Method for screening of miscibility-reducing agent for carbon dioxide flooding

Abstract

The present application relates to the technical field of oil and gas field development, and particularly relates to a method for screening a miscibility-enhancing agent molecular sieve for carbon dioxide flooding to reduce minimum miscibility pressure, which comprises obtaining a molecular data set and screening to obtain a first candidate molecular set; selecting molecules with wide structural coverage and large difference in the first candidate molecular set to obtain a molecular synthesizability scoring model, and using the molecular synthesizability scoring model to predict the molecules in the first candidate molecular set to obtain a second candidate molecular set; obtaining the real value of the MMP reduction range of existing miscibility-enhancing agent molecules to construct an initial label data set, which is used to obtain a miscibility-enhancing agent molecular performance prediction model; using the miscibility-enhancing agent molecular performance prediction model to predict the molecules in the second candidate molecular set, obtaining the reduction range prediction value of each molecule in the second candidate molecular set, and obtaining a plurality of miscibility-enhancing agent molecules with excellent performance. By using the method, the screening efficiency and prediction accuracy of the miscibility-enhancing agent molecules can be greatly improved on the basis of ensuring the synthesizability of the miscibility-enhancing agent molecules.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development technology, and specifically to a molecular screening method for a mixing aid to reduce minimum miscibility pressure for carbon dioxide flooding. Background Technology

[0002] As conventional oil and gas resources gradually enter the development stage with high water cut and high recovery rates, improving oil recovery has become an important research direction in the oil and gas development field. Among them, carbon dioxide enhanced oil recovery (COLOR) technology has been widely used in recent years due to its dual advantages of improving oil recovery and reducing carbon dioxide emissions through sequestration.

[0003] Under supercritical conditions, carbon dioxide possesses strong diffusion capabilities, low viscosity, and good solubility, enabling it to penetrate reservoir pores and interact with crude oil through mass transfer, thereby reducing crude oil viscosity and improving displacement efficiency. Therefore, supercritical carbon dioxide flooding has become one of the important development technologies for low-permeability, tight, and unconventional oil reservoirs.

[0004] In actual oil displacement processes, the miscibility between supercritical carbon dioxide and crude oil has a significant impact on oil displacement efficiency. Minimum miscibility pressure (MMP) is typically used as an important indicator for evaluating the miscibility of carbon dioxide with crude oil. When the reservoir pressure reaches or exceeds the minimum miscibility pressure, carbon dioxide can form a dynamic miscibility with crude oil, thereby improving oil recovery. When the reservoir pressure is below the minimum miscibility pressure, it is difficult to form a miscible state, and the displacement efficiency decreases.

[0005] In actual oilfield development, some reservoirs have low reservoir pressures, making it difficult to meet the minimum miscibility pressure required for carbon dioxide and crude oil to form a miscible phase. This hinders miscible displacement and thus affects oil recovery. However, by adjusting the minimum miscibility pressure of the system using mixing agents, a miscible state can be achieved or approached under existing reservoir pressure conditions without significantly increasing the injection pressure. This improves displacement and increases oil recovery. Simultaneously, it reduces equipment investment, operating costs, and wellbore and formation safety risks associated with high-pressure injection. Therefore, developing mixing agent molecules that can effectively reduce the minimum miscibility pressure and their efficient screening methods has significant engineering application value. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the background art mentioned above, and to provide a method for screening mixing agent molecules that reduce minimum miscibility pressure for carbon dioxide flooding. This method can rapidly and efficiently screen mixing agent molecules that can significantly reduce minimum miscibility pressure from a large number of compounds while ensuring molecular syntheticability, thereby improving the efficiency of mixing agent development, reducing experimental costs, and providing reliable technical support for improving the recovery rate of supercritical carbon dioxide flooding.

