Method for encoding ionic liquid characteristics based on graph neural networks

Abstract

The application relates to the technical field of ion liquid property prediction, and provides a method for encoding ion liquid characteristics based on a graph neural network, which comprises the following steps: 1) collecting SMILES names and physical and chemical property data of ion liquids; 2) constructing molecular graphs of cations and anions of the ion liquids to generate node and edge information; 3) fusing the node and edge information by using a graph neural network to generate molecular characteristics of the ion liquids; and 4) performing model training and model testing on the generated molecular characteristics with the physical and chemical property data of the ion liquids as a target. The application provides an ion liquid characteristic graph encoding generation technology which fuses atomic attributes and structural attributes, realizes direct fusion of atomic characteristics and structural characteristics of ion liquid components, and is simple in feature composition. A feature training model constructed based on the application can achieve better property prediction effect under the condition of consuming less computing resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ion liquid property prediction, and particularly relates to a method for encoding ion liquid characteristics based on a graph neural network. BACKGROUND

[0002] As a new type of electrolyte solution, lubricant and gas absorbent, ion liquid has great application prospects in the fields of energy, friction and environmental protection. Predicting the physical and chemical properties of ion liquid through an AI model can efficiently realize the industrial application of ion liquid. Reasonably constructing the feature set of ion liquid is a key step to affect the performance of the AI model. Ion liquid is composed of cations and anions. Each type of cation (or anion) has different atomic composition and molecular structure. Fusing these atomic compositions and molecular structures is a prerequisite for effectively representing the characteristics of ion liquid. Currently, the features used for constructing the AI model of ion liquid can be mainly divided into the following four categories: (1) constructing features based on the overall properties of ion liquid; (2) constructing features based on the constituent atoms of ion liquid; (3) constructing features based on the structural information of ion liquid; and (4) constructing features based on the environment of ion liquid. The specific technical solutions are as follows:

[0003] (1) Constructing features based on the overall properties of ion liquid

[0004] Features are constructed based on the cation, anion or ion liquid as a whole without considering the details such as molecular structure. Common descriptors include: cation type, anion type, whether containing a ring structure, ion liquid mass, cation charge, anion charge, carbon atom number, nitrogen atom number, ultraviolet spectrum wavelength, hydrogen bond number, etc. Common cation types include: imidazole, pyrrole, pyridine, quaternary ammonium salt, quaternary phosphonium salt, guanidine, etc. Common anion types include: halide, alkali salt, fluorine-containing salt, amino acid, ester, halide, etc. The cation type, anion type and whether containing a ring structure are converted into numerical type descriptors through zero-one encoding; the ion liquid mass, cation charge, anion charge, ultraviolet spectrum wavelength and hydrogen bond number can be calculated; and the carbon atom number and nitrogen atom number can be obtained through statistical calculation. The above several overall descriptors are selected for combination to form constraints for the feature representation of ion liquid and serve as input features of the AI model.

[0005] (2) Constructing features based on the constituent atoms of ion liquid

[0006] The feature construction method refines to the constituent atoms of the ionic liquid, and takes the attributes of the constituent atoms as feature descriptors. Common descriptors include: element symbol, atomic mass, atomic charge, hybridization type, atomic energy parameter, atomic size, whether a hydrogen bond acceptor, whether a hydrogen bond donor, whether an atom on a ring, etc. Element symbols, whether a hydrogen bond acceptor, whether a hydrogen bond donor, whether an atom on a ring, etc. Text type descriptors are converted into numerical type descriptors recognizable by a computer through zero-one coding and other methods; atomic mass, atomic charge, hybridization type, atomic energy parameter, atomic size can be obtained by calculation. Select some or all of the above atomic attribute features for combination as the features of the ionic liquid for AI model training.

