Ionic liquid viscosity prediction method based on molecular descriptor deep ensemble model

Abstract

This invention relates to a method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors. The method acquires molecular structure information, temperature, pressure, and viscosity data of the ionic liquid, calculates and processes molecular descriptors, constructs a convolutional neural network (CNN) basic model and a multilayer perceptron (MLP) basic model, generates out-of-fold prediction values ​​through 5-fold cross-validation, and constructs a meta-feature matrix. A linear regression model is used as the meta-learner to form a deep ensemble model. After Bayesian optimization and validation on an independent test set, the predicted viscosity value of the ionic liquid is output, and the SHAP method is used to identify key influencing features. This method improves prediction accuracy, stability, and interpretability, and is suitable for rapid prediction of ionic liquid viscosity and screening of low-viscosity candidate systems.

Description

Technical Field

[0001] This invention belongs to the fields of ionic liquid property prediction, machine learning-assisted chemical process design, and green solvent screening, specifically involving a method for predicting the viscosity of ionic liquids based on a deep integration model of molecular descriptors. Background Technology

[0002] Ionic liquids are a class of liquid salts composed of cations and anions. They are characterized by low vapor pressure, good thermal stability, high structural designability, and tunable solubility, and can be applied to processes such as extraction and separation, catalytic reactions, electrochemical energy storage, gas absorption, and biomass conversion. Because cations, anions, and their substituents can form numerous combinations, the transport properties of different ionic liquids vary significantly.

[0003] Viscosity is a crucial transport property affecting the engineering applications of ionic liquids. Higher viscosity reduces mixing and mass transfer efficiency, increases energy consumption in pumping and pipeline transportation, and hinders process scale-up. Therefore, rapidly obtaining the viscosity of ionic liquids under given temperature and pressure conditions is of practical significance for screening low-viscosity candidate systems.

[0004] Existing methods for obtaining viscosity mainly include experimental determination, empirical correlation, functional group contribution, molecular simulation, and data-driven prediction. Experimental determination requires samples and time; the applicability of empirical correlation and functional group contribution methods is affected by the system and parameters used; molecular simulation has high computational costs; although traditional machine learning methods can improve prediction efficiency, they are difficult to fully learn the complex nonlinear relationship between high-dimensional molecular descriptors and temperature and pressure.

[0005] Furthermore, while single convolutional neural network models can extract local correlation features between molecular descriptors through convolution operations, their ability to comprehensively express complex nonlinear relationships between molecular descriptors, temperature, and pressure, among other global variables, remains limited. Multilayer perceptrons (MLPs) possess strong global nonlinear mapping capabilities, enabling them to learn the overall relationship between high-dimensional descriptor features and thermodynamic conditions. Therefore, how to further introduce a global nonlinear learning branch into the convolutional neural network model and characterize the learning capabilities of different features through ensemble methods is crucial for improving the accuracy and stability of ionic liquid viscosity prediction.

[0006] Therefore, developing a method for predicting the viscosity of ionic liquids based on a deep integration model of molecular descriptors is of great significance for reducing experimental measurement costs, improving the accuracy of ionic liquid viscosity prediction, and supporting the screening of low-viscosity ionic liquids. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the viscosity of ionic liquids based on a deep integrated model of molecular descriptors, in order to solve the problems of high experimental measurement cost of ionic liquid viscosity, large prediction bias of single model, insufficient utilization of complementary information of different models, and limited interpretability of prediction results.

