Method for predicting solubility of carbon dioxide in eutectic solvent based on machine learning model

Abstract

The invention relates to the technical field of chemical engineering and machine learning, and provides a method for predicting the solubility of carbon dioxide in a deep-eutectic solvent based on a machine learning model, and the method comprises the steps: collecting CO2 solubility data, carrying out the coding and feature descriptor conversion of a deep-eutectic solvent molecular structure through an open-source chemoinformatics software package RDKit, and constructing a data set. Four algorithms of XGBoost, random forest, a deep neural network and a convolutional neural network are adopted to construct a model, and hyper-parameters are optimized through grid search and k-fold cross validation. After the model is trained, the prediction performance is evaluated, and the optimal model is selected for CO2 solubility prediction. The method improves prediction efficiency and accuracy, has strong adaptability and generalization ability, and provides an effective tool for formula optimization of the eutectic solvent and improvement of CO2 capture efficiency.

Description

Technical Field

[0001] The present invention relates to the field of chemical engineering and machine learning technology, and in particular to a method for predicting the solubility of carbon dioxide in a low eutectic solvent based on a machine learning model. Background Art

[0002] The capture and storage technology of carbon dioxide is one of the important means to cope with global warming, and it is of key significance to achieve the goal of "carbon peak and carbon neutrality". In the process of CO2 capture, the selection of absorbent is crucial. Among them, deep eutectic solvents (DES) have become a hot topic of research due to their environmental friendliness, strong designability, and high solubility in CO2. However, there are a series of cumbersome problems in experimentally determining the solubility of CO2 in different deep eutectic solvents one by one, including complex experimental conditions, time-consuming operating procedures, and high experimental costs. In addition, due to the wide variety of low eutectic solvents and their complex molecular structures, the difficulty and complexity of experimental determination are further increased.

[0003] Traditional experimental methods are not only inefficient, but also difficult to fully cover all possible deep eutectic solvent combinations and reaction conditions, which limits the widespread application of deep eutectic solvents in the field of CO2 capture. Therefore, there is an urgent need for an efficient, economical and reliable method to predict the solubility of CO2 in different deep eutectic solvents to accelerate the exploration and development of excellent deep eutectic solvents.

[0004] As an important branch of artificial intelligence, machine learning has shown great potential in the field of chemical property prediction in recent years. Machine learning models can identify and utilize nonlinear relationships and complex patterns in data, transcend the limitations of traditional statistical methods, and effectively process multivariate and high-dimensional data. In addition, machine learning models also have strong adaptability and generalization capabilities, and can make accurate predictions in unseen situations, providing a new method for the extensive prediction of chemical properties.

[0005] The existing technology often uses a single model without comparing the performance of different models in predicting the solubility of CO2 in low eutectic solvents. Different machine learning models will show different performance when processing specific types of data and tasks. For example, some models may be better at handling linear relationships, while others are good at capturing complex nonlinear relationships. By trying multiple models, you can find the model that performs best in the specific task of predicting CO2 solubility.

[0006] Different machine learning models also have different generalization capabilities. Generalization capability refers to the ability of a model to perform well on new data. When predicting the solubility of CO2 in deep eutectic solvents, the generalization capability of the model is particularly important because it directly affects the model's prediction accuracy under different reaction conditions.

[0007] There are many types of deep eutectic solvents with complex molecular structures. The solubility of different solvents in CO2 may be affected by many factors. The use of multiple machine learning models can more comprehensively adapt to different types of deep eutectic solvents and their interactions with CO2, improving the accuracy and reliability of predictions.

[0008] Different machine learning models have different focuses in capturing data features. For example, random forest (RF) implements classification or regression tasks by building multiple decision trees, and is good at processing interactive data; while deep neural network (DNN) and convolutional neural network (CNN) can automatically extract and learn deep features in data, and are suitable for processing highly complex data. By combining these models, we can make full use of their respective advantages and improve the accuracy and stability of predictions.

[0009] Based on this, using different machine learning models to predict the solubility of CO2 in low eutectic solvents is of great significance for improving prediction accuracy, stability and generalization ability, verifying the performance of different algorithms on specific tasks and dispersing prediction risks. Summary of the invention

[0010] In order to solve the problems existing in the background technology, the present invention provides a method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model, which comprises the following steps:

[0011] S1, data collection: CO2 solubility data points with a solubility range of 0.0014-0.0326 were collected, covering at least 13 deep eutectic solvents. The data set was randomly divided into a training set and a test set, with the training set accounting for 80% of the total data and the test set accounting for 20% of the total data;

[0012] S2, molecular structure coding of deep eutectic solvent;

[0013] S3, model construction, training and evaluation;

[0014] S4. Result analysis and model selection: compare the prediction performance of different models and select the optimal model;

[0015] S5. Use the trained optimal model to predict the solubility of CO2 in low eutectic solvents under different conditions.

[0016] In a preferred embodiment, in step S1, the temperature range of the data set is 293.15 to 323.15 K, and the pressure range is 0.515 to 5.842 bar.

[0017] In a preferred solution, the specific steps of step S2 include:

[0018] S21.Select tools:

[0019] The open source cheminformatics software package RDKit was selected as a tool for obtaining the structure encoding of deep eutectic solvents; RDKit is an open source Python cheminformatics toolkit that is widely used for processing and calculating chemical structure data;

[0020] S22. Get the structure code:

[0021] Take the molecular structure of the low eutectic solvent as input; these molecular structures can be obtained from various chemical databases, literature or experimental data; use various functions and classes of RDKit to process molecular structures, including reading SMILES strings and parsing SDF files;

[0022] The built-in algorithm of RDKit generates a corresponding structure code for each deep eutectic solvent; the structure code is a digital representation of the molecular structure, which is convenient for computer processing and analysis;

[0023] S23. Conversion feature descriptor:

[0024] The generated structural codes are further converted into feature descriptors. This step is to simplify the complex molecular structure information into a series of numerical features so that the machine learning algorithm can understand and process it. The structural code of each deep eutectic solvent is converted into 280 feature descriptors. These descriptors comprehensively cover multiple aspects of the molecule, ensuring that the complexity and diversity of the molecular structure can be fully expressed.