[0007] (III) Constructing features based on the structural information of ionic liquids

[0008] The structural information of the ionic liquid refers to the bonding relationship between the cation and the anion, which is an important factor affecting the properties of the ionic liquid. The same constituent atoms often exhibit different physical and chemical forms due to different bonding modes. Currently, the structural information of the ionic liquid is usually constructed by an indirect method, such as using Cartesian coordinates of the ionic liquid, internal coordinates of the ionic liquid, and distances between constituent atoms of the ionic liquid. Cartesian coordinates and internal coordinates can represent the three-dimensional spatial information of the constituent atoms, and the distance between the constituent atoms describes the structure of the ionic liquid in one dimension. Both can be obtained by calculation.

[0009] (IV) Constructing features based on the environment of the ionic liquid

[0010] The external environment of the ionic liquid mainly refers to temperature and pressure. The same ionic liquid will exhibit different physical and chemical properties under different temperature and pressure conditions. Temperature and pressure are usually used as necessary external environmental features as inputs to the AI model, and both can be determined by experimental methods.

[0011] The existing ionic liquid feature construction methods mainly have the following shortcomings: (1) The methods of constructing features based on the overall properties of the ionic liquid and the constituent atoms of the ionic liquid require a large number of features to improve the performance of the model. The original feature dimension can reach hundreds, making it difficult for model training and parameter adjustment; (2) The methods of constructing features based on the structural information of the ionic liquid still use indirect methods such as atomic coordinates and atomic distances, which cannot describe the bonding information, bonding properties, and mutual influence between bonded atoms; (3) Due to the different total number of atoms contained in different ionic liquids, the above feature construction methods require padding of the original features before they can be used for model training.

[0012] Generally, the existing ionic liquid feature construction method lacks the fusion processing of atomic, molecular attribute features and structure information features, resulting in the dependence on a large number of features to improve the performance of the model. The feature construction process is complex, time-consuming, and the model training is difficult, and the parameter adjustment is time-consuming. SUMMARY

[0013] The present application aims to solve at least one of the above technical problems, and provides a method for encoding ionic liquid features based on a graph neural network. The present application provides a ionic liquid feature graph encoding generation technology that fuses atomic attributes and structural attributes, which realizes the direct fusion of ionic liquid composition atomic features and structural features: using a graph neural network to construct the molecular graph of the cation and anion of the ionic liquid, and fusing the features of the constituent atoms and the bonding information between the atoms into molecular features through information transmission technology, linearly combining the molecular features of the cation, the molecular features of the anion, and the environmental features of the ionic liquid as the features of the ionic liquid, and training a machine learning model for property prediction. The technical solution provided by the present application is as follows:

[0014] The present application provides a method for encoding ionic liquid features based on a graph neural network, comprising the following steps:

[0015] 1) Collecting the SMILES name and physical and chemical property data of the ionic liquid;

[0016] 2) Constructing the molecular graph of the cation and anion of the ionic liquid to generate node and edge information;

[0017] 3) Graph neural network fusion encoding node and edge information to generate molecular features of the ionic liquid;

[0018] 4) Model training on the generated molecular features with the ionic liquid physical and chemical property data prediction as the target.

[0019] Preferably, in step 1), the physical and chemical properties of the ionic liquid include the density, viscosity, conductivity, critical temperature, critical pressure, and Henry coefficient of the ionic liquid; and the physical and chemical property data of the ionic liquid is derived from the ionic liquid database ILThermo v2.0.

[0020] Preferably, step 2) is specifically: constructing a molecular graph for the cation and anion of the ionic liquid respectively to generate node information for describing the constituent atoms and edge information for describing the bonding relationship between the atoms.

[0021] Preferably, in step 2), the chemical informatics software package RDKit is used to construct the molecular graph for the cation and anion of the ionic liquid respectively.

[0022] Preferably, the node information selects the element characteristics of the constituent atoms as descriptors, and converts them into a feature matrix recognizable by a computer through zero-one coding; the edge information includes both the characteristics of the edge and the index of the edge, the characteristics of the edge take single bond, double bond, triple bond, and ring bond as descriptors, and are converted into a feature matrix recognizable by a computer through zero-one coding, and the index of the edge is used to record the atom number information of the bonded atoms.

