A method for screening of metal-organic framework adsorbents for sulfur hexafluoride recovery regeneration
Patent Information
- Application Number
- CN202611009976.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-09-29
AI Technical Summary
[0004]为解决以上的技术问题，本发明提供了一种用于六氟化硫回收再生金属有机框架吸附剂筛选方法，以解决现有技术筛选周期长、计算成本高、预测准确性不足的问题，实现从十万级以上MOFs候选库中快速、精准地识别具有优异吸附/分离性能的候选材料
[0025]（1）筛选效率极高：通过几何初筛（PLD约束）将候选材料规模压缩至原来的22.3%，再通过机器学习预测代替高成本的GCMC模拟，筛选效率提升数千倍，将传统需要数月的工作缩短至数小时；
Smart Images
Figure CN122842779A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of gas adsorption separation and materials genetic engineering, specifically to a method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration. Background Technology
[0002] Sulfur hexafluoride (SF6) is widely used in core power equipment such as gas-insulated switchgear due to its excellent insulation and arc-quenching properties. However, SF6 has an extremely high global warming potential (approximately 23,900 times that of CO2) and is strictly controlled under the Kyoto Protocol and environmental regulations of various countries. Therefore, the recycling and reuse of SF6 has significant environmental and economic value.
[0003] In the SF6 regeneration process, it is necessary to simultaneously remove various impurities such as H2O, SO2, H2S, and CO2. Existing processes typically rely on multiple adsorbents (such as alumina and molecular sieves) for stepwise treatment, which is cumbersome, material matching depends on experience, and has high operating costs. Metal-organic frameworks (MOFs) are ideal sources of high-performance adsorbents due to their high specific surface area and tunable pore structure. However, the number of MOF candidate materials is extremely large (more than 140,000 structures have been disclosed). If only molecular-by-molecule simulations (such as grand canonical Monte Carlo simulations) or experimental testing are used, the cycle can take months or even years, resulting in high computational and experimental costs. Therefore, there is an urgent need for a method that can quickly and accurately screen high-performance adsorbents suitable for SF6 regeneration from a massive number of MOF structures. Existing pure GCMC high-throughput screening of all materials consumes a huge amount of computational power; traditional single machine learning screening lacks a comprehensive evaluation index that takes into account both impurity adsorption capacity and SF6 separation selectivity, resulting in insufficient prediction accuracy; and the lack of pre-emptive pore size geometric constraints leads to a large number of invalid materials participating in the simulation calculation, further lengthening the screening cycle. Summary of the Invention
[0004] To address the above technical problems, this invention provides a method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration. This method solves the problems of long screening cycles, high computational costs, and insufficient prediction accuracy in existing technologies, enabling rapid and accurate identification of candidate materials with excellent adsorption / separation performance from a candidate library of over 100,000 MOFs.
[0005] The technical solution is as follows: A method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration, the key points of which include the following steps:
[0006] S1: Merge the data from the CoRE-MOF 2019 database and the hMOF database, calculate the pore limit diameter (PLD) of each metal-organic framework from the merged data and screen them to build a preliminary screening database;
[0007] S2: Calculate the structural feature descriptor for each metal-organic framework in the initial screening database, and randomly select some metal-organic frameworks to carry out five-component mixed gas adsorption simulation, calculate the comprehensive performance index TSQ, and construct a dataset with the structural feature descriptor as input and TSQ as label.
[0008] S3: Divide the dataset and perform feature selection. Use an ensemble tree model as the base model for training and optimization to obtain the TSQ prediction model.
[0009] S4: Use the TSQ prediction model to predict and rank the TSQ values of the remaining metal-organic frameworks in the initial screening database, and select the top N metal-organic frameworks as adsorbents.
[0010] Preferably, in step S1, high-throughput analysis software for the pore structure properties of porous materials is used to calculate the pore limit diameter (PLD) of each metal-organic framework in batches based on the Voronoi decomposition algorithm, and the metal-organic frameworks are screened to retain metal-organic framework structures with a pore limit diameter (PLD) between 2.8 Å and 5.5 Å, forming a preliminary screening database.
[0011] Preferably, the lower limit of the channel limiting diameter PLD is 2.8 Å, which corresponds to the dynamic diameter of H2O molecules, and the upper limit is 5.5 Å, which corresponds to the dynamic diameter of SF6 molecules.
