Method for predicting solubility of hydrogen sulfide gas in ionic liquid

Abstract

The invention discloses a method for predicting the solubility of hydrogen sulfide gas in ionic liquid. The method comprises the following steps of data collection, data preprocessing, feature engineering, model establishment and prediction. The method is carried out through two ways: in the former way, a series of characteristics which are generated based on an Rdkit molecular fingerprint method and comprise operation parameters P and T and a solubility data set are input into a random forest model for prediction; and the gas solubility direct prediction is carried out on the molecular descriptor characteristics, the operation parameters P and T and the solubility data set by using new characteristic combination parameters generated by an AutoFeatRegressor method. By utilizing a machine learning method, prediction of the solubility of H2S gas in the ionic liquid at different temperatures and different pressure intensities is finally realized, and the development period of screening an ionic liquid absorbent to treat hydrogen sulfide is remarkably shortened.

Description

Technical Field

[0001] The invention relates to a method for predicting the solubility of hydrogen sulfide gas in an ionic liquid, and relates to a method for quickly predicting the solubility by using Rdkit to generate molecular descriptors, using Feature Selector to perform feature screening, and using Random Forest to predict the solubility or AutoFeatRegressor to obtain a feature combination calculation formula. Background Art

[0002] Hydrogen sulfide (H2S) is a colorless, highly toxic acidic gas. It is an inorganic compound and a byproduct released during the production of many drugs. The emission of hydrogen sulfide will cause air pollution, affect the ecological environment, and pose a serious threat to human health. Currently, the commonly used methods for removing hydrogen sulfide in industry are mainly chemical absorption and physical adsorption. Although these methods are easy to operate and stable in operation, they have low adsorption selectivity and high raw material consumption, and are not the best way to remove hydrogen sulfide.

[0003] Ionic liquids (ILs) are low vapor pressure, non-flammable molten salts at room temperature with excellent physical and chemical properties. They are suitable for multifunctional applications such as solvents, electrolytes and catalysts. Their high solubility and low volatility are conducive to reducing equipment scale and environmental pollution, and their economic efficiency is better than traditional methods.

[0004] Machine Learning (ML) is a subfield of artificial intelligence that focuses on developing algorithms that can automatically learn from data. This approach enables computers to automatically extract knowledge from experience (data), optimize themselves, and make accurate predictions on new data. Using ML technology, ionic liquids with high absorption performance for hydrogen sulfide can be predicted and screened.

[0005] During the project implementation, we collected and sorted out the solubility data of hydrogen sulfide gas in ionic liquids at certain temperatures and pressures. We built a prediction model using machine learning algorithms such as random forest, Lasso regression, SVR and MLP. By evaluating multiple sets of performance indicators of the model, random forest was selected as the final algorithm due to its superior fitting effect. This method significantly reduced the development cycle of screening ionic liquid absorbents for treating hydrogen sulfide. Summary of the invention

[0006] Purpose of the invention: The present invention aims to provide a method based on machine learning Rdkit to generate molecular descriptors as features, use Feature Selector for feature screening, and use AutoFeatRegressor for feature regression and random forest to predict the solubility of hydrogen sulfide gas in various ionic liquids at different temperatures and pressures, providing a convenient and feasible prediction method.

[0007] Technical solution: The present invention provides a method for predicting the solubility of hydrogen sulfide gas in an ionic liquid, the method comprising the following steps:

[0008] (1) Data collection: Collect the solubility of H2S gas in different ionic liquids at different temperatures and pressures, and organize the SMILES expressions of ionic liquid anions and cations;

[0009] (2) Data preprocessing: normalize the data according to the molar solubility of H2S gas in different ionic liquids, process the outliers in the data, and use visualization methods to view the distribution of the data;

[0010] (3) Feature engineering: Based on the collected SMILES expressions of ionic liquids, the Rdkit molecular fingerprint method is used to generate molecular descriptors as features. All features including operating parameters P and T are screened using Feature Selector, and single feature values ​​are eliminated to obtain the final data set that can be used for model prediction. The important features in the data set are calculated and analyzed.

