Method for screening and designing doped carbon-based CO2 adsorbent based on machine learning of multi-scale simulation driving

Abstract

The invention discloses a method for screening and designing a doped carbon-based CO2 adsorbent based on multi-scale simulation driving machine learning. The method comprises the following steps: S1, constructing a model; s2, extracting a free volume descriptor Vf of the slit hole; s3, CO2 adsorption characteristic calculation; s4, feature value extraction; s5, constructing a machine learning training set; s6, training a machine learning model; s7, screening a machine learning model; s8, predicting the configuration of the doped carbon-based adsorbent; and S9, guiding experimental synthesis. According to the method, a free volume descriptor is introduced into a machine learning training model for the first time so as to reflect the size of a CO2 accessible adsorption space in a micropore and doping point location coupling mode. According to the method, density functional-giant regular Monte Carlo-molecular dynamics multi-scale calculation data is used as a training set to drive machine learning, so that the complex coupling effect in the carbon-based adsorbent is comprehensively considered, and accurate prediction and screening of the functional structure matching mode of the high-performance carbon-based adsorbent are realized.

Description

Technical Field

[0001] The present invention belongs to flue gas CO 2 The field of adsorption capture technology relates to a method for directional design of carbon-based adsorbents, specifically to a method for machine learning screening and design of doped carbon-based CO based on multi-scale simulation-driven 2 Adsorbent method. Background Art

[0002] CO 2 Capture Utilization and Storage (CCUS) is seen as a key way to achieve CO 2 The only technical approach to near-zero emissions is CO 2 The capture process has the highest cost and the largest energy consumption, and is the key link in determining system cost and energy efficiency. 2 Physical adsorption capture process has CO 2 Easy to recycle, adsorbent can be recycled harmlessly, CO 2 The adsorption and desorption temperature window matches the flue gas environment and is regarded as the next generation of low-cost, low-energy CO 2 Among various adsorbent materials, carbon materials have the advantages of wide raw material sources, low cost, and good tolerance to flue gas environment. 2 There is huge application potential in the field of adsorption capture.

[0003] Achieving CO 2 The key to efficient adsorption and capture lies in the precise and directional design and preparation of high-performance adsorbents. For carbon-based adsorbents, microporous structure and doping sites are the key factors that determine their CO 2 The key factor of adsorption performance. At present, researchers mainly adjust the parameters of the carbonization-activation process through trial and error, and try to prepare high-performance adsorbent materials by simply adjusting the pores or doping environment of carbon materials. However, the micropores and doping sites in the carbon-based adsorbents used in practice are interdependent, and the combination between them will affect CO 2 At the same time, the flue gas contains N 2 , H 2 O and other impurities. The interaction between impurities and carbon-based adsorbents can also profoundly affect CO 2 The interdependence between the micropore space and doping sites of the above carbon-based adsorbents, combined with the complexity of flue gas components, has a great influence on the adsorption of CO 2 Adsorption results in nonlinear and complex coupling effects that go beyond the effects of pores or doping alone. This means that carbon-based CO is difficult to achieve through conventional trial-and-error experiments. 2 The precise and targeted design of adsorbents requires a new method for the targeted design of carbon-based adsorbents that takes the aforementioned coupling effects into consideration.

[0004] Recently, a high-throughput computational screening method based on grand canonical Monte Carlo simulation and machine learning has become a new paradigm for guiding the development of high-performance adsorbents. Currently, researchers have adopted machine learning methods driven by grand canonical Monte Carlo simulation data to design ordered porous adsorbents such as MOFs (Nature 2019, 576, 253-256, CN118609683A, CN115169220A). However, for carbon-based adsorbents used for flue gas CO 2 In the adsorption and capture, the structural characteristics of the highly coupled microporous space and doping sites in the carbon-based adsorbent, combined with the complex components of the flue gas, often simultaneously trigger multi-scale effects such as surface polarization, molecular accessible space occupation, and optimal adsorption pore size shift. Machine learning models based on single-scale simulation data alone are difficult to accurately reflect the above multi-scale effects, hindering the screening and directional design of high-performance carbon-based adsorbents. Summary of the Invention

[0005] The present invention provides a method for screening and designing doped carbon-based CO 2 adsorbents driven by multi-scale simulation, which screens and designs doped carbon-based adsorbents through a machine learning method that couples density functional theory calculations, grand canonical Monte Carlo, and molecular dynamics simulation data.

