A method for screening CO2 solubilizing eutectic solvents based on quantum chemical descriptors and machine learning

By combining quantum chemical descriptors with machine learning, a multi-dimensional feature space is constructed, which solves the problems of high experimental cost and insufficient generalization of prediction models in the screening of eutectic solvents. This enables efficient screening and accurate prediction of eutectic solvents for CO2 capture and provides microscopic mechanism guidance for DESs design.

CN123120077APending Publication Date: 2026-11-17XINJIANG UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510622925.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing screening methods for eutectic solvents suffer from high experimental costs, insufficient generalization of prediction models, and unclear microscopic mechanisms, resulting in low CO2 capture efficiency.

Method used

By combining quantum chemical descriptors with machine learning, a multi-dimensional feature space is constructed. By collecting a wide range of experimental data, a quantum chemical descriptor system is built, the feature set is optimized, and a multi-paradigm machine learning model is trained to achieve efficient and accurate prediction and screening of DESs performance.

Benefits of technology

It significantly reduces DESs screening time from months to hours, improves prediction accuracy and generalization, provides micro-mechanism guidance for DESs design, and reduces the development cost of DESs for CO2 capture.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120077A_ABST
    Figure CN123120077A_ABST
Patent Text Reader

Abstract

This invention relates to a method for screening CO2-solubilizing eutectic solvents (DESs) based on quantum chemical descriptors and machine learning, belonging to the field of carbon capture technology. This method overcomes the technical bottlenecks of high experimental costs and insufficient model generalization in traditional DES development by integrating dynamic electronic structure analysis and multi-paradigm machine learning prediction methods. For the first time, a multi-dimensional quantum chemical descriptor system covering the dynamic synergistic effect of hydrogen bond networks, three-dimensional spatial steric hindrance, and surface potential distribution is constructed, and a high-discrimination multi-dimensional feature space is built by combining experimental parameters. An ensemble learning algorithm is used to train a high-precision prediction model, achieving accurate prediction of CO2 solubility over a wide temperature and pressure range. Interpretable machine learning is used to analyze and reveal the charge transfer and electrostatic complementarity mechanisms of the solvation process, guiding the design of DES molecules. Finally, high-performance solvent combinations are obtained through high-throughput computational screening. Compared with traditional experimental trial-and-error methods, this method significantly reduces prediction errors (e.g., RMSE is reduced by more than 80%) and shortens the screening cycle from months to hours, providing an efficient computational tool for the rational design of CO2 capture solvents.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of carbon capture technology, specifically involving a method for screening eutectic solvents (DESs) that combines quantum chemical descriptor construction with machine learning model optimization. By constructing a multi-dimensional feature space and a high-performance prediction model, it achieves accurate prediction of CO2 solubility and efficient screening of solvent systems. Background Technology

[0002] Eutectic solvents (DESs) are considered potential materials for CO2 capture due to their advantages such as adjustability, low cost, and low volatility. However, their development faces the following core challenges: High experimental trial-and-error costs: Traditional DESs screening relies on a "synthesis-testing" cycle, which takes months or even years and requires significant human and material resources; Insufficient accuracy of prediction models: Existing σ-profile-based models (such as the literature DOI:10.1016 / j.cej.2021.131567) only describe intermolecular interactions through static charge distribution, failing to capture the dynamic hydrogen bond network synergy of hydrogen bond donors (HBDs) and acceptors (HBAs) in DESs, the influence of the HBA / HBD molar ratio, and the steric hindrance effect of functional groups, resulting in poor model generalization; Unclear microscopic mechanisms: Traditional models rely on simplified molecular features (such as single polarity parameters), failing to reveal the microscopic interaction mechanisms such as electron transfer and electrostatic complementarity between CO2 and DESs, making it difficult to guide the rational design of DESs.

[0003] Therefore, there is an urgent need for a DESs screening method that can integrate the dynamic electronic properties of molecules, cover a wide range of operating conditions, and has high prediction accuracy, in order to accelerate the development of DESs for CO2 capture. Summary of the Invention

[0004] Technical problems to be solved

[0005] This invention aims to address the problems of high experimental costs, insufficient generalization of prediction models, and unclear microscopic mechanisms in the development of existing DESs. It provides a quantum chemical descriptor-machine learning driven method for efficient screening of CO2-capturing eutectic solvents. By integrating dynamic electronic structure features with a wide range of experimental data, it achieves efficient and accurate prediction and screening of DESs performance.