[0007] To achieve the above objectives, the present invention provides a method for screening molecules of mixing aids for reducing minimum miscibility pressure in carbon dioxide flooding, the method comprising: S1. Obtain the molecular dataset and filter the molecules in the molecular dataset to obtain the first candidate molecule set; S2. Select molecules with broad structural coverage and high diversity from the first candidate molecule set to construct a syntheticity scoring model training molecule set. Then, obtain the syntheticity information of molecules in the syntheticity scoring model training molecule set through retrosynthetic analysis, construct a syntheticity label dataset, and then train a molecular syntheticity scoring model based on the syntheticity label dataset. Finally, use the molecular syntheticity scoring model to predict the molecules in the first candidate molecule set and filter out molecules that do not meet the syntheticity prediction criteria to obtain a second candidate molecule set. S3. Obtain the true value of the MMP reduction of existing mixing agent molecules through indoor experiments, and after mapping them one by one with existing mixing agent molecules, construct an initial label dataset. Then, construct multiple Gaussian process regression single models with different kernel functions based on the active learning framework, and input the initial label dataset into each single model for training. Finally, the trained single models are fused by hyperparameter optimization and Bayesian averaging to obtain the mixing agent molecule performance prediction model. S4. The molecular performance prediction model of the blending agent is used to predict the molecules in the second candidate molecule set, and the predicted value and prediction variance of the MMP reduction of each molecule in the second candidate molecule set are obtained. Then, the expected improvement value of each molecule is calculated, and the molecules are sorted from high to low according to the expected improvement value. The molecules with the highest expected improvement value are selected for functional group distribution statistical analysis. Then, combined with the molecular syntheticability score, raw material acquisition cost and industrial application feasibility, several high-performance blending agent molecules are obtained.

[0008] Preferably, the method for screening blending agents that reduce minimum miscibility pressure for carbon dioxide flooding further includes: step S5, referring to step S3, conducting indoor experiments on multiple high-performance blending agents, and then adding them to the initial label dataset after corresponding them one-to-one with the actual values ​​of MMP reduction obtained from the indoor experiments. Then, based on the expanded initial label dataset, a new performance prediction model for blending agents is generated, and finally, step S4 is returned to obtain multiple high-performance blending agents again.

[0009] Preferably, in step S1, the data source for the molecular dataset is a public chemical database, specifically the PubChem public chemical database.

[0010] Preferably, in step S1, screening the molecules within the molecular dataset specifically includes: Molecules within the aforementioned molecular dataset are screened for molecular weight, functional groups, and unwanted atoms. Molecular weight screening retains molecules with a molecular weight below 500. Functional group screening removes molecules containing one or more of the following functional groups: cyano, nitro, organophosphorus, thiocyanate, aldehyde, hydrazine, nitrosamine, peroxide, azide, epoxy three-membered ring, azacyclopropane, anhydride, acyl chloride, isocyanate, diazo, imine, carboxylic acid, carboxylate, β-carbonyl ketone, β-carbonyl ester, heterocyclic structures containing nitrogen / oxygen / sulfur / phosphorus atoms, and primary and secondary amino groups. Undesirable atom screening removes molecules containing one or more of the following atoms: mercury, lead, cadmium, arsenic, selenium, chromium, titanium, silver, nickel, cobalt, copper, thallium, barium, antimony, gold, chlorine, bromine, iodine, and fluorine.

[0011] Preferably, in step S2, the step of selecting molecules with broad structural coverage and high diversity within the first candidate molecule set to construct a syntheticity scoring model training molecule set specifically includes: First, based on the functional group category, the first candidate molecule set is divided into eleven major categories: aromatic ring, carbonyl, ether, alcohol hydroxyl, carbon-carbon double bond, ester, sulfur-containing functional group, amine, amide, siloxane and others. Then, based on the intermolecular similarity to characterize the degree of molecular structural similarity, a clustering method is used to screen the molecules in the eleven major categories, retaining molecules with an intermolecular similarity of less than 0.3, and finally removing duplicate molecules to construct a syntheticity scoring model training molecule set.

[0012] Preferably, in step S2, the synthesizability information of molecules in the syntheticity scoring model training molecule set obtained by retrosynthetic analysis specifically includes: whether there is an achievable retrosynthetic path, the total number of synthetic paths, and the number of synthetic steps corresponding to each synthetic path.

