Rare earth electrolyte solution activity coefficient prediction method based on machine learning

Abstract

The invention discloses a rare earth electrolyte solution activity coefficient prediction method based on machine learning, and belongs to the technical field of rare earth electrolyte. The method comprises the following steps: acquiring target rare earth historical electrolyte solution data, executing a preprocessing operation, and dividing the target rare earth historical electrolyte solution data into a training set and a test set according to a preset proportion; constructing a Debye-shock item by using the obtained training set; introducing a high-order interaction item in the Pitzer model into a Debye-shock item, and constructing a thermodynamic parameter prediction model input item based on the training electrolyte solution data set; after the input item is input into the thermodynamic parameter prediction model, the activity coefficient of the target rare earth is output from the model output layer through nonlinear transformation; and verifying and adjusting the trained model based on test electrolyte solution data. Compared with an experimental measurement method, the method is quicker; compared with an existing machine algorithm, the accuracy is higher; the obtaining time of the thermodynamic parameters is shortened, and the obtaining cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of rare earth electrolytes, and in particular to a method for predicting the activity coefficient of a rare earth electrolyte solution based on machine learning. Background Art

[0002] Rare earth electrolytes are widely used in numerous fields, such as rare earth metallurgy and fuel cells. Accurately understanding the thermodynamic parameters of rare earth electrolytes is crucial for optimizing their performance, improving production efficiency, and reducing costs. The most important thermodynamic parameter is the activity coefficient. Existing techniques for determining the thermodynamic parameters of rare earth electrolytes primarily rely on experimental measurement methods.

[0003] However, experimental measurement methods are often costly, time-consuming, and complex. Furthermore, with the continuous emergence of new rare earth electrolyte materials, traditional methods are unable to quickly, efficiently, and accurately obtain their thermodynamic parameters, which in turn slows down the material development process. Therefore, there is an urgent need for a machine learning-based method for predicting the thermodynamic parameters of rare earth electrolytes. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method for predicting the activity coefficient of rare earth electrolyte solutions based on machine learning.

[0005] To implement the above technology, the specific steps are as follows:

[0006] S1. Obtain historical electrolyte solution data of the target rare earth through open source data, and perform preprocessing operations on the acquired data; divide the preprocessed data into a training electrolyte solution dataset and a test electrolyte solution dataset according to a preset ratio;

[0007] The target rare earth historical electrolyte solution data include: element type, content, temperature and pressure of the target rare earth solution;

[0008] Preprocessing operations remove outliers; Z-score normalization is performed on the solution concentration and temperature; each rare earth element type is converted into a binary vector using one-hot encoding, where only the corresponding element position in the vector is 1 and the rest are 0, making it easier to process and distinguish different rare earth element types and facilitating subsequent accurate thermodynamic parameter prediction; missing value processing, that is, using KNN interpolation processing for sparse solution concentration data.

[0009] S2. Constructing Debye-Hückel terms from the obtained training electrolyte solution dataset for low-concentration electrolyte solution analysis;

[0010] The Debye-Hückel term is expressed as follows:

[0011]

[0012] Where I represents the ionic strength of the solution; B represents a constant related to the dielectric constant of the solvent and temperature; and a represents the radius of the ion.

[0013] S3. Introducing the high-order interaction terms in the Pitzer model into the Debye-Hückel terms and constructing the input terms of the input layer of the thermodynamic parameter prediction model based on the training electrolyte solution dataset;

[0014] The high-order interaction terms in the Pitzer model are used to enhance the generalization ability of the model to high-concentration, multi-component systems; the high-order interaction terms include: the two-body interaction parameter β (0) and concentration-dependent correction parameter β (1) ;

[0015] The element type of the target rare earth solution in the training electrolyte dataset is selected, and the radius of the element is obtained based on the element type; the target rare earth electrolyte solution temperature is also selected;

[0016] XGBoost is used as the thermodynamic parameter prediction model;

[0017] The inputs of the thermodynamic parameter prediction model include: Debye-Hückel term, high-order interaction term (β (0) and β (1) ), the element radius of the target rare earth, the element concentration of the target rare earth and the target rare earth electrolyte temperature.