[0006] Technical solution

[0007] The screening method of this invention is implemented through the following process: First, CO2 solubility data of a wide range of DESs are systematically collected and preprocessed; second, a quantum chemical descriptor system covering dynamic electronic interactions, three-dimensional spatial effects, and surface potential heterogeneity is constructed and calculated; next, the feature set is optimized through feature engineering; finally, high-performance DESs are trained and screened based on a multi-paradigm machine learning model. The specific steps are as follows:

[0008] S100 Experimental Data Collection and Preprocessing

[0009] Collect data on the solubility of CO2 in eutectic solvents that meet the following conditions: DESs have a defined chemical structure and the solubility unit is uniformly mol·mol⁻¹. -1 The data underwent rigorous validation. To optimize distribution symmetry, the solubility values ​​in the final dataset underwent logarithmic transformation preprocessing.

[0010] S200 Quantum Chemical Descriptor Construction and Computation

[0011] An enhancement system containing 48 types of quantum chemical descriptors was constructed, covering: (1) dynamic electronic interactions (HOMO-LUMO band gap, orbital delocalization index, etc.); (2) three-dimensional spatial effects (molecular size, quadrupole moment, etc.); and (3) surface potential heterogeneity (electrostatic potential extrema, polar surface area, etc.). The specific calculation process was as follows: molecular structures were obtained using Avogadro or PubChem, fully optimized using DFT software such as ORCA, and single-point energies were calculated with an accuracy no lower than B3LYP-D3(BJ) / def2-TZVP. Descriptors were extracted from the wavefunction file. Molecular structure optimization was performed under gas-phase conditions without introducing implicit solvent models to preserve intrinsic interactions between DES components.

[0012] S300 Feature Engineering and Dataset Construction

[0013] Forty-eight quantum chemical descriptors were combined with experimental parameters (temperature, pressure, molar ratio) to form an initial 51-dimensional feature set. Through a three-stage screening process (variance filtering, high-correlation removal, and missing value screening), 39 key features were retained, and the maximum cross-correlation coefficient was reduced from 0.93 to 0.71, improving feature independence.

[0014] S400 Machine Learning Model Training

[0015] A multi-paradigm framework including linear regression, random forest, XGBoost, CatBoost, and MLP was adopted. The training and test sets were split in an 8:2 ratio, and hyperparameters were optimized using 10-fold cross-validation. Among them, CatBoost (number of trees = 1000, L2 regularization = 3) performed best, with R on the test set. 2 ≥0.99, RMSE≤0.12.

[0016] S500 Model Validation and Performance Evaluation

[0017] The model performance was verified using multi-dimensional methods such as prediction error plots, residual plots, and Taylor plots: the predicted points on the CatBoost test set were closely distributed near the y=x line (R0). 2=0.9921), the residuals are concentrated in the range of -0.25 to 0.25 (standard deviation 0.1013), which is significantly better than LR and other models in terms of generalization.

[0018] S600 SHAP Feature Importance Analysis

[0019] The SHAP algorithm reveals that pressure has a bidirectional effect on solubility (SHAP±0.77); the orbital delocalization index of HBA (ODI_Mean_1<20V) dominates charge transfer; the polar surface area of ​​HBD optimizes solvation entropy; and the quadrupole moment of HBD >7 enhances adsorption, providing theoretical guidance for DES design.

[0020] S700 High-Performance DESs Screening

[0021] Based on typical CO2 capture conditions (298.15 K, 1600 kPa, molar ratio 1:2 / 1:4), 1088 combinations of 34 HBAs and 32 HBDs were predicted, and high-performance DESs were selected.

[0022] The “rigorous validation” mentioned in step S100 includes experimental repeatability ≥95% and error ≤5% (such as NIST database validation standards);

[0023] The "chemical software" mentioned in step S200 includes, but is not limited to, Avogadro 1.2.0 and GaussView 6.0; the "DFT software" includes, but is not limited to, ORCA 5.3.0, Gaussian 16, and MaterialsStudioDMol3; the basis sets and functionals include the r2SCAN-3c basis set (structure optimization), the B3LYP-D3(BJ) functional, and the def2-TZVP basis set (single-point energy calculation); the "wavefunction analysis software" includes, but is not limited to, Multiwfn 3.8dev and TopMoD.

[0024] The "three-stage screening" mentioned in step S300 specifically includes: variance filtering (removing constant features with variance < 0.01), high correlation removal (removing feature pairs with Pearson correlation coefficient |ρ| > 0.90), and missing value screening (excluding features with missing value ratio > 10%).

