Search

CN-122024955-A - Method and system for predicting properties of attributable lithium ion battery material

CN122024955ACN 122024955 ACN122024955 ACN 122024955ACN-122024955-A

Abstract

The application provides a method and a system for predicting the properties of a lithium ion battery material, wherein the method comprises the steps of coding a chemical formula of the material to obtain an implicit structural feature Z, extracting physical and chemical statistical features of the material to form an artificial feature vector X, conducting chemical semantic guided feature-by-feature linear modulation on the artificial feature vector X by utilizing the implicit structural feature Z to obtain an enhanced modulation feature X ', converting the modulation feature X' into a sequence form containing [ CLS ] token, inputting the sequence form into an encoder network to learn global dependency relationship among features to generate a global characterization vector of the material, classifying or regressing based on the global characterization vector, outputting a material property prediction result, and realizing the attributive interpretation of the prediction result by combining an attention mechanism and gradient analysis. The application obviously enhances the physical and chemical expression capacity of the form characteristics and improves the prediction generalization performance.

Inventors

  • ZHANG SHUJIAN
  • WEI ZHIYUN

Assignees

  • 上海灵纭科技有限公司

Dates

Publication Date
20260512
Application Date
20260123

Claims (10)

  1. 1. A method for predicting properties of a attributable lithium ion battery material, comprising: Coding the chemical formula of the material to obtain an implicit structural feature Z; extracting physical and chemical statistical characteristics of the material to form an artificial characteristic vector X; Performing chemical semantic guided feature-by-feature linear modulation on the artificial feature vector X by utilizing the implicit structural feature Z to obtain an enhanced modulation feature X'; Converting said modulation signature X' into a sequence form comprising [ CLS ] token; Inputting the sequence form into an encoder network to learn global dependency relationships among features and generating global characterization vectors of materials; And classifying or regressing based on the global characterization vector, outputting a material property prediction result, and realizing attribution interpretation of the prediction result by combining an attention mechanism and gradient analysis.
  2. 2. The method for predicting properties of a lithium ion battery material according to claim 1, wherein, The feature-by-feature linear modulation of the artificial feature vector X by using the implicit structural feature Z to obtain an enhanced modulation feature X', including: respectively inputting the implicit structural features Z into two parallel feedforward neural networks to generate scaling parameters gamma and translation parameters beta; based on the scaling parameter gamma and the translation parameter beta, affine transformation is carried out on X to obtain a modulation characteristic X', specifically: ; Wherein +.is the multiplication by element, As a learnable scaling coefficient, the method is used for stabilizing an initial state and regulating and controlling modulation intensity, and realizes dynamic guidance of chemical knowledge on the logarithmic characteristics.
  3. 3. The method of claim 1, wherein said converting the modulation signature X' into a sequence form comprising a [ CLS ] token comprises: assigning an independent learnable vector to each feature dimension of the modulation feature X And by linear transformation The feature token is obtained and the feature token is obtained, Embedding a column index for reserving column position information, j being the j-th dimension of a modulation feature X', j from 1 to d, d representing the dimension of an adjustment feature; At all The front end of the formed sequence is added with a learnable [ CLS ] identifier token to form the final input sequence form.
  4. 4. A method of predicting properties of a attributable lithium ion battery material in accordance with claim 3, wherein said [ CLS ] identifier token is a learnable, preset special embedded vector automatically evolved in training into a semantic representation capable of aggregating global information for generating an overall feature representation of the material.
  5. 5. The method of claim 1, wherein the encoder network is a transducer encoder comprising an N-layer stacked self-attention encoding structure; Each layer of structure comprises a multi-head attention sub-layer and a feedforward network sub-layer, wherein: the multi-head attention sub-layer comprises a multi-head attention unit, a discarding layer and an addition and normalization layer, wherein the addition and normalization layer residual is connected with the output of the previous sub-layer; The feedforward network sub-layer comprises a feedforward unit, a discarding layer and an addition and normalization layer, wherein the addition and normalization layer residual is connected with the output of the previous sub-layer; The input of the transducer encoder is in a sequence form comprising [ CLS ] token, the N layers of transducer encoder are used for processing, the characteristics corresponding to the [ CLS ] are extracted, and the final output, namely the global characterization vector, is obtained through a pooling layer and a linear layer in sequence.
  6. 6. The method of claim 1, wherein the chemical formula of the material is encoded by CrabNet model, the encoder network adopts a transducer encoder, the feature-by-feature linear modulation adopts FiLM mechanism; The CrabNet model, fiLM mechanism, and transducer encoder constitute a CrabNet-FiLM-transducer model, which is trained by end-to-end back propagation, and the loss function includes a task loss term and FiLM modulation regularization term: ; Wherein, the For the task loss term, γ is the scaling parameter, β is the translation parameter, Is a regularized term weight coefficient.
  7. 7. The method for predicting properties of a lithium ion battery material according to claim 6, wherein, The CrabNet model is based on self-supervision pre-training, has general characterization capability on the chemical formula of the material, and can effectively capture the element interaction rule in the unseen combination.
  8. 8. The method of claim 1, wherein classifying or regressing based on the global characterization vector, outputting a material property prediction result, and implementing an explanation of the attribution of the prediction result in combination with a mechanism of attention and a gradient analysis, comprises: classifying or regressing the global characterization vector, and outputting a prediction result of the material property of the target lithium ion battery; Constructing a task loss function based on the difference between the prediction result and the real label; Calculating the gradient of each physical and chemical feature dimension in the artificial feature vector X relative to the predicted output through a back propagation algorithm based on the task loss function; weighting the response intensity of each feature by using the amplitude of the gradient to generate a feature importance map or a sequencing list; meanwhile, by combining attention weights in the encoder network, the dependency relationship and interaction paths among different feature dimensions are analyzed, the feature parameters which are most critical to prediction are identified, and multi-level attribution interpretation of model prediction behaviors is realized.
  9. 9. The method of any one of claims 1-8, wherein the implicit structural features include bond formation tendencies, electronegativity combinations, valence state preferences, and stability perception, the artificial features include average atomic weight, electronegativity differences, density, and lattice constants, and the prediction results include band gap, formation energy, ion migration barrier, and stability.
  10. 10. A system for predicting properties of a material for a lithium ion battery, comprising: The chemical embedding module is used for encoding the chemical formula of the material to obtain an implicit structural characteristic Z; the artificial extraction module is used for extracting physical and chemical statistical characteristics of the materials to form an artificial characteristic vector X; The characteristic modulation module is used for carrying out characteristic-by-characteristic linear modulation of chemical semantic guidance on the artificial characteristic vector X by utilizing the implicit structural characteristic Z to obtain an enhanced modulation characteristic X'; the sequence construction module is used for converting the modulation characteristic X' into a sequence form containing [ CLS ] token; the global modeling module inputs the sequence form into an encoder network to learn global dependency relationships among features and generate a global characterization vector of the material; And the prediction attribution module is used for classifying or regressing based on the global characterization vector, outputting a material property prediction result, and realizing attribution interpretation of the prediction result by combining an attention mechanism and gradient analysis.

