Customized model construction method for electrolyte coulombic efficiency prediction

Abstract

The invention discloses a customized model construction method for electrolyte coulombic efficiency prediction, and relates to the technical field of new energy batteries, and the method comprises the following steps: obtaining a mole percentage based on an original electrolyte data set, and converting all formula names in an electrolyte to obtain standardized electrolyte component data; pre-training a molecular morphological model by using a chemical corpus, inputting standardized electrolyte component data, and outputting chemical characteristics of each component in the electrolyte; based on the chemical characteristics of each component in the electrolyte, the coulombic efficiency prediction model is generated by using the molar percentage weighted summation and combining the mathematical theory neural network. The electrolyte coulombic efficiency prediction method shows remarkable advantages in the aspect of electrolyte coulombic efficiency prediction, and subtle differences such as geometric arrangement of elements in an isomeride can be captured by comprehensively considering chemical molecular structure information, so that electrolyte characteristics and battery performance are predicted more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of electrolyte prediction and analysis, and in particular to a method for constructing a customized model for predicting electrolyte coulombic efficiency. Background Art

[0002] The discovery and optimization of new energy batteries are crucial to mitigating the impact of climate change and promoting sustainable development, as they have great potential to accelerate decarbonization in all economic sectors. Improving Coulombic Efficiency (CE) is the key to adopting new energy batteries with high energy density. Modern new energy batteries usually use liquid electrolytes composed of one or more organic solvents and salt additives. The formulation has a significant impact on the Coulombic efficiency. However, the combination of electrolyte components and ratios is extremely large, and the search space is about 10 18 , and there is a lack of recognized mathematical analysis solutions. This complexity makes it difficult for traditional experimental methods to fully explore all possible formulas. Chemical researchers have to verify them one by one through time-consuming experiments, which is not only inefficient but also costly. Therefore, in order to significantly reduce the workload of researchers and accelerate the development of new high-performance new energy batteries, it is urgent to develop innovative technologies that can accurately predict the coulombic efficiency in a large number of electrolyte formulas, thereby providing strong support for the rapid iteration and optimization of battery technology and promoting the development of the entire new energy industry.

[0003] Traditional electrolyte Coulombic efficiency prediction technology is mainly based on machine learning methods, which are implemented through supervised learning models such as regression. These models take the selective characteristics of electrolyte components (such as redox potential, dielectric constant, ionic conductivity and elemental composition) as input and output the predicted Coulombic efficiency. In recent years, large language models have made significant progress in understanding complex chemical representations. Currently, more effective technical solutions usually adopt a two-stage framework: first pre-training on a large unlabeled dataset, and then fine-tuning on specific downstream tasks; for example, the MultiModal-MoLFormer uses a large amount of chemical information learned from an unlabeled corpus in the pre-training stage, and allows multiple molecules in the Simplified Molecular Input Line EntrySystem (SMILES) format as input to predict the Coulombic efficiency of electrolytes.

[0004] However, the existing technical solutions still have some limitations. First, the regression supervised learning model needs to intelligently select input features and calculation methods during the prediction process, which may lead to error propagation problems in model predictions, such as ignoring the structural information of chemical molecules, making the model unable to fully capture the key relationship between molecular structure and performance; second, although the multimodal molecular morphology converter can expand the data set by changing the order of electrolyte components, increasing the number of samples from 147 to 27266, the limited diversity may lead to overfitting during training, and the prediction performance of the model still needs to be improved. The root mean square error (RMSE) on the test set is only 0.195, which indicates that the model may have large prediction biases in practical applications; these problems highlight the need to develop a more accurate and robust electrolyte Coulombic efficiency prediction model.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] 1. Technical issues to be resolved In view of the shortcomings of the prior art, the present invention provides a customized model construction method for the prediction of coulombic efficiency of electrolytes, which has the advantages of only requiring limited experimental data to teach the model to produce high-quality outputs and being able to predict coulombic efficiency in a large number of electrolyte formulas, greatly reducing the workload of staff, thereby solving the problems of error propagation in existing model predictions, ignoring the structural information of chemical molecules, and limited sample diversity that may lead to overfitting in the model training process, and the prediction performance of the model still needs to be improved.