[0008] Technical solution

[0009] This invention first obtains the molecular structure information, temperature, pressure, and viscosity experimental values ​​of ionic liquids. Based on the molecular structure information, a molecular descriptor is calculated, and the obtained data is divided into a training set and an independent test set. Based on the training set, feature selection rules, missing value imputation parameters, and standardization parameters are determined to form the model input features. Subsequently, a convolutional neural network (CNN) basic model and a multilayer perceptron (MLP) basic model are constructed. Cross-validation within the training set and Bayesian optimization are used to determine the hyperparameters of the basic models. Five-fold cross-validation is used to generate out-of-fold prediction values ​​for the training samples from the two basic models. These out-of-fold prediction values ​​are combined to form a meta-feature matrix. The viscosity experimental values ​​are used as supervised labels to train the meta-learner, constructing a deep ensemble model. After the model is determined, the basic models are retrained using all training set data, and the predictive performance of the deep ensemble model is evaluated using the independent test set. For the ionic liquid to be predicted, basic prediction values ​​are obtained through the retrained basic models, and then the meta-learner integrates these basic prediction values ​​to output the predicted viscosity value of the ionic liquid.

[0010] The present invention is achieved through the following technical solution.

[0011] A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors includes the following steps: (1) Obtain ionic liquid viscosity data and construct an ionic liquid viscosity database containing 11,226 ionic liquid viscosity records; (2) Use the MoleculeDescriptors module in the RDKit framework to calculate descriptors, generating a total of 208 molecular descriptors. Divide the data containing the original molecular descriptors, temperature, pressure and viscosity experimental values ​​into 80% training set and 20% test set. (3) Preprocess the training set features, including deleting descriptors with a missing ratio of more than 60%, containing only a single value or low variance, and retaining one of the highly correlated features with an absolute value of correlation coefficient greater than 0.98; fill the remaining missing values ​​with the mean of the training set, and standardize them according to the mean and standard deviation of the training set, finally retaining 138 effective descriptor features; (4) Convolutional neural network basic model and multilayer perceptron basic model are constructed based on the input features of the model, respectively. Both basic models use mean squared error as the loss function and use Adam optimizer to update parameters. (5) The hyperparameters of the convolutional neural network base model and the multilayer perceptron base model are determined by Bayesian optimization. The optimal value of the hidden layer dimension of the multilayer perceptron is 241. The number of output channels of the first convolutional layer of the convolutional neural network is 32, the number of output channels of the second convolutional layer is 64, and the optimal value of the convolutional kernel size is 5. The dropout rate of the base model is 0.004 and the learning rate is 0.001. (6) The training set is divided into multiple non-overlapping subsets. One subset is selected as the validation subset and the remaining subsets are selected as the in-fold training subsets. The base models are trained respectively, and out-of-fold prediction values ​​are generated for the validation subset. (7) The hyperparameters of the convolutional neural network base model and the multilayer perceptron base model are determined by Bayesian optimization. The out-of-place prediction values ​​generated by the neural network basic model and the multilayer perceptron basic model are combined to form a meta-feature matrix. The meta-learner is trained with the meta-feature matrix as input and the corresponding viscosity experimental value as the supervision label to form a deep ensemble model. (8) The convolutional neural network basic model and the multilayer perceptron basic model are retrained using all training set data. The model input features of the independent test set are input into the two retrained basic models respectively. The obtained basic prediction values ​​are combined according to the column correspondence consistent with the meta-feature matrix and then input into the meta-learner to obtain the viscosity prediction value of the independent test set and evaluate the prediction performance of the deep ensemble model. (9) The model input features of the ionic liquid to be predicted are input into the retrained convolutional neural network basic model and the multilayer perceptron basic model respectively. The obtained basic prediction values ​​are combined according to the column correspondence consistent with the meta-feature matrix and then input into the meta-learner to output the viscosity prediction value of the ionic liquid to be predicted. (10) The contribution values ​​of temperature, pressure and each molecular descriptor to the output result of the deep ensemble model are calculated using the SHAP method, and the key features affecting the viscosity prediction result of the ionic liquid are identified according to the mean absolute SHAP value.

[0012] As a further optimization of the present invention: the basic model of the convolutional neural network includes at least one one-dimensional convolutional layer, activation function layer, dropout layer and fully connected regression layer, for learning local combination patterns in the input features of the model.