Description

Method and system for predicting properties of attributable lithium ion battery material Technical Field The application relates to the technical field of material informatics and machine learning, in particular to a method and a system for predicting the properties of a attributable lithium ion battery material. Background Material informatics is used as a leading field of intersection of material science and artificial intelligence, and new material design paradigms are being remodeled. In recent years, researchers can realize rapid prediction of material properties and high-throughput screening with low cost by means of machine learning and deep learning models. However, the current mainstream machine learning framework has significant limitations in modeling material form features. Conventional material property prediction methods include first principles computation methods based on DFT and MD, and semi-empirical models. Although the method has high modeling precision in the atomic scale, the method has the problems of high calculation cost, limited expandability and the like. For this reason, machine learning models based on tabular data, such as XGBoost, tabPFN and TabTransformer, have emerged in recent years, which can rapidly predict material properties, but practical applications have the following core bottlenecks: 1. The method lacks of chemical semantics and physical consistency, the traditional form features are difficult to reflect atomic layer information, and the predicted result often shows non-physical behaviors. 2. The one-sided performance and redundancy of the feature expression are that the existing form features ignore microscopic interaction information, feature drift is easy to occur, and the generalization performance of a complex system is poor. 3. The manual characteristic engineering has strong dependence and poor mobility, the table model input depends on manual characteristic design, and different material systems are difficult to migrate. 4. And the chemical information and the structural information are split, wherein the existing form features are not coded to have chemical hidden semantics, and a form feature model and a chemical embedding model are difficult to fuse. 5. The over-fitting and under-fitting under the small sample data are that experimental data are limited, the characteristic dimension of the table is high, the redundancy is strong, and the traditional machine learning method is unstable in multi-element complex system tasks. In response to the above problems, various improvements have been proposed in recent years, including attention mechanisms, feature interaction modeling, and multimodal fusion framework. However, some of the problems are overcome and these methods still have significant limitations: although the form Transformer model (such as TabTransformer, SAINT) makes a certain progress in the modeling of the inter-feature dependence, the attention mechanism lacks the special chemical semantic support in the material field, the learned latent space still stays at the statistical level, and the true physicochemical rule is difficult to embody. Meanwhile, although the chemical modeling network (such as CrabNet) can capture the bond formation characteristics and electronegativity distribution among elements, the chemical modeling network cannot be fused with experimental characteristics or structural parameters, so that chemical and form information is split, and the generalization capability and physical interpretation of the model are limited. Disclosure of Invention In view of the shortcomings/drawbacks of the prior art, it is an object of the present application to provide a method and system for predicting properties of lithium ion battery materials that are attributable. In a first aspect of the present application, there is provided a method of predicting properties of a lithium ion battery material, comprising: Coding the chemical formula of the material to obtain an implicit structural feature Z; extracting physical and chemical statistical characteristics of the material to form an artificial characteristic vector X; Performing chemical semantic guided feature-by-feature linear modulation on the artificial feature vector X by utilizing the implicit structural feature Z to obtain an enhanced modulation feature X'; Converting said modulation signature X' into a sequence form comprising [ CLS ] token; Inputting the sequence form into an encoder network to learn global dependency relationships among features and generating global characterization vectors of materials; And classifying or regressing based on the global characterization vector, outputting a material property prediction result, and realizing attribution interpretation of the prediction result by combining an attention mechanism and gradient analysis. Optionally, the feature-by-feature linear modulation of the artificial feature vector X by using the implicit structural feature Z to