[0006] The object of the present invention is achieved through the following technical solutions:

[0007] A method for screening and designing doped carbon-based CO 2 adsorbents driven by multi-scale simulation, comprising the following steps:

[0008] Step S1, model construction: construct a slit pore carbon model with typical doping sites coupled with microporous spaces of different sizes, and optimize the geometric configuration of the slit pore carbon model using density functional theory;

[0009] Step S2, extraction of slit pore free volume descriptor V f Extract: calculate the slit pore free volume descriptor V f ;

[0010] Step S3, CO 2 Adsorption property calculation: use the grand canonical Monte Carlo method and the molecular dynamics method to calculate the adsorption thermodynamic properties and kinetic properties of CO 2 in the adsorbate respectively;

[0011] Step S4, eigenvalue extraction: extract the eigenvalues affecting the CO 2 adsorption performance based on the multi-scale simulation calculation results of Step S3: pore size, nitrogen atom content, oxygen atom content, electronegativity, dipole moment, charge, H 2 O volume fraction, N2 Volume fraction;

[0012] Step S5, Construction of machine learning training set: Divide the density functional-grand canonical Monte Carlo-molecular dynamics multi-scale calculation data into a training set and a test set, use the training set to train a machine learning model, and use the test set to evaluate the machine learning performance of the model;

[0013] Step S6, Machine learning model training: Use the slit pore free volume descriptor V f and the eigenvalue extracted in step S4 as input variables, and the CO 2 adsorption capacity as the target variable, iteratively train the machine learning model using the training set, and normalize the input slit pore free volume descriptor V f and the eigenvalue extracted in step S4 to improve the generalization ability of the model;

[0014] Step S7, Machine learning model screening: Find the optimal machine learning model applicable to CO 2 adsorption prediction based on prediction accuracy, and find the eigenvalue that dominates CO 2 adsorption based on the feature interaction analysis method;

[0015] Step S8, Configuration prediction of doped carbon-based adsorbent: Based on the eigenvalue that dominates CO 2 adsorption, predict the configuration of the doped carbon-based adsorbent with the optimal CO 2 adsorption performance;

[0016] Step S9, Guiding experimental synthesis: Based on the predicted configuration of the doped carbon-based adsorbent with the optimal CO 2 adsorption performance, directionally synthesize high-performance carbon-based adsorption materials suitable for flue gas CO 2 capture.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] (1) Compared with the traditional experimental trial-and-error method, the design method proposed by the present invention can more accurately and efficiently guide the preparation of high-performance carbon-based adsorbents. At present, the design and preparation of carbon-based adsorbents mainly rely on the experimental trial-and-error method, adjusting the parameters of the carbonization-activation process to obtain the optimal preparation conditions. Such methods are time-consuming and costly. More importantly, the multi-scale functional elements in the carbon-based adsorbent are interdependent, and it is difficult to decouple the complex coupling effect between different functional elements only through simple trial-and-error experiments, and it is impossible to achieve the precise and directional preparation of high-performance adsorbents. By introducing machine learning methods, the present invention decouples the non-linear superposition effect between multi-dimensional variables, and can effectively determine the key features that determine CO 2 adsorption performance, thereby realizing the directional design and preparation of high-performance doped carbon-based adsorbents.

[0019] (2) The machine learning training set proposed by the present invention is based on multi-scale simulation calculation results. Compared with the conventional machine learning models that only rely on grand canonical Monte Carlo simulation data, it can more comprehensively reflect the multi-scale effects caused by the coupling of micropore space and doping sites in carbon-based adsorbents, and achieve high-precision prediction and screening. Currently, for the machine learning prediction models of porous adsorbents, the calculation results of single-scale grand canonical Monte Carlo simulation are mainly used as the training set. However, the actual carbon-based adsorbents have amorphous structural characteristics. In particular, the interdependent relationship between micropore space and doping sites will lead to complex multi-scale effects. The machine learning method that only relies on grand canonical Monte Carlo simulation is difficult to achieve accurate prediction of adsorbents because it cannot comprehensively consider this multi-scale effect. The present invention proposes to drive machine learning with density functional-grand canonical Monte Carlo-molecular dynamics multi-scale calculation data, so as to comprehensively consider the complex coupling effects in carbon-based adsorbents and achieve accurate prediction and screening of the functional structure matching modes of high-performance carbon-based adsorbents.