[0023] Preferably, the index of the edge adopts COO storage format to record the atom number information of the bonded atoms.

[0024] Preferably, the index of the edge is divided into two rows, the first row is used to record the atom number information of the starting end of the bonded atoms, and the second row is used to record the atom number information of the ending end of the bonded atoms.

[0025] Preferably, in step 3), the graph neural network fuses the encoding information of the nodes and the edges, specifically: an AttentiveFP graph neural network framework is built, AttentiveFP is divided into two steps when fusing the node information and the edge information: the first step is an atom encoding stage, respectively taking each constituent atom as the center, and according to the index of the edge, the information of the surrounding atoms is fused into the center atom; the second step is a molecule encoding stage, a center representing a cation or an anion is virtually set, all the encoded atoms are connected to the center, the center gathers all the atom characteristics representing the characteristics of the cation or the anion, and the characteristics include the element properties of the constituent atoms and the bonding properties.

[0026] Preferably, in step 3), the AttentiveFP graph neural network framework is built by using a PYG software package.

[0027] Preferably, in step 3), the generated characteristics of the ionic liquid are specifically: after generating the cation and anion characteristics of the element properties of the constituent atoms and the bonding properties, the cation characteristics, the anion characteristics, and the environmental characteristics are linearly combined into the molecular characteristics of the ionic liquid.

[0028] Preferably, in step 4), the model training is: the molecular characteristics of the ionic liquid are one-to-one corresponding to the physical and chemical properties to form a data set, the data set is divided into a training set and a test set; the ionic liquid data on the training set is input into a machine learning model for training to obtain a trained model; the algorithm of the machine learning model includes: a random forest algorithm (Random Forest, RF), an XGBoost algorithm, a decision tree algorithm (Decision Tree, DT), and a gradient boosting regression tree algorithm (Gradient Boost Regression Tree, GBRT).

[0029] Preferably, step 4) further comprises model testing: a test set is used to verify the effect of the model, the molecular characteristics of the ionic liquid on the test set are input into the trained model for property prediction, the model algorithm includes random forest algorithm, XGBoost algorithm, decision tree algorithm, gradient regression tree algorithm, etc., according to the model evaluation parameters such as the determination coefficient, the mean square error, the mean absolute error, the prediction performance of the model is obtained, and the promotion effect of the molecular characteristics of the ionic liquid based on the graph neural network on the model performance is verified.

[0030] Preferably, step 4) further comprises model optimization: according to the evaluation parameters such as the determination coefficient, the mean square error, and the mean absolute error, the hyperparameter space of the prediction model such as random forest, XGBoost, decision tree, and gradient regression tree is optimized, so that the value of the determination coefficient is as close to 1 as possible, and the values of the mean square error and the mean absolute error are as close to 0 as possible.

[0031] Compared with the prior art, the present application can achieve the following beneficial effects:

[0032] 1) The feature composition is simple, only the element name and bonding information of the constituent atom are needed, avoiding the cumbersome feature generation process of the method of constructing features based on the overall properties of ionic liquids and the method of constructing features based on the constituent atoms of ionic liquids, and avoiding the feature completion operation of both methods.

[0033] 2) The atomic properties and structural properties are directly fused, avoiding the insufficient expression of structural information in the method of constructing features based on the structural information of ionic liquids.

