capacity of battery materials. For example, two neurons are used to output the predicted results of specific energy and specific capacity respectively, ensuring that the output format matches the actual needs.

[0170] In summary, an RNN recurrent neural network model structure suitable for predicting specific energy and specific capacity of battery materials is constructed, which includes an input layer, a hidden layer, a fully connected layer, and an output layer.

[0171] The input layer is responsible for receiving preprocessed data, performing vectorization processing, and converting the data into a format suitable for model calculation.

[0172] The hidden layer processes the temporal dependencies in the input sequence through the hidden state, remembering the information of the previous time step and passing it to the next time step, thereby adjusting the output of the current time step. The hidden layer of the RNN contains one or more hidden units (or hidden neurons), which are closely connected to the input and output layers through carefully designed weight matrices and bias terms.

[0173] Each hidden unit in the hidden layer receives input from the input layer and the hidden state from the previous time step, and then calculates the hidden state for the current time step.

[0174] At each time step of the RNN, the hidden state is calculated according to a specific formula:

[0175]

[0176] where h t represents the hidden state at time step t;

[0177] σ is a nonlinear activation function (such as tanh, ReLU, etc.). By reasonably selecting the activation function, the nonlinear processing ability of the model is enhanced;

[0178] x t is the input at time step t;

[0179] W hh is the weight matrix from the previous hidden state to the current hidden state;

[0180] h t-1 is the hidden state of the previous time step;

[0181] is the weight matrix input to the hidden state;

[0182] b h is the bias term for the hidden state.

[0183] Figure 4 This is a diagram of a recurrent neural network (RNN). The following are the meanings of the elements in the diagram and the principles of the model:

[0184] Left: The basic unit structure of RNN.

[0185] x represents the input vector; U is the weight matrix input to the hidden layer; S represents the hidden layer state; W is the weight matrix from hidden layer to hidden layer, reflecting the RNN's memory of time series information; V is the weight matrix from hidden layer to output layer; o is the output vector.

[0186] Right: This is the form of RNN after expansion by time step.

[0187] x t-1 、x t 、x t+1 are the inputs at time t-1, t, and t+1 respectively; S t-1 、S t 、S t+1 It is the hidden layer state at the corresponding moment. The hidden layer state at the previous moment will participate in the calculation at the current moment to capture the long-term dependency of sequence data. t-1 、o t 、o t+1 is the output at the corresponding moment. “Unfold” means expanding the loop structure in the time dimension to facilitate understanding of its operation process at different time steps.

[0188] Calculation process:

[0189] At each time step t, the hidden layer state S t The calculation depends on the current input x t and the previous hidden layer state S t-1 , the formula is S t =f(Ux t +WS t-1), where f is the activation function. Output o t By the hidden layer state S t By o t =VS t In this way, RNN can process data with time series characteristics, such as data on battery material performance changing over time.

[0190] Step 2B: Set model hyperparameters.

[0191] In this embodiment, the hidden layer size, learning rate, sequence length, batch size, loss function, and optimization algorithm are 256, 0.001, 32, 64, mean square error loss function (MSELoss), and Adam optimization algorithm, respectively.

[0192] Determining model hyperparameters is a complex but critical process that requires comprehensive consideration of multiple factors and multiple experimental verifications.

[0193] 1) The choice of hidden layer size is closely related to the dimension and complexity of the data.

[0194] If the input data dimension is high (such as containing a large number of crystal structure-related features) and the relationship between the data is complex (such as the complex impact of interactions between different atoms on performance), a larger hidden layer is required to learn and express these complex relationships, but an overly large hidden layer can easily lead to overfitting problems.

[0195] If the performance of battery materials is affected by multiple factors and the relationship is complex, the task complexity is high.

[0196] In this embodiment, an initial range is first set based on experience or reference values ​​of similar tasks, and then the performance of the model under different hidden layer sizes (such as mean square error, accuracy, etc.) is evaluated on the validation set to select the hidden layer size with the best performance.