[0011] (4) Establishing a model for prediction: Method 1: Using the random forest method in machine learning, the data set obtained in step (3) is divided to establish a solubility prediction model for hydrogen sulfide in different ionic liquids, and the operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are input into the solubility prediction model for data training, verification and / or prediction; or, Method 2: The molecular descriptor features, operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are combined with new feature parameters generated by the AutoFeatRegressor method to directly predict the gas solubility.

[0012] Preferably, in step (1), the collection method includes consulting literature.

[0013] Preferably, in step (1), anions and cations of different ionic liquids are input into InDraw software to obtain SMILES expressions of different anions and cations.

[0014] Preferably, in step (2), data preprocessing includes the following steps:

[0015] (I) collating the collected solubility data, converting all the solubility data into molar solubility data values, converting the obtained molar solubility data values ​​into units to obtain gas solubility percentages, normalizing the obtained gas solubility percentages to obtain solubility values ​​between 0 and 1, and verifying the normalized solubility data values ​​to ensure that the data are authentic and valid;

[0016] (II) performing abnormal processing on the collected data. If the value of a characteristic variable under a certain dissolution condition exceeds or is lower than 100% of the range of the remaining data values ​​in the collected data set, the abnormal data value is selected for removal;

[0017] (III) After confirming that all the data are correct, visualize the data by drawing a scatter plot, box plot, or line graph to intuitively present the relationship between the feature value and the target value for further analysis.

[0018] More preferably, in step (I), the normalization process uses OmicShare Tools, Python, R language or Excel software.

[0019] More preferably, in step (II), Isolation Forest is introduced by python to perform an anomaly detection algorithm on the collected data to process abnormal data.

[0020] More preferably, in step (III), the visualization process uses matplotlib, yellowbrick or R language in Python.

[0021] Preferably, in step (3), feature engineering includes the following steps:

[0022] (I) The processed data were analyzed using the RDKit module, and the Rdkit molecular fingerprint was produced using the CalcDescriptors method. The molecular feature descriptors of the compounds were obtained using the MolFromSmiles method in combination with the collected ionic liquid SMILES.

[0023] (II) Organize the molecular feature descriptor data and fill missing values ​​with 0 to ensure that the data is complete and usable;

[0024] (III) Combine the molecular feature descriptors and the operation parameters P and T, use Feature Selector to perform feature screening, remove single feature values, and retain the remaining important features;

[0025] (IIII) The SHAP library was used to plot the importance of different characteristic variables on the solubility of H2S gas in ionic liquids, and independence and dependence analysis of important features was performed based on the results.

[0026] Preferably, in step (4), the machine learning method described in method 1 is random forest (RF), and the data set obtained in step (3) is divided into a training data set and a test data set in a ratio of 7:3.

[0027] Preferably, in step (4), the method described in method 2 uses the AutoFeatRegressor method in the autofeatselect library to combine feature values ​​to obtain feature parameters, and the new feature parameters are formed by combining the molecular descriptor and the operating parameters P and T through a mathematical formula.

[0028] Preferably, in step (4), the parameters used to evaluate the predictive ability of the quantitative prediction model are: decision coefficient R2, mean absolute error MAE and mean square error MSE, and the distribution is shown in formulas (1)-(3):

[0029]

[0030] Among them, Y i is the actual value, is the predicted value, is the average of the actual values, and N represents the number of samples;

[0031] Among them, MAE and MSE are error indicators. The smaller the value, the smaller the prediction error and the higher the model stability. R2 is an indicator for evaluating the goodness of fit. The closer the value is to 1, the better the model fit.

[0032] R2 stands for the coefficient of determination, also known as goodness of fit or coefficient of determination, which can be used to judge the fitting performance of the model. However, it also has certain limitations. When there are too many independent variables, R2 will be too high due to overfitting of the model and lose its reference value. Therefore, the error indicators MAE and MSE are introduced. MAE can measure the average value of the absolute error between the predicted value and the true value, and intuitively reflects the actual situation of the predicted value error. MSE averages the sum of the squares of the absolute errors and is more sensitive to situations with larger errors.