[0012] Preferably, in step S2, the structural feature descriptor of the metal-organic framework is calculated using high-throughput analysis software for the pore structure properties of porous materials. The structural feature descriptor includes the channel limit diameter PLD, the maximum cavity diameter LCD, the global cavity diameter GLD, the framework density Density, the accessible surface area ASA, the mass specific surface area GASA, the accessible pore volume APV, the unit cell volume, the porosity AVF, the volume specific surface area, the pore surface area, and the absolute pore volume.
[0013] A portion of metal-organic framework materials were randomly selected from the initial screening candidate library. The adsorption behavior of these materials on a five-component mixed gas containing H2O, SO2, H2S, CO2, and SF6 was simulated using the giant canonical Monte Carlo (GCMC) method. The overall performance index (TSQ) of the simulated metal-organic framework was then calculated.
[0014] A dataset is constructed using structural feature descriptors as input features and the calculated TSQ value as the performance label.
[0015] Preferably, TSQ is a comprehensive value of the total adsorption amount of impurity gases and the relative SF6 adsorption selectivity of impurities, and its calculation formula is as follows:
[0016]
[0017] In the formula, This represents the sum of the equilibrium adsorption capacities of the target impurity gas containing H2O, SO2, H2S, and CO2. The adsorption selectivity of the mixed impurity gas relative to SF6.
[0018] Preferably, in step S3, the dataset is divided into a training set and a test set, features are selected from the dataset, an integrated tree model is used as the base model for training, the hyperparameters of the selected base model are optimized, and the TSQ prediction model is trained using the optimized hyperparameters and the feature-selected training set.
[0019] Preferably, the feature selection involves calculating the F-statistic and mutual information value of each input feature and performance label to obtain two rankings, then using the rank average method to merge the two rankings and removing features with an average rank lower than a preset threshold.
[0020] Preferably, the features retained after feature selection include: accessible pore volume (APV), porosity (AVF), mass surface area (GASA), skeleton density (Density), volumetric surface area, pore surface area, pore limit diameter (PLD), accessible surface area (ASA), global cavity diameter (GLD), and maximum cavity diameter (LCD).
[0021] Preferably, the base model includes XGBoost or CatBoost, and the hyperparameters of the base model are optimized using a grid search combined with k-fold cross-validation.
[0022] Preferably, the method further includes a simulation verification step;
[0023] This step uses GCMC simulation to verify the TSQ value, compares it with the predicted TSQ value, calculates the average absolute relative error between the predicted and simulated TSQ values, verifies the reliability of the predicted value, and selects the optimal adsorbent by combining it with the actual engineering constraints.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0025] (1) Extremely high screening efficiency: The size of candidate materials is reduced to 22.3% of the original size through geometric primary screening (PLD constraint), and machine learning prediction replaces the high-cost GCMC simulation, improving screening efficiency by thousands of times and shortening the traditional work that takes months to hours.
[0026] (2) High prediction accuracy: By using the optimized XGBoost or CatBoost model, combined with feature selection and hyperparameter optimization, the model determination coefficient R² can reach more than 0.95, and the average absolute relative error between the predicted value and the GCMC simulation verification is less than 5%, which meets the engineering pre-screening accuracy requirements.
[0027] (3) Strong generalization ability: The model maintains stable prediction ability in the low, medium and high TSQ ranges. The Spearman rank correlation coefficient between the predicted ranking and the simulated ranking is higher than 0.9, which ensures the reliability of the screening results.
[0028] (4) Closed-loop verifiable: The present invention forms a closed-loop process of “molecular simulation-machine learning prediction-simulation verification”, avoiding the risk of pure black box prediction. The selected candidate materials have all undergone independent simulation verification and have clear physical meaning.
[0029] (5) Strong engineering practicality: It can automatically identify the key structural features that determine high-performance adsorbents and provide clear priority paths for experimental synthesis. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the overall process of high-throughput screening of SF6 recovery and regeneration adsorbent based on machine learning in this invention.
[0031] Figure 2 The flowchart shows the initial screening logic based on PLD.
[0032] Figure 3 Here are schematic diagrams of the crystal structures of the 15 high-performance MOFs materials selected in Example 1;
[0033] Figure 4 This diagram illustrates the comparison of computing power and cycle time between the traditional pure GCMC high-throughput screening method (HTCS) and the HTCS+machine learning screening scheme of this invention. Detailed Implementation
[0034] The present invention will be further described below with reference to the embodiments and accompanying drawings.