[0025] S24. Data preprocessing:

[0026] The generated feature descriptors are standardized and normalized to ensure comparability between different features and eliminate the impact of dimension; check whether there are outliers or missing values ​​in the feature descriptors, and perform corresponding processing to fill in missing values ​​or remove abnormal data.

[0027] In a preferred solution, step S3 includes step S31 of using four machine learning algorithms to build a model, wherein the four machine learning algorithms are: XGBoost algorithm, random forest RF algorithm, deep neural network DNN algorithm and convolutional neural network CNN algorithm, wherein:

[0028] The XGBoost algorithm uses the extreme gradient boosting tree method to enhance classification capabilities by integrating multiple CART regression trees. XGBoost organically combines multiple CART regression trees to form a classifier. By continuously adding trees, each new tree corrects the prediction error of the previous tree. The model partitions the decision tree through a greedy algorithm and approximates the loss function quadratically through Taylor expansion. At the same time, regularization terms are introduced to control model complexity and prevent overfitting.

[0029] The random forest RF algorithm achieves classification and regression functions by fusing multiple decision trees. RF constructs multiple decision trees through self-service sampling, and only uses some features for node segmentation during the generation of each tree to enhance the diversity and anti-overfitting ability of the model. The output of RF is the average of the prediction results of all trees.

[0030] The deep neural network (DNN) algorithm processes complex nonlinear relationships through a multi-layer neural network structure. DNN contains an input layer, multiple hidden layers, and an output layer. Each layer consists of multiple neurons connected by weights and biases. DNN optimizes parameters through the back-propagation algorithm and gradient descent method to minimize prediction errors.

[0031] The convolutional neural network (CNN) algorithm extracts features from images or data through convolutional layers and pooling layers. CNN simulates the processing mechanism of the visual cortex of the human brain and uses local receptive field, weight sharing and pooling operation characteristics to effectively reduce network complexity and stabilize learning effects. It extracts features by sliding the convolution kernel on the input data and passes them to subsequent layers for processing.

[0032] In a preferred solution, step S31 further includes the following process:

[0033] The grid search method combined with k-fold cross-validation is used to optimize the model's hyperparameters: the grid search method involves traversing the preset hyperparameter space to find the optimal hyperparameter combination; the k-fold cross-validation involves dividing the data set into k subsets, selecting k-1 subsets as training sets each time, and the remaining 1 subset as a validation set, repeating k times to evaluate the generalization ability of the model; the key hyperparameters to be optimized include: the number of decision trees: controls the number of trees in the model, affecting the complexity and prediction accuracy of the model; the maximum depth of the decision tree: controls the maximum number of layers of the tree, and too deep a tree can lead to overfitting; the learning rate: in the optimization algorithm, controls the step size of each weight update, affecting the learning speed and stability of the model.

[0034] In a preferred solution, the step S3 further includes a model training and evaluation process S32, wherein:

[0035] S321, Model training: Use the training set to adjust the model parameters and develop a prediction model; by continuously optimizing the loss function, the model can accurately predict the solubility of CO2 in low eutectic solvents;

[0036] S322 Model Evaluation: Evaluate the performance of the model through the test set; use mean square error (MSE), mean absolute error (MAE) and determination coefficient (R) 2 As an evaluation indicator, it can comprehensively evaluate the prediction accuracy and generalization ability of the model.

[0037] In a preferred solution, the specific process of step S321 includes:

[0038] The model is trained using the training subset data. During the training process, the model calculates the predicted value based on the input data and the current parameters, and calculates the error between the predicted value and the true value, i.e., the loss. Then, the error is passed from the output layer to the input layer layer by layer through the back propagation algorithm, and the model parameters are updated according to optimization algorithms such as gradient descent to reduce the error. This process is repeated many times until the performance of the model on the validation subset is no longer significantly improved, or the preset upper limit of training rounds is reached. After the training is completed, the trained model parameters are saved for subsequent prediction or further analysis.

[0039] In a preferred solution, the specific process of step S322 includes:

[0040] Use the test set data divided in the initial stage as the input for evaluating the model;

[0041] Use the trained model to predict the test set data and calculate the error between the predicted value and the true value; use the mean square error MSE, mean absolute error MAE and determination coefficient R 2 As an evaluation indicator;

[0042] Mean square error (MSE): measures the average of the squares of the difference between the predicted value and the true value. The smaller the MSE, the more accurate the prediction.

[0043] Mean absolute error (MAE): measures the average of the absolute values ​​of the difference between the predicted value and the true value. Similarly, the smaller the MAE, the more accurate the prediction.

[0044] Coefficient of determination R 2 : Measures the extent to which the model explains the variability, R 2 The closer the value is to 1, the better the model fits the data.

[0045] In the preferred solution, the specific process of step S4 is as follows:

[0046] S41 compares the MSE, MAE and R of four models XGBoost, RF, DNN and CNN on the test set 2 value:

[0047] XGBoost: Very low MSE and MAE, R 2 Close to 1, indicating that the XGBoost model has good predictive ability;

[0048] RF: Although the MSE and MAE of the RF model are also lower, R 2 The value is slightly lower than that of XGBoost and CNN, indicating that its prediction accuracy is slightly inferior;

[0049] DNN: The DNN model also performs well on the test set, but is slightly inferior to XGBoost and CNN.

[0050] CNN: The best performance among all models, with MSE and MAE almost 0, R 2 The value is extremely high, showing strong predictive ability;

[0051] After analysis and comparison, the CNN model was finally selected as the final prediction model.