[0033] Preferably, in step (4), the prediction method 1 includes inputting the molecular descriptor characteristic value, pressure and temperature of the target ionic liquid into a random forest program to predict the solubility of hydrogen sulfide in the ionic liquid at the temperature and pressure; method 2 includes inputting the molecular descriptor characteristic value, pressure and temperature of the ionic liquid into AutoFeatRegressor to obtain a solubility formula formed by new feature combination parameters, and predicting the solubility of hydrogen sulfide in the ionic liquid at the temperature and pressure.

[0034] Principle of the invention: The present invention uses Rdkit to generate molecular descriptors based on the SMILES expressions of anions and cations of different ionic liquids, and uses temperature and pressure as features to predict gas solubility, in order to find ionic liquids with higher solubility in hydrogen sulfide gas, improve the long and laborious development cycle of ionic liquids in the absorption field, and provide a scientific approach to design ionic liquids.

[0035] Different from traditional methods of hydrogen sulfide removal such as adsorption removal, membrane separation and wet scrubbing, the present invention introduces ionic liquids to absorb hydrogen sulfide, generates a series of Rdkit molecular fingerprints for SMILES of different ionic liquids, and uses machine learning methods to predict their solubility at different temperatures and pressures.

[0036] Beneficial effects: Compared with the existing hydrogen sulfide absorption method, the present invention has the following significant advantages:

[0037] The method of the present invention uses ionic liquid to treat hydrogen sulfide gas at low cost, high efficiency and is environmentally friendly. The use of two machine learning methods to establish and predict models for a large amount of solubility data collected effectively reduces the blindness of ionic liquid design, thereby reducing raw material consumption and capital costs, and shortening the development cycle of high-efficiency ionic liquid absorbents. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of a method for predicting the solubility of hydrogen sulfide gas based on various ionic liquids at different temperatures and pressures according to the present invention;

[0039] Figure 2 The molar mass of ionic liquid anions and cations, operating conditions P, T and solubility data distribution diagram;

[0040] Figure 3 To explain the feature importance ranking of random forest models using SHAP;

[0041] Figure 4 Independence dependency graph for important features;

[0042] Figure 5 Comparison of predictions and true values ​​and error distribution diagram for training and testing of the random forest model in method 1;

[0043] Figure 6 Comparison of predictions and true values ​​and error distribution of Autofeatregressor training and testing in method 2. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below according to the flow chart.

[0045] Example 1

[0046] This embodiment provides a prediction method for predicting the solubility of various ionic liquids in hydrogen sulfide gas based on the Rdkit molecular fingerprint generation method at different temperatures and pressures. The flow chart is as follows: Figure 1 As shown, the following steps are included:

[0047] Step 1, data collection:

[0048] (1) By consulting the literature and collating the actual measurement data in the laboratory, the solubility data of different ionic liquids for hydrogen sulfide gas at different temperatures and pressures were collected: including the structure of the anion, structure of the cation, SMILES expression, molar mass (Molecular Wt), temperature (T), pressure (P), and gas solubility (Solubility) of different ionic liquids to form a data set.

[0049] (2) The solubility data were sorted out and the solubility data of different units in the literature were converted into mol / mol, that is, the number of moles of hydrogen sulfide dissolved per mole of ionic liquid. The molar solubility data were then converted into the percentage of hydrogen sulfide gas in the ionic liquid and normalized using Excel software to obtain a solubility value between 0 and 1. The obtained data information was preliminarily verified to ensure that the data was authentic and valid.

[0050] Step 2, data preprocessing:

[0051] (1) After introducing Isolation Forest in Python to perform anomaly detection on the collected data, we removed duplicate ionic liquids and those with small data volumes (if a characteristic variable value under a certain dissolution condition exceeds or is lower than 100% of the range of the remaining data values, the abnormal data value is removed). Finally, we collected a total of 1908 hydrogen sulfide gas solubility data values ​​of 49 different ionic liquids at different temperatures and pressures.

[0052] (2) The distribution range of the collected variables, anion molar mass, cation molar mass, temperature, pressure, and solubility, was visualized using matplotlib in Python. The specific distribution is as follows Figure 2 shown.