[0020] (3) The present invention first proposes a free volume descriptor to reflect the steric effect caused by the coupling of micropore space and doping sites in doped carbon-based adsorbents. The machine learning model constructed based on this significantly improves the prediction accuracy compared with the traditional model that only considers pore size and surface chemical environment. For the steric effect caused by the coupling of doping sites and micropore space in carbon-based adsorbents, the present invention first introduces the free volume descriptor into the machine learning training model to reflect the size of the accessible adsorption space of CO 2 under the coupling mode of micropores and doping sites. The results of the examples show that after introducing the free volume eigenvalue into the machine learning model, the R 2 between the predicted value and the actual value increases from 0.687 to 0.934, while the root mean square error decreases from 0.597 to 0.253. Based on the prediction results of this model, a nitrogen and oxygen co-doped carbon-based adsorbent with an expanded pore size is prepared directionally. At room temperature and normal pressure, the CO 2 adsorption capacity is as high as 4 mmol g -1 , which is 130% higher than that of the carbon-based adsorbent prepared by the conventional experimental trial-and-error method. Description of the Drawings

[0021] Figure 1 Flow chart of machine learning screening and design of doped carbon-based CO 2 adsorbents driven by multi-scale simulation;

[0022] Figure 2 Training results of the machine learning model considering the free volume descriptor;

[0023] Figure 3 CO at room temperature of the N / O co-doped carbon-based adsorbent with an expanded pore size prepared under the guidance of the machine learning model considering the free volume descriptor 2Adsorption isotherm;

[0024] Figure 4 is the training result of the machine learning model without considering the free volume descriptor;

[0025] Figure 5 is the machine learning model without considering the free volume descriptor, guiding the preparation of the N / O co-doped carbon-based adsorbent for CO 2 adsorption isotherm at room temperature. Detailed implementation manners

[0026] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered within the protection scope of the present invention.

[0027] The present invention provides a method for screening and designing doped carbon-based CO 2 adsorbent driven by multi-scale simulation, as Figure 1 shown, the method includes the following steps:

[0028] Step S1, model construction: construct a slit pore carbon model coupling typical doping sites with different-sized microporous spaces, and optimize the geometric configuration of the slit pore carbon model using density functional theory.

[0029] In this step, the size of the slit pore in the slit pore carbon model is 0.5 - 2.0 nm, and the doped heteroatoms include individual and combined forms of N, O, B, and S, and 0 < heteroatom concentration ≤ 20 at%.

[0030] Step S2, extraction of the slit pore free volume descriptor V f : Calculate the slit pore free volume descriptor V f .

[0031] In this step, the slit pore free volume descriptor V f is the ratio of the accessible volume of gas molecules in the slit pore to the total volume.

[0032] Step S3, CO 2 adsorption characteristic calculation: Use the grand canonical Monte Carlo method and the molecular dynamics simulation method to calculate the adsorption thermodynamic characteristics and kinetic characteristics of CO 2 in the adsorbate respectively.

[0033] In this step, the adsorbate gas components are a mixture of CO 2 , N 2 and H 2 O, where 0 < CO 2 concentration ≤ 100 vol%, 0 ≤ H2 The concentration of O is ≤ 10 vol%, and the remaining gas is N 2 , and the ambient temperature is 298 K.

[0034] In this step, the grand canonical Monte Carlo method is specifically set as follows: The Metropolis algorithm is used to randomly select the movement of particles. In each GCMC simulation, the movement distribution is set to 20% exchange, 20% conformation, 40% rotation, and 20% translation. The maximum sizes of the translation angle and the rotation angle are and 5°, respectively. The equilibrium step size and the total step size are 1,000,000 and 10,000,000 respectively, and the simulation adopts a fully rigid model.

[0035] In this step, the molecular dynamics simulation method is specifically set as follows: The Dreiding force field based on the Lennard-Jones potential function is adopted. The interaction with the potential field generated by atoms includes the surface of the slit pore carbon model and the gas; Random initial atomic velocities are adopted, and the Nose–Hoover temperature control system corresponding to the simulation temperature and the Maxwell–Boltzmann distribution are used. A 1000 ps EMD simulation is carried out in the NVT ensemble with a time step of 1 fs.