[0052] In the preferred solution, the specific process of step S5 is as follows:

[0053] Detailed steps for using the trained optimal model CNN to predict the solubility of CO2 in deep eutectic solvents under different conditions:

[0054] S51. Prepare to input data:

[0055] Collect or generate the corresponding input data according to the conditions that need to be predicted; these data include the structural code SC of the low eutectic solvent, the partial pressure of carbon dioxide PCO2 and the reaction temperature T; the structural code needs to use the chemical informatics software package RDKit to convert the molecular structure of the low eutectic solvent into a numerical form; ensure that the format of the input data is consistent with the data format used when training the model so that the model can parse and process it correctly;

[0056] S52. Data preprocessing:

[0057] For missing values ​​or outliers contained in the input data, appropriate processing is performed to fill in missing values ​​or remove outliers, and the input data is normalized or standardized to ensure that the prediction performance of the model is stable under different conditions;

[0058] S53, Model loading and prediction:

[0059] Load the trained optimal CNN model and input the prepared input data into the loaded CNN model; the model calculates and outputs the predicted CO2 solubility value based on the input low eutectic solvent structure code, carbon dioxide partial pressure and reaction temperature;

[0060] S54. Interpretation and application of results:

[0061] Interpret the CO2 solubility values ​​predicted by the model and evaluate their accuracy and reliability under specific conditions; apply the prediction results to the actual scenario of optimizing the formulation of low eutectic solvents and improving CO2 capture efficiency.

[0062] The beneficial effects achieved by the present invention are:

[0063] The present invention accurately describes the structural characteristics of deep eutectic solvent molecules and predicts the solubility of CO2 under different conditions through machine learning algorithms. Compared with the traditional method of one-by-one experimental determination, the prediction efficiency and accuracy are significantly improved. Through models such as XGBoost, RF, and CNN, the R 2 The values ​​all reached above 0.950, and the CNN model even reached 0.991.

[0064] The machine learning model of the present invention has strong adaptability and generalization ability, and can make accurate predictions in unseen situations. This means that the method can not only be applied to known deep eutectic solvents, but also effectively predict the solubility of new deep eutectic solvents in CO2, making it possible to predict a wider range of chemical properties.

[0065] The present invention uses four different machine learning algorithms, XGBoost, random forest, deep neural network and convolutional neural network, and selects the optimal CNN model by comparing their performance on the test set. This multi-algorithm comparison method improves the reliability and stability of prediction.

[0066] The present invention can quickly predict the solubility of CO2 in deep eutectic solvents under different conditions, providing strong support for the exploration and development of excellent deep eutectic solvents. It is helpful to accelerate the research and development of new deep eutectic solvents and improve the CO2 capture efficiency.

[0067] The present invention has great advantages in processing multivariate and high-dimensional data, identifying nonlinear relationships and complex patterns in the data, and can cope with complex and changeable experimental conditions such as different temperatures, pressures and solution concentrations to achieve accurate solubility prediction.

[0068] The present invention is a low-cost prediction scheme, which avoids the cumbersome and time-consuming experimental determination process and reduces the research cost. The training and prediction process of the machine learning model is relatively simple, which reduces the consumption of experimental materials and labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of a method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model;

[0070] Figure 2 It is the DES database generation structure encoding process;

[0071] Figure 3 It is a schematic diagram of the machine learning algorithm - XGBoost;

[0072] Figure 4 It is the principle diagram of machine learning algorithm - RF;

[0073] Figure 5It is the principle diagram of machine learning algorithm - DNN;

[0074] Figure 6 It is the principle diagram of machine learning algorithm - CNN;

[0075] Figure 7 It is grid search-k-fold cross validation;

[0076] Figure 8 It is the regression result of the XGBoost model training set and test set;

[0077] Fig. 9 It is the regression result of RF model training set and test set;

[0078] Fig.10 It is the regression result of CNN model training set and test set;

[0079] Fig.11 It is the regression result of DNN model training set and test set. DETAILED DESCRIPTION

[0080] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are merely illustrative, and the present invention is not limited to the various structures recorded in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0081] Reference Figure 1-Figure 11 A method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model of the present invention comprises the following steps:

[0082] S1, data collection: CO2 solubility data points with a solubility range of 0.0014-0.0326 were collected, covering at least 13 deep eutectic solvents. The data set was randomly divided into a training set and a test set, with the training set accounting for 80% of the total data and the test set accounting for 20% of the total data;

[0083] S2, molecular structure coding of deep eutectic solvent;

[0084] S3, model construction, training and evaluation;

[0085] S4. Result analysis and model selection: compare the prediction performance of different models and select the optimal model;

[0086] S5. Use the trained optimal model to predict the solubility of CO2 in low eutectic solvents under different conditions.

[0087] In a preferred embodiment, in step S1, the temperature range of the data set is 293.15 to 323.15 K, and the pressure range is 0.515 to 5.842 bar.

[0088] In a preferred solution, the specific steps of step S2 include:

[0089] S21.Select tools:

[0090] The open source cheminformatics software package RDKit was selected as a tool for obtaining the structure encoding of deep eutectic solvents; RDKit is an open source Python cheminformatics toolkit that is widely used for processing and calculating chemical structure data;

[0091] S22. Get the structure code:

[0092] Take the molecular structure of the low eutectic solvent as input; these molecular structures can be obtained from various chemical databases, literature or experimental data; use various functions and classes of RDKit to process molecular structures, including reading SMILES strings and parsing SDF files;

[0093] The built-in algorithm of RDKit generates a corresponding structure code for each deep eutectic solvent; the structure code is a digital representation of the molecular structure, which is convenient for computer processing and analysis;

[0094] S23. Conversion feature descriptor:

[0095] The generated structural codes are further converted into feature descriptors. This step is to simplify the complex molecular structure information into a series of numerical features so that the machine learning algorithm can understand and process it. The structural code of each deep eutectic solvent is converted into 280 feature descriptors. These descriptors comprehensively cover multiple aspects of the molecule, ensuring that the complexity and diversity of the molecular structure can be fully expressed.

[0096] S24. Data preprocessing:

[0097] The generated feature descriptors are standardized and normalized to ensure comparability between different features and eliminate the impact of dimension; check whether there are outliers or missing values ​​in the feature descriptors, and perform corresponding processing, such as filling missing values ​​or removing abnormal data.