[0053] Step 3, feature engineering:

[0054] (1) Based on the anion SMILES expression and cation SMILES expression of the ionic liquid collected in step 1, the data were analyzed and processed using the RDKit module, and the Rdkit molecular fingerprint was generated using the MolFromSmiles method to obtain a new data set containing 208 molecular descriptor features. The molecular descriptor data was sorted and missing values ​​were filled with 0 to ensure that the data was complete and available.

[0055] (2) The newly generated molecular descriptor features were preliminarily screened using the Feature selector library in Python software (combining molecular descriptors and operating parameters P and T), and 68 single-variable molecular descriptor features were removed, leaving 140 molecular descriptor features.

[0056] (3) Calculate feature importance:

[0057] The collected data sets were used to calculate and plot the order of the importance of different molecular descriptor feature variables on the solubility of H2S gas in ionic liquids using the SHAP library of Python software. The results are shown in Figure 3 .

[0058] like Figure 3 As shown, under different dissolution conditions, temperature and pressure have a greater impact on the solubility of hydrogen sulfide gas. Temperature and pressure are the key factors affecting solubility, which is consistent with the objective laws of reality. Under the same temperature and pressure, the solubility mainly depends on the structure of different ionic liquids, that is, the generated molecular descriptors.

[0059] (4) Perform independence and dependency analysis on important features based on the ranking results of feature importance:

[0060] like Figure 4 As shown, partial_dependence, pdp, info_plots and PartialDependenceDisplay modules are introduced based on Figure 3 The 20 important features that affect the solubility of H2S gas in ionic liquids are analyzed for independence and dependence one by one. When analyzing the effect of one feature on the solubility of H2S gas, the values ​​of other features will be marginalized. Figure 4 As shown in the figure, when the line in the figure is upward, the solubility value will increase, and vice versa; when the image tends to be flat, it means that the feature has little effect on solubility at this time. Step 4, build a model:

[0061] (4) Establishing a model for prediction: Method 1: Using the random forest method in machine learning, the data set obtained in step (3) is divided to establish a solubility prediction model for hydrogen sulfide in different ionic liquids, and the operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are input into the solubility prediction model for data training, verification and / or prediction; or, Method 2: The molecular descriptor features, operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are combined with new feature parameters generated by the AutoFeatRegressor method to directly predict the gas solubility.

[0062] Method 1: First, the obtained data set is divided into a training set and a test set with a ratio of 7:3. The random forest (RF) method is used to construct a quantitative prediction model of whether the Rdkit molecular fingerprint method is contained. A series of characteristics of different ionic liquids at different temperatures and pressures are substituted for prediction.

[0063] And compare the changes in prediction accuracy of different methods (LASSO, MLP, SVR).

[0064] The model was established through machine learning methods and the decision coefficient R2, mean absolute error MAE and mean absolute error MSE were calculated to evaluate the model.

[0065] The results are shown in Table 1 and Figure 5 It was found that the various values ​​of the random forest model showed that the model had good stability and prediction accuracy.

[0066] Table 1 R2, MAE and MSE of random forest model

[0067]

[0068] like Figure 5 As shown, under the condition of containing Rdkit molecular descriptors, the R2 of the test set of H2S gas solubility in ionic liquids based on step 4 is 0.9999, proving that the established quantitative prediction model is acceptable.

[0069] The method described in method 2 uses the AutoFeatRegressor method in the autofeatselect library to combine feature values ​​to obtain feature parameters. The new feature parameters are composed of molecular descriptors and operating parameters P and T through mathematical formulas.