[0197] The validation set is a portion of data separated from the original dataset. This data can be used to adjust hyperparameters, monitor model training status, and perform preliminary model evaluation.

[0198] The validation set is the same reasonable data used when testing.

[0199] 2) The learning rate is initially set to 0.001 and needs to be dynamically adjusted during training according to the convergence of the model.

[0200] If the model converges too slowly at the beginning of training, you can increase the learning rate appropriately (such as adjusting it to 0.01) to speed up the model's learning speed. If the model is found to be oscillating (the loss function value fluctuates violently within a certain range) or non-converging during training, reduce the learning rate (such as to 0.0001 or even smaller) to allow the model to converge stably.

[0201] In this embodiment, the learning rate that best suits the current task is determined by adopting a strategy of using a larger learning rate for rapid exploration at the beginning of training and then gradually reducing the learning rate as training progresses (such as learning rate decay technology), or by comparing the convergence speed and final performance of the model under different learning rates through multiple experiments.

[0202] 3) The optimization of length sequence should be based on the time scale of battery material performance changes and data characteristics.

[0203] "Timescale" refers to the length of time or period over which changes in battery material performance occur. For example, battery charge-discharge cycle performance data can exhibit long cycles with regular performance changes over extended periods, or rapid data changes with strong short-term dependencies. These time-related characteristics are manifestations of the timescale.

[0204] “Data characteristics” refer to the attributes of battery material-related data, including:

[0205] Data representation and structure: For example, crystal structure data represents atomic coordinates and unit cell parameters in vector form, and the input layer needs to be designed with reception and processing capabilities accordingly.

[0206] The changing pattern of data: such as how fast physical property data (voltage, current, specific energy, etc.) changes over time, whether it is periodic, etc.

[0207] Statistical characteristics of the data: such as whether there are outliers, the discreteness or continuity of the data, volatility (the stability of data changes), etc., which will affect processing methods such as normalization and sequence length selection.

[0208] Dependencies between data: For example, the degree of dependency between previous and subsequent time steps in time series data determines the model’s learning strategy for long-term or short-term dependencies.

[0209] The combination of the two can help optimize model parameters (such as sequence length) so that the model can better learn the relationship between battery material performance and data.

[0210] For example, for charge and discharge cycle performance data, if the cycle period is long and the performance changes show a certain pattern over a long period of time, a longer sequence length (such as 100 time steps or more) can be selected so that the model can fully learn this long-term dependency: if the data changes quickly and the short-term dependency is strong, the sequence length can be appropriately shortened (such as about 30 time steps), while ensuring that the selected sequence length can cover the key information in the data.

[0211] In this example, we conduct in-depth data analysis to understand the characteristic time periods of battery material performance changes (e.g., the periodicity of battery charge and discharge cycle performance). We also experimentally verify the performance of the model at different sequence lengths, taking into account hardware resources (e.g., memory, computing power, etc.) and training time requirements, to determine the optimal sequence length setting. For example, when hardware resources are limited, it may be necessary to appropriately shorten the sequence length while ensuring that critical time-dependent information is not lost.

[0212] The steps to determine the optimal sequence length setting are as follows:

[0213] 1. Analyze the temporal characteristics of data: Conduct in-depth research on the temporal variations of battery material performance, such as the characteristic periods of charge and discharge cycles, the speed of performance changes, and dependencies (long-term or short-term dependencies). This will clarify the temporal distribution of key information in the data and set a preliminary range for the sequence length.

[0214] 2. Evaluate hardware and time constraints: Consider hardware resources (memory, computing power) and training time requirements to determine the acceptable range of sequence lengths. For example, when memory and computing power are limited, prioritize shorter sequences. When training time is tight, avoid selecting overly long sequences to reduce computational time.

[0215] 3. Design multiple comparative experiments: Select multiple representative sequence lengths (such as 30, 50, and 100 time steps), construct datasets and input models for training respectively, and ensure that other training parameters (learning rate, batch size, etc.) remain consistent to isolate the impact of sequence length on model performance.