[0098] The step S3 includes a step S31 of using four machine learning algorithms to construct a model, wherein the four machine learning algorithms are: XGBoost algorithm, random forest RF algorithm, deep neural network DNN algorithm and convolutional neural network CNN algorithm, wherein:

[0099] The XGBoost algorithm uses the extreme gradient boosting tree method to enhance classification capabilities by integrating multiple CART regression trees. XGBoost organically combines multiple CART regression trees to form a classifier. By continuously adding trees, each new tree corrects the prediction error of the previous tree. The model partitions the decision tree through a greedy algorithm and approximates the loss function quadratically through Taylor expansion. At the same time, regularization terms are introduced to control model complexity and prevent overfitting.

[0100] The random forest RF algorithm achieves classification and regression functions by fusing multiple decision trees. RF constructs multiple decision trees through self-service sampling, and only uses some features for node segmentation during the generation of each tree to enhance the diversity and anti-overfitting ability of the model. The output of RF is the average of the prediction results of all trees.

[0101] The deep neural network (DNN) algorithm processes complex nonlinear relationships through a multi-layer neural network structure. DNN contains an input layer, multiple hidden layers, and an output layer. Each layer consists of multiple neurons connected by weights and biases. DNN optimizes parameters through the back-propagation algorithm and gradient descent method to minimize prediction errors.

[0102] The convolutional neural network (CNN) algorithm extracts features from images or data through convolutional layers and pooling layers. CNN simulates the processing mechanism of the visual cortex of the human brain and uses local receptive field, weight sharing and pooling operation characteristics to effectively reduce network complexity and stabilize learning effects. It extracts features by sliding the convolution kernel on the input data and passes them to subsequent layers for processing.

[0103] The step S31 also includes the following process:

[0104] The grid search method combined with k-fold cross-validation is used to optimize the model's hyperparameters: the grid search method involves traversing the preset hyperparameter space to find the optimal hyperparameter combination; the k-fold cross-validation involves dividing the data set into k subsets, selecting k-1 subsets as training sets each time, and the remaining 1 subset as a validation set, repeating k times to evaluate the generalization ability of the model; the key hyperparameters to be optimized include: the number of decision trees: controls the number of trees in the model, affecting the complexity and prediction accuracy of the model; the maximum depth of the decision tree: controls the maximum number of layers of the tree, and too deep a tree can lead to overfitting; the learning rate: in the optimization algorithm, controls the step size of each weight update, affecting the learning speed and stability of the model.

[0105] The step S3 also includes a model training and evaluation process S32, wherein:

[0106] S321, Model training: Use the training set to adjust the model parameters and develop a prediction model; by continuously optimizing the loss function, the model can accurately predict the solubility of CO2 in low eutectic solvents;

[0107] S322 Model Evaluation: Evaluate the performance of the model through the test set; use mean square error (MSE), mean absolute error (MAE) and determination coefficient (R) 2 As an evaluation indicator, it can comprehensively evaluate the prediction accuracy and generalization ability of the model.

[0108] In a preferred solution, the specific process of step S321 includes:

[0109] The model is trained using the training subset data. During the training process, the model calculates the predicted value based on the input data and the current parameters, and calculates the error between the predicted value and the true value, i.e., the loss. Then, the error is passed from the output layer to the input layer layer by layer through the back propagation algorithm, and the model parameters are updated according to optimization algorithms such as gradient descent to reduce the error. This process is repeated many times until the performance of the model on the validation subset is no longer significantly improved, or the preset upper limit of training rounds is reached. After the training is completed, the trained model parameters are saved for subsequent prediction or further analysis.

[0110] The specific process of step S322 includes:

[0111] Use the test set data divided in the initial stage as the input for evaluating the model;

[0112] Use the trained model to predict the test set data and calculate the error between the predicted value and the true value; use the mean square error MSE, mean absolute error MAE and determination coefficient R 2 As an evaluation indicator;

[0113] Mean square error (MSE): measures the average of the squares of the difference between the predicted value and the true value. The smaller the MSE, the more accurate the prediction.

[0114] Mean absolute error (MAE): measures the average of the absolute values ​​of the difference between the predicted value and the true value. Similarly, the smaller the MAE, the more accurate the prediction.

[0115] Coefficient of determination R 2 : Measures the extent to which the model explains the variability, R 2 The closer the value is to 1, the better the model fits the data.

[0116] The specific process of step S4 is as follows:

[0117] S41 compares the MSE, MAE and R of four models XGBoost, RF, DNN and CNN on the test set 2 value:

[0118] XGBoost: Very low MSE and MAE, R 2 Close to 1, indicating that the XGBoost model has good predictive ability;

[0119] RF: Although the MSE and MAE of the RF model are also lower, R 2 The value is slightly lower than that of XGBoost and CNN, indicating that its prediction accuracy is slightly inferior;

[0120] DNN: The DNN model also performs well on the test set, but is slightly inferior to XGBoost and CNN.

[0121] CNN: The best performance among all models, with MSE and MAE almost 0, R 2 The value is extremely high, showing strong predictive ability;

[0122] After analysis and comparison, the CNN model was finally selected as the final prediction model.

[0123] In the preferred solution, the specific process of step S5 is as follows:

[0124] Detailed steps for using the trained optimal model CNN to predict the solubility of CO2 in deep eutectic solvents under different conditions:

[0125] S51. Prepare to input data:

[0126] Collect or generate the corresponding input data according to the conditions that need to be predicted; these data include the structural code SC of the low eutectic solvent, the partial pressure of carbon dioxide PCO2 and the reaction temperature T; the structural code needs to be converted into a numerical form using a chemical informatics software package such as RDKit; ensure that the format of the input data is consistent with the data format used when training the model so that the model can parse and process it correctly;

[0127] S52. Data preprocessing:

[0128] For missing values ​​or outliers contained in the input data, appropriate processing is performed to fill in missing values ​​or remove outliers, and the input data is normalized or standardized to ensure that the prediction performance of the model is stable under different conditions;

[0129] S53, Model loading and prediction:

[0130] Load the trained optimal CNN model and input the prepared input data into the loaded CNN model; the model calculates and outputs the predicted CO2 solubility value based on the input low eutectic solvent structure code, carbon dioxide partial pressure and reaction temperature;

[0131] S54. Interpretation and application of results:

[0132] Interpret the CO2 solubility values ​​predicted by the model and evaluate their accuracy and reliability under specific conditions; apply the prediction results to the actual scenario of optimizing the formulation of low eutectic solvents and improving CO2 capture efficiency.