[0036] Step S4, eigenvalue extraction: Based on the multi-scale simulation calculation results of step S3, extract the eigenvalues that affect the CO 2 adsorption performance. The eigenvalues include: (1) pore size; (2) nitrogen atom content; (3) oxygen atom content; (4) electronegativity; (5) dipole moment; (6) charge; (7) H 2 O volume fraction; (8) N 2 volume fraction.

[0037] In this step, the electronegativity of the model is calculated using the formula for group electronegativity:

[0038]

[0039] In the formula, N G represents the total number of atoms in the group, q represents the total atomic charge of the group, n is the total number of atoms in the model, n i is the number of i atoms in the model, is the electronegativity of the i atom before bonding.

[0040] In this step, the Mulliken charge of the geometrically optimized aromatic carbon cluster is calculated using density functional theory;

[0041] In this step, the dipole moment of the geometrically optimized aromatic carbon cluster is calculated using density functional theory.

[0042] Step S5, Construction of Machine Learning Training Set: Divide the multi-scale simulation calculation results in Step S3 into a training set and a test set, use the training set to train the machine learning model, and use the test set to evaluate the machine learning performance of the model.

[0043] In this step, the ratio of the training set to the test set is 9:1.

[0044] Step S6, Training of Machine Learning Model: Use the free volume descriptor V of the slit pore f and the eigenvalue extracted in Step S4 as input variables, and the CO 2 adsorption amount as the target variable, and iteratively train the machine learning model using the training set, and normalize the free volume descriptor V of the input slit pore f and the eigenvalue extracted in Step S4 to improve the generalization ability of the model.

[0045] In this step, the method of training the machine learning model is to introduce the Gradient Boosting Regression (GBR), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbor Algorithm (KNN) or Random Forest (RF) algorithm to train the CO 2 adsorption amount.

[0046] Step S7, Screening of Machine Learning Model: Find the optimal machine learning model applicable to CO 2 adsorption prediction based on prediction accuracy, and find the eigenvalue that dominates CO 2 adsorption based on the feature interaction analysis method.

[0047] In this step, use the regression coefficient R 2 , Root Mean Square Error RMSE as evaluation indicators to find the optimal machine learning model applicable to CO 2 adsorption prediction, where:

[0048] The calculation formula of the regression coefficient R 2 is:

[0049]

[0050] In the formula, N represents the number of data.

[0051] The calculation formula of the Root Mean Square Error RMSE is:

[0052]

[0053] In the formula, Y i actual and Y i pred represent the actual value and the predicted value respectively, represents the actual average value.

[0054] In this step, the feature interaction analysis method uses the SHapley Additive explanation (SHAP) value to deeply explain the contribution of each feature value. The SHAP value represents the average incremental contribution φj of each feature j, and the calculation formula of the SHAP value is as follows:

[0055]

[0056] In the formula, M represents the set containing all features; S is any subset of features other than feature j, |S| represents the number of features in set S, and |M| represents the total number of all features; f(S) represents that the model makes predictions only by observing the features within set S; f(S∪{j}) represents the predicted value when feature j is incorporated into set s, and |S|!(|M| - |S| - 1)! is used as the weight to balance sets of different sizes. A machine learning model is trained using the training set, and the performance of the model is evaluated using the test set.

[0057] Step S8, prediction of the configuration of the doped carbon-based adsorbent: Based on the eigenvalue that dominates CO 2 adsorption, predict the configuration of the doped carbon-based adsorbent with the optimal CO 2 adsorption performance.

[0058] Step S9, guiding experimental synthesis: Based on the predicted configuration of the doped carbon-based adsorbent with the optimal CO 2 adsorption performance, directionally synthesize a high-performance carbon-based adsorbent material suitable for flue gas CO 2 capture, and verify the experimental adsorption data and the theoretically predicted data.

[0059] In this step, the method of directional synthesis is the microwave-assisted activation method, and a carbon-based adsorbent with the required pore size and heteroatom distribution is obtained by changing the activation conditions.