[0007] (II) Technical solution In order to achieve the above advantages of only needing limited experimental data to teach the model to produce high-quality output, predicting coulombic efficiency in a large number of electrolyte formulations, and greatly reducing the workload of the staff, the specific technical solutions adopted by the present invention are as follows: A customized model construction method for predicting the coulombic efficiency of electrolytes includes the following steps: S1. Based on the original electrolyte data set, obtain the molar percentage and convert all the formula names in the electrolyte to obtain the standardized electrolyte composition data; S2. Use the chemical corpus to pre-train the molecular morphology model, input the standardized electrolyte composition data, and output the chemical characteristics of each component in the electrolyte; S3. Based on the chemical characteristics of each component in the electrolyte, the weighted sum of mole percentage is used and combined with mathematical theoretical neural network to generate a coulombic efficiency prediction model.

[0008] Furthermore, the original electrolyte dataset includes the solvent volume and salt molarity of each electrolyte, and the standardized electrolyte composition data includes the elemental composition and structural information of all chemical molecules in the electrolyte formula.

[0009] Furthermore, based on the original electrolyte data set, obtaining the mole percentage and converting all the formula names in the electrolyte to obtain the standardized electrolyte composition data includes the following steps: S11. Using the principle of stoichiometry, convert the solvent volume and salt molarity into mole percentage; S12, based on the simplified molecular input linear representation rule, convert all chemical molecule names in the electrolyte formula to obtain mapping format names; S13. Establish standardized electrolyte composition data according to the mole percentage and the mapping format name.

[0010] Furthermore, the chemical corpus includes a public chemical database and a zinc database, and the pre-trained molecular morphology model includes optimizing relative position encoding using a self-attention mechanism.

[0011] Furthermore, using the chemical corpus, pre-training the molecular morphology model, and inputting the standardized electrolyte composition data, outputting the chemical characteristics of each component in the electrolyte includes the following steps: S21. Pre-train the molecular morphogen model using unlabeled molecular sequences in the public chemical database and zinc database; S22, based on the pre-trained molecular morphogen model, inputting standardized electrolyte composition data; S23. According to the input standardized electrolyte composition data, the chemical characteristics of each component in the electrolyte are extracted and output.

[0012] Furthermore, based on the chemical characteristics of each component in the electrolyte, the weighted sum of mole percentages is used and combined with a mathematical theoretical neural network to generate a Coulombic efficiency prediction model, which includes the following steps: S31, generating a characteristic vector of the electrolyte by using a weighted summation method based on the chemical characteristics and molar percentage of each component in the electrolyte; S32, establishing a Coulomb efficiency prediction model based on the characteristic vector of the electrolyte using a mathematical theory neural network; S33. Use pruning technology and function fitting method to optimize the Coulomb efficiency prediction model.

[0013] Furthermore, based on the chemical characteristics and molar percentage of each component in the electrolyte, the expression of the characteristic vector of the electrolyte is generated by using the weighted summation method:

[0014] Where x represents the overall characteristic vector of the electrolyte; ei represents the molecular vector of the i-th component in the electrolyte; n represents the number of electrolyte components; p i It represents the molar percentage of the i-th component in the electrolyte.

[0015] Furthermore, according to the characteristic vector of the electrolyte, a Coulomb efficiency prediction model is established using a mathematical theory neural network, including the following steps: S321, using a linear layer to obtain a dimensionally reduced input vector according to the characteristic vector of the electrolyte; S322. Construct a Coulomb efficiency prediction model based on a multi-layer mathematical theory neural network structure; S323. Based on the Coulomb efficiency prediction model, input the input vector after dimensionality reduction, and output the Coulomb efficiency prediction result.