[0025] The "multi-paradigm framework" described in step S400 is implemented based on the Scikit-learn 1.2.2 library; the CatBoost hyperparameters are configured as follows: number of trees = 1000, L2 regularization strength = 3, boundary count = 128; test set R. 2 =0.9921, RMSE=0.1152.

[0026] The "multidimensional evaluation method" described in step S500 is implemented using tools such as Matplotlib 3.5.1 and Seaborn 0.11.2, and includes prediction error plot, residual plot, Taylor plot and target plot.

[0027] The “SHAP algorithm” mentioned in step S600 is based on the SHAP0.41.0 library. Key conclusions include: the mean absolute value of the pressure SHAP is 0.77, the SHAP value of the HBA orbital delocalization index (ODI_Mean_1<20eV) is positive, and the SHAP value is significantly positive when the HBD quadrupole moment is >7.

[0028] Beneficial effects

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) More comprehensive data coverage and preprocessing: 2287 sets of data from 55 HBAs and 40 HBDs were collected, covering a wide range of conditions including molar ratio of 1:1–1:16, temperature of 293.15–353.15 K, and pressure of 0.26–76.2 bar; the data distribution was optimized by logarithmic transformation to improve the stability of model training.

[0031] (2) The quantum descriptor system is more accurate: By using three types of enhanced descriptors, namely dynamic electron interaction (HOMO-LUMO band gap, ODI), three-dimensional spatial effect (Mol_Size_L, quadrupole moment), and surface potential heterogeneity (ESPmin / ESPmax, Polar_Area), the problems of conventional descriptors ignoring dynamic hydrogen bond network and spatial steric hindrance are solved, and the microscopic interaction mechanism between DES and CO2 is captured more accurately.

[0032] (3) Significantly improved model performance and generalization: The CatBoost model achieves R on the test set. 2 =0.9921, RMSE=0.1152, AARD%=2.76%, compared with conventional linear fitting (R² = 0.9921, RMSE=0.1152, AARD%=2.76%), which is better than conventional linear fitting (R² = 0.9921, RMSE=0.1152, AARD%=2.76%). 2 =0.7800, RMSE=0.6095) The error was reduced by 80.1%, compared with the random forest model (R 2 =0.9762, RMSE=0.1620) decreased by 42.5%; verifying its robustness to a wide range of DES systems.

[0033] (4) Microscopic mechanism interpretability: SHAP analysis clarified the key driving factors of CO2 solubility (such as the orbital delocalization index of HBA regulating charge transfer and the polar surface area of ​​HBD optimizing solvation entropy), providing theoretical guidance for DES molecular design.

[0034] (5) Efficient screening of DES combinations: 1088 DES combinations are predicted by the model, and high-performance DES can be quickly located. The screening time is shortened from several months in traditional experiments to hours, which significantly reduces the development cost of DESs for CO2 capture. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0036] Figure 1 This is an overall flowchart of the CO2 solubilization eutectic solvent screening method based on quantum chemical descriptors and machine learning described in this invention.

[0037] Figure 2 This is a histogram showing the CO2 solubility distribution of the experimental dataset of this invention.

[0038] Figure 3 This is a Pearson correlation heatmap (after feature screening) between the quantum chemical descriptor of this invention and CO2 solubility.

[0039] Figure 4 This is a scatter plot comparing the prediction performance of different machine learning models (LR, RF, XGBoost, CatBoost, MLP) of the present invention on CO2 solubility in DESs.

[0040] Figure 5 This invention uses Taylor diagrams and target diagrams to evaluate the performance of the model across multiple metrics.

[0041] Figure 6 This is a graph showing the importance analysis of SHAP features based on the CatBoost model in this invention.

[0042] Figure 7 This is a heatmap showing the solubility prediction of 1088 DES combinations at 298.15 K and 1600 kPa according to the present invention.

[0043] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Example 1: Data Collection and Preprocessing

[0046] Solubility data for CO2 absorption by DESs were screened through a systematic literature review, meeting the following criteria: the DESs have a defined chemical structure, and the solubility unit is uniformly expressed as mol·mol⁻¹. -1 Experimental data were validated using the NIST database or through peer review. The final dataset covers a wide range of conditions, including datasets of 55 HBAs and 40 HBDs. To optimize model training stability, the solubility values ​​underwent logarithmic transformation preprocessing (e.g., logarithmic transformation). Figure 2 (As shown).

