Method and system for predicting property of attributable lithium ion battery material

Abstract

The invention provides an attributable lithium ion battery material property prediction method and system, and the method comprises the steps: coding a chemical formula of a material, and obtaining an implicit structure feature Z; extracting physical and chemical statistical characteristics of the material to form an artificial characteristic vector X; utilizing the implicit structure feature Z to carry out chemical semantic guided feature-by-feature linear modulation on the artificial feature vector X to obtain an enhanced modulation feature X '; converting the modulation characteristic X'into a sequence form containing [CLS] token; inputting the sequence form into an encoder network to learn a global dependency relationship between features, and generating a global representation vector of the material; and performing classification or regression based on the global representation vector, outputting a material property prediction result, and realizing attributable interpretation of the prediction result in combination with an attention mechanism and gradient analysis. The physical and chemical expression ability of the table features is remarkably enhanced, and the prediction generalization performance is improved.

Description

Technical Field

[0001] This application relates to the fields of materials informatics and machine learning, and specifically to an attributable method and system for predicting the properties of lithium-ion battery materials. Background Technology

[0002] Materials informatics, as a cutting-edge field at the intersection of materials science and artificial intelligence, is reshaping the paradigm of new materials design. In recent years, researchers have been able to achieve rapid prediction and high-throughput screening of material properties at low cost by using machine learning and deep learning models. However, current mainstream machine learning frameworks have significant limitations in modeling material table features.

[0003] Traditional methods for predicting material properties include first-principles calculations based on DFT and MD, as well as semi-empirical models. While these methods offer high accuracy at the atomic scale, they suffer from high computational costs and limited scalability.

[0004] To address this, machine learning models based on tabular data, such as XGBoost, TabPFN, and TabTransformer, have emerged in recent years, which can quickly predict material properties. However, their practical applications face the following core bottlenecks: 1. Lack of consistency between chemical semantics and physics: Traditional tabular features are difficult to reflect atomic-level information, and prediction results often exhibit non-physical behavior.

[0005] 2. One-sidedness and redundancy of feature representation: Existing table features ignore micro-interaction information, are prone to feature drift, and have poor generalization performance for complex systems.

[0006] 3. Strong reliance on manual feature engineering and poor transferability: The tabular model input relies on manual feature design, making it difficult to transfer to different material systems.

[0007] 4. Separation of chemical and structural information: Existing table features do not encode the implicit semantics of chemical formulas, making it difficult to integrate table feature models with chemical embedding models.

[0008] 5. Overfitting and underfitting under small sample data: Experimental data is limited, table features have high dimensionality and redundancy, and traditional machine learning methods are unstable in multi-element complex system tasks.

[0009] To address the aforementioned issues, scholars have proposed various improvement schemes in recent years, including the introduction of attention mechanisms, feature interaction modeling, and multimodal fusion frameworks. However, while some problems have been overcome, these methods still have significant limitations: While tabular Transformer-type models (such as TabTransformer and SAINT) have made some progress in modeling dependencies between features, their attention mechanisms lack the chemical semantic support specific to the materials science field. The latent space they learn remains at the statistical level and fails to reflect real physicochemical laws. Meanwhile, chemical modeling networks (such as CrabNet), although capable of capturing bonding features and electronegativity distributions between elements, cannot be integrated with experimental features or structural parameters, resulting in a disconnect between chemical and tabular information and limiting the model's generalization ability and physical interpretability. Summary of the Invention

[0010] In view of one of the deficiencies in the prior art, the purpose of this application is to provide an attributable method and system for predicting the properties of lithium-ion battery materials.

[0011] A first aspect of this application provides a method for predicting the properties of attributable lithium-ion battery materials, comprising: The chemical formula of the material is encoded to obtain the implicit structural feature Z; Extract the physicochemical statistical characteristics of the material to form an artificial feature vector X; The artificial feature vector X is subjected to chemical semantic-guided feature-by-feature linear modulation using the implicit structural feature Z to obtain the enhanced modulation feature X′. The modulation feature X′ is converted into a sequence containing the [CLS] token; The sequence is input into the encoder network to learn the global dependencies between features and generate a global representation vector of the material. Based on the global representation vector, classification or regression is performed to output the predicted material properties. Attention mechanism and gradient analysis are combined to achieve the attributable explanation of the predicted results.