[0060] Example 1:

[0061] Construct slit pore carbon models with different pore sizes ranging from 0.5 to 2.0 nm, and modify typical N and O heteroatoms in the models, where 0 < the concentration of heteroatoms ≤ 20 at%. Use density functional theory to perform geometric optimization on the constructed models, and extract the free volume and charge distribution of the models. Use grand canonical Monte Carlo simulation and molecular dynamics simulation methods to calculate the adsorption behavior of CO 2 in the constructed models at a partial pressure of 15 kPa and a temperature of 298 K. For the obtained adsorption capacity data, use pore size, free volume, nitrogen atom content, oxygen atom content, N 2 vol%, H 2 O vol%, electronegativity, dipole moment, and charge as eigenvalue for machine learning training and prediction. The results of the machine learning training are as Figure 2 shown, from Figure 2It can be seen that the R between the predicted value and the actual value 2 is as high as 0.939, and the root mean square error is only 0.243, indicating that the machine learning model based on multi-scale simulation data-driven can accurately predict the CO 2 adsorption performance of doped carbon-based adsorbents.

[0062] Example 2:

[0063] Based on the prediction results of the machine learning model driven by multi-scale simulation data in Example 1, a N / O co-doped microporous carbon material with expanded pore size was prepared by a microwave-assisted KOH activation method. Specifically, a certain mass of pulverized coal, KOH and different masses of melamine were mixed evenly, dried and then heated at a microwave power of 200 W for 10 minutes. After washing away the residual impurities with 1 M hydrochloric acid, a N / O co-doped microporous carbon material was obtained. Physicochemical structure analysis shows that it contains 3 at% of N atoms and 5 at% of oxygen atoms, and the pore size is concentrated at 0.8 nm. The adsorption isotherm of the prepared N / O co-doped microporous carbon with expanded pore size at 298 K is as Figure 3 shown, and the adsorption capacity at 1 bar is 4 mmol g -1 , further indicating that the machine learning model based on multi-scale simulation data-driven can realize the directional design and preparation of high-performance doped carbon-based adsorbents for flue gas CO 2 adsorption and capture.

[0064] Comparative Example 1:

[0065] The difference between this comparative example and Example 1 is that the free volume V f descriptor is not considered in the machine learning training, and only the pore size, nitrogen atom content, oxygen atom content, N 2 vol%, H 2 O vol%, electronegativity, dipole moment and charge are used as eigenvalue for machine learning training and prediction. The machine learning training results are as Figure 4 shown. It can be found from Figure 4 that the R between the predicted value and the actual value 2 is only 0.687, while the root mean square error is as high as 0.597, indicating that the machine learning model without considering the free volume descriptor cannot accurately predict the CO 2 adsorption performance of doped carbon-based adsorbents.

[0066] Comparative Example 2:

[0067] Based on the prediction results of the machine learning model in Comparative Example 1, an N / O co-doped microporous carbon material was prepared by the ammonia physical activation method. Specifically, the deashed coal was placed in a horizontal tube furnace, and ammonia water was injected at a rate of 140 μL / min with a jet pump, and nitrogen gas at 80 mL / min was used as the gas carrier. After activation at 900 °C for 1 h, nitrogen-doped microporous carbon was obtained. Physicochemical structure analysis showed that the content and distribution of its N and O heteroatoms were the same as those in Example 2, and the pore size was concentrated at 0.7 nm. The adsorption isotherm of the prepared N / O co-doped microporous carbon with expanded pore size at 298 K is as Figure 5 shown, and the adsorption capacity at 1 bar is 1.74 mmol / g -1 , which is only 40% of the carbon-based adsorbent with expanded pore size prepared in Example 2. This further shows that a machine learning model that does not consider free volume cannot predict and screen high-performance doped carbon-based CO 2 adsorbents.