[0013] As a further optimization of the present invention: the basic model of the multilayer perceptron includes an input layer, at least one fully connected hidden layer, an activation function layer, a dropout layer, and a regression output layer, which are used to learn the global nonlinear mapping relationship between molecular descriptors, temperature, and pressure and viscosity.

[0014] As a further optimization of the present invention, the method also includes using the SHAP method to calculate the contribution values ​​of temperature, pressure and each molecular descriptor to the output results of the deep integrated model, in order to identify key features affecting the viscosity of ionic liquids.

[0015] Beneficial effects

[0016] (1) This invention utilizes molecular descriptors to characterize the composition, topology, charge, electronic state and fragment structure of ionic liquids, and combines temperature and pressure to form a unified model input feature, which can simultaneously characterize the influence of ionic liquid molecular structure and external thermodynamic conditions on viscosity.

[0017] (2) The present invention integrates the prediction results of the basic model of convolutional neural network and the basic model of multilayer perceptron by using a meta-learner, which can reduce the prediction bias of a single model and improve the accuracy, stability and generalization ability of ionic liquid viscosity prediction.

[0018] (3) The present invention combines the SHAP method to identify key features that affect viscosity prediction results, thereby improving the structural interpretability of the model prediction results.

[0019] (4) The present invention can be used for rapid prediction of the viscosity of ionic liquids and preliminary screening of low viscosity candidate ionic liquid systems, reducing the workload of experimental measurement and the cost of candidate system screening. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an ionic liquid viscosity prediction method based on a deep integration model of molecular descriptors according to the present invention. Detailed Implementation Example 1

[0021] A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors includes the following steps: A dataset of 11,226 ionic liquid viscosity records was obtained. Each record included molecular structure information, temperature, pressure, and viscosity experimental values. Two-dimensional molecular descriptors were calculated based on the ionic liquid structure information. The dataset was divided into training and independent test sets at a ratio of 80% and 20%, respectively. Descriptors with a missing value ratio exceeding 60%, descriptors containing only a single value, or descriptors with low variance were removed. For highly correlated features with an absolute correlation coefficient greater than 0.98, one feature was retained. The remaining missing values ​​were filled using the mean of the training set and standardized based on the mean and standard deviation of the training set, resulting in 138 effective descriptor features.

[0022] Molecular descriptors, temperature, and pressure are combined to form the model input features, and the same feature processing method is used in the model training and prediction stages. Based on the model input features, a convolutional neural network base model and a multilayer perceptron base model are constructed respectively; both base models use mean squared error as the loss function, use the Adam optimizer to update parameters, and determine the target hyperparameter combination through Bayesian optimization. Example 2

[0023] A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors includes the following steps: The training set is divided into five non-overlapping subsets. For each base model, one subset is selected sequentially as the validation subset, and the remaining four subsets are used as the in-fold training subsets. This ensures that each training sample receives an out-of-fold prediction value generated by a model not trained on that sample. Following the order of the training samples, the out-of-fold prediction values ​​of the convolutional neural network base model and the multilayer perceptron base model are used as two columns of the meta-feature matrix. A linear regression meta-learner is trained using the meta-feature matrix as input and viscosity experiment values ​​as supervision labels. After the parameters of the base models and the meta-learner are determined, both base models are retrained using all the training data. For the sample to be predicted, the retrained base models first generate base prediction values, then these base prediction values ​​are combined according to the column order of the meta-feature matrix and input into the meta-learner to output the final viscosity prediction value. Example 3

[0024] A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors includes the following steps: Using the same training and test set partitioning, the same molecular descriptors, the same target value processing method, and the same evaluation metric, we compared single convolutional neural network models and deep ensemble models. On a unified target value scale, the R-value of the single convolutional neural network model... 2 The R-values ​​are 0.961, MAE is 0.146, and RMSE is 0.319; the R-values ​​of the deep ensemble model are... 2The accuracy was 0.980, MAE was 0.091, and RMSE was 0.231, with MAE and RMSE decreasing by approximately 37.7% and 27.6%, respectively. These results demonstrate that integrating complementary prediction information from two base models using a meta-learner can improve prediction accuracy and stability. Example 4