[0012] Optionally, the step of performing chemical semantic-guided feature-by-feature linear modulation on the artificial feature vector X using the implicit structural feature Z to obtain the enhanced modulation feature X′ includes: The implicit structural features Z Two parallel feedforward neural networks are input separately to generate scaling parameter γ and translation parameter β; Based on the scaling parameter γ and the translation parameter β, an affine transformation is performed on X to obtain the modulation feature X′, specifically: Where ⊙ represents element-wise multiplication. As a learnable scaling factor, it is used to stabilize the initial state and regulate the modulation intensity, enabling dynamic guidance of numerical characteristics by chemical knowledge.

[0013] Optionally, converting the modulation feature X′ into a sequence containing the [CLS] token includes: Assign an independent learnable vector to each feature dimension of the modulation feature X'. And through linear transformation Obtain the feature token. For column index embedding, used to preserve column position information, j is the j-th dimension of the modulation feature X', j ranges from 1 to d, and d represents the dimension of the modulation feature; In all The constructed sequence is prefixed with a learnable [CLS] identifier token to form the final input sequence.

[0014] Optionally, the [CLS] identifier token is a learnable, pre-defined special embedding vector that automatically evolves during training into a semantic representation capable of aggregating global information, used to generate an overall feature expression of the material.

[0015] Optionally, the encoder network is a Transformer encoder, comprising an N-layer stacked self-attention coding structure; Each layer consists of a multi-head attention sublayer and a feedforward network sublayer, wherein: The multi-head attention sublayer includes a multi-head attention unit, a dropout layer, and a summation and normalization layer; the residual of the summation and normalization layer is connected to the output of the previous sublayer. The feedforward network sublayer includes a feedforward unit, a dropout layer, and a summation and normalization layer; the residual of the summation and normalization layer is connected to the output of the previous sublayer. The input to the Transformer encoder is a sequence containing the [CLS] token. After being processed by N layers of Transformer encoders, the features corresponding to [CLS] are extracted, and then passed through pooling layers and linear layers to obtain the final output, which is the global representation vector.

[0016] Optionally, the CrabNet model is used to encode the chemical formula of the material; the encoder network is a Transformer encoder; and the feature-by-feature linear modulation adopts the FiLM mechanism. The CrabNet model, FiLM mechanism, and Transformer encoder constitute the CrabNet-FiLM-Transformer model. This model is trained via end-to-end backpropagation, and its loss function includes a task loss term and a FiLM modulation regularization term. in, Here, γ is the task loss term, β is the scaling parameter, and β is the translation parameter. This is the weight coefficient for the regularization term.

[0017] Optionally, the CrabNet model is based on self-supervised pre-training and has the ability to represent the chemical formulas of materials in a general way, which can effectively capture the interaction rules of elements in unseen combinations.

[0018] Optionally, the step of performing classification or regression based on the global representation vector to output material property prediction results, and combining attention mechanisms and gradient analysis to achieve attributable explanation of the prediction results, includes: The global characterization vector is classified or regressed to output the predicted properties of the target lithium-ion battery material. Based on the difference between the prediction result and the true label, a task loss function is constructed; The gradient of each physicochemical feature dimension in the artificial feature vector X with respect to the predicted output is calculated using the backpropagation algorithm. The response intensity of each feature is weighted using the gradient magnitude to generate a feature importance map or ranking list; Simultaneously, by combining the attention weights in the encoder network, the dependencies and interaction paths between different feature dimensions are analyzed, the most critical feature parameters for prediction are identified, and a multi-level attributable explanation for the model's predictive behavior is achieved.

[0019] Optionally, the implicit structural features include bonding tendency, electronegativity combination, valence state preference and stability perception, the artificial features include average atomic weight, electronegativity difference, density and lattice constant, and the prediction results include band gap, formation energy, ion migration barrier and stability.