[0047] Example 2: Quantum Chemical Descriptor Calculation

[0048] The molecular structures of HBA and HBD were drawn using chemical software (such as Avogadro) or obtained from public databases (such as PubChem). Geometric optimization (including dispersion correction) was then performed using DFT software (such as ORCA) to ensure structural accuracy. After optimization, single-point energies were calculated at a high-precision DFT level (such as B3LYP-D3 / def2-TZVP). Wavefunction analysis tools (such as Multiwfn) were used to extract 48 types of quantum chemical descriptors (covering dynamic electronic properties, three-dimensional spatial effects, and surface potential heterogeneity), comprehensively characterizing the microscopic interaction mechanism between DES and CO2.

[0049] Example 3: Feature Engineering and Dataset Construction

[0050] Forty-eight quantum chemical descriptors were combined with experimental parameters (temperature T / K, pressure P / bar, HBA / HBD molar ratio R) to form an initial descriptor set containing 51 features. To eliminate multicollinearity and redundant information, a three-stage feature selection process was employed: first, constant features were removed by using a variance threshold; second, highly collinear feature pairs were removed by Pearson correlation analysis (|ρ|>0.90); and finally, descriptors with a missing value ratio >10% were excluded. After optimization, 39 features were retained, and the maximum cross-correlation coefficient decreased from 0.93 to 0.71 (e.g., ...). Figure 3 As shown in the figure, it significantly improves feature independence and reduces the risk of model overfitting.

[0051] Example 4: Machine Learning Model Development and Training

[0052] A multi-paradigm prediction framework, including Linear Regression (LR), Random Forest (RF), XGBoost (XGB), CatBoost (CAT), and Multilayer Perceptron (MLP), is implemented using the Scikit-learn library, covering three major paradigms: traditional statistical learning, ensemble tree models, and neural networks. The dataset is divided into a training set (1835 sets) and a test set (452 ​​sets) in an 8:2 ratio. Hyperparameters are optimized using a 10-fold cross-validation grid search on the training set (the test set is used only for final evaluation). The specific hyperparameter configurations for each model are as follows: RF: number of trees = 100, maximum depth = 20, minimum number of leaf samples = 1; XGBoost: number of trees = 300, learning rate = 0.3, subsampling rate = 0.9; CatBoost: number of trees = 1000, L2 regularization strength = 3, boundary count = 128; MLP: hidden layer size = (128, 64), activation function = ReLU, learning rate = 0.005, maximum number of iterations = 200; LR uses default parameters.

[0053] Example 5: Model Validation and Performance Evaluation

[0054] Model validation employs a multidimensional statistical framework, comprehensively evaluating the prediction accuracy and generalization ability of each model through prediction error plots, residual plots, Taylor plots, and target plots. (Prediction error plot...) Figure 4 The results show that the predicted points (blue dots) of the CatBoost model on the test set are closely distributed near the ideal fitted line of y = x (black dashed line), and R0 is... 2 The value reached 0.9921, which is significantly better than LR(R). 2 =0.7800), RF(R) 2 Models such as (=0.9762); residual analysis ( Figure 4 Further analysis shows that CatBoost's residuals are concentrated in the range of -0.25 to 0.25, with a standard deviation of only 0.1013, a 72% reduction compared to LR (standard deviation 0.3620), and superior to XGBoost (standard deviation 0.1410) and MLP (standard deviation 0.1280). Taylor plot ( Figure 5 In the target graph, CatBoost had the highest correlation coefficient (0.9939) and the lowest RMSE (0.1363); Figure 5 In the data analysis, the predicted standard deviation perfectly matched the experimental value (1.2342), and it occupied the innermost loop of the correlation and RMSE indices, verifying its accuracy and stability in a wide range of DES systems. Furthermore, compared with other models and descriptors, the quantum chemical descriptor proposed in this invention exhibits higher accuracy (see Table 1). In summary, this invention, through a complete data-descriptor-model-screening process, achieves efficient prediction of CO2 solubility by DESs and rapid screening of high-performance materials, providing a reliable tool for CO2 capture technology.