[0216] 4. Verify model performance: Evaluate the model's performance indicators (such as loss value, accuracy, prediction accuracy, etc.) under different sequence lengths on the validation set, observe the model's ability to capture time-dependent information, and determine whether key information (such as performance mutation points and periodic characteristics) is lost.

[0217] 5. Comprehensive Trade-Offs and Determination: Compare experimental results. If longer sequences significantly improve model performance and the hardware allows, choose longer sequences. If shorter sequences are acceptable and better meet hardware / time constraints, prioritize shorter sequences. Ultimately, determine the optimal sequence length that strikes a balance between information integrity, computational efficiency, and model performance.

[0218] Through the above steps, we can systematically determine the optimal sequence length that is consistent with the data characteristics and adapts to the hardware and training requirements, ensuring that the model effectively learns time dependencies while taking into account training efficiency and performance.

[0219] 4) The choice of batch size should be based on the size of the dataset and the hardware resource configuration.

[0220] If the dataset is large and the hardware memory is sufficient, you can choose a larger batch size (such as 128 or 256), which can improve computing efficiency and reduce training time. If hardware resources are limited (such as small memory), you need to choose a smaller batch size (such as 32 or 64). However, a batch size that is too small may cause the model to update too frequently, affecting convergence stability. Therefore, a trade-off needs to be made between the two.

[0221] In this example, the appropriate batch size is selected based on the dataset size, hardware resources, and model stability requirements. The number of layers is related to the model's depth and abstraction capabilities. Increasing the number of layers allows the model to learn more complex feature representations, but also increases model complexity and training difficulty.

[0222] In summary, the hidden layer size directly affects the expressiveness and complexity of the model. If the hidden layer size is too small, the model may not be able to fully learn the complex relationships in the data, resulting in low prediction accuracy. On the other hand, if the hidden layer size is too large, it is easy to cause overfitting, causing the model to perform well on the training data but reduce its ability to generalize to new data. This also increases training time and computing resource consumption.

[0223] Set the hidden layer size based on task complexity and data size. Generally speaking, more complex tasks require larger hidden layers, but excessively large hidden layers can lead to overfitting and increased training time. Therefore, it's important to strike a balance between task requirements and model performance to find the most appropriate hidden layer size.

[0224] Learning rate: The learning rate is an important parameter that controls the update step size of model parameters. If the learning rate is too large, the model may not converge to the optimal solution during training, or may even diverge. If the learning rate is too small, the training process will be too slow, requiring a large amount of time and computing resources.

[0225] For RNNs and other types of neural networks, the learning rate is typically set between 0.001 and 0.1. However, this range is not absolute and depends on the complexity of the task, the size of the dataset, and the model architecture. Experimentation and analysis are required to determine a learning rate that allows the model to converge quickly and stably.

[0226] Sequence length optimization: Choosing an appropriate sequence length depends on the specific task and dataset characteristics. This may require multiple experiments and validations, comparing model performance with different sequence lengths to find the optimal sequence length setting. Also, factors such as computing resources and training time must be considered to achieve a balance between performance and efficiency.

[0227] Batch size, number of layers: The batch size determines the number of data samples used for each model parameter update. A larger batch size can improve computational efficiency but may slow down model convergence. A smaller batch size increases the frequency of model updates but may introduce more noise.

[0228] Choose the right batch size and number of layers based on the specific task, dataset characteristics, and hardware resources. The batch size affects memory usage and convergence speed during model training, while the number of layers determines the model's depth and expressiveness. Consider these factors comprehensively and choose the most appropriate parameter settings for the task and environment.

[0229] Loss function and optimization algorithm selection: Choose an appropriate loss function to measure the difference between the model's predicted value and the true value. Based on the task type (e.g., regression or classification), select an appropriate loss function, such as mean squared error (for regression) or cross-entropy loss (for classification). Also, choose an effective optimization algorithm to adjust model parameters to minimize the loss function value, such as stochastic gradient descent and its variants (e.g., Adagrad, Adadelta, Adam, etc.), to ensure efficient model learning and optimization.