Claims

1. A method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning, characterized in that The method includes the following steps: Step S1, model construction: Construct a slit pore carbon model with the coupling of typical doping sites and micropore spaces of different sizes, and optimize the geometric configuration of the slit pore carbon model using density functional theory; Step S2: Slit hole free volume descriptor V f Extraction: Calculation of the slit pore free volume descriptor V under the spatial coupling of doping sites and micropores f ; Step S3, CO2 adsorption property calculation: Use the grand canonical Monte Carlo method and the molecular dynamics method to calculate the adsorption thermodynamic properties and kinetic properties of CO2 in the adsorbate respectively; Step S4, eigenvalue extraction: Extract the eigenvalues affecting the CO2 adsorption performance based on the multi-scale simulation calculation results of Step S3: pore size, nitrogen atom content, oxygen atom content, electronegativity, dipole moment, charge, H2O volume fraction, N2 volume fraction; Step S5, machine learning training set construction: Divide the density functional-grand canonical Monte Carlo-molecular dynamics multi-scale calculation data into a training set and a test set, use the training set to train the machine learning model, and use the test set to evaluate the machine learning performance of the model; Step S6: Machine learning model training: using the slit pore free volume descriptor V f The feature value extracted in step S4 is used as the input variable and the CO2 adsorption amount is used as the target variable. The machine learning model is iteratively trained using the training set, and the input slit pore free volume descriptor V is used. f The feature values ​​extracted in step S4 are normalized to improve the generalization ability of the model; Step S7, machine learning model screening: Find the optimal machine learning model applicable to CO2 adsorption prediction based on prediction accuracy, and find the eigenvalues dominating CO2 adsorption based on the feature interaction analysis method; Step S8, doped carbon-based adsorbent configuration prediction: Predict the configuration of the doped carbon-based adsorbent with the optimal CO2 adsorption performance based on the eigenvalues dominating CO2 adsorption; Step S9, guiding experimental synthesis: Based on the predicted configuration of the doped carbon-based adsorbent with the optimal CO2 adsorption performance, directionally synthesize a high-performance carbon-based adsorption material suitable for flue gas CO2 capture.

2. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In the said Step S1, the size of the slit pore in the slit pore carbon model is 0.5 - 2.0 nm, and the doped heteroatoms include individual and combined forms of N, O, B, and S, and 0 < the concentration of heteroatoms ≤ 20 at%.

3. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In step S2, the slit hole free volume descriptor V f It is the ratio of the accessible volume of gas molecules in the slit hole to the total volume.

4. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In the said Step S3, the adsorbate gas components are a mixture of CO2, N2, and H2O, where 0 < the concentration of CO2 ≤ 100 vol%, 0 ≤ the concentration of H2O ≤ 10 vol%, the remaining gas is N2, and the ambient temperature is 298K.

5. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In step S3, the Grand Canonical Monte Carlo method is specifically set as follows: the movement of particles is randomly selected using the Metropolis algorithm, and the movement distribution in each GCMC simulation is set to exchange 20%, conformation 20%, rotation 40%, and translation 20%, and the maximum sizes of the translation angle and rotation angle are respectively and 5°, the equilibrium step size and total step size are 1000000 and 10000000 respectively, and the simulation adopts a fully rigid model; the molecular dynamics simulation method is specifically set as follows: a Dreiding force field based on the Lennard-Jones potential function is used, and the interaction with the potential field generated by the atoms includes the slit hole carbon model surface and the gas; random initial atomic velocities are used, and the Nose–Hoover temperature control system and Maxwell-Boltzmann distribution corresponding to the simulation temperature are used. An EMD simulation of 1000ps is performed in the NVT ensemble, and the time step is 1fs.

6. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In the said Step S4, the electronegativity of the model is calculated using the formula for group electronegativity: Where N G represents the total number of atoms in the group, q represents the total number of atomic charges in the group, n is the total number of atoms in the model, and n i is the number of atoms i in the model, is the electronegativity of atom i before bonding; The Mulliken charge of the aromatic carbon cluster after geometric configuration optimization is calculated using density functional theory; The dipole moment of the aromatic carbon cluster after geometric configuration optimization is calculated using density functional theory.

7. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In the said Step S6, the method of training the machine learning model is to introduce gradient boosting regression, extreme gradient boosting, K-nearest neighbor algorithm, or random forest algorithm to train the CO2 adsorption capacity.

8. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In step S7, the regression coefficient R 2 , root mean square error RMSE is used as the evaluation index to find the optimal machine learning model for CO2 adsorption prediction.

9. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In the said Step S7, the feature interaction analysis method uses SHAP values to deeply explain the contribution of each eigenvalue. The SHAP value represents the average incremental contribution φj of each feature j. The SHAP value calculation formula is as follows: In the formula, M represents the set containing all features; S is any feature subset other than feature j, |S| represents the number of features in set S, |M| represents the total number of all features; f(S) represents the prediction of the model by only observing the features within set S; f(S∪{j}) represents the predicted value when feature j is incorporated into set s, and S!(M - S - 1)! is used as the weight to balance sets of different sizes.

10. The method for screening and designing doped carbon-based CO2 adsorbents based on multi-scale simulation-driven machine learning according to claim 1, characterized in that In step S9, the method of directional synthesis is a microwave-assisted activation method, and the carbon-based adsorbent with the desired pore size and heteroatom distribution is obtained by changing the activation conditions.

Images