[0018] S4. After the input items are input into the thermodynamic parameter prediction model, the activity coefficient of the target rare earth is output from the model output layer through nonlinear transformation to complete the model training;

[0019] The expression of nonlinear transformation is as follows:

[0020]

[0021] Where A represents a constant related to solvent and temperature; z + 、z - represents the charge number of positive ions and negative ions respectively; I represents the ionic strength of the solution; B represents a constant related to the dielectric constant of the solvent and temperature; a represents the radius of the ion; β (0) represents the two-body interaction parameter; β (1) represents the concentration-dependent correction parameter; γ ± represents the activity coefficient;

[0022] The hidden layer inputs the output result into the output layer through the ReLU function; the activation function uses the ReLU function to avoid the gradient vanishing problem and improve the performance of the thermodynamic parameter prediction model;

[0023] The thermodynamic parameter prediction model uses the mean square error (MSE) as the loss function to measure the difference between the activity coefficient and the historical thermodynamic parameters;

[0024] The thermodynamic parameter prediction model uses the stochastic gradient descent algorithm to adjust the parameters of the machine learning model until the training error reaches the minimum.

[0025] S5. Validate and adjust the trained model based on test electrolyte solution data;

[0026] Validation includes: using 5-fold cross validation to ensure model generalization;

[0027] The adjustments were made by using L2 regularization to prevent overfitting, interpreting feature importance through SHAP analysis, and calculating the model error on the test data until MSE ≤ 0.08.

[0028] Beneficial effects of the present invention:

[0029] The present invention introduces the Debye-Huckel term for the analysis of low-concentration electrolyte solutions, and simultaneously introduces the higher-order interaction term in the Pitzer model to correct the effects of ion pairs and complexes at high concentrations. The thermodynamic parameter prediction model is used to quickly and accurately predict the thermodynamic parameters of rare earth electrolytes, thereby greatly shortening the time and cost of obtaining the thermodynamic parameters.

[0030] The method of the present invention is not limited by experimental conditions and can predict the thermodynamic parameters of rare earth electrolytes under various complex conditions, providing strong support for the application of rare earth electrolytes in different scenarios. In addition, the system of the present invention has good scalability. With the continuous accumulation and updating of data, the thermodynamic parameter prediction model can be further optimized, improving the accuracy and reliability of the model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below with reference to specific embodiments.

[0033] like Figure 1 As shown in Figure 1, a method for predicting the activity coefficient of rare earth electrolyte solutions based on machine learning is used to predict the activity coefficient of LaCl3 at 298K and a concentration of 2.5 mol / L. The steps are as follows:

[0034] S1. Obtain historical electrolyte solution data of the target rare earth through open source data, and perform preprocessing operations on the acquired data; divide the preprocessed data into a training electrolyte solution dataset and a test electrolyte solution dataset according to a preset ratio;

[0035] In this embodiment, open source data includes: historical experimental records, research reports, and experimental databases (such as NIST and ICDD); the preset ratio is 8:2;

[0036] The target rare earth historical electrolyte solution data include: element type, content, temperature and pressure of the target rare earth solution;

[0037] The preprocessing operations include: removing outliers; performing Z-score normalization on the solution concentration and temperature; using one-hot encoding to convert each rare earth element species into a binary vector, where only the corresponding element position in the vector is 1 and the rest are 0, making it easier to process and distinguish different rare earth element species and facilitating subsequent accurate thermodynamic parameter prediction; and missing value processing, that is, using KNN interpolation processing on sparse solution concentration data.