[0016] Furthermore, based on the Coulomb efficiency prediction model, the input vector after dimensionality reduction is input, and the expression of the output Coulomb efficiency prediction result is:

[0017] Where, CE represents the predicted result of Coulomb efficiency; Ф j represents the function matrix corresponding to the j-th layer of the mathematical theory neural network; L represents the total number of layers of the multi-layer mathematical theory neural network; ° represents function composition, that is, taking the output of one function as the input of the next function; σ1 represents the activation function of the linear layer; W1 represents the weight matrix of the linear layer; x represents the overall eigenvector of the electrolyte.

[0018] Furthermore, using pruning technology and function fitting method, the optimization of the Coulomb efficiency prediction model includes the following steps: S331. Use pruning technology to optimize the network structure of the Coulomb efficiency prediction model; S332. Based on the network structure of the pruned Coulomb efficiency prediction model, a multidimensional continuous function is fitted using a one-variable function combination method; S333. Using logarithmic transformation, adjust the Coulomb efficiency prediction result output by the Coulomb efficiency prediction model.

[0019] (III) Beneficial effects Compared with the prior art, the present invention provides a customized model construction method for predicting the coulombic efficiency of electrolyte, which has the following beneficial effects: (1) The present invention demonstrates significant advantages in predicting the Coulombic efficiency of electrolytes. Compared with traditional machine learning models, the root mean square error of the present invention is reduced by about 80%. This huge improvement is due to its comprehensive consideration of chemical molecular structure information, rather than relying solely on simple features such as elemental composition. This in-depth feature extraction capability enables the present invention to capture subtle differences such as the geometric arrangement of elements in isomers, thereby more accurately predicting electrolyte properties and battery performance.

[0020] (2) In the comparison of deep learning models, the present invention achieved the most advanced performance, with an accuracy improvement of at least 25.7% compared with other models. This outstanding performance is attributed to the combination of multiple innovations, including the use of MoLFormer to accurately represent chemical molecules, the use of formula-level weighted average pooling, and the use of the advantages of KAN networks in the scientific field. It is particularly noteworthy that the present invention can achieve high-precision predictions by fine-tuning only more than one hundred samples, while other existing technologies require a large amount of data amplification. This high efficiency greatly reduces the workload of chemical researchers in experimental verification in the vast electrolyte search space.

[0021] (3) In addition, the present invention further improves the efficiency and generalization ability of the model by pruning the KAN network. The pruning operation reduces the number of neurons in the middle layer from 30 to 17, effectively removing nodes and connections that have little impact on the prediction. This not only simplifies the model structure, reduces the amount of calculation, speeds up the reasoning speed, but also reduces the risk of overfitting while maintaining the high performance of the model. This optimization makes the present invention more efficient and reliable in practical applications, and provides a strong tool support for the prediction and optimization of electrolyte performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 is a schematic flow chart of a method for building a customized model for predicting coulombic efficiency of electrolyte according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the coulombic efficiency prediction model phase of the customized model building method for predicting the coulombic efficiency of electrolyte according to an embodiment of the present invention; Figure 3 2. It is a schematic diagram of a coulombic efficiency prediction model framework of a customized model building method for predicting coulombic efficiency of electrolyte according to an embodiment of the present invention; Figure 4 is a network comparison diagram before and after pruning of a customized model building method for electrolyte coulombic efficiency prediction according to an embodiment of the present invention; Figure 5 is an LCE output distribution diagram of training data based on formula component counting according to a customized model building method for electrolyte coulombic efficiency prediction according to an embodiment of the present invention; Figure 6 It is a specific operational implementation diagram of a customized model building method for predicting coulombic efficiency of electrolyte according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0025] According to an embodiment of the present invention, a customized model building method for predicting the coulombic efficiency of electrolyte is provided.