[0020] A second aspect of this application provides an attributable lithium-ion battery material property prediction system, comprising: Chemical embedding module: Encodes the chemical formula of the material to obtain implicit structural features Z; Artificial extraction module: Extracts the physicochemical statistical characteristics of the material to form an artificial feature vector X; Feature modulation module: The artificial feature vector X is subjected to chemical semantic-guided feature-by-feature linear modulation using the implicit structural feature Z to obtain the enhanced modulation feature X′; Sequence construction module: Converts the modulation feature X′ into a sequence form containing the [CLS] token; Global modeling module: Inputs the sequence into the encoder network to learn the global dependencies between features and generate a global representation vector of the material; Prediction attribution module: Based on the global representation vector, it performs classification or regression, outputs the predicted material properties, and combines attention mechanism and gradient analysis to realize the attributable explanation of the predicted results.

[0021] The attributable lithium-ion battery material property prediction method provided in this application deeply integrates the extracted chemical embedding vector (i.e., implicit structural features) with tabular features (i.e., artificial feature vectors), introduces a modulation-by-modulation mechanism to enhance the physicochemical expressive power of traditional tabular features, realizes dynamic injection of chemical semantics, and combines an encoder network to establish global dependency modeling across feature dimensions, significantly improving prediction generalization performance.

[0022] Other technical effects resulting from the additional features will be further illustrated in the relevant examples. Attached Figure Description

[0023] Other features, objectives, and advantages of this application will become more apparent from the following detailed description of non-limiting examples with reference to the accompanying drawings: Figure 1 A flowchart illustrating an attributable lithium-ion battery material property prediction method according to an exemplary example; Figure 2 This is a schematic diagram of the structure of an encoder network according to an exemplary example; Figure 3 This is a schematic diagram of the structure of an attributable lithium-ion battery material property prediction system according to an exemplary example. Detailed Implementation

[0024] The present application will be described in detail below with specific examples. These examples will help those skilled in the art to further understand the present application, but do not limit the application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all fall within the scope of protection of the present application. Parts not described in detail in the following examples can be implemented using existing technology.

[0025] In existing technologies, attention mechanisms lack chemical semantic support from the materials science domain, and the learning potential space remains only at the statistical level, failing to reflect real physicochemical laws. Meanwhile, chemical modeling networks (such as CrabNet) can capture elemental bonding features and electronegativity distributions, but cannot integrate them with experimental features or structural parameters, resulting in a separation between chemical and tabular information, limiting the model's generalization ability and physical interpretability. Based on these problems, this application provides an attributable method for predicting the properties of lithium-ion battery materials to address these issues.

[0026] Reference Figure 1 As shown in one example of this application, a method for predicting the properties of lithium-ion battery materials that can be attributed to specific factors is provided, comprising the following steps: S100 encodes the chemical formula of the material to obtain the implicit structural feature Z; Specifically, implicit structural features include bonding tendency, electronegativity combination, valence state preference, and stability perception, etc. S200 extracts the physicochemical statistical characteristics of the material to form an artificial feature vector X.

[0027] Artificial features include average atomic weight, electronegativity difference, density, and lattice constant.

[0028] S300, using implicit structural features Z to perform chemical semantic-guided feature-by-feature linear modulation on artificial feature vector X, to obtain enhanced modulation features X′; S400 converts the modulation feature X′ into a sequence containing the [CLS] token; S500 inputs the sequence into the encoder network to learn the global dependencies between features and generate a global representation vector of the material. S600 performs classification or regression based on global representation vectors, outputs material property prediction results, and combines attention mechanism and gradient analysis to achieve attributable explanation of the prediction results.

[0029] Specifically, the predictions include band gap, formation energy, ion migration barrier, and stability.

[0030] In the above examples of this application, the extracted chemical embedding vectors (i.e., implicit structural features) are deeply integrated with tabular features (i.e., artificial feature vectors), and a modulation-by-modulation mechanism is introduced to enhance the physicochemical expressive power of traditional tabular features, realize the dynamic injection of chemical semantics, and establish global dependency modeling across feature dimensions by combining encoder network, which significantly improves prediction generalization performance.

[0031] To obtain more accurate prior chemical knowledge, in some specific embodiments of this application, S100, the chemical formula of the material is encoded to obtain implicit structural features Z, which can be achieved by the following steps: S101, Determine the material dataset.