[0230] In the embodiment, the stochastic gradient descent algorithm SGD and its variants (such as Adagrad, Adadelta, Adam, etc.) are selected. LSTM introduces a gating mechanism and cell state to preserve long-term information, thereby alleviating the gradient vanishing problem. GRU is a simplified version of LSTM with faster training speed and similar performance. These variants have achieved remarkable success in fields such as natural language processing and speech recognition.

[0231] The loss function is used to measure the difference between the model's predicted value and the true value. Common loss functions include mean squared error (MSE), cross entropy loss function, and logarithmic loss function. In the embodiment, the mean squared error is selected as the loss function to minimize the squared error between the model's predicted value and the true value. The optimization algorithm is used to adjust the parameters of the model to minimize it.

[0232] Specifically, different combinations of parameters, such as hidden layer size, learning rate, sequence length, batch size, and number of layers, can be set. The optimal hyperparameter combination can be determined by training the model and evaluating its performance. This invention enables rapid and accurate prediction of the specific energy and capacity of battery materials, providing strong support for the research, development, and design of battery materials.

[0233] Step 3: Model training and optimization.

[0234] Step 3A: Dataset division.

[0235] Dividing the preprocessed data into training and test sets in a ratio of 8:2 is a common and reasonable choice. During the division process, it is necessary to ensure the randomness and representativeness of the data to avoid the impact of uneven data distribution on model evaluation. For example, random sampling can be used to shuffle the data and then distribute it to the training and test sets in proportion. At the same time, to further improve the generalization ability of the model and the accuracy of the evaluation, you can also consider introducing a validation set, such as dividing the dataset into a training set (60%), a validation set (20%), and a test set (20%). The validation set is used to monitor the performance of the model during training, and to adjust the model parameters and training strategy in a timely manner to prevent overfitting.

[0236] The test set is used for the post-training evaluation phase of model training and optimization. After model training, the test set is used to verify the model's generalization and prediction accuracy on unseen data. This ensures that the model has not simply memorized the training data but has truly learned the patterns in the data, thereby objectively evaluating the model's final performance. In this article, the separate test set is included when the dataset is partitioned to prepare for this evaluation phase and ensure the model's reliability in real-world applications.

[0237] In summary, the dataset is divided into training set, validation set and test set to facilitate model evaluation and optimization during the training process.

[0238] Step 3B: Model training.

[0239] When using the training set to train the RNN recurrent neural network model, the stochastic gradient descent optimization algorithm and the mean square error loss function are adopted.

[0240] The stochastic gradient descent algorithm updates the model parameters by randomly selecting a small batch of data samples each time to calculate the gradient, which can speed up the training and avoid falling into the local optimal solution to a certain extent.

[0241] The mean squared error loss function is suitable for regression problems and can accurately measure the difference between the model's predicted value and the true value.

[0242] During training, model parameters are continuously adjusted to gradually fit the model to the data. Measures are taken to prevent overfitting, such as regularization techniques (L1 and L2 regularization) and dropout, to ensure that the model has good generalization capabilities and can accurately predict new data. Grid search, random search, or Bayesian optimization are used to find the optimal hyperparameter combination, continuously optimizing model performance and improving generalization and predictive stability.

[0243] Specifically, based on the calculated gradient direction and magnitude, the weight matrix and bias terms are updated according to the learning rate step size. When the model's loss function value no longer decreases, training is stopped. However, it should be noted that "no longer decreases" here requires comprehensive consideration of multiple factors, such as the trend and magnitude of the loss function value changes over several consecutive training rounds (epochs), to avoid stopping training too early, which may result in the model not reaching the optimal state, or stopping training too late, which may waste computing resources. For example, a threshold can be set. When the decline in the loss function value within five consecutive epochs is less than the threshold, the model is considered to have converged and training is stopped; or when the loss function value fluctuates within a certain range and the fluctuation magnitude is less than the set threshold, it can also be considered to have converged.