[0013] Preferably, in step S2, constructing the syntheticity label dataset specifically includes: The molecules in the training set of the syntheticity scoring model are encoded using extended connectivity fingerprints to obtain the molecular coding features of each molecule, which are then used as input to the piecewise function. The syntheticity score of each molecule is then calculated using the piecewise function. Finally, the molecular coding features of each molecule are mapped one-to-one with its syntheticity score to construct a syntheticity label dataset. The formula for calculating the piecewise function is:

[0014] in, The score represents the syntheticity score for a single molecule; This represents the total number of synthetic pathways for a single molecule. The maximum number of synthetic pathways for all molecules in the molecular set is trained for the syntheticity scoring model. This represents the minimum number of synthesis steps required for a single molecule. The maximum number of synthesis steps for all molecules in the molecular set is trained to create a syntheticity scoring model. The minimum number of synthesis steps required to train the syntheticity scoring model for all molecules in the set.

[0015] Preferably, in step S3, obtaining the true value of the MMP reduction of the existing mixing agent molecules through indoor experiments specifically includes: a. Sample physical property testing: Determine the density and viscosity of shale oil in the target reservoir under different formation temperature and pressure conditions to provide basic physical property parameters for interfacial tension calculation; b. Interfacial tension test: Using an interfacial tensiometer, the equilibrium interfacial tension of the pure CO2-shale oil blank system and the CO2-shale oil experimental system with added mixing agent were determined by the pendant drop method under different pressure conditions. The test was conducted with the actual formation temperature of the target reservoir as a constant temperature, and a pressure gradient covering the CO2 critical pressure to the target formation pressure was set. After the system was kept at a constant temperature at each pressure point until it was completely in equilibrium, the morphological data of 3-5 parallel droplets were collected, and the average value was taken as the interfacial tension value at that pressure point. c. MMP Calculation and Performance Evaluation: Using the test pressure as the abscissa and the equilibrium interfacial tension as the ordinate, a linear relationship between the pressure and interfacial tension in the high-pressure section is fitted. The linear extrapolation is then performed to the pressure corresponding to the interfacial tension being 0. Using the MMP of the pure CO2-shale oil blank system as a benchmark, the relative percentage reduction of MMP in the experimental system of the additive is calculated to obtain the true value of the MMP reduction corresponding to the additive molecule.

[0016] Preferably, in step S3, the kernel function includes: radial basis function kernel, radial basis function-periodic composite kernel, rational quadratic kernel, Matern-3 / 2 kernel, and paired kernel.

[0017] Preferably, the step of fusing multiple trained single models through hyperparameter optimization and Bayesian averaging to obtain a predictive model for the molecular properties of the mixing agent specifically includes: For each Gaussian process regression single model, the optimization objective is to maximize the logarithmic marginal likelihood function of the training set data. The maximum likelihood estimation method is used to automatically optimize the hyperparameters of the model kernel function, obtaining the optimal hyperparameter combination for each single model. Then, based on the Bayesian model averaging framework, the normalized fusion weights of each single model are calculated using the Bayesian formula based on the logarithmic marginal likelihood values ​​of each optimized single model. The predicted values ​​of MMP reduction magnitude of multiple single models are then weighted and summed according to the obtained weights, finally obtaining the molecular performance prediction model of the mixing agent.

[0018] Compared with the prior art, the beneficial effects of this invention are as follows: the method can rapidly and efficiently screen out mixing agent molecules that can significantly reduce minimum miscibility pressure from a large number of compounds while ensuring molecular syntheticity, thereby improving the efficiency of mixing agent development, reducing experimental costs, and providing reliable technical support for supercritical carbon dioxide flooding to improve oil recovery. Attached Figure Description

[0019] Figure 1 This is a flowchart of the molecular screening method for a mixing aid to reduce minimum miscibility pressure for carbon dioxide flooding according to the present invention. Figure 2 This is a group distribution diagram of the 5000 molecules with the best potential for improving blending performance, selected based on the expected improvement value after the performance prediction model of the first blending agent molecule in Example 1 was used. Detailed Implementation

[0020] The following provides a detailed description of specific embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the present invention.