[0133] Example 1: A method for predicting the solubility of CO2 in a low eutectic solvent based on a machine learning model is introduced below through a specific example. The specific steps are as follows:

[0134] (1) Machine Learning Dataset Establishment

[0135] A total of 362 CO2 solubility data points were collected in this study, with a carbon dioxide solubility range of 0.0014-0.0326 (a total of 13 low eutectic solvents), a temperature range of 293.15 to 323.15K, a pressure range of 0.515 to 5.842bar, 13 hydrogen bond donors, including 1,2-propylene glycol, phenol, levulinic acid, 1,4-butanediol, furfuryl alcohol, etc., and 1 hydrogen bond acceptor, namely choline chloride. The low eutectic solvent dataset is randomly divided into a training set (about 80% of the total data) and a test set (about 20% of the total data). The training set is used to adjust hyperparameters and develop models, and the test set is mainly used to evaluate the performance of the developed model. These low eutectic solvents are of various types, laying a solid foundation for the development of carbon dioxide equilibrium solubility prediction models. Figure 1 The distribution of carbon dioxide partial pressure and temperature is shown. These detailed data will provide sufficient support for a more in-depth study of the effect of deep eutectic solvents on the equilibrium solubility of carbon dioxide.

[0136] (2) Determination of molecular description of deep eutectic solvent

[0137] Adding structural encoding is a basic strategy in the machine learning process, which helps to convert the complex molecular structure of low eutectic solvents into numerical data that can be processed by machine learning algorithms. This conversion can accurately represent the molecular structure. This study used the open source chemical informatics software package RDKit to obtain these structural encodings. RDKit provides a series of powerful functions for effectively processing the molecular structure of low eutectic solvents. Using RDKit, we successfully converted the structural encoding of each low eutectic solvent into 280 features, providing a basis for an in-depth and comprehensive description of the structural encoding of low eutectic solvents. For the specific process, see Figure 2 .

[0138] (3) Principles of XGBoost, RF, CNN and DNN

[0139] XGBoost (EXtremeGradientBoosting), that is, the extreme gradient boosting tree, is one of the Boosting algorithms. The basic concept of the Boosting algorithm is to organically combine multiple weak classifiers to form a powerful classifier. The key to this combination is that there is a close dependency between individual learners, and their generation is sequential, that is, a serialized generation method. XGBoost uses a boosting tree model that achieves powerful classification capabilities by integrating multiple tree models, especially the CART (Classification and Regression Tree) regression tree model. In short, XGBoost combines multiple CART regression trees to form an efficient and high-performance classifier.

[0140] The core structure of XGBoost is based on decision trees, which play the role of "weak learners" in the model and together constitute the overall framework of XGBoost. These decision trees follow a certain sequential logic during the construction process: the generation of subsequent decision trees will refer to the prediction results of the previous tree. Specifically, the prediction bias generated by the previous decision tree is taken into account in order to pay more attention to the training samples that the previous tree failed to accurately classify. Subsequently, based on this adjusted sample distribution, the system will train the next decision tree. This design enables XGBoost to continuously optimize the prediction effect during the continuous learning process. The principle is as follows Figure 3 shown.

[0141] The learning process of the XGBoost model is: based on the objective function, derive the "required weight of each leaf node", "the amount of information obtained after node partitioning" and "the eigenvalue importance ranking function". First, the current decision tree is partitioned using a greedy algorithm to obtain the optimal solution; on this basis, the corresponding features are selected by calculating the objective function (and the above-mentioned "information gain after partitioning"). A split point solution method based on the "eigenvalue importance ranking function" of the "weight quantile" is proposed. The weight of the corresponding leaf node is determined based on the "each leaf node should be weighted" mentioned above.

[0142] XGBoost is an improved gradient boosting algorithm that applies Newton's method to the maximum value of the loss function and expands it to quadratic through Taylor, which greatly improves the accuracy of the calculation. In addition, in order to control the complexity of the model and prevent overfitting, XGBoost introduces a regularization term in the loss function. The objective function of this method mainly includes two main components, namely the loss based on the gradient boosting algorithm and the regularization term that constrains the complexity of the model. This design enables XGBoost to have better generalization ability while maintaining efficient learning. Loss function definition formula (1):

[0143]

[0144] Where n is the number of samples of the training function, l is the loss of a single sample, assuming it is a convex function, y′ i is the model’s predicted value for the training sample, y i is the true value of the training sample. The regularization term defines the complexity of the model, see formula (2):

[0145]

[0146] Among them, γ and μ are manually set parameters, ω is the vector formed by the values ​​of all leaf nodes of the decision tree, and T is the number of leaf nodes.

[0147] RF (Randomforest), also known as random forest, has both classification and regression functions. Its core lies in the ingenious integration of multiple decision trees. Figure 4 As shown, when constructing RF, first perform n self-sampling on the original data set (n tree is the total number of decision trees in RF), thus generating n custom sample sets. Then classify each self-sampling set to form a decision tree. When the algorithm splits each node, it first selects the optimal split point instead of selecting from all the features. This double randomness ensures that each decision tree in the random forest has its own uniqueness, so that it has a good ability to resist over-simulation. The generation and calculation process of random forest is as follows Figure 4 .

[0148] In the regression problem, the output of the RF signal is the average of the prediction values ​​of each tree. Among them, the number of decision trees and the size of the feature subset are two key hyperparameters that determine its prediction accuracy. Generally speaking, as the number of decision trees increases, the prediction accuracy of RF will gradually improve, but too many decision trees will also burden the training and calculation speed of the model.