[0070] Method 2: Use the AutoFeatRegressor method in the autofeatselect library to process the 140 selected features together with temperature and pressure, and generate 51 new features composed of linear combinations through the Autofeatregressor method, as shown below:

[0071] ['log(MolMR) / T','P*log(Chi3n)','sqrt(P)*PEOE_VSA7','sqrt(P)*SlogP_VSA1',

[0072] 'PEOE_VSA1**3 / MolLogP','PEOE_VSA12**3*log(P)','FractionCSP3**3*sqrt(P)',

[0073] 'sqrt(PEOE_VSA6)*SlogP_VSA1**3','VSA_EState9**3*sqrt(fr_unbrch_alkane)','sqrt(P)*qed**2',

[0074] 'SMR_VSA3**2 / Ipc','MolLogP*fr_NH0**3','P**2*PEOE_VSA12**3','sqrt(P)*exp(fr_COO)',

[0075] 'PEOE_VSA12**3*VSA_EState9','SlogP_VSA1**3*sqrt(fr_NH0)',

[0076] 'PEOE_VSA1**3*log(FpDensityMorgan1)','1 / sqrt(P)','Kappa2**3*P','MolMR**2*sqrt(P)',

[0077] 'P**2*sqrt(PEOE_VSA1)','MinAbsPartialCharge**3*P','sqrt(P)*exp(HallKierAlpha)',

[0078] 'sqrt(VSA_EState2) / SMR_VSA1','sqrt(PEOE_VSA12)*sqrt(PEOE_VSA8)',

[0079] 'sqrt(SMR_VSA3)*sqrt(VSA_EState2)','P**2*fr_ether','SMR_VSA3**3 / P','qed**3 / HallKierAlpha',

[0080] 'sqrt(PEOE_VSA11)*log(P)','BCUT2D_LOGPHI**3*SlogP_VSA1**2',

[0081] 'fr_Ndealkylation1*exp(MolLogP)','Kappa2**3*exp(fr_unbrch_alkane)',

[0082] 'sqrt(PEOE_VSA8)*Abs(VSA_EState5)','VSA_EState9*log(MinAbsEStateIndex)',

[0083] 'sqrt(VSA_EState2)*exp(fr_unbrch_alkane)','PEOE_VSA8 / P','P**3*fr_NH1','PEOE_VSA5 / P',

[0084] 'log(P) / TPSA','P**3 / MolLogP','PEOE_VSA14**3 / P','T**2*log(SMR_VSA1)',

[0085] 'EState_VSA4*SlogP_VSA12**3','PEOE_VSA7**3*VSA_EState8**3','MinAbsEStateIndex**3*log(P)','sqrt(PEOE_VSA8)*sqrt(fr_NH0)','sqrt(Kappa3)*Abs(HallKierAlpha)',

[0086] 'sqrt(PEOE_VSA8)*sqrt(PEOE_VSA9)','PEOE_VSA7**3*exp(fr_unbrch_alkane)',

[0087] 'PEOE_VSA2**2*log(MinAbsEStateIndex)'].

[0088] The new characteristic parameter is formed by combining the molecular descriptor and the operating parameters P and T through a mathematical formula, and the prediction formula for the new characteristic value is obtained as follows:

[0089] 0.5833502156347767+13.101708*log(MolMR) / T+0.587576*SMR_VSA3**2 / Ipc+0.122159*

[0090] sqrt(VSA_EState2) / SMR_VSA1+0.099528*P2*fr_ether+0.094410*sqrt(P)*qed2-0.080557*

[0091] qed3 / HallKierAlpha+0.069115*sqrt(P)*exp(HallKierAlpha)+0.044121*log(P) / TPSA+0.040159*

[0092] fr_Ndealkylation1*exp(MolLogP)+0.030509*FractionCSP33*sqrt(P)