[0032] Specifically, material datasets containing chemical formulas and basic table features are collected based on material databases (such as Materials Project, AFLOW, and OQMD). Only the chemical formula and target property are needed, where the target covers a variety of properties such as formation energy, band gap, thermal conductivity, and elastic modulus.

[0033] S102, input the chemical formula of each sample into the pre-trained CrabNet model, and extract its intermediate embedding vector before the final regression layer (ResidualNetwork). This vector can form a latent space that can be used for chemical semantic expression, thereby obtaining the implicit structural features Z of the material.

[0034] Specifically, the implicit structural feature Z contains the elemental bonding characteristics and electronic structure, providing prior chemical knowledge for subsequent fusion. This feature is typically 512-dimensional.

[0035] The above examples in this application construct a multi-task chemical formula-property dataset based on a large-scale materials database and use a pre-trained CrabNet model to extract intermediate embedding vectors, effectively capturing prior chemical knowledge such as elemental composition and bonding rules, and significantly improving the expressive power of implicit features and cross-property generalization performance.

[0036] It should be noted that the CrabNet model is self-supervised pre-trained on large-scale chemical formula data, which has a good ability to encode chemical prior knowledge. It can stably output high-quality implicit structural features Z in small sample scenarios, thereby improving the prediction accuracy and cross-system generalization ability of downstream tasks.

[0037] Similarly, based on the material dataset determined above, in some specific embodiments of this application, S200, extracting the physicochemical statistical characteristics of the material to form an artificial feature vector X can be achieved by the following steps: S201 uses Alphamat to extract handcrafted features (physicochemical statistical features) for each formula, including: atomic fraction, average atomic weight of constituent elements, average and range of atomic radii, average and range of electronegativity, average number of valence electrons in each orbital, and other numerical features. Specifically, Alphamat is an open-source Python toolkit for materials informatics that can quickly and automatically extract a large number of hand-crafted physicochemical features, also known as "tabular features," from the chemical formula of materials (such as LiCoO2).

[0038] S202, X is standardized to eliminate differences in the dimensions of different features, forming a feature matrix that can be used for subsequent fusion.

[0039] In the above examples of this application, multidimensional physicochemical statistical features are extracted using the Alphamat system and standardized to effectively eliminate dimensional differences, construct a structured feature matrix, improve feature comparability and model input stability, and lay a data foundation for subsequent fusion of chemical semantics and tabular features.

[0040] To enhance the physical consistency of feature representation, this application introduces the FiLM (Feature-wise Linear Modulation) mechanism to achieve dynamic injection of chemical semantics. In some specific embodiments of this application, in S300, the artificial feature vector X is subjected to feature-wise linear modulation guided by chemical semantics using implicit structural features Z to obtain the enhanced modulated feature X′. This can be achieved through the following steps: S301, implicit structural features Z Two parallel feedforward neural networks are input to generate scaling parameter γ and translation parameter β. The modulation amplitude is controlled within the range of [-1, 1] by parameter constraints to ensure physical stability.

[0041] Specifically, the two parallel feedforward neural networks refer to two independent small MLPs that share the input Z and output γ and β respectively, which are used to realize dynamic feature reweighting in FiLM modulation and are key to chemical semantic injection table features.

[0042] S302, Perform an affine transformation on X based on the scaling parameter γ and the translation parameter β: Where ⊙ represents element-wise multiplication. As a learnable scaling factor, it is used to stabilize the initial state and regulate the modulation intensity, enabling dynamic guidance of numerical characteristics by chemical knowledge. and The value is determined by Dynamic determination enables the dimensional adjustment of table features by chemical semantics, thereby giving numerical features chemical context sensitivity.

[0043] The above-mentioned example of this application realizes the deep fusion of chemical latent embedding and table features. Through the FiLM modulation mechanism, scaling and offset parameters are dynamically generated using the chemical latent features Z extracted by CrabNet, and feature-by-feature affine transformation is performed on the table features to realize the explicit guidance of chemical semantics on physical parameters.

[0044] To achieve global dependency modeling across feature dimensions, in some specific embodiments of this application, S400, the modulation feature X′ is converted into a sequence containing the [CLS] token; this can be achieved by the following steps: S401, the modulation feature X′ is mapped to a token representation by a column embedding mechanism.