[0021] This invention provides a method for screening molecules of mixing aids that reduce minimum miscibility pressure for carbon dioxide flooding, such as... Figure 1 As shown, the molecular screening method for mixing aids to reduce minimum miscibility pressure for carbon dioxide flooding includes: S1. Obtain the molecular dataset and filter the molecules in the molecular dataset to obtain the first candidate molecule set; S2. Select molecules with broad structural coverage and high diversity from the first candidate molecule set to construct a syntheticity scoring model training molecule set. Then, obtain the syntheticity information of molecules in the syntheticity scoring model training molecule set through retrosynthetic analysis, construct a syntheticity label dataset, and then train a molecular syntheticity scoring model based on the syntheticity label dataset. Finally, use the molecular syntheticity scoring model to predict the molecules in the first candidate molecule set and filter out molecules that do not meet the syntheticity prediction criteria to obtain a second candidate molecule set. S3. Obtain the true value of the MMP reduction of existing mixing agent molecules through indoor experiments, and after mapping them one by one with existing mixing agent molecules, construct an initial label dataset. Then, construct multiple Gaussian process regression single models with different kernel functions based on the active learning framework, and input the initial label dataset into each single model for training. Finally, the trained single models are fused by hyperparameter optimization and Bayesian averaging to obtain the mixing agent molecule performance prediction model. S4. Using a molecular performance prediction model for mixing aids, predict the MMP reduction magnitude and variance of each molecule in the second candidate molecule set. Then, calculate the expected improvement value for each molecule (the expected improvement value is used to quantify the performance improvement potential of the candidate molecule relative to the currently verified best-performing mixing aid molecule). Sort the molecules according to their expected improvement values ​​from high to low. Select the top-ranked molecules and perform functional group distribution statistical analysis. Combined with the molecule's syntheticity score, raw material acquisition cost, and feasibility for industrial application, obtain 10-50, preferably 10-30, high-performance mixing aid molecules. Selecting only 10-50 high-performance mixing aid molecules each time reduces experimental time and cost while ensuring the accuracy of the mixing aid molecule prediction. Typically, obtaining the true MMP reduction magnitude of 20 mixing aid molecules through laboratory experiments takes 1-2 days.

[0022] According to the above technical solution, the method for screening mixing agents that reduce minimum miscibility pressure for carbon dioxide flooding can quickly and efficiently screen mixing agent molecules that can significantly reduce minimum miscibility pressure from a large number of compounds while ensuring molecular syntheticability. This improves the efficiency of mixing agent development, reduces experimental costs, and provides reliable technical support for enhancing oil recovery in supercritical carbon dioxide flooding.

[0023] In practical applications, if it is desired to further improve the mixing performance of the obtained high-performance mixing agent molecules, in some preferred embodiments, the screening method for mixing agent molecules that reduce minimum miscibility pressure for carbon dioxide flooding further includes: step S5, referring to step S3, conducting indoor experiments on multiple high-performance mixing agent molecules, and adding them to the initial label dataset after corresponding them one-to-one with the true values ​​of MMP reduction obtained from the indoor experiments, and then regenerating the mixing agent molecule performance prediction model based on the expanded initial label dataset, and finally returning to step S4 to obtain multiple high-performance mixing agent molecules again.

[0024] In this embodiment of the invention, the true values ​​of the MMP reduction of the first batch of high-performance mixing agent molecules are obtained by referring to the indoor experiments in step S3, and used to expand the initial label dataset to regenerate a new mixing agent molecule performance prediction model. Then, based on the new mixing agent molecule performance prediction model, step S4 is executed to obtain a new batch of high-performance mixing agent molecules. This process is repeated. In practical applications, this can effectively improve the mixing performance of the final high-performance mixing agent molecules. The specific number of cycles can be set according to actual needs, preferably to meet the requirement of reducing the minimum miscibility pressure for actual carbon dioxide flooding.

[0025] In some preferred embodiments, in step S1, the data source for the molecular dataset is a public chemical database, specifically the PubChem public chemical database.

[0026] Specifically, the number of molecules obtained from the PubChem public chemical database should not be too small, preferably in the tens of thousands or more, so as to avoid the molecular dataset containing too few molecular samples, which would affect the accuracy of subsequent analysis.