[0093] -0.027555*P3*fr_NH1+0.014171*sqrt(P)*exp(fr_COO)+0.013512*MinAbsPartialCharg3*P+0.007587*MinAbsEStateInde3*log(P) -0.006654*1 / sqrt(P)+0.004988*P*log(Chi3n)+0.004818*sqrt(SMR_VSA3)*sqrt(VSA_EState2)+0.004620*sqrt(PEOE_VSA12)*sqr t(PEOE_VSA8)+0.003584*sqrt(P)*SlogP_VSA1-0.003532*sqrt(Kappa3)*Abs(HallKierAlpha)+0.002804*sqrt(PEOE_VSA11)*log(P )-0.002421*T+0.001849*sqrt(PEOE_VSA8)*sqrt(PEOE_VSA9)+0.001635*sqrt(P)*PEOE_VSA7+0.001548*sqrt(PEOE_VSA8)*sqrt(fr_ NH0)-0.000931*PEOE_VSA5 / P-0.000901*sqrt(PEOE_VSA8)*Abs(VSA_EState5)+0.000737*sqrt(VSA_EState2)*exp(fr_unbrch_alka ne)-0.000606*VSA_EState9*log(MinAbsEStateIndex)-0.000453*MolLogP*fr_NH03+0.000098*PEOE_VSA22*log(MinAbsEStateIndex )+0.000067*BCUT2D_LOGPHI3*SlogP_VSA12+0.000061*P2*sqrt(PEOE_VSA1)+0.000032*SlogP_VSA3*sqrt(fr_NH0)+0.000030*VSA_E State93*sqrt(fr_unbrch_alkane)-0.000017*PEOE_VSA123*log(P)+0.000011*EState_VSA4*SlogP_VSA123-0.000010*PEOE_VSA8 / P.

[0094] The decision coefficient R2, mean absolute error MAE and mean absolute error MSE were calculated to evaluate the prediction formula. The results are shown in Table 2 and Figure 6 .

[0095] Table 2 R2 and MSE of Autofeatregressor method

[0096]

[0097] like Figure 6 As shown, the R2 of the test set using the Autofeatregressor method to predict the solubility of H2S gas in ionic liquids is 0.9510, proving that the established prediction formula is acceptable.

[0098] The parameters used to evaluate the prediction ability of the quantitative prediction model in the above two methods are the decision coefficient R2, the mean absolute error MAE and the mean square error MSE as shown in formulas (1)-(3) respectively:

[0099]

[0100] Among them, Y i is the actual value, is the predicted value, is the average of the actual values, and N represents the number of samples.

[0101] Among them, MAE and MSE are error indicators. The smaller the value, the smaller the prediction error and the higher the model stability. R2 is an indicator for evaluating the goodness of fit. The closer the value is to 1, the better the model fit.

[0102] R2 stands for the coefficient of determination, also known as goodness of fit or coefficient of determination, which can be used to judge the fitting performance of the model. However, it also has certain limitations. When there are too many independent variables, R2 will be too high due to overfitting of the model and lose its reference value. Therefore, the error indicators MAE and MSE are introduced. MAE can measure the average value of the absolute error between the predicted value and the true value, and intuitively reflects the actual situation of the predicted value error. MSE averages the sum of the squares of the absolute errors and is more sensitive to situations with larger errors.

[0103] Step 5, prediction:

[0104] Analysis of the results of actual values ​​and predicted values: Using the random forest model, based on the visualization between the predicted values ​​and the actual values, the prediction accuracy of the model for the solubility of H2S gas in ionic liquids was clarified.

[0105] The prediction results show that the predicted values ​​obtained by introducing a series of molecular descriptor features generated by the Rdkit molecular fingerprint method are basically consistent with the actual values, proving that this method is suitable for the prediction of H2S gas solubility and has good prediction performance.

Claims

1. A method for predicting the solubility of hydrogen sulfide gas in ionic liquids, characterized in that: The prediction method comprises the following steps: (1) Data collection: Collect the solubility of H2S gas in different ionic liquids at different temperatures and pressures, and organize the SMILES expressions of ionic liquid anions and cations; (2) Data preprocessing: normalize the data according to the molar solubility of H2S gas in different ionic liquids, process the outliers in the data, and use visualization methods to view the distribution of the data; (3) Feature engineering: Based on the collected SMILES expressions of ionic liquids, the Rdkit molecular fingerprint method is used to generate molecular descriptors as features. All features including operating parameters P and T are screened using Feature Selector, and single feature values ​​are eliminated to obtain the final data set that can be used for model prediction. The important features in the data set are calculated and analyzed. (4) Establishing a model for prediction: Method 1: Using the random forest method in machine learning, the data set obtained in step (3) is divided to establish a solubility prediction model for hydrogen sulfide in different ionic liquids, and the operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are input into the solubility prediction model for data training, verification and / or prediction; or, Method 2: The molecular descriptor features, operating parameters P, T and solubility data set generated based on the Rdkit molecular fingerprint method are combined with new feature parameters generated by the AutoFeatRegressor method to directly predict the gas solubility.