[0045] Specifically, an independent learnable vector W is assigned to each feature dimension of the modulation feature X'. j And through linear transformation of the token j =xj ·W j +E j Obtain the feature token, E j For column index embedding, used to preserve column position information, j is the j-th dimension of the modulation feature X′.

[0046] S402, add a learnable CLS identifier token to the first segment of all ordered token sequences in S401.

[0047] Specifically, the CLS vector itself is a trainable parameter that corresponds to each Token. Each Token has its own CLS, and it is automatically updated during training to better aggregate global information.

[0048] The above examples in this application map modulation features to token sequences and introduce [CLS] tags through column embedding mechanism to complete sequence modeling, effectively preserving feature semantics and positional information, and providing a structured input foundation for Transformer global dependency learning.

[0049] In some specific embodiments of this application, the S500 encoder network employs a Transformer encoder, such as... Figure 2 As shown, it includes an N-layer stacked self-attention encoding structure. Each layer consists of a multi-head attention sublayer and a feed-forward network sublayer. The multi-head attention sublayer includes a multi-head attention unit, a dropout layer, and an add & normalize layer. The residual of the add & normalize layer is connected to the output of the previous sublayer. The feed-forward network sublayer includes a feed-forward unit, a dropout layer, and an add & normalize layer. The residual of the add & normalize layer is connected to the output of the previous sublayer. The input of the Transformer encoder is a sequence containing the [CLS] token. After processing by the N-layer Transformer encoder, the features corresponding to [CLS] are extracted. Then, the features are passed through a pooling layer and a linear layer to obtain the final output, which is the global representation vector.

[0050] The above examples in this application effectively capture nonlinear relationships between features through multi-layer Transformer residual stacking and adaptive attention head design, enhance the deep expressive power of the model, realize the gradual fusion from local interaction to global dependency, and improve the modeling stability and generalization performance under small sample conditions.

[0051] In order to obtain the global representation vector, in some specific embodiments of this application, S500, the sequence is input into the encoder network to learn the global dependencies between features and generate the global representation vector of the material, which can be achieved by the following steps: S501: Obtain the feature token sequence containing the [CLS] identifier and input it into the Transformer encoder; S502: Compute dependencies between features in a multi-head self-attention architecture.

[0052] Specifically, a multi-head self-attention mechanism is used to model the bidirectional interaction between all tokens, enabling each feature to not only perceive its own attributes but also pay attention to other related features (such as the influence of "electronegativity difference" on "lattice constant"), thereby capturing complex nonlinear coupling effects.

[0053] S503: Update the context representation of the [CLS] token using attention weights.

[0054] Specifically, in each layer of self-attention computation, the [CLS] token actively queries (attends to) each feature token, and integrates key information according to its attention weight, gradually building a holistic understanding focused on the prediction task, realizing the transition from local features to global semantics.

[0055] S504: Optimizes global feature representations step-by-step across multiple Transformer stacks.

[0056] Specifically, as the network depth increases, the hidden state of the [CLS] token undergoes multiple residual connections and normalization operations, continuously absorbing high-level semantic information, and completing feature abstraction and aggregation from shallow to deep in the N-layer encoder.

[0057] S505: Output the [CLS] token from the final layer as the global representation vector of the material.

[0058] Specifically, the hidden state corresponding to the [CLS] token in the output of the last Transformer layer is extracted as the final global representation vector. This vector has fused all input features and their interaction history, which can effectively represent the overall characteristics of the material. It is suitable for subsequent classification or regression tasks and supports attribution analysis based on gradients and attention.

[0059] For example: In step S501, it is assumed that a certain material has obtained a modulation feature vector through the previous steps. ′=[ 1′, 2′, 3′, 4′], corresponding to average atomic weight, electronegativity difference, density, and lattice constant, respectively, are used to map each feature dimension to a feature token through a column embedding mechanism. A learnable [CLS] identifier token is added to the beginning of the sequence, where the initial [CLS] token is an independent trainable embedding vector. CLS ∈ It does not contain any specific physicochemical numerical information before being input into the Transformer encoder; In S502, the multi-head self-attention mechanism includes [ CLS , 1, 2, 3, [4] Sequence modeling features and their dependencies; In S503, the [CLS] token serves as a query node, and the representations of each feature token are weighted and aggregated according to attention weights to update its context representation; In S504, as multiple Transformers are stacked, the hidden state of the [CLS] token absorbs higher-level semantic information step by step through residual connections and normalization. Finally, in S505, the hidden state of the [CLS] token output from the last layer is extracted as the global representation vector of the material, which is used for subsequent classification or regression prediction and supports attribution analysis.