[0026] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, the method for building a customized model for predicting the coulombic efficiency of electrolyte according to an embodiment of the present invention includes the following steps: S1. Based on the original electrolyte data set, obtain the molar percentage and convert all the formula names in the electrolyte to obtain the standardized electrolyte composition data; S2. Use the chemical corpus to pre-train the molecular morphology model, input the standardized electrolyte composition data, and output the chemical characteristics of each component in the electrolyte; S3. Based on the chemical characteristics of each component in the electrolyte, the weighted sum of mole percentage is used and combined with mathematical theoretical neural network to generate a coulombic efficiency prediction model.

[0027] Specifically, Figure 2 As shown in the figure, the Coulombic efficiency prediction model consists of two stages, the pre-training of the general chemical model and the fine-tuning of the downstream field data. The construction of the model includes three steps: (1) data processing; (2) pre-training based on the chemical corpus; (3) inputting the pre-training results into the mathematical theoretical neural network and completing the model training.

[0028] Specifically, first, a molecular formulator model is used as an encoder. In this embodiment, the open source MoLFormer is used as an encoder, which obtains molecular-level embedding by pre-training an efficient Transformer encoder on a large unlabeled chemical corpus. The input of MoLFormer conforms to SMILES (Simplified Molecular Input Line Entry System), which can simultaneously represent the elemental composition and structural information of chemical molecules. In the Coulomb efficiency prediction model constructed by the present invention, all solvents and salts in the electrolyte are independently input into MoLFormer to extract the chemical characteristics of each component. Then, the labeled embeddings of all SMILES molecules are multiplied by the corresponding proportions and weighted summed to obtain the characteristic vector of the electrolyte; finally, the characteristic vector of the electrolyte is input into the mathematical theory neural network (Kolmogorov-Arnold Networks, KAN) to predict the Coulomb efficiency.

[0029] It should be noted that KAN (Kolmogorov-Arnold Networks) is a neural network architecture based on mathematical theory. It aims to approximate multi-dimensional input functions by decomposing complex functions. It is suitable for regression of high-dimensional data and modeling of complex systems. Compared with MLP (Multilayer Perceptron), it has learnable edge activation functions and node summation operations. Our model further proves that in the scientific field, almost all knowledge in language models is learned during pre-training, and only limited experimental data is needed to teach the model to produce high-quality output.

[0030] In one embodiment, the original electrolyte dataset includes the solvent volume and salt molarity of each electrolyte, and the standardized electrolyte composition data includes the elemental composition and structural information of all chemical molecules in the electrolyte formula.

[0031] In one embodiment, based on the original electrolyte data set, obtaining the mole percentage and converting all the formula names in the electrolyte to obtain standardized electrolyte composition data includes the following steps: S11. Using the principle of stoichiometry, convert the solvent volume and salt molarity into mole percentage; S12, based on the simplified molecular input linear representation rule, convert all chemical molecule names in the electrolyte formula to obtain mapping format names; S13. Establish standardized electrolyte composition data according to the mole percentage and the mapping format name.

[0032] Specifically, first, in the dataset, we convert the molar concentrations of solvents and salts into molar ratios, and map the names of chemical molecules into SMILES format.

[0033] Specifically, the dataset based on lithium / copper half-cell used in the present invention includes the solvent volume and salt molarity of each electrolyte, emphasizing the role of elemental composition in predicting coulombic efficiency (CE). Due to the error in the approximate value of mass ratio or volume ratio, the present invention converts the molar concentration (mol / L) into molar percentage (%) as the input of the coulombic efficiency prediction model (CEAK). In addition, we also convert the names of all formulas in the electrolyte into SMILES format, making it easier for MoLFormer to output its corresponding embedding.