[0027] As a concrete example, the method of retrieving molecules from the PubChem public chemical database is automated using a Python script written based on the RDKitcheminformatics toolkit.

[0028] In some other preferred embodiments, step S1, screening the molecules within the molecular dataset, specifically includes: Molecules within the molecular dataset are screened by molecular weight, functional group, and unwanted atoms.

[0029] Specifically, the molecular weight screening retains molecules with a molecular weight below 500; the functional group screening eliminates molecules containing one or more of the following functional groups: cyano, nitro, organophosphorus, thiocyanate, aldehyde, hydrazine, nitrosamine, peroxide, azide, epoxy three-membered ring, azacyclopropane, anhydride, acyl chloride, isocyanate, diazo, imine, carboxylic acid, carboxylate, β-carbonyl ketone, β-carbonyl ester, heterocyclic structures containing nitrogen / oxygen / sulfur / phosphorus atoms, primary and secondary amino groups; and the unwanted atom screening eliminates molecules containing one or more of the following atoms: mercury, lead, cadmium, arsenic, selenium, chromium, titanium, silver, nickel, cobalt, copper, thallium, barium, antimony, gold, chlorine, bromine, iodine, and fluorine.

[0030] In this embodiment of the invention, by using the above method to perform preliminary screening on the obtained molecules, the resulting multiple high-performance mixing agent molecules can meet the application requirements of the target reservoir.

[0031] In some preferred embodiments, step S2, which involves selecting molecules with broad structural coverage and high diversity within the first candidate molecule set to construct a syntheticity scoring model training molecule set, specifically includes: First, based on the functional group category, the first candidate molecule set is divided into eleven major categories: aromatic ring, carbonyl, ether, alcohol hydroxyl, carbon-carbon double bond, ester, sulfur-containing functional group, amino, amide, siloxane and others. Then, based on the intermolecular similarity to characterize the degree of molecular structural similarity, the Butina clustering method is used to screen the molecules in the eleven major categories, retaining molecules with an intermolecular similarity of less than 0.3, and selecting molecules with a similarity of less than 0.2. Finally, duplicate molecules are removed, and a syntheticity scoring model is constructed to train the molecule set.

[0032] In this embodiment of the invention, by further screening the molecules in the first candidate molecule set to construct a syntheticity scoring model training molecule set, the syntheticity of the multiple high-performance blending agent molecules can be effectively guaranteed in practical applications.

[0033] In some other preferred embodiments, in step S2, the synthesizability information of molecules in the syntheticity scoring model training molecule set obtained by retrosynthetic analysis specifically includes: whether there is an achievable retrosynthetic path, the total number of synthetic paths, and the number of synthetic steps corresponding to each synthetic path.

[0034] More preferably, in step S2, the construction of the syntheticity label dataset specifically includes: The molecules in the training set of the syntheticity scoring model are encoded using ECFP4 extended connectivity fingerprints to obtain molecular coding features (1024-dimensional vectors) of each molecule, which are then used as input to the piecewise function. The syntheticity score of each molecule is then calculated through the piecewise function. Finally, the molecular coding features of each molecule are mapped one-to-one with its syntheticity score to construct a syntheticity label dataset. The formula for calculating the piecewise function is:

[0035] in, The score represents the syntheticity score for a single molecule; This represents the total number of synthetic pathways for a single molecule. The maximum number of synthetic pathways for all molecules in the molecular set is trained for the syntheticity scoring model. This represents the minimum number of synthesis steps required for a single molecule. The maximum number of synthesis steps for all molecules in the molecular set is trained to create a syntheticity scoring model. The minimum number of synthesis steps required to train the syntheticity scoring model for all molecules in the set.

[0036] In this embodiment of the invention, by further obtaining the syntheticity scores of molecules in the training molecular set of the syntheticity scoring model based on retrosynthetic analysis and piecewise functions, and using them to construct a syntheticity label dataset, it is possible to further and effectively ensure that the multiple high-performance blending agent molecules obtained in the end have good syntheticity and reduce the difficulty of synthesis.

[0037] Specifically, the AiZynthfinder tool was used to obtain the syntheticity information of molecules in the training molecule set of the syntheticity scoring model by limiting the search time for a single molecule to 4 minutes.