[0060] This example introduces a Transformer-based deep coding structure to achieve dynamic aggregation and contextualized modeling of multidimensional physicochemical features of materials. It effectively overcomes the problem that traditional tabular models struggle to depict complex dependencies between features. Furthermore, it constructs a global representation with strong interpretive potential through the [CLS] token mechanism, improving the model's generalization ability, prediction accuracy, and physical consistency in small-sample, multi-element systems.

[0061] In some specific embodiments of this application, S600, based on the global representation vector, classification or regression is performed through a fully connected layer to output the predicted results of the target lithium-ion battery material properties, and the attributable explanation of the predicted results is achieved by combining attention and gradient analysis. This can be achieved through the following steps: S601 performs classification or regression processing on the global characterization vector and outputs the prediction results of the target lithium-ion battery material properties; S602, Based on the difference between the prediction result and the true label, construct the task loss function; S603 uses the backpropagation algorithm to calculate the gradient of each physicochemical feature dimension (such as average atomic weight, electronegativity difference, ionic radius ratio, etc.) in the artificial feature vector X relative to the model output. The gradient reflects the sensitivity of the feature to the prediction result - the larger the absolute value of the gradient, the more significant the small change of the feature has on the output. S604 uses gradient magnitude to weight the response intensity of each feature, generating a visual feature importance map or a feature list sorted by contribution, intuitively showing which input features dominate the current prediction decision; S605, combined with the attention weight matrix generated by the multi-head self-attention mechanism in the Transformer encoder, analyzes the dependencies and interaction paths between different feature dimensions, such as identifying whether "electronegativity difference" strongly focuses on "ionic radius ratio", thereby revealing nonlinear synergistic effects; S606 ultimately integrates gradient significance and attention association patterns to identify the single feature and its combination that have the greatest influence on the prediction results, thereby enabling multi-level and interpretable attribution analysis of the model's predictive behavior.

[0062] In other words, the most influential and critical feature parameters need to be identified through gradient saliency and attention correlation.

[0063] Example: Suppose we predict the performance of a novel cathode material The band gap is .

[0064] To explain this prediction result: Gradient analysis showed that “electronegativity difference” and “ionic radius ratio” had the highest gradient magnitudes, indicating that these two features are most sensitive to band gap prediction. Generate a ranking list based on feature importance, displayed in a table: Rank 1: Poor electronegativity (contribution 45%) 2nd: Ion radius ratio (38%) 3rd: Lattice constant (10%) Attention analysis revealed that in the Transformer multilayer structure, the electronegativity difference tokens consistently pay attention to the ion radius ratio tokens, and the attention weights are higher than average, indicating that the model believes there is a strong coupling relationship between the two.

[0065] Then it can be determined that the band gap of the material is mainly caused by The strength of the bond's ionicity (manifested as a large difference in electronegativity) and the degree of lattice distortion (affected by the ratio of ionic radii) are jointly determined.

[0066] In the above examples of this application, property prediction is achieved through fully connected layers, and by combining gradient backpropagation and attention response weighting, feature importance maps and ranking lists are generated, which effectively reveal the basis of model decision-making and improve the interpretability and scientific credibility of prediction results.

[0067] In some specific embodiments of this application, the CrabNet model is used to encode the chemical formula of the material; the encoder network uses the Transformer encoder; and the feature-by-feature linear modulation uses the FiLM mechanism. The CrabNet model, FiLM mechanism, and Transformer encoder constitute the CrabNet-FiLM-Transformer model. The model is trained through end-to-end backpropagation, and the loss function includes a task loss term and a FiLM modulation regularization term. in, Here, γ is the task loss term, β is the scaling parameter, and β is the translation parameter. These are the weight coefficients for the regularization term. The regularization term is used to suppress overmodulation and prevent the model from deviating from physical laws.