[0055] Example 6: SHAP Feature Importance Analysis

[0056] To reveal the nonlinear relationship and microscopic mechanism between quantum chemical descriptors and CO2 solubility, the SHAP0.41.0 algorithm was used to interpret the optimal CatBoost model (results are shown in...). Figure 6 Global feature contribution analysis shows that pressure (P) has the strongest bidirectional regulatory effect on solubility (SHAP value ±0.77), and the negative contribution at high temperature (-0.15) is consistent with the exothermic characteristics of CO2 dissolution. The quantum chemical descriptors of HBA (such as the mean orbital delocalization index ODI_Mean_1 and the negative surface area Neg_Surface_Area_1) dominate the prediction logic, while the polar surface area (Polar_Area_2) and local electron affinity variance (LEA_Var_2) of HBD affect solubility through a geometric-electrostatic synergistic effect. Dynamic trend analysis further shows that when ODI_Mean_1 of HBA is <20 eV, the concentration of electron density enhances charge transfer with CO2, while electron dispersion leads to weakened adsorption after exceeding 20 eV; when LEA_Var_2 of HBD is >0.02, the dispersion of adsorption sites reduces solubility; when the quadrupole moment of HBD is >7, it complements the inherent quadrupole moment of CO2, significantly improving adsorption efficiency. These dynamic relationships provide key theoretical guidance for the design of DES molecules.

[0057] Application Example 1: High-performance DESs screening

[0058] Based on the developed model and descriptors, and using typical pre-combustion CO2 capture parameters (temperature 293–313 K, pressure 1000–3000 kPa), 298.15 K, 1600 kPa, and molar ratios of 1:2 and 1:4 were selected as prediction conditions to predict the solubility of 1088 DES combinations of 34 HBAs and 32 HBDs (prediction heatmap shown). Figure 7 (As shown). Heatmap analysis identified high-performance DES combinations: the optimal HBAs were methyltrioctylammonium chloride ([MTTA][Cl]), tetrabutylammonium chloride ([TBAC][Cl]), and tetrabutylammonium bromide ([TBAB][Br]), and the optimal HBDs were lauric acid and decanoic acid; inefficient combinations included HBAs such as betaine hydrochloride ([Bet][Cl]) and benzyltrimethylammonium chloride ([BTMA][Cl]), and HBDs such as glycerol. This screening method rapidly locates high-performance DESs using a model, avoiding the high costs of traditional "synthesis-testing" processes and significantly accelerating the industrial development of DESs for CO2 capture.

[0059] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A quantum chemical descriptor-machine learning driven method for screening eutectic solvents for CO2 capture, characterized in that, Includes the following steps: S100. Experimental Data Collection and Preprocessing: Collect CO2 solubility data of eutectic solvents (DESs) with well-defined chemical structures and rigorous verification. The data are from public databases (NIST CODATA) and peer-reviewed literature datasets. The solubility data are preprocessed (including but not limited to logarithmic transformation) to form an optimized dataset. S200, Quantum Chemical Descriptor Construction and Computation: Static multidimensional quantum chemical descriptors for hydrogen bond acceptor (HBA) and donor (HBD) were obtained through quantum chemical calculations (using B3LYP-D3(BJ) / def2-TZVP computational level). These descriptors characterize the microscopic interaction between DES and CO2 through the following synergistic effects: (a) Molecular size information: number of atoms (AtomNum), molecular weight (Weight), and van der Waals volume (vdW_Volume) for HBA and HBD; (b) Molecular orbital characteristics: Highest occupied molecular orbital energy (HOMO), lowest unoccupied molecular orbital energy (LUMO), and HOMO-LUMO gap (HOMO_LUMO_Gap) for HBA and HBD; (c) Orbital dispersion index class: HOMO orbital dispersion (ODI_HOMO), LUMO orbital dispersion (ODI_LUMO), and mean orbital dispersion (ODI_Mean) for HBA and HBD; (d) Molecular geometry: Molecular major axis dimension (Mol_Size_L), molecular minor axis dimension (Mol_Size_S), molecular symmetry ratio (Length_Ratio); (e) Multipole moment distribution class: quadrupole moment (Quadrupole_Moment), dipole moment (Dipole_Moment), and polar surface area (Polar_Area) of HBA and HBD; (f) Surface potential characteristics: minimum surface electrostatic potential (ESPmin), maximum surface electrostatic potential (ESPmax), polar area (Polar_Area), negative surface area (Neg_Surface_Area); (g) Local electronic properties: average local ionization energy (ALIE_Ave), local electron affinity variance (LEA_Var), electron density gradient (ED_Gradient); The quantum chemical calculations were performed using ORCA software to optimize molecular structures and Multiwfn tool to extract wavefunction features. S300, Feature Engineering and Dataset Construction: The quantum chemical descriptors are combined with experimental condition parameters (such as temperature, pressure, HBA / HBD molar ratio) to form an initial feature set, and key feature subsets are formed through feature screening (including but not limited to variance filtering, high correlation elimination, and missing value screening); S400, Machine Learning Model Training: A multi-model framework including CatBoost, XGBoost ensemble learning algorithms and multilayer perceptron (MLP) neural network is adopted. The feature selection strategy is adjusted by introducing time-dependent parameters, and the hyperparameters are configured as follows: number of trees = 1000, L2 regularization = 3, boundary count = 128. S500, Model Validation and Performance Evaluation: Validate the model's prediction accuracy and generalization ability through multidimensional methods (including but not limited to prediction error analysis, residual test, Taylor plot); S600, Interpretability Analysis: Based on interpretable algorithms (such as SHAP), the predictive logic of the analytical model is revealed to expose the influence mechanism of key features on CO2 solubility; S700, high-performance DESs screening: Based on preset operating conditions, the solubility of candidate DES combinations is predicted, and the optimal solvent system is screened through visual analysis (such as heatmaps). S800, Dynamic Update Mechanism: When the amount of newly added experimental data reaches 10% of the total dataset, the quantum chemical descriptor recalculation and model retraining process is triggered.