[0034] Specifically, Figure 2 As shown in the figure, the name of each component in the electrolyte (2.54M LiFSI DME-BTFE (1:4w / w)) is converted into SMILES format (in this example, DME is converted to COCCOC), and the molarity is also converted to molar ratio (in this example, 2.77mol / L is converted to 25.67%). This data processing method can better capture the relationship between the solvent and salt in the electrolyte, thereby improving the ability of the model to predict CE.

[0035] In one embodiment, the chemical corpus includes a public chemical database and a zinc database, and the pre-trained molecular morphology model includes optimizing relative position encoding using a self-attention mechanism.

[0036] In one embodiment, using a chemical corpus, pre-training a molecular morphology model, and inputting standardized electrolyte composition data, outputting the chemical characteristics of each component in the electrolyte includes the following steps: S21. Pre-train the molecular morphogen model using unlabeled molecular sequences in the public chemical database and zinc database; Specifically, the chemical corpus includes a public chemical database and a zinc database. In this embodiment, 10% of the Zinc database and 10% of the PubChem database are selected.

[0037] It should be noted that the Zinc database is a free compound database containing more than millions of commercially available compounds. These compounds can be directly used for virtual screening and drug discovery research; the PubChem database is a public chemical database maintained by the National Center for Biotechnology Information under the National Institutes of Health, which contains a very wide range of compound information.

[0038] S22, based on the pre-trained molecular morphogen model, inputting standardized electrolyte composition data; S23. According to the input standardized electrolyte composition data, the chemical characteristics of each component in the electrolyte are extracted and output.

[0039] Specifically, the present invention adopts the molecular formulator model as the basic model. In this embodiment, the open source MoLFormer is used, which is a transformation encoder widely used in chemical language representation. Then the basic model is pre-trained on SMILES sequences of 1.1 billion unlabeled molecules from PubChem and ZINC datasets; Figure 3 As shown, MoLFormer is a large-scale masked language model that processes inputs through a series of blocks, alternating between self-attention and feed-forward connections. Unlike most Transformer architectures, MoLFormer optimizes relative encoding by using an improved version of the RoFormer attention mechanism. In the above embodiment, the relative position embedding introduced by rotation enables the model to effectively capture position information, thereby improving the performance of molecular property prediction. During the model construction process, all solvents and salts in the electrolyte represented by SMILES are independently input into MoLFormer.

[0040] In one embodiment, based on the chemical characteristics of each component in the electrolyte, using mole percentage weighted summation and combining mathematical theory neural network, generating a coulombic efficiency prediction model includes the following steps: S31, generating a characteristic vector of the electrolyte by using a weighted summation method based on the chemical characteristics and molar percentage of each component in the electrolyte; S32, establishing a Coulomb efficiency prediction model based on the characteristic vector of the electrolyte using a mathematical theory neural network; S33. Use pruning technology and function fitting method to optimize the Coulomb efficiency prediction model.

[0041] In one embodiment, based on the chemical characteristics and molar percentage of each component in the electrolyte, a weighted summation method is used to generate an expression for a characteristic vector of the electrolyte:

[0042] Where x represents the overall characteristic vector of the electrolyte; e i represents the molecular vector of the i-th component in the electrolyte; n represents the number of electrolyte components; p i It represents the molar percentage of the i-th component in the electrolyte.

[0043] Specifically, in order to obtain the characteristics of the electrolyte formula input into the model, the present invention converts the chemical characteristics of each component in the above electrolyte into i Multiply by its mole percentage p i And sum them up, the resulting eigenvector is shown in the above expression.

[0044] In one embodiment, according to the characteristic vector of the electrolyte, using a mathematical theory neural network, establishing a Coulomb efficiency prediction model includes the following steps: S321, using a linear layer to obtain a dimensionally reduced input vector according to the characteristic vector of the electrolyte; S322. Construct a Coulomb efficiency prediction model based on a multi-layer mathematical theory neural network structure; S323. Based on the Coulomb efficiency prediction model, input the input vector after dimensionality reduction, and output the Coulomb efficiency prediction result.