[0068] It is worth noting that the model learns through pre-trained chemical embeddings in few-shot learning. To achieve transfer learning, CrabNet's pre-trained weights are used as prior chemical knowledge to improve the model's generalization performance under sparse data conditions.

[0069] The examples described in this application achieve robust prediction under sparse data. They employ pre-trained CrabNet to provide prior chemical knowledge and combine regularization constraints with a Transformer residual architecture to mitigate overfitting, improving model convergence and stability under small sample conditions. Furthermore, the entire model possesses cross-system generalization capabilities: based on a pretrain-finenetune strategy, it first learns a general chemical representation on large-scale material data and then transfers it to the target system to fine-tune the overall network, achieving knowledge transfer from known materials to new systems.

[0070] Based on the same technical concept, other embodiments of this application provide an attributable lithium-ion battery material property prediction system 100, such as... Figure 3 As shown, it includes: Chemical embedding module 110: Encodes the chemical formula of the material to obtain implicit structural features Z; Artificial Extraction Module 120: Extracts the physicochemical statistical characteristics of the material to form an artificial feature vector X; Feature modulation module 130: Uses implicit structural features Z to perform chemical semantic-guided feature-wise linear modulation on artificial feature vector X to obtain enhanced modulation features X′; Sequence building module 140: Converts modulation feature X′ into a sequence form containing [CLS] token; Global modeling module 150: Inputs the sequence into the encoder network to learn the global dependencies between features and generate a global representation vector of the material; Prediction Attribution Module 160: Based on the global representation vector, it performs classification or regression, outputs the predicted material properties, and combines attention mechanism and gradient analysis to achieve the attributable interpretation of the prediction results.

[0071] The specific implementation techniques of each module / unit in the above examples of this application can be referred to the corresponding steps of the attributable lithium-ion battery material property prediction method in the above examples, and will not be repeated here.

[0072] Each of the preferred features in the above examples can be used individually in any example, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the examples can be implemented using existing technologies.

[0073] To verify the effectiveness of CrabNet-FiLM-Transformer, our model (CrabNet-FiLM-Transformer model) was compared with several mainstream methods, including TabPFN, CatBoost, and XGBoost models. In the classification task of predicting the bandgap of lithium-ion battery electrode materials, CrabNet-FiLM-Transformer significantly improved the model's prediction accuracy, with an ACC improvement of 5% compared to the TabPFN model.

[0074] The terms "comprising" and "having," and any variations thereof, used in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A method for predicting the properties of lithium-ion battery materials that can be attributed to specific factors, characterized in that, include: The chemical formula of the material is encoded to obtain the implicit structural feature Z; Extract the physicochemical statistical characteristics of the material to form an artificial feature vector X; The artificial feature vector X is subjected to chemical semantic-guided feature-by-feature linear modulation using the implicit structural feature Z to obtain the enhanced modulation feature X′. The modulation feature X′ is converted into a sequence containing the [CLS] token; The sequence is input into the encoder network to learn the global dependencies between features and generate a global representation vector of the material. Based on the global representation vector, classification or regression is performed to output the predicted material properties. Attention mechanism and gradient analysis are combined to achieve the attributable explanation of the predicted results.

2. The attributable lithium-ion battery material property prediction method according to claim 1, characterized in that, The step of performing chemical semantic-guided feature-wise linear modulation on the artificial feature vector X using the implicit structural feature Z to obtain the enhanced modulation feature X′ includes: The implicit structural features Z Two parallel feedforward neural networks are input separately to generate scaling parameter γ and translation parameter β; Based on the scaling parameter γ and the translation parameter β, an affine transformation is performed on X to obtain the modulation feature X′, specifically: ； Where ⊙ represents element-wise multiplication. As a learnable scaling factor, it is used to stabilize the initial state and regulate the modulation intensity, enabling dynamic guidance of numerical characteristics by chemical knowledge.

3. The method for predicting the attributable properties of lithium-ion battery materials according to claim 1, characterized in that, The step of converting the modulation feature X′ into a sequence containing the [CLS] token includes: Assign an independent learnable vector to each feature dimension of the modulation feature X′. And through linear transformation Obtain the feature token. For column index embedding, used to preserve column position information, j is the j-th dimension of the modulation feature X′, j ranges from 1 to d, and d represents the dimension of the modulation feature; In all The constructed sequence is prefixed with a learnable [CLS] identifier token to form the final input sequence.