[0035] like Figure 1 As shown, a method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration includes the following steps:
[0036] S1: Merge the data from the CoRE-MOF 2019 database and the hMOF database, calculate the pore limit diameter (PLD) of each metal-organic framework from the merged data and screen them to build a preliminary screening database;
[0037] S2: Calculate the structural feature descriptor for each metal-organic framework in the initial screening database, and randomly select some metal-organic frameworks to carry out five-component mixed gas adsorption simulation, calculate the comprehensive performance index TSQ, and construct a dataset with the structural feature descriptor as input and TSQ as label.
[0038] S3: Divide the dataset and perform feature selection. Use an ensemble tree model as the base model for training and optimization to obtain the TSQ prediction model.
[0039] S4: Use the TSQ prediction model to predict and rank the TSQ values of the remaining metal-organic frameworks in the initial screening database, and select the top N metal-organic frameworks as adsorbents.
[0040] S5: Simulation verification. GCMC simulation is used to verify the TSQ simulation value, which is then compared with the predicted TSQ value to calculate the average absolute relative error, verifying the reliability of the predicted value. Combined with the actual engineering constraints, the optimal adsorbent is determined.
[0041] Example 1
[0042] S1: Geometric preliminary screening, constructing a preliminary screening database;
[0043] like Figure 2 As shown, by merging data from the CoRE-MOF 2019 database and the hMOF database, 142,234 initial MOF structures were obtained. High-throughput analysis software for the pore structure properties of porous materials (such as Zeo++) was used to calculate the pore limit diameter (PLD) of each metal-organic framework in batches based on the Voronoi decomposition algorithm. The metal-organic frameworks were then screened, and MOFs with a pore limit diameter (PLD) of 2.8 Å to 5.5 Å were retained. Finally, 31,765 materials were retained, with a screening rate of 77.7%.
[0044] S2: Calculate the structural feature descriptor and TSQ value to construct the dataset;
[0045] For the 31,765 materials retained in step S1, structural feature descriptors were calculated using high-throughput analysis software for porous material pore structure properties (such as Zeo++). These structural feature descriptors include the channel limit diameter PLD, the maximum cavity diameter LCD, the global cavity diameter GLD, the framework density Density, the accessible surface area ASA, the mass specific surface area GASA, the accessible pore volume APV, the unit cell volume, the porosity AVF, the volumetric surface area, the pore surface area, and the absolute pore volume.
[0046] From the 31,765 retained materials, 1,500 materials were randomly selected. The adsorption of a five-component mixed gas containing H2O, SO2, H2S, CO2 and SF6 was simulated using simulation software (such as RASPA2.0) at 298 K and 1 bar. The total number of simulation steps was 12 million, with the first 6 million steps for equilibration and the last 6 million steps for sampling.
[0047] The comprehensive performance index TSQ of the simulated material is calculated. TSQ is a combined value of the total adsorption amount of impurity gases and the relative SF6 adsorption selectivity of impurities. The calculation formula is as follows:
[0048]
[0049] In the formula, This represents the sum of the equilibrium adsorption capacities of the target impurity gas containing H2O, SO2, H2S, and CO2. The adsorption selectivity of the mixed impurity gas relative to SF6;
[0050] The obtained 12 structural feature descriptors are used as input features, and the TSQ value is the performance label to construct a supervised learning dataset.
[0051] S3: Build a TSQ prediction model using the dataset;
[0052] The dataset in step S2 is divided into training and testing sets in a 2:1 ratio, and XGBoost is selected as the base model.
[0053] Feature selection is performed on the dataset, and the F-statistic and mutual information value of each input feature and performance label are calculated. The F-statistic represents the linear distinguishing ability between the feature and the performance label through analysis of variance, and the mutual information represents the degree of linear and non-linear correlation between the two based on information entropy. The larger the values of the two indicators, the higher the contribution of the input feature to the performance prediction. Two types of feature rankings are generated according to the values of the two indicators from large to small. The rank average method is used to fuse the two rankings to obtain the comprehensive feature ranking. Features with an average rank lower than a preset threshold are removed, and the complete feature selection is completed.