[0025] A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors includes the following steps: The SHAP method is used to interpret and analyze the target deep ensemble model, calculating the contributions of temperature, pressure, and molecular descriptors to the model output, and identifying key influencing features based on the mean absolute SHAP value. For the ionic liquid to be predicted, the feature list, fill values, mean, and standard deviation determined during the training phase are used to form the model input features. After obtaining the viscosity prediction value, the candidate systems can be sorted from low to high predicted viscosity.

Claims

1. A method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors, characterized in that, Includes the following steps: (1) Obtain multiple ionic liquid viscosity data and construct a viscosity property database. The ionic liquid viscosity data includes ionic liquid molecular structure information, temperature, pressure and viscosity experimental values. (2) Calculate 208 molecular descriptors based on the molecular structure information, divide the data containing the original molecular descriptors, temperature, pressure and viscosity experimental values ​​into training set and independent test set, determine feature screening rules, missing value filling parameters and standardization parameters based on the training set, retain 138 effective molecular descriptor features after feature preprocessing, and combine the preprocessed molecular descriptors, temperature and pressure to form model input features; (3) Based on the cross-validation results within the training set, Bayesian optimization is used to determine the hyperparameters of the convolutional neural network base model and the multilayer perceptron base model. The convolutional neural network base model and the multilayer perceptron base model both use mean square error as the loss function and the Adam optimizer is used to update the model parameters. (4) Five-fold cross-validation is performed using the convolutional neural network basic model and the multilayer perceptron basic model after the hyperparameters are determined. The out-of-fold prediction values ​​generated by the two basic models for each training sample are obtained respectively. According to the correspondence of the training samples, the out-of-fold prediction values ​​generated by the two basic models are combined to form a meta-feature matrix. The meta-learner is trained with the meta-feature matrix as input and the corresponding viscosity experiment value as the supervision label to construct a deep ensemble model. (5) Retrain the convolutional neural network base model and the multilayer perceptron base model using all training set data, and evaluate the prediction performance of the deep ensemble model using the test set; (6) Input the model input features of the ionic liquid to be predicted into the retrained convolutional neural network base model and multilayer perceptron base model respectively to obtain the base prediction value. Then, combine the obtained base prediction values ​​in the same order as the meta-feature matrix and input them into the meta-learner to output the ionic liquid viscosity prediction value.

2. The method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors according to claim 1, characterized in that, The feature preprocessing includes: deleting descriptors with a missing proportion exceeding 60%, deleting descriptors containing only a single value or low variance, and retaining one of the highly correlated features with a correlation coefficient absolute value greater than 0.98; filling the remaining missing values ​​with the training set mean, and standardizing them based on the training set mean and standard deviation.

3. The method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors according to claim 1, characterized in that: The deep ensemble model was determined through Bayesian optimization. When the hidden layer dimension of the multilayer perceptron is 241, the number of output channels of the first convolutional layer of the convolutional neural network is 32, the number of output channels of the second convolutional layer is 64, the optimal value of the convolutional kernel size is 5, the dropout rate is 0.004, and the learning rate is 0.001, the model is optimal.

4. The method for predicting the viscosity of ionic liquids based on a deep ensemble model of molecular descriptors according to claim 1, characterized in that, The target deep ensemble model has a determination coefficient R² of 0.980, a mean absolute error (MAE) of 0.091, and a root mean square error (RMSE) of 0.231 on the independent test set. Compared with a single convolutional neural network model, the target deep ensemble model reduces the mean absolute error (MAE) by 37.7% and the root mean square error (RMSE) by 27.6%.

Images