[0149] In RF, data that are not included in the self-sampling set are given a special name - "out-of-bag data (OOBdata)". Therefore, this project proposes a method to correct hyper parameters based on the prediction error of out-of-bag data (OOBerror). Simply put, a lower out-of-bag error indicates that the model has better generalization performance, that is, better prediction performance for unknown data. For regression tasks, the mean square error of out-of-bag data (OOBMSE) can be calculated by the following formula (3):

[0150]

[0151] Among them, OOB is the mean of the prediction values ​​of the decision tree for the i-th out-of-bag data in RF, yi is the experimental value. Moreover, because out-of-bag data is used, a function in random forest can be used to evaluate the relative importance of a variable. When the data of an input variable is replaced and the other input variables remain unchanged, the relative importance of each input variable to the response variable (output value) can be calculated by calculating the change in the prediction error of the out-of-bag data. The relative importance of variables calculated by RF can be used for feature screening, input variable contribution assessment, etc. For the prediction model, the relative importance of the variable is reflected in the degree of decrease in the prediction accuracy of the model when the data of the variable is missing.

[0152] A deep neural network (DNN) is a multi-layer neural network that forms a deep structure by connecting multiple neurons together. DNN is a form of artificial neural network that contains multiple neural network layers, usually including an input layer, multiple hidden layers, and an output layer. Each layer is composed of neurons that are connected together by weights. DNN can be used to process various types of data, including images, text, speech, etc. In a deep neural network, the number of hidden layers and the number of neurons in each layer can depend on the specific task and data type. The training process of DNN usually uses the backpropagation algorithm and gradient descent optimization method to continuously adjust the parameters of the neural network to minimize the prediction error and loss function. Its principle is Figure 5 shown.

[0153] DNN has multiple nonlinear mapping feature transformations and can fit highly complex functions. Compared with shallow modeling, deep modeling can more carefully and efficiently represent actual complex nonlinear problems. Deep neural networks have achieved remarkable results in many fields, such as image classification, speech recognition, natural language processing, etc. However, deep neural networks also face some challenges, such as over-fitting, gradient disappearance, and computing resources.

[0154] exist Figure 5 In the neural network, the input layer is mainly responsible for receiving the original data and converting it into a form that the neural network can process efficiently. The core of DNN (deep neural network) is the hidden layer. The hidden layer contains multiple neurons, and the number and level of these neurons can be flexibly adjusted according to the needs of specific tasks and the characteristics of the input data. Between layers, a fully connected design is adopted, and any neuron in the i-th layer will establish a connection with every neuron in the i+1-th layer. The output layer is responsible for converting the results of the neural network after multi-layer processing into a human-readable output form.

[0155] In DNN, each layer contains multiple neurons, from the input layer to the hidden layer to the output layer. Each neuron receives information from the previous layer and passes it to the next layer. In the hidden layer, each neuron passes information through the connection with the neurons in the next layer, and this process also involves the adjustment of weights and biases. The learning process of DNN usually adopts the back-propagation algorithm and gradient descent optimization method. The back-propagation algorithm calculates the error between the output layer and the expected output, and propagates the error back to the previous hidden layer, so as to continuously adjust the weight and bias parameters, so that the output of the neural network is closer and closer to the expected output.

[0156] In general, the basic structure of DNN is a multi-layer neural network, which extracts and transforms data features through multiple hidden layers to obtain output results. This structure can handle complex nonlinear problems and has strong expressiveness and flexibility.

[0157] Convolutional Neural Networks (CNN) is an important research direction of deep learning. Due to its unique convolution calculation and deep feedback structure, it occupies an important position in artificial intelligence research. The network has a strong representation learning ability and can classify the input information according to a certain level, so it is also called "translation invariant neural network".

[0158] The research history of CNN can be traced back to the 1980s and 1990s. Time delay networks and LeNet-5 are early representatives of convolutional neural networks. After entering the 21st century, with the rise of deep learning theory and the improvement of computing equipment performance, CNN has ushered in rapid development and is widely used in many fields such as computer vision and natural language processing.

[0159] The design of CNN is inspired by the visual perception mechanism of organisms, and supports both supervised and unsupervised learning. The convolution kernel parameter sharing in the hidden layer and the sparsity of the inter-layer connection enable CNN to efficiently learn grid features, such as pixels and audio data, at a low computational cost, with stable results and without the need for complex feature engineering processes.

[0160] CNN is also particularly suitable for processing image-related machine learning problems. CNN simulates the image processing mechanism of the human brain's visual cortex and adopts characteristics such as local receptive field, weight sharing and pooling operations to effectively reduce the complexity of the network model and the number of weights. This enables CNN to process high-dimensional input data, such as image data, and has achieved remarkable success in tasks such as image recognition and classification. The basic structure of CNN mainly includes convolutional layers, activation functions, pooling layers and fully connected layers. Its principles are as follows: Figure 6As shown in the figure. During the training phase, the convolutional neural network uses the back propagation algorithm and the gradient descent method to update the weights, thereby improving the learning performance of the neural network. CNN has broad application prospects in image recognition, object detection, image segmentation, natural language processing, etc., and is a hot topic in current deep learning research.

[0161] Four models in this study: XGBoost, random forest, convolutional neural network, and deep neural network. The input variables are molecular structure feature values, and the output variables are predicted values ​​of the mole fraction solubility of CO2 in low eutectic solvents. Once the hyperparameters are determined, the performance of the machine model can be evaluated. In order to accurately evaluate the prediction performance of the machine learning model for the solubility of CO2 in low eutectic solvents, various evaluation indicators are used, including mean square error (MSE), mean absolute error (MAE), determination coefficient (R 2 ). The calculation formulas (4)-(6) of these indicators are as follows:

[0162]

[0163] The predicted and experimental values ​​of CO2 solubility in deep eutectic solvents are denoted by y t and f t , represents the average solubility of CO2 in the deep eutectic solvent, and N represents the number of samples.