[0045] Specifically, the present invention adopts an efficient and interpretable KAN network to replace MLP, and its learnable activation function on the edge and the sum operation on the edge enable it to have a more powerful feature extraction capability, which can capture more complex and deeper patterns in the data, rather than just a simple linear or nonlinear combination like MLP.

[0046] Specifically, because it is necessary to reduce the dimension of the feature vector embedding, thereby simplifying the structure of the KAN network and making parameter adjustment easier, the present invention introduces a linear layer. In this embodiment, the feature vector x of the electrolyte is input into the L-layer KAN to predict the coulomb efficiency CE, and the expression is shown in the expression for outputting the coulomb efficiency prediction result.

[0047] In one embodiment, based on the Coulomb efficiency prediction model, the input vector after dimensionality reduction is input, and the expression for outputting the Coulomb efficiency prediction result is:

[0048] Where, CE represents the predicted result of Coulomb efficiency; Ф j represents the function matrix corresponding to the j-th layer of the mathematical theory neural network; L represents the total number of layers of the multi-layer mathematical theory neural network; ° represents function composition, that is, taking the output of one function as the input of the next function; σ1 represents the activation function of the linear layer; W1 represents the weight matrix of the linear layer; x represents the overall eigenvector of the electrolyte.

[0049] In one embodiment, optimizing the Coulomb efficiency prediction model using pruning technology and function fitting method includes the following steps: S331. Use pruning technology to optimize the network structure of the Coulomb efficiency prediction model; Specifically, since the KAN network decomposes a multivariate function into a combination of a set of unary functions, and some unimportant nodes or connections that contribute little to the final result may appear in practical applications, the present invention adopts pruning to reduce these redundant parts.

[0050] Specifically, the present invention removes nodes or connections that have little impact on model prediction through pruning operations, thereby simplifying the model structure, reducing the amount of calculation, accelerating the reasoning speed, and reducing the risk of overfitting without significantly affecting the performance of the model.

[0051] S332. Based on the network structure of the pruned Coulomb efficiency prediction model, a multidimensional continuous function is fitted using a one-variable function combination method; Specifically, KAN can effectively capture the complexity and nonlinear characteristics of functions by decomposing complex multi-dimensional inputs into simple unary functions and combining them layer by layer, providing powerful expression capabilities, thereby achieving accurate fitting of multi-dimensional continuous functions; the present invention further uses KAN for function fitting, which improves the interpretability of the model, thereby finding general rules for electrolyte CE prediction.

[0052] S333. Using logarithmic transformation, adjust the Coulomb efficiency prediction result output by the Coulomb efficiency prediction model.

[0053] Specifically, in order to numerically amplify the output change caused by the change in electrolyte, the present invention converts the coulomb efficiency CE into a logarithmic format, and the expression is LCE=-log(1-CE); wherein LCE represents the predicted result of the coulomb efficiency of the adjusted output; CE represents the predicted result of the coulomb efficiency.

[0054] In order to facilitate understanding of the above technical solution of the present invention, a data set based on a lithium / copper half-cell is taken as an example for specific description as follows: The framework of the Coulomb efficiency prediction model is as follows Figure 3 As shown in the figure, 2.54M LiFSI DME-BTFE (1:4w / w) describes a lithium-ion battery electrolyte with a specific formula, that is, each liter of solution contains 2.54 moles of LiFSI as a lithium salt, and the solvent is a mixture of DME and BTFE in a mass ratio of 1:4; the PoolingLayer in the figure performs a flattening and summing operation, multiplying each molecule embedding vector by its molar percentage and summing them; the Kolmogorov-ArnoldNetwork (KAN) in the figure is used to predict the Coulombic efficiency of the electrolyte.

[0055] The proposed method is compared with some strong baseline methods, which are divided into two groups: machine learning and deep learning. We conducted comparative experiments to evaluate the effectiveness of each method in the task of predicting Coulomb efficiency. The root mean square error (RMSE) results of all models in the dataset are shown in Table 1.