[0038] S2. Constructing Debye-Hückel terms from the obtained training electrolyte solution dataset for low-concentration electrolyte solution analysis;

[0039] The Debye-Hückel term is expressed as follows:

[0040]

[0041] Where I represents the solution ionic strength; B represents a constant related to the solvent dielectric constant and temperature; a represents the radius of the ion;

[0042] In this embodiment, the prediction object is LaCl3 solution, and I=3×2.0=6.0.

[0043] S3. Introducing the high-order interaction terms in the Pitzer model into the Debye-Hückel terms and constructing the input terms of the input layer of the thermodynamic parameter prediction model based on the training electrolyte solution dataset;

[0044] The number of neurons in the input layer is consistent with the number of input items to ensure that the data enters the input layer completely and accurately;

[0045] The higher-order interaction terms in the Pitzer model are used to enhance the generalization ability of the model to high-concentration, multi-component systems; the higher-order interaction terms include: the two-body interaction parameter β and the concentration-dependent correction parameter β 1 ;

[0046] The element type of the target rare earth solution in the training electrolyte dataset is selected, and the radius of the element is obtained based on the element type; the target rare earth electrolyte solution temperature is also selected;

[0047] In this embodiment, La 3 + Radius is Cl - The radius is The target rare earth electrolyte solution temperature is 298K;

[0048] XGBoost is used as the thermodynamic parameter prediction model;

[0049] The inputs of the thermodynamic parameter prediction model include: Debye-Hückel term, high-order interaction term (β (0) and β (1) ), the element radius of the target rare earth, the element concentration of the target rare earth and the target rare earth electrolyte temperature.

[0050] S4. After the input items are input into the thermodynamic parameter prediction model, the activity coefficient of the target rare earth is output from the model output layer through nonlinear transformation to complete the model training;

[0051] The nonlinear transformation process is completed in the hidden layer of the thermodynamic parameter prediction model;

[0052] The expression of nonlinear transformation is as follows:

[0053]

[0054] Where A represents a constant related to solvent and temperature; z + 、z - Represents positive ions (La 3 +) and negative ions (Cl - ) charge number; I represents the solution ionic strength; B represents a constant related to the solvent dielectric constant and temperature; a represents the radius of the ion; β (0) represents the two-body interaction parameter, β (0) The range is generally 0.10-0.35 in rare earth electrolytes (high charge leads to enhanced electrostatic effect), and in this embodiment it is 0.2; (1) represents the concentration-dependent correction parameter, β (1) Range: usually -1.2-0.8, in this embodiment, -0.05; γ ± represents the activity coefficient;

[0055] The hidden layer inputs the output result into the output layer through the ReLU function; the activation function uses the ReLU function to avoid the gradient vanishing problem and improve the performance of the thermodynamic parameter prediction model;

[0056] The number of neurons in the output layer is the same as the number of output items; in the present invention, the output item is the activity coefficient;

[0057] The thermodynamic parameter prediction model uses the mean square error (MSE) as the loss function to measure the difference between the activity coefficient and the historical thermodynamic parameters;

[0058] The thermodynamic parameter prediction model uses the stochastic gradient descent algorithm to adjust the parameters of the machine learning model until the training error reaches the minimum.

[0059] S5. Validate and adjust the trained model based on test electrolyte solution data;

[0060] Validation includes: using 5-fold cross validation to ensure model generalization;

[0061] The adjustments were made by using L2 regularization to prevent overfitting, interpreting feature importance through SHAP analysis, and calculating the model error on the test data until MSE ≤ 0.08.

[0062] In this embodiment, XGBoost predicts the activity coefficient γ ± =0.67, the experimental value is 0.68 (error 2.9%); compared with the activity coefficient γ predicted by the Pitzer model ± =0.62 (error 10.1%), showing the advantages of the present invention. At high concentrations (>1 mol / L), the root mean square error (RMSE) of the present invention is 56% lower than that of the Pitzer model, and the coefficient of determination (R 2 ) is improved by 12%, and the calculation speed is 60 times faster than the Pitzer model.