[0038] In some other preferred embodiments, step S2, which involves training a molecular syntheticity scoring model based on the syntheticity label dataset, specifically includes: Using the molecular coding features of each molecule in the syntheticity label dataset as input and their corresponding syntheticity scores as output, a molecular syntheticity scoring model is obtained by training with a random forest algorithm.

[0039] Specifically, before model training, the molecules in the syntheticity label dataset are randomly divided into training and test sets in a 9:1 ratio. The training set is used for model fitting and hyperparameter optimization, while the test set is used to verify the model's generalization ability. A 5-fold cross-validation combined with a grid search method is used to automatically optimize the core hyperparameters of the random forest model. The optimized hyperparameter range includes: number of decision trees (n_estimators) of 100-1000, maximum tree depth (max_depth) of 3-30, minimum number of samples for node splits (min_samples_split) of 2-20, minimum number of samples for leaf nodes (min_samples_leaf) of 1-10, and maximum number of split features (max_features) of 0.3-0.8. The cross-validation decision coefficient R0 is used as the benchmark. 2 Maximizing the desired outcome, the optimal combination of hyperparameters was selected and fitted to obtain the optimal molecular syntheticity scoring model. The final model was validated on a test set, and its test set determination coefficient R0 was [value missing]. 2 The accuracy is ≥0.90, and the mean absolute error (MAE) is ≤0.07, which meets the accuracy requirements for subsequent batch molecular synthesizability prediction.

[0040] In some other preferred embodiments, in step S2, the screening of molecules that fail to meet the syntheticity prediction standard specifically means screening out molecules whose syntheticity value predicted by the syntheticity scoring model is lower than 0.8, thereby further ensuring that the multiple high-performance blending agent molecules obtained in the end have good syntheticity.

[0041] In some preferred embodiments, step S3 involves obtaining the true value of the MMP reduction of existing mixing agent molecules through laboratory experiments, specifically including: a. Sample property testing: First, the density and viscosity of the target reservoir shale oil under different formation temperature and pressure conditions are determined using an oven and pressure vessel to provide basic physical property parameters for interfacial tension calculation; b. Interfacial tension test: A high-temperature and high-pressure interfacial tension meter was used to determine the equilibrium interfacial tension of the pure CO2-shale oil blank system and the CO2-shale oil experimental system with added mixing agent under different pressure conditions using the pendant drop method. The test was conducted with the actual formation temperature of the target reservoir as a constant temperature, and a pressure gradient covering the CO2 critical pressure to the target formation pressure was set. After the system was kept at a constant temperature at each pressure point until it was completely in equilibrium, the morphological data of 3-5 parallel droplets were collected, and the average value was taken as the interfacial tension value at that pressure point. c. MMP Calculation and Performance Evaluation: Using the test pressure as the abscissa and the equilibrium interfacial tension value as the ordinate, a linear relationship between the pressure and interfacial tension in the high-pressure section is fitted. The linear extrapolation is performed to the pressure corresponding to the interfacial tension being 0, which is the minimum miscibility pressure (MMP) of the system. Then, using the MMP of the pure CO2-shale oil blank system as a benchmark, the relative percentage reduction of MMP in the mixed agent experimental group system is calculated to obtain the true value of the MMP reduction corresponding to the mixed agent molecule.

[0042] In this embodiment of the invention, the true values ​​of the MMP reduction of existing blending agent molecules are obtained through indoor experiments. After constructing an initial labeled dataset, this dataset is then used to build a performance prediction model for the blending agent molecules. In practical applications, this effectively ensures the accuracy of the obtained prediction model. The existing blending agent molecules can be directly obtained from existing blending agents, and the number of existing blending agent molecules used in each indoor experiment should not be too large, preferably 10-50, more preferably 10-30, thereby controlling experimental time and cost while ensuring the accuracy of model prediction.

[0043] In some specific implementations, before step S3, the method further includes: encoding the molecular structure of the existing mixing agent molecules and the molecules in the second candidate molecule set using ECFP4 extended connectivity fingerprinting, generating a 1024-dimensional vector; then, using principal component analysis (PCA) to reduce the dimensionality of the