[0034] 3) The feature training model constructed based on the present application can accurately predict the physical and chemical properties of ionic liquids, and can achieve better property prediction effect with less computing resources. In four experimental models, the overall prediction effect of the test set is as follows: the determination coefficient is greater than 0.98, the mean square error is less than 0.06, and the mean absolute error is less than 0.12. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The flow chart of the method for encoding the features of ionic liquids based on the graph neural network of the preferred embodiment of the present application;

[0036] Figure 2 The schematic diagram of the molecule of the preferred embodiment of the present application;

[0037] Figure 3 The schematic diagram of the atom coding of the preferred embodiment of the present application;

[0038] Figure 4 The schematic diagram of the molecule coding of the preferred embodiment of the present application;

[0039] Figure 5The ion liquid molecular feature constructed based on the graph neural network of the preferred embodiment of the application has a conductivity prediction effect on four algorithm models. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0041] As Figures 1 to 4 described, the preferred embodiment of the application provides a method for encoding ion liquid features based on a graph neural network, comprising the following steps:

[0042] (1) Ion liquid data collection

[0043] The SMILES name and physical and chemical property data of the ion liquid are derived from the ion liquid database ILThermo v2.0 (https: / / ilthermo.boulder.nist.gov / ), and the physical and chemical properties of the ion liquid include the density, viscosity, conductivity, critical temperature, critical pressure, Henry coefficient and the like of the ion liquid. Taking the conductivity property prediction as an example, the ion liquid conductivity properties and the corresponding cation and anion SMILES names reported in the literature are mainly collected, and the total data is 5715, involving 183 cations and 85 anions, and the temperature range is 208.15K-528.55K.

[0044] (2) Ion liquid molecular graph construction

[0045] The chemical information software package RDKit (https: / / rdkit.org / ) is used to construct a molecular graph for the cation and anion of the ion liquid, including node information for describing constituent atoms and edge information for describing the bonding relationship between atoms.

[0046] For the node information, the application only selects the element characteristics of the constituent atoms as the descriptor, which is converted into a feature matrix recognizable by a computer through a zero-one encoding method, as shown in Figure 2 For i atoms and j atoms, the node information is identified as X i ,X j . For the edge information, it includes edge features and edge indexes. Among them, the edge features take single bond, double bond, triple bond and ring bond as descriptors, which are converted into a feature matrix recognizable by a computer through a zero-one encoding method, as shown in Figure 2 The edge information between i atoms and j atoms is eij The index of the edge is used to record the atom serial number information of the bond, adopts the COO storage format, and is divided into two lines: the first line is used for recording the atom serial number of the starting end of the bond, and the second line is used for recording the atom serial number of the ending end of the bond, as shown in the following formula: Figure 2 The index of the edge in the application is represented by I. In the application, the unordered graph is constructed for the cation and the anion, that is, the direction of the bond is not distinguished, and the attribute information can be transmitted between the two atoms of the bond.

[0047] (3) Generation of ionic liquid features

[0048] In the application, the graph neural network is adopted to deeply fuse the node information and the edge information according to the molecular graph of the cation and the anion of the ionic liquid, so as to generate the cation and anion features containing the attribute of the constituent atom element and the bonding property. The cation feature, the anion feature and the environmental feature (temperature, pressure) are linearly combined into the molecular feature of the ionic liquid.

[0049] In the application, the AttentiveFP graph neural network framework is built by using the PYG software package. When the node information and the edge information are fused, AttentiveFP is divided into two steps: the first step is the atom encoding stage, the information of the surrounding atoms is fused into the center atom (hi) according to the index of the edge, respectively taking each constituent atom as the center. Figure 3 hi is the center atom, hj is the connected surrounding atom, and the message transmission is to transmit the information of the surrounding atom hj to the center atom hi; the atom information hi (0) is obtained after T times of message transmission T ); the second step is the molecular encoding stage, a virtual center representing the cation or the anion is set, all the encoded atoms have a connection relationship with the virtual center, and the information transmission is performed (as shown in the following formula: Figure 4 The center collects all the atom features to represent the features of the cation (or the anion). The cation feature, the anion feature and the environmental feature of the ionic liquid are linearly combined to form the molecular feature of the ionic liquid.