[0054] The features retained after feature selection include: accessible pore volume (APV), porosity (AVF), surface area (GASA), skeleton density (Density), volumetric surface area, pore surface area, pore limit diameter (PLD), accessible surface area (ASA), global cavity diameter (GLD), and maximum cavity diameter (LCD).
[0055] In the XGBoost model, a grid search was performed on the hyperparameters n_estimators (300, 500), max_depth (6, 8), learning_rate (0.05, 0.1), subsample (0.8, 0.9), colsample_bytree (0.8, 0.9), and min_child_weight (1, 3), followed by 5-fold cross-validation. The optimal hyperparameter combination was finally determined to be n_estimators=300, max_depth=6, learning_rate=0.05, subsample=0.9, colsample_bytree=0.8, and min_child_weight=3.
[0056] The final model is trained on the training set using the optimal hyperparameters and preserved features, resulting in the TSQ prediction model. The evaluation result of this prediction model on the test set is R. 2 =0.9517, MAE=27.21, RMSE=35.27.
[0057] S4: Predict and sort the TSQ values of the remaining materials in the initial screening database;
[0058] The TSQ prediction model was used to perform batch prediction of the TSQ of the remaining 30,265 materials in the initial screening data. The predicted TSQ values were then sorted in descending order, and the top 15 materials were selected as candidate adsorbents. The specific types of these 15 materials are as follows: Figure 3 As shown. In this embodiment, the top three materials are: hMOF-5065611 (predicted TSQ≈2479), hMOF-5065701 (predicted TSQ≈2460), and SETDUS_clean (predicted TSQ≈2270).
[0059] S5: Simulation Verification and Engineering Evaluation
[0060] Independent GCMC simulations were performed on the 15 candidate adsorbents selected in step S4, increasing the simulation steps to 20 million, with the first 10 million steps reaching equilibrium. The results showed that the simulated TSQ values of the 15 candidate adsorbents ranged from 1723.80 to 2553.71, significantly higher than the typical level for ordinary MOFs (usually below 1000), and the top three materials selected in the simulation were consistent with the predicted top three materials and their order.
[0061] Thirty materials were stratified and randomly sampled from the middle region (500 ≤ TSQ < 1500) and the low region (TSQ < 500) of the predicted TSQ values. These were then combined with 14 other materials from the 15 candidate materials (excluding those already in the training set using SETDUS_clean), resulting in a total of 44 materials. The predicted and simulated values were independently validated, with an average absolute relative error of 4.96%. 70.5% of the sample errors were less than 5%, and the Spearman rank correlation coefficient was 0.91, ensuring the reliability of the selection results.
[0062] In practical engineering applications, suitable adsorbents are recommended based on factors such as synthesis records and cost control. In this embodiment, SETDUS_clean is derived from the CoRE-MOF experimental database, has existing synthesis records, and is a cost-controllable copper-based metal-organic framework material. Therefore, SETDUS_clean is the preferred experimental material.
[0063] Example 2
[0064] This embodiment is basically the same as Embodiment 1, except that the base model used in step S3 is CatBoost. The optimal hyperparameter combination is obtained using grid search and 5-fold cross-validation, specifically: iterations=300, depth=8, learning_rate=0.1, l2_leaf_reg=3, subsample=0.8. After optimization, the model achieves R²=0.9372, MAE=31.75, and RMSE=40.22 on the test set. The batch prediction results for the remaining materials are highly consistent with Embodiment 1, with over 80% overlap among the top 15 materials, verifying the robustness of this invention.
[0065] like Figure 4 As shown, while traditional high-throughput computational screening of HTCS yields accurate results, it suffers from significant drawbacks such as long processing times and high computational costs. This invention, however, integrates HTCS with machine learning to construct a TSQ prediction model. By leveraging the model's rapid predictive capabilities, the scale of high-precision simulations can be reduced. Without sacrificing the reliability of the final data, this significantly shortens the MOF material screening cycle and saves computational resources, making it a more efficient and economical high-throughput R&D strategy for porous materials.
[0066] Furthermore, this invention employs optimized XGBoost or CatBoost models, combined with feature selection and hyperparameter optimization, achieving a model determination coefficient R² of over 0.95. The average absolute relative error between the predicted values and GCMC simulation verification is less than 5%, demonstrating high prediction accuracy that meets engineering pre-screening accuracy requirements. This invention forms a closed-loop process of "molecular simulation - machine learning prediction - simulation verification," exhibiting strong engineering applicability.