[0056] Table 1

[0057] First, it can be seen that the method of the present invention performs significantly better than the machine learning model on the data set. Compared with linear regression, random forest, boosting and bagging, the root mean square error of CEAK in predicting CE is reduced by 81.2%, 80.9%, 81.3% and 81.1% respectively; the performance of these methods is limited by whether the feature selection is appropriate and complete. For example, the existing technology uses the elemental composition of the liquid electrolyte as an input feature. This technology successfully reveals that reducing the oxygen content in the solvent is crucial to obtaining a higher CE, but ignores the structural information of the chemical molecules (such as the geometric arrangement of elements in isomers), and this subtle difference has been shown to affect the electrolyte properties and battery performance.

[0058] Secondly, it can be found that among all deep learning models on the same data set, the method of the present invention achieves SOTA performance in the task of predicting the coulombic efficiency of electrolytes. In particular, compared with other models, the model constructed by the present invention has at least 25.7% SOTA performance improvement. Although the present invention uses a two-stage registration framework, it benefits from the precise representation of chemical molecules by MoLFormer, the weighted average pooling at the recipe level, and the performance advantages of KAN in the scientific field; in addition, MultiModal-MoLFormer uses data augmentation to expand the training set to 27,266 samples.

[0059] Next, the present invention prunes KAN. Before pruning, KAN has 30 neurons in the middle layer, but after pruning, it has only 17 neurons. The network comparison before and after pruning is shown in the figure below: Figure 4 As shown in the figure, the upper figure is the network diagram before pruning, and the lower figure is the network diagram after pruning; it can be seen that through the pruning operation, those nodes or connections that have little impact on model prediction are removed, thereby simplifying the model structure, reducing the amount of calculation, speeding up the reasoning speed, and reducing the risk of overfitting, which will not significantly affect the performance of the model.

[0060] Finally, in this embodiment, the present invention uses a data set based on lithium / copper half-cells collected from some chemical literature, in which the electrolyte formula consists of 2 to 7 components described by SMILES. In order to numerically amplify the output changes caused by electrolyte changes, CE is converted to logarithmic format (LCE=-log(1-CE)); we deleted an entry with repeated ratios but different measurement methods, and corrected some errors in ratios and molecular information; thus, 149 electrolyte formulas were finally obtained as training sets and 13 entries as test sets. Figure 5As shown, the box plot in the figure clearly shows the LCE output distribution of the training data based on the formula ingredient counts. In the figure, the outer whiskers represent the minimum and maximum values, the center line represents the median, the colored boxes represent the 25th to 75th percentiles of the data, and the outer data points are outliers in the data. In addition, our model only needs to be fine-tuned on more than a hundred samples to achieve more accurate predictions. Predicting the coulombic efficiency of the electrolyte through our method can greatly reduce the workload of chemical researchers to conduct experimental verification one by one in the huge electrolyte search space.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A customized model construction method for predicting the coulombic efficiency of electrolyte, characterized in that: The customized model construction method for electrolyte coulombic efficiency prediction includes the following steps: S1. Based on the original electrolyte data set, obtain the molar percentage and convert all the formula names in the electrolyte to obtain the standardized electrolyte composition data; S2. Use the chemical corpus to pre-train the molecular morphology model, input the standardized electrolyte composition data, and output the chemical characteristics of each component in the electrolyte; S3. Based on the chemical characteristics of each component in the electrolyte, the weighted sum of mole percentage is used and combined with mathematical theoretical neural network to generate a coulombic efficiency prediction model.

2. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 1, characterized in that: The original electrolyte data set includes the solvent volume and salt molarity of each electrolyte, and the standardized electrolyte composition data includes the elemental composition and structural information of all chemical molecules in the electrolyte formula.

3. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 2, characterized in that: The method of obtaining the molar percentage based on the original electrolyte data set and converting all formula names in the electrolyte to obtain standardized electrolyte composition data includes the following steps: S11. Using the principle of stoichiometry, convert the solvent volume and salt molarity into mole percentage; S12, based on the simplified molecular input linear representation rule, convert all chemical molecule names in the electrolyte formula to obtain mapping format names; S13. Establish standardized electrolyte composition data according to the mole percentage and the mapping format name.

4. The method for constructing a customized model for predicting the coulombic efficiency of electrolyte according to claim 1, characterized in that: The chemical corpus includes a public chemical database and a zinc database, and the pre-trained molecular morphology model includes optimizing relative position encoding using a self-attention mechanism.

5. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 4, characterized in that: The method of using a chemical corpus to pre-train a molecular morphology model and inputting standardized electrolyte composition data to output the chemical characteristics of each component in the electrolyte includes the following steps: S21. Pre-train the molecular morphogen model using unlabeled molecular sequences in the public chemical database and zinc database; S22, based on the pre-trained molecular morphogen model, inputting standardized electrolyte composition data; S23. According to the input standardized electrolyte composition data, the chemical characteristics of each component in the electrolyte are extracted and output.

6. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 1, characterized in that: The method of generating a Coulomb efficiency prediction model based on the chemical characteristics of each component in the electrolyte, using weighted sum of mole percentages, and combining mathematical theoretical neural networks includes the following steps: S31, generating a characteristic vector of the electrolyte by using a weighted summation method based on the chemical characteristics and molar percentage of each component in the electrolyte; S32, establishing a Coulomb efficiency prediction model based on the characteristic vector of the electrolyte using a mathematical theory neural network; S33. Use pruning technology and function fitting method to optimize the Coulomb efficiency prediction model.

7. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 6, characterized in that: Based on the chemical characteristics and molar percentage of each component in the electrolyte, the expression for generating the characteristic vector of the electrolyte using the weighted summation method is: ； Where x represents the overall characteristic vector of the electrolyte; e i represents the molecular vector of the i-th component in the electrolyte; n represents the number of electrolyte components; p i It represents the molar percentage of the i-th component in the electrolyte.

8. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 6, characterized in that: The method of establishing a Coulomb efficiency prediction model based on the characteristic vector of the electrolyte and using a mathematical theory neural network comprises the following steps: S321, using a linear layer to obtain a dimensionally reduced input vector according to the characteristic vector of the electrolyte; S322. Construct a Coulomb efficiency prediction model based on a multi-layer mathematical theory neural network structure; S323. Based on the Coulomb efficiency prediction model, input the input vector after dimensionality reduction, and output the Coulomb efficiency prediction result.

9. A customized model construction method for predicting the coulombic efficiency of electrolyte according to claim 8, characterized in that: The Coulomb efficiency prediction model is based on the input vector after dimensionality reduction, and the expression for outputting the Coulomb efficiency prediction result is: ； Where, CE represents the predicted result of Coulomb efficiency; Ф j represents the function matrix corresponding to the j-th layer of the mathematical theory neural network; L represents the total number of layers of the multi-layer mathematical theory neural network; ° represents function composition, that is, taking the output of one function as the input of the next function; σ1 represents the activation function of the linear layer; W1 represents the weight matrix of the linear layer; x represents the overall eigenvector of the electrolyte.

10. The method for constructing a customized model for predicting the coulombic efficiency of electrolyte according to claim 6, characterized in that: The method of optimizing the Coulomb efficiency prediction model by using pruning technology and function fitting method comprises the following steps: S331. Use pruning technology to optimize the network structure of the Coulomb efficiency prediction model; S332. Based on the network structure of the pruned Coulomb efficiency prediction model, a multidimensional continuous function is fitted using a one-variable function combination method; S333. Using logarithmic transformation, adjust the Coulomb efficiency prediction result output by the Coulomb efficiency prediction model.

Images