[0050] (4) Effect test of ionic liquid features

[0051] The molecular characteristics of the ionic liquid are corresponded to the physical and chemical properties to form a data set, and the data set is divided into a training set and a test set. The ionic liquid data on the training set is input into the machine learning model for training to obtain a trained model. The effect test of the model is completed on the test set: first, the molecular characteristics of the ionic liquid in the test set are input into the trained machine learning model for property prediction, and the advantages and disadvantages of the prediction performance are judged through the evaluation parameters such as the coefficient of determination, mean square error and mean absolute error, which are used to evaluate the promotion effect of the ionic liquid characteristics based on the graph neural network on the model performance. In the present application, four common machine learning models are selected for training and testing, which are random forest algorithm (Random Forest, RF), XGBoost algorithm (extreme gradient boosting), decision tree algorithm (Decision Tree, DT) and gradient boosting regression tree algorithm (Gradient Boost Regression Tree, GBRT). According to the numerical value of the evaluation parameter, the hyperparameter space of RF, XGBoost, DT and GBRT is adjusted, so that the value of the coefficient of determination is close to 1 on the training set and the test set, and the values of the mean square error and the mean absolute error are close to 0, the model is optimized, and the best performance model is obtained.

[0052] The conductivity of the ionic liquid is predicted to test the effect of the ionic liquid characteristics. The specific steps and prediction results are as follows:

[0053] (1) From the ionic liquid thermodynamic property database "ILThermo v2.0", select conductivity as the label, download the corresponding SMILES file and conductivity data, and after data arrangement and cleaning, a total of 5715 ionic liquid conductivity data.

[0054] (2) According to the SMILES files of the cations and anions of the ionic liquid, the molecular graphs of the cations and anions are constructed, the node characteristics of the molecular graphs are composed of the element properties of the constituent atoms, and the edge characteristics are composed of single bond, double bond, triple bond, ring bond and other descriptors. Based on the AttentiveFP graph neural network, the node characteristics and edge characteristics are fused to form the molecular characteristics of the cations and anions. The molecular characteristics of the cations and anions, temperature and pressure are linearly combined to form the characteristics of the ionic liquid, which are corresponded to the conductivity to form a data set.

[0055] (3) The data set is divided into a training set and a test set according to 9:1, that is, 90% of the data is randomly extracted as the training set, and the remaining 10% of the data is as the test set. XGBoost, decision tree, random forest, gradient boosting regression tree algorithm are selected to train the model, which is recorded as GNN+XGBoost, GNN+DT, GNN+RF, GNN+GBRT. Through the hyperparameter space of the above algorithm, the model is optimized. When the determination coefficient of the model on the training set and the test set is greater than 0.98, and the mean square error and the mean absolute error are less than 0.15, it is determined that the model parameter optimization is successful, and the optimized model is output.

[0056] (4) The conductivity prediction results of the four models on the training set and the test set under the ionic liquid characteristic conditions constructed by the application are as shown in Figure 5 It can be seen from Figure 5 that the ionic liquid characteristics constructed by the method of the application can significantly improve the prediction effect of the conductivity of ionic liquids under the premise of using only simple descriptors.

[0057] (5) Further obtained, the R2, MSE and MAE of the ionic liquid characteristics constructed by the application on the test set are as shown in Table 1.

[0058] Table 1. Prediction effect of ionic liquid characteristics constructed by the application on the test set

[0059]

[0060] It can be seen from the table that the determination coefficient on the test set of the four models using the characteristics of the application is greater than 0.98, the mean square error is less than 0.06, and the average absolute error is less than 0.12, which indicates that the ionic liquid characteristics constructed based on the application have excellent effect on the improvement of the prediction performance of the model.

[0061] The application has been tested, simulated and used, and the results prove that the effect of the application is good, the feature construction is more simple and efficient, and the conductivity property of ionic liquids can be accurately predicted.