4. The attributable lithium-ion battery material property prediction method according to claim 3, characterized in that, The [CLS] identifier token is a learnable, pre-defined special embedding vector that automatically evolves during training into a semantic representation that can aggregate global information, used to generate the overall feature expression of the material.

5. The method for predicting the attributable properties of lithium-ion battery materials according to claim 1, characterized in that, The encoder network is a Transformer encoder, which includes an N-layer stacked self-attention coding structure; Each layer consists of a multi-head attention sublayer and a feedforward network sublayer, wherein: The multi-head attention sublayer includes a multi-head attention unit, a dropout layer, and a summation and normalization layer; the residual of the summation and normalization layer is connected to the output of the previous sublayer. The feedforward network sublayer includes a feedforward unit, a dropout layer, and a summation and normalization layer; the residual of the summation and normalization layer is connected to the output of the previous sublayer. The input to the Transformer encoder is a sequence containing the [CLS] token. After being processed by N layers of Transformer encoders, the features corresponding to [CLS] are extracted, and then passed through pooling layers and linear layers to obtain the final output, which is the global representation vector.

6. The method for predicting the attributable properties of lithium-ion battery materials according to claim 1, characterized in that, The chemical formula of the material is encoded using the CrabNet model; the encoder network uses a Transformer encoder; and the feature-wise linear modulation uses the FiLM mechanism. The CrabNet model, FiLM mechanism, and Transformer encoder constitute the CrabNet-FiLM-Transformer model. This model is trained via end-to-end backpropagation, and its loss function includes a task loss term and a FiLM modulation regularization term. ； in, Here, γ is the task loss term, β is the scaling parameter, and β is the translation parameter. This is the weight coefficient for the regularization term.

7. The method for predicting the attributable properties of lithium-ion battery materials according to claim 6, characterized in that, The CrabNet model is based on self-supervised pre-training and has the ability to represent the chemical formulas of materials in a general way. It can effectively capture the interaction patterns of elements in unseen combinations.

8. The method for predicting the attributable properties of lithium-ion battery materials according to claim 1, characterized in that, The process of classifying or regressing based on the global representation vector to output material property prediction results, and combining attention mechanisms and gradient analysis to achieve attributable explanation of the prediction results, includes: The global characterization vector is classified or regressed to output the predicted properties of the target lithium-ion battery material. Based on the difference between the prediction results and the true labels, a task loss function is constructed; Based on the task loss function, the gradient of each physicochemical feature dimension in the artificial feature vector X relative to the predicted output is calculated using the backpropagation algorithm. The response intensity of each feature is weighted using the magnitude of the gradient to generate a feature importance map or a ranking list; Simultaneously, by combining the attention weights in the encoder network, the dependencies and interaction paths between different feature dimensions are analyzed, the most critical feature parameters for prediction are identified, and a multi-level attributable explanation for the model's predictive behavior is achieved.

9. The method for predicting the attributable properties of lithium-ion battery materials according to any one of claims 1-8, characterized in that, The implicit structural features include bonding tendency, electronegativity combination, valence state preference and stability perception; the artificial features include average atomic weight, electronegativity difference, density and lattice constant; and the prediction results include band gap, formation energy, ion migration barrier and stability.

10. An attributable lithium-ion battery material property prediction system, characterized in that, include: Chemical embedding module: Encodes the chemical formula of the material to obtain implicit structural features Z; Artificial extraction module: Extracts the physicochemical statistical characteristics of the material to form an artificial feature vector X; Feature modulation module: The artificial feature vector X is subjected to chemical semantic-guided feature-by-feature linear modulation using the implicit structural feature Z to obtain the enhanced modulation feature X′; Sequence construction module: Converts the modulation feature X′ into a sequence form containing the [CLS] token; Global modeling module: Inputs the sequence into the encoder network to learn the global dependencies between features and generate a global representation vector of the material; Prediction attribution module: Based on the global representation vector, it performs classification or regression, outputs the predicted material properties, and combines attention mechanism and gradient analysis to realize the attributable explanation of the predicted results.

Images