[0244] Step 3C: Model evaluation and validation.

[0245] Use the validation set to verify and evaluate the trained model and calculate the model's prediction accuracy and generalization ability.

[0246] Prediction accuracy can be measured by a variety of indicators, such as precision, recall, F1 score (for classification tasks), mean square error (MSE), mean absolute error (MAE), coefficient of determination (R'), etc.

[0247] The mean square error is the average of the squares of the differences between the predicted value and the true value, which can reflect the overall error level of the model prediction;

[0248] The mean absolute error directly calculates the average of the absolute values ​​of the differences between the predicted value and the true value, which more intuitively reflects the size of the prediction error;

[0249] The coefficient of determination R indicates the proportion of the variance explained by the model to the total variance, and its value range is between 0 and 1. The closer it is to 1, the better the model fitting effect.

[0250] Fine-tune model parameters and training strategies based on the evaluation results, such as adjusting the learning rate, increasing or decreasing the number of hidden layers, and optimizing activation functions.

[0251] Common activation functions include ReLU, Sigmoid, and Tanh. The ReLU function has a linear, unsaturated property and can accelerate the model training process; the Sigmoid function can map the output value to the interval (0, 1) and is suitable for binary classification problems; the Tanh function is a variant of the Sigmoid function that can map the output value to the interval (-1.1). In the present invention, either tanh or ReLU can be selected as the activation function according to specific needs.

[0252] By calculating these indicators, we can comprehensively evaluate the performance of the model on new data (test set), determine whether the model can accurately predict the specific energy and capacity of battery materials, and the generalization ability of the model, that is, whether the model can be applied to different types of battery material data, not just the data in the training set.

[0253] To mitigate the vanishing and exploding gradient problem in RNNs, variants such as LSTM or GRU can be used. Techniques such as gradient clipping can also be used to limit the maximum gradient. Strategies such as grid search, random search, or Bayesian optimization can be used to find the optimal hyperparameter combination.

[0254] Step 4: Predict the specific energy and specific capacity of battery materials.

[0255] The battery material data to be predicted is preprocessed and then input into the trained RNN recurrent neural network model. After the forward propagation of the model, the predicted values ​​of specific energy and specific capacity are calculated.

[0256] When inputting the battery material data to be predicted into the trained RNN recurrent neural network model, the predicted data must first undergo the same preprocessing steps as the training data, including data cleaning (removing possible outliers and erroneous data) and serialization processing (mapping the data to the [0.1] interval, etc.) to ensure that the data format and feature distribution are consistent with the training data.

[0257] Here, “battery material data to be predicted” refers to the crystal structure data and physical property data of the battery to be predicted.

[0258] The preprocessed data is then organized and input into the model’s input layer. The model receives data through the input layer, processes temporal dependencies and extracts feature information through the hidden layer, and finally obtains the predicted values ​​of specific energy and specific capacity at the output layer.

[0259] Figure 2 This is a flowchart of the RNN neural network model processing. This is a flowchart of the neural network model processing for battery capacity estimation. The following is an explanation of each step:

[0260] 1. Data collection: Collect battery-related data from various sources, such as battery chemical composition, charge and discharge current, voltage, temperature, usage time and other information. This data is the basis for subsequent analysis.

[0261] 2. Data preprocessing: Cleaning and transforming the collected data. For example, removing outliers, filling missing values, and normalizing or standardizing the data to meet the model input requirements.

[0262] 3. Model construction and parameter initialization.

[0263] 4. Hidden Layer Feature Extraction: The hidden layer of the neural network is used to extract features from the preprocessed data. Nonlinear factors are introduced through activation functions (such as ReLU, or Rectified Linear Unit, with the formula f(x) = max(0, x)), allowing the model to learn more complex relationships in the data.

[0264] 5. Flatten the feature map: Convert the multi-dimensional feature map output by the hidden layer into a one-dimensional vector so that it can be input into the subsequent fully connected layer for processing.