[0062] Finally, it should be noted that: the above only describes the preferred embodiments of the application and is not used to limit the application, although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A method of encoding ionic liquid features based on graph neural networks, characterized in that, The method comprises the following steps: 1) collecting SMILES names and physical and chemical property data of ionic liquids; 2) constructing molecular graphs of cations and anions of ionic liquids to generate node and edge information; 3) fusing and encoding the node and edge information by using a graph neural network to generate molecular features of the ionic liquid; The graph neural network fuses and encodes the node and edge information, specifically: an AttentiveFP graph neural network framework is built, and AttentiveFP is divided into two steps when fusing node information and edge information: the first step is an atomic encoding stage, and the information of surrounding atoms is fused into the center atom according to the index of the edge, with each component atom as the center; the second step is a molecular encoding stage, and a center representing a cation or anion is virtually set, and all the encoded atoms are connected to the center, and the center gathers all the atomic features representing the features of the cation or the anion, and the features include element attributes and bonding properties of the component atoms; the generated features of the ionic liquid are linearly combined into molecular features of the ionic liquid after generating the cation and anion features of the element attributes and bonding properties of the component atoms; 4) training a model by using the generated molecular features as a target to predict the physical and chemical property data of the ionic liquid.

2. The method for encoding ionic liquid features based on graph neural networks according to claim 1, characterized in that, In step 1), the physical and chemical properties of the ionic liquid include the density, viscosity, conductivity, critical temperature, critical pressure and Henry coefficient of the ionic liquid; and the physical and chemical property data of the ionic liquid is obtained from an ionic liquid database ILThermo v2.

0.

3. The method for encoding ionic liquid features based on graph neural networks according to claim 1, characterized in that, In step 2), the molecular graphs of the cations and anions of the ionic liquid are constructed respectively to generate node information for describing the component atoms and edge information for describing the bonding relationship between the atoms.

4. The method for encoding ionic liquid features based on graph neural networks according to claim 3, characterized in that, The node information selects element features of the component atoms as descriptors, and converts them into a feature matrix recognizable by a computer through a zero-one coding mode; the edge information includes two aspects of edge features and edge indexes, the edge features take single bond, double bond, triple bond and ring bond as descriptors, and are converted into a feature matrix recognizable by a computer through a zero-one coding mode, and the edge indexes are used to record the bonding atom serial number information.

5. The method for encoding ionic liquid features based on graph neural networks according to claim 4, characterized in that, The edge indexes are divided into two rows, the first row is used to record the bonding starting atom serial number, and the second row is used to record the bonding ending atom serial number.

6. The method for encoding ionic liquid characteristics based on graph neural networks according to claim 1, characterized in that, In step 4), the model training is performed by one-to-one correspondence between the molecular features of the ionic liquid and the physical and chemical properties to form a data set, and the data set is divided into a training set and a test set; the ionic liquid data on the training set is input into a machine learning model for training to obtain a trained model; and the algorithm of the machine learning model includes a random forest algorithm, an XGBoost algorithm, a decision tree algorithm and a gradient boosting regression tree algorithm.

7. The method of encoding ionic liquid characteristics based on graph neural networks according to claim 6, characterized in that, The step 4) further comprises model testing: a test set is used to verify the effect of the model, the molecular characteristics of the ionic liquid on the test set are input into the trained model to predict the properties, the model algorithm comprises a random forest algorithm, an XGBoost algorithm, a decision tree algorithm and a gradient regression tree algorithm, and the prediction performance of the model is obtained according to the model evaluation parameters including a determination coefficient, a mean square error and a mean absolute error.

8. The method of encoding ionic liquid characteristics based on graph neural networks according to claim 7, characterized in that, The step 4) further comprises model optimization: according to the model evaluation parameters, the hyperparameter space of the machine learning model algorithm is combined and optimized to obtain the best model; the standard of the model evaluation parameters is that the closer the value of the determination coefficient to 1, the better the performance of the model; the closer the values of the mean square error and the mean absolute error to 0, the better the performance of the model.

Images