[0164] (4) Hyperparameter Optimization

[0165] As a crucial preliminary preparation, the optimization of hyperparameters aims to ensure that the model can capture the complexity of data features while being simple enough to avoid overfitting. In the process of hyperparameter optimization, special attention is paid to the key parameters such as estimator, maximum depth, maximum features and learning rate. In order to find the potential optimal hyperparameter combination, a grid search method based on automatic spatial exhaustiveness is introduced to conduct detailed exploration within a preset range. This method uses a grid search based on automatic spatial exhaustiveness to explore the potential combination of optimal hyperparameters within the initial range. Figure 7 As shown, the process is systematically carried out with the complementation of a k-fold cross validation (CV) strategy.

[0166] During this process, 20% of the data is designated for testing, while the rest of the data is uniformly divided into k = 5 mutually exclusive subsets, called fold 1 to fold k, each containing the same number of samples. Each fold is used as a validation set in turn, while the remaining k-1 folds constitute the training set, and finally k iterations of training and validation are performed. Finally, the accuracy of the model is verified using the reserved test set to ensure a comprehensive evaluation of its performance. The optimization results are shown in Table 1-4.

[0167] (5) XGBoost, RF, CNN and DNN model regression results

[0168] XGBoost, RF, CNN and DNN models were used to predict the solubility of carbon dioxide in deep eutectic solvents, and the MSE, MAE and R 2 as a performance evaluation indicator. Figure 8-11 The regression plots of the prediction results are shown, and Table 5 summarizes the detailed results of the XGBoost, RF, CNN, and DNN model evaluations.

[0169] Depend on Figure 8-11 It can be seen that all four models can accurately predict the solubility of CO2 in deep eutectic solvents, and R 2 The CNN model has the best prediction effect on the test set, with an MSE as low as 0.000 and a MAE of 0.000. 2 The model maintains strong performance even on the training set, with an MSE of 0.000, MAE of 0.000, and R 2 The second best model in the test set is the XGBoost model, with an MSE as low as 0.000 and a MAE of 0.001. 2 The MSE of the training set is 0.984, the MAE is 0.000, and its R 2 As high as 0.997. In contrast, the RF model is worse on the test set, but R 2 It also reached 0.950, MSE reached 0.000, and MAE was 0.001. CNN showed the best performance in predicting carbon dioxide solubility due to its ability to more effectively utilize the spatial characteristics of data, automatically extract features, model complex relationships, optimize parameters and generalization capabilities, reduce dependence on data preprocessing and feature engineering, and better model structure and implementation.

[0170] The predictive performance using structural encoding as input is achieved by extracting the core features and relationships of the data and representing them in a more compact, low-dimensional form. This dimensionality reduction process helps reduce the complexity of the model and prevent overfitting while retaining enough information for accurate predictions. Therefore, the new structural encoding representation method can effectively predict the solubility of CO2 in low eutectic solvents. And among the XGBoost, RF, CNN, and DNN models that use structural encoding as input, the CNN model has the best fitting effect in predicting the solubility of CO2 in low eutectic solvents.

[0171] Table 1-5 is as follows:

[0172] Table 1 XGBoost model hyperparameter optimization results

[0173]

[0174]

[0175] Table 2 RF model hyperparameter optimization results

[0176]

[0177] Table 3 CNN model hyperparameter optimization results

[0178]

[0179] Table 4 DNN model hyperparameter optimization results

[0180]

[0181] Table 5 Evaluation results of XGBoost, RF, CNN and DNN models

[0182]

[0183] At this point, a method for predicting the solubility of CO2 in low eutectic solvents based on a machine learning model has been successfully completed.

[0184] 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 method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model, characterized in that: It includes the following steps: S1, data collection: CO2 solubility data points with a solubility range of 0.0014-0.0326 were collected, covering at least 13 deep eutectic solvents. The data set was randomly divided into a training set and a test set, with the training set accounting for 80% of the total data and the test set accounting for 20% of the total data; S2, deep eutectic solvent molecular structure coding; S3, model construction, training and evaluation; S4. Result analysis and model selection: compare the prediction performance of different models and select the optimal model; S5. Use the trained optimal model to predict the solubility of CO2 in low eutectic solvents under different conditions.

2. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 1, characterized in that: In step S1, the temperature range of the data set is 293.15 to 323.15 K, and the pressure range is 0.515 to 5.842 bar.

3. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 2, characterized in that: The specific steps of step S2 include: S21.Select tools: The open source cheminformatics software package RDKit was selected as a tool for obtaining the structure encoding of deep eutectic solvents; RDKit is an open source Python cheminformatics toolkit that is widely used for processing and calculating chemical structure data; S22. Get the structure code: Take the molecular structure of the low eutectic solvent as input; these molecular structures can be obtained from various chemical databases, literature or experimental data; use various functions and classes of RDKit to process molecular structures, including reading SMILES strings and parsing SDF files; The built-in algorithm of RDKit generates a corresponding structure code for each deep eutectic solvent; the structure code is a digital representation of the molecular structure, which is convenient for computer processing and analysis; S23. Transformation feature descriptor: The generated structural codes are further converted into feature descriptors. This step is to simplify the complex molecular structure information into a series of numerical features so that the machine learning algorithm can understand and process it. The structural code of each deep eutectic solvent is converted into 280 feature descriptors. These descriptors comprehensively cover multiple aspects of the molecule, ensuring that the complexity and diversity of the molecular structure can be fully expressed. S24. Data preprocessing: The generated feature descriptors are standardized and normalized to ensure comparability between different features and eliminate the impact of dimension. Check whether there are outliers or missing values ​​in the feature descriptors, fill in missing values ​​or remove abnormal data.

4. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 1, characterized in that: The step S3 includes a step S31 of using four machine learning algorithms to construct a model, wherein the four machine learning algorithms are: XGBoost algorithm, random forest RF algorithm, deep neural network DNN algorithm and convolutional neural network CNN algorithm, wherein: The XGBoost algorithm uses the extreme gradient boosting tree method to enhance classification capabilities by integrating multiple CART regression trees. XGBoost organically combines multiple CART regression trees to form a classifier. By continuously adding trees, each new tree corrects the prediction error of the previous tree. The model partitions the decision tree through a greedy algorithm and approximates the loss function quadratically through Taylor expansion. At the same time, regularization terms are introduced to control model complexity and prevent overfitting. The random forest RF algorithm achieves classification and regression functions by fusing multiple decision trees. RF constructs multiple decision trees through self-service sampling, and only uses some features for node segmentation during the generation of each tree to enhance the diversity and anti-overfitting ability of the model. The output of RF is the average of the prediction results of all trees. The deep neural network (DNN) algorithm processes complex nonlinear relationships through a multi-layer neural network structure. DNN contains an input layer, multiple hidden layers, and an output layer. Each layer consists of multiple neurons connected by weights and biases. DNN optimizes parameters through the back-propagation algorithm and gradient descent method to minimize prediction errors. The convolutional neural network (CNN) algorithm extracts features from images or data through convolutional layers and pooling layers. CNN simulates the processing mechanism of the visual cortex of the human brain and uses local receptive field, weight sharing and pooling operation characteristics to effectively reduce network complexity and stabilize learning effects. It extracts features by sliding the convolution kernel on the input data and passes them to subsequent layers for processing.

5. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 4, characterized in that: The step S31 also includes the following process: The grid search method combined with k-fold cross-validation is used to optimize the model's hyperparameters: the grid search method involves traversing the preset hyperparameter space to find the optimal hyperparameter combination; the k-fold cross-validation involves dividing the data set into k subsets, selecting k-1 subsets as training sets each time, and the remaining 1 subset as a validation set, repeating k times to evaluate the generalization ability of the model; the key hyperparameters to be optimized include: the number of decision trees: controls the number of trees in the model, affecting the complexity and prediction accuracy of the model; the maximum depth of the decision tree: controls the maximum number of layers of the tree, and too deep a tree can lead to overfitting; the learning rate: in the optimization algorithm, controls the step size of each weight update, affecting the learning speed and stability of the model.

6. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 4, characterized in that: The step S3 also includes a model training and evaluation process S32, wherein: S321, Model training: Use the training set to adjust the model parameters and develop a prediction model; by continuously optimizing the loss function, the model can accurately predict the solubility of CO2 in deep eutectic solvents; S322 Model Evaluation: Evaluate the performance of the model through the test set; use mean square error (MSE), mean absolute error (MAE) and determination coefficient (R) 2 As an evaluation indicator, it can comprehensively evaluate the prediction accuracy and generalization ability of the model.

7. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 6, characterized in that: The specific process of step S321 includes: The model is trained using the training subset data. During the training process, the model calculates the predicted value based on the input data and the current parameters, and calculates the error between the predicted value and the true value, i.e., the loss. Then, the error is passed from the output layer to the input layer layer by layer through the back propagation algorithm, and the model parameters are updated according to optimization algorithms such as gradient descent to reduce the error. This process is repeated many times until the performance of the model on the validation subset is no longer significantly improved, or the preset upper limit of training rounds is reached. After the training is completed, the trained model parameters are saved for subsequent prediction or further analysis.

8. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 6, characterized in that: The specific process of step S322 includes: Use the test set data divided in the initial stage as the input for evaluating the model; Use the trained model to predict the test set data and calculate the error between the predicted value and the true value; use the mean square error MSE, mean absolute error MAE and determination coefficient R 2 As an evaluation indicator; Mean square error (MSE): measures the average of the squares of the difference between the predicted value and the true value. The smaller the MSE, the more accurate the prediction. Mean absolute error (MAE): measures the average of the absolute values ​​of the difference between the predicted value and the true value. Similarly, the smaller the MAE, the more accurate the prediction. Coefficient of determination R 2 : Measures the extent to which the model explains the variability, R 2 The closer the value is to 1, the better the model fits the data.

9. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 8, characterized in that: The specific process of step S4 is as follows: S41 compares the MSE, MAE and R of four models XGBoost, RF, DNN and CNN on the test set 2 value: XGBoost: Very low MSE and MAE, R 2 Close to 1, indicating that the XGBoost model has good predictive ability; RF: Although the MSE and MAE of the RF model are also lower, R 2 The value is slightly lower than that of XGBoost and CNN, indicating that its prediction accuracy is slightly inferior; DNN: The DNN model also performs well on the test set, but is slightly inferior to XGBoost and CNN. CNN: The best performance among all models, with MSE and MAE almost 0, R 2 The value is extremely high, showing strong predictive ability; After analysis and comparison, the CNN model was finally selected as the final prediction model.

10. The method for predicting the solubility of carbon dioxide in a deep eutectic solvent based on a machine learning model according to claim 9, characterized in that: The specific process of step S5 is as follows: Detailed steps for using the trained optimal model CNN to predict the solubility of CO2 in deep eutectic solvents under different conditions: S51. Prepare to input data: Collect or generate corresponding input data according to the conditions that need to be predicted; these data include the structural code SC of the low eutectic solvent, the partial pressure of carbon dioxide PCO2 and the reaction temperature T; the structural code uses the chemical informatics software package RDKit to convert the molecular structure of the low eutectic solvent into a numerical form; ensure that the format of the input data is consistent with the data format used when training the model so that the model can parse and process it correctly; S52. Data preprocessing: For missing values ​​or outliers contained in the input data, appropriate processing is performed to fill in missing values ​​or remove outliers, and the input data is normalized or standardized to ensure that the prediction performance of the model is stable under different conditions; S53, Model loading and prediction: Load the trained optimal CNN model and input the prepared input data into the loaded CNN model; the model calculates and outputs the predicted CO2 solubility value based on the input low eutectic solvent structure code, carbon dioxide partial pressure and reaction temperature; S54. Interpretation and application of results: Interpret the CO2 solubility values ​​predicted by the model and evaluate their accuracy and reliability under specific conditions; apply the prediction results to the actual scenario of optimizing the formulation of low eutectic solvents and improving CO2 capture efficiency.

Images