2. The prediction method according to claim 1, characterized in that: In step (1), the collection method includes consulting literature.

3. The prediction method according to claim 1, characterized in that: In step (1), anions and cations of different ionic liquids are input into the InDraw software to obtain SMILES expressions of different anions and cations.

4. The prediction method according to claim 1, characterized in that: In step (2), data preprocessing includes the following steps: (I) collating the collected solubility data, converting all the solubility data into molar solubility data values, converting the obtained molar solubility data values ​​into units to obtain gas solubility percentages, normalizing the obtained gas solubility percentages to obtain solubility values ​​between 0 and 1, and verifying the normalized solubility data values ​​to ensure that the data are authentic and valid; (II) performing abnormal processing on the collected data. If the value of a characteristic variable under a certain dissolution condition exceeds or is lower than 100% of the range of the remaining data values ​​in the collected data set, the abnormal data value is selected for removal; (III) After confirming that all the data are correct, visualize the data by drawing a scatter plot, box plot, or line graph to intuitively present the relationship between the feature value and the target value for further analysis.

5. The prediction method according to claim 4, characterized in that: In step (I), the normalization process is performed using OmicShare Tools, Python, R language or Excel software. In step (II), the IsolationForest in Python is used to perform an anomaly detection algorithm on the collected data to process the abnormal data. In step (III), the visualization process is performed using matplotlib, yellowbrick in Python or R language.

6. The prediction method according to claim 1, characterized in that: In step (3), feature engineering includes the following steps: (I) The processed data were analyzed using the RDKit module, and the Rdkit molecular fingerprint was produced using the CalcDescriptors method. The molecular feature descriptors of the compounds were obtained using the MolFromSmiles method in combination with the collected ionic liquid SMILES. (II) Organize the molecular feature descriptor data and fill missing values ​​with 0 to ensure that the data is complete and usable; (III) Combine the molecular feature descriptors and the operation parameters P and T, use Feature Selector to perform feature screening, remove single feature values, and retain the remaining important features; (IIII) The SHAP library was used to plot the importance of different characteristic variables on the solubility of H2S gas in ionic liquids, and independence and dependence analysis of important features was performed based on the results.

7. The prediction method according to claim 1, characterized in that: In step (4), the machine learning method described in method 1 is random forest (RF), and the data set obtained in step (3) is divided into a training data set and a test data set in a ratio of 7:

3.

8. The prediction method according to claim 1, characterized in that: In step (4), the method described in method 2 uses the AutoFeatRegressor method in the autofeatselect library to combine feature values ​​to obtain feature parameters, where the new feature parameters are formed by combining the molecular descriptor and the operating parameters P and T through a mathematical formula.

9. The prediction method according to claim 1, characterized in that: In step (4), the parameters used in method 1 to evaluate the predictive ability of the quantitative prediction model are: decision coefficient R2, mean absolute error MAE and mean square error MSE, and the distribution is shown in formulas (1)-(3): Among them, Y i is the actual value, is the predicted value, is the average of the actual values, and N represents the number of samples; Among them, MAE and MSE are error indicators. The smaller the value, the smaller the prediction error and the higher the model stability. R2 is an indicator for evaluating the goodness of fit. The closer the value is to 1, the better the model fit.

10. The prediction method according to claim 1, characterized in that: In step (4), the prediction method 1 includes inputting the molecular descriptor characteristic value, pressure and temperature of the target ionic liquid into the random forest program to predict the solubility of hydrogen sulfide in the ionic liquid at the temperature and pressure; the method 2 includes inputting the molecular descriptor characteristic value, pressure and temperature of the ionic liquid into the random forest program to predict the solubility of hydrogen sulfide in the ionic liquid at the temperature and pressure; The descriptor eigenvalue, pressure and temperature are input into AutoFeatRegressor to obtain the solubility formula formed by the new feature combination parameters. Predict the solubility of hydrogen sulfide in the ionic liquid at this temperature and pressure.

Images