[0265] 6. Fully connected layer processing: Each neuron in the fully connected layer is connected to all neurons in the previous layer, and the flattened features are further processed. The activation function is selected according to the task type, such as the Sigmoid function (f(x) = \frac{1}{1+e^{-x}}) for classification tasks or the Softmax function (for multi-classification, which converts the outputs of multiple neurons into a probability distribution). For regression tasks such as battery capacity estimation, linear activation (i.e., no additional nonlinear transformations) may be used.

[0266] 7. The output layer gives the estimated value of battery capacity: After processing by the previous layers, the output layer outputs the final battery capacity estimation result.

[0267] 8. Evaluation and Adjustment: Compare the model's predicted battery capacity with the actual value to assess the model's accuracy. If the estimate is inaccurate, adjustments may be needed to the neural network structure (e.g., increasing or decreasing the number of hidden layers or neurons) or training parameters (e.g., learning rate, number of iterations).

[0268] For example, for a new lithium-ion battery cathode material, its crystal structure data and related physical property data are pre-processed and then fed into the model. Based on previously learned battery material performance patterns, the model calculates and outputs predicted specific energy and specific capacity values ​​for the material. These predicted values ​​provide an important reference for battery material research and development, design, and performance evaluation, helping researchers quickly screen and optimize battery materials and improve the overall level of battery technology.

[0269] At the same time, the prediction results are analyzed and evaluated to verify the accuracy and reliability of the model.

[0270] The prediction results of different battery materials can be compared and analyzed to determine the applicability of the model.

[0271] The prediction results should be recorded and analyzed, and compared with known similar materials or theoretical expected values ​​to further verify the accuracy and reliability of the model. If the prediction results are significantly different from the expectations, it is necessary to further check whether there are problems in the data quality, model parameter settings, and model structure, so as to make targeted improvements and optimizations.

[0272] For example, the model can compare the difference between the predicted and true values, calculate error metrics (such as mean square error and mean absolute error), and plot the predicted and true values ​​in scatter plots or line graphs for intuitive observation. By conducting a detailed comparative analysis of the prediction results for different battery materials, the model's applicability and limitations can be determined. For example, for battery materials with special structures or compositions, the model's prediction accuracy may be lower, requiring further research and improvement. This analysis provides valuable reference for subsequent model optimization and battery material research and development.

[0273] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing: Collect relevant data on battery materials from the Materials Project database, including: chemical composition, crystal structure data, and physical property data of the materials; physical property data include specific energy and specific capacity; After data preprocessing, the data used to train the RNN model is determined, including crystal structure data, specific energy and specific capacity; Step 2: RNN model structure design: The RNN model consists of an input layer, a hidden layer, a fully connected layer, and an output layer connected in sequence; Design the number of neurons in the input layer based on the characteristics of the crystal structure data; Design the number of hidden units in the hidden layer according to the complexity of the wafer structure; Design the specific energy and specific capacity prediction values ​​of battery materials as the output of the output layer; Step 3: RNN model training and optimization; Step 4: Predict the specific energy and specific capacity of battery materials.

2. The method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project according to claim 1, characterized in that: Step 1 determines the data used to train the RNN model, including the following steps: Taking into account voltage, current, temperature, time series characteristics, crystal structure data, specific energy and specific capacity; Use filtering method, wrapping method or embedding method for feature selection.

3. The method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project according to claim 1, characterized in that: Step 2 also includes the step of setting hyperparameters, which include hidden layer size, learning rate, sequence length, and batch size.

4. The method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project according to claim 3, characterized in that: The steps for optimizing the hidden layer size include setting an initial range, then evaluating the performance of the model under different hidden layer sizes on the validation set, and selecting the hidden layer size with the best performance.

5. The method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project according to claim 3, characterized in that: The method for optimizing the sequence length specifically includes: determining the sequence length according to changes in battery material performance.

6. The method for predicting specific energy and specific capacity of battery materials based on RNN and Materials Project according to claim 1, characterized in that: In step 3, the loss function selected is mean square error, and the optimization algorithm is stochastic gradient descent algorithm.

Images