[0063] In summary, the present invention can use the thermodynamic parameter prediction model to quickly and accurately predict the thermodynamic parameters of rare earth electrolytes, thereby greatly shortening the acquisition time and cost of thermodynamic parameters. The method of the present invention is not limited by experimental conditions and can predict the thermodynamic parameters of rare earth electrolytes under various complex conditions, providing strong support for the application of rare earth electrolytes in different scenarios. In addition, the system of the present invention has good scalability. With the continuous accumulation and updating of data, the thermodynamic parameter prediction model can be further optimized to improve the accuracy and reliability of model prediction.

Claims

1. A method for predicting the activity coefficient of rare earth electrolyte solutions based on machine learning, characterized in that: The following steps are involved: S1. Obtain historical rare earth electrolyte solution data of the target through open source data, and perform preprocessing operations on the acquired data; The target rare earth historical electrolyte solution data include: element type, content, temperature and pressure of the target rare earth solution; The preprocessed data is divided into a training electrolyte solution data set and a test electrolyte solution data set according to a preset ratio; S2. Constructing Debye-Huckel terms from the obtained training electrolyte solution dataset for low-concentration electrolyte solution analysis; S3. Introducing the high-order interaction terms in the Pitzer model into the Debye-Hückel terms and constructing the input terms of the input layer of the thermodynamic parameter prediction model based on the training electrolyte solution dataset; The higher-order interaction terms in the Pitzer model include: (0) and concentration-dependent correction parameter β (1) ; The input items include: the concentration of the target rare earth solution, the temperature of the target rare earth solution, the Debye-Huckel term, the two-body interaction parameter β (0) , concentration-dependent correction parameter β (1) and the ionic radius of the target rare earth in solution; S4. After the input items are input into the thermodynamic parameter prediction model, the activity coefficient of the target rare earth is output from the model output layer through nonlinear transformation to complete the model training; S5. Verify and adjust the trained model based on the test electrolyte solution data.

2. The method for predicting the activity coefficient of a rare earth electrolyte solution based on machine learning according to claim 1, wherein: The method obtains target rare earth historical electrolyte solution data through open source data and performs preprocessing operations on the obtained data. The preprocessing operations include: removing outliers; performing Z-score normalization on the concentration and temperature of the solution; encoding each rare earth element using one-hot encoding; and using KNN interpolation processing on sparse solution concentration data.

3. The method for predicting the activity coefficient of a rare earth electrolyte solution based on machine learning according to claim 1, wherein: The expression for constructing the Debye-Huckel term from the obtained training electrolyte solution dataset is as follows: Where I represents the ionic strength of the solution; B represents a constant related to the dielectric constant of the solvent and temperature; and a represents the radius of the ion.

4. The method for predicting the activity coefficient of a rare earth electrolyte solution based on machine learning according to claim 1, wherein: After the input items are input into the thermodynamic parameter prediction model, the activity coefficient of the target rare earth is output from the model output layer through nonlinear transformation. The expression of nonlinear transformation is as follows: Where A represents a constant related to solvent and temperature; z + 、z - represents the charge number of positive ions and negative ions respectively; I represents the ionic strength of the solution; B represents a constant related to the dielectric constant of the solvent and temperature; a represents the radius of the ion; β (0) represents the two-body interaction parameter; β (1) represents the concentration-dependent correction parameter; γ ± represents the activity coefficient.

5. The method for predicting the activity coefficient of a rare earth electrolyte solution based on machine learning according to claim 1, wherein: The validation and adjustment of the trained model based on the test electrolyte solution data includes: using 5-fold cross-validation to ensure model generalization; adjusting the model by using L2 regularization to prevent overfitting; interpreting feature importance through SHAP analysis; and calculating the error of the model on the test electrolyte solution data until the mean square error is ≤0.08.

Images