2. The screening method according to claim 1, characterized in that, The solubility data described in step S100 covers a wide range of operating conditions (including but not limited to molar ratios of 1:1 to 1:16, temperatures of 293.15 to 353.15 K, and pressures of 0.26 to 76.2 bar) and has been rigorously verified (repeatability ≥ 95%, relative error ≤ 5%).

3. The screening method according to claim 1, characterized in that, The 48 types of descriptors mentioned in step S200 improve prediction accuracy through the following synergistic mechanism: (a) The solvation cavity size and CO2 permeation resistance of DESs are jointly characterized by molecular size information and surface potential characteristics; (b) The regulation of charge transfer by dynamic recombination of hydrogen bond networks through the synergistic quantification of orbital dispersion index and local electronic properties; (c) The multipole moment distribution class and the molecular orbital property class together reveal the electrostatic complementary effect between DESs components. The quantum chemical calculations employ a computational level no lower than B3LYP-D3(BJ) / def2-TZVP (or wB97M-V / def2-TZVP, etc.), and molecular structures are obtained and optimized using software such as Avogadro and Gaussian, while wavefunction features are extracted using tools such as Multiwfn.

4. The screening method according to claim 1, characterized in that, The feature selection in step S300 includes: first stage removing constant features with variance <0.01; second stage removing highly collinear feature pairs with Pearson correlation coefficient |ρ|>0.90; third stage excluding descriptors with missing values ​​>10%, and finally retaining 39 key features.

5. The screening method according to claim 1, characterized in that, The range of grid search parameters for the CatBoost model described in step S400 is: number of trees 500-1500, L2 regularization strength 1-5, learning rate 0.01-0.

3. The optimal parameters are selected by cross-validation AUC≥0.

98.

6. The screening method according to claim 1, characterized in that, The model validation described in step S500 involves plotting the prediction error (test set R) using Matplotlib and Origin tools. 2 ≥0.99, RMSE≤0.12), residual plot (standard deviation≤0.1013) and Taylor plot (correlation coefficient≥0.9939).

7. The screening method according to claim 1, characterized in that, The interpretability analysis described in step S600 reveals that: the mean absolute value of the pressure parameter SHAP is ≥0.77; the mean orbital dispersion of HBA (ODI_Mean<20eV) promotes charge transfer; and HBD quadrupole moment >7 enhances CO2 adsorption.

8. The screening method according to claim 1, characterized in that, The preset operating conditions mentioned in step S700 include a temperature of 293.15-353.15K, a pressure of 1000-3000kPa, and an HBA / HBD molar ratio of 1:2 to 1:

4. The optimal system (such as a DES with HBA being methyltrioctylammonium chloride and HBD being lauric acid) is selected from 1088 DES combinations through heatmap visualization.

9. The screening method according to claim 1, characterized in that, The preset threshold mentioned in step S800 is 10% of the total dataset.

10. A computer-readable storage medium, characterized in that, The device stores program instructions for performing the method according to any one of claims 1-9, which, when executed by a processor, implement an automated processing flow for quantum chemical descriptor calculation, feature engineering optimization, and machine learning prediction.