[0067] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. Those skilled in the art, under the guidance of the present invention, can make various similar representations without departing from the spirit and claims of the present invention, and such modifications all fall within the protection scope of the present invention.
Claims
1. A method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration, characterized in that: Includes the following steps: S1: Merge the data from the CoRE-MOF 2019 database and the hMOF database, calculate the pore limit diameter (PLD) of each metal-organic framework from the merged data and screen them to build a preliminary screening database; S2: Calculate the structural feature descriptor for each metal-organic framework in the initial screening database, and randomly select some metal-organic frameworks to carry out five-component mixed gas adsorption simulation, calculate the comprehensive performance index TSQ, and construct a dataset with the structural feature descriptor as input and TSQ as label. S3: Divide the dataset and perform feature selection. Use an ensemble tree model as the base model for training and optimization to obtain the TSQ prediction model. S4: Use the TSQ prediction model to predict and rank the TSQ values of the remaining metal-organic frameworks in the initial screening database, and select the top N metal-organic frameworks as adsorbents.
2. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 1, characterized in that: In step S1, high-throughput analysis software for the pore structure properties of porous materials is used to calculate the pore limit diameter (PLD) of each metal-organic framework in batches based on the Voronoi decomposition algorithm. The metal-organic frameworks are then screened, and metal-organic framework structures with a pore limit diameter (PLD) between 2.8 Å and 5.5 Å are retained to form a preliminary screening database.
3. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 2, characterized in that: The lower limit of the channel limiting diameter PLD, 2.8 Å, corresponds to the dynamic diameter of the H2O molecule, and the upper limit, 5.5 Å, corresponds to the dynamic diameter of the SF6 molecule.
4. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 1, characterized in that: In step S2, the structural feature descriptor of the metal-organic framework is calculated using high-throughput analysis software for the pore structure properties of porous materials. The structural feature descriptor includes the channel limit diameter PLD, the maximum cavity diameter LCD, the global cavity diameter GLD, the framework density Density, the accessible surface area ASA, the mass specific surface area GASA, the accessible pore volume APV, the unit cell volume, the porosity AVF, the volume specific surface area, the channel surface area, and the absolute pore volume. A portion of metal-organic framework materials were randomly selected from the initial screening candidate library. The adsorption behavior of these materials on a five-component mixed gas containing H2O, SO2, H2S, CO2, and SF6 was simulated using the giant canonical Monte Carlo (GCMC) method. The overall performance index (TSQ) of the simulated metal-organic framework was then calculated. A dataset is constructed using structural feature descriptors as input features and the calculated TSQ value as the performance label.
5. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 4, characterized in that: TSQ is a combined value of the total adsorption amount of impurity gases and the relative SF6 adsorption selectivity of impurities. Its calculation formula is as follows: In the formula, This represents the sum of the equilibrium adsorption capacities of the target impurity gas containing H2O, SO2, H2S, and CO2. The adsorption selectivity of the mixed impurity gas relative to SF6.
6. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 4, characterized in that: In step S3, the dataset is divided into a training set and a test set. Feature selection is performed on the dataset, and an ensemble tree model is used as the base model for training. The hyperparameters of the selected base model are optimized, and the TSQ prediction model is trained using the optimized hyperparameters and the feature-selected training set.
7. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 6, characterized in that: The feature selection process involves calculating the F-statistic and mutual information value of each input feature and performance label to obtain two rankings. The two rankings are then merged using the rank average method, and features with an average rank lower than a preset threshold are removed.
8. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 7, characterized in that: The features retained after feature selection include: accessible pore volume (APV), porosity (AVF), surface area (GASA), skeleton density (Density), volumetric surface area, pore surface area, pore limit diameter (PLD), accessible surface area (ASA), global cavity diameter (GLD), and maximum cavity diameter (LCD).
9. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 8, characterized in that: The base model includes XGBoost or CatBoost, and the hyperparameters of the base model are optimized using a grid search combined with k-fold cross-validation.
10. The method for screening metal-organic framework adsorbents for sulfur hexafluoride recovery and regeneration according to claim 1, characterized in that: The method also includes a simulation verification step; This step uses GCMC simulation to verify the TSQ value, compares it with the predicted TSQ value, calculates the average absolute relative error between the predicted and simulated TSQ values, verifies the reliability of the predicted value, and selects the optimal adsorbent by combining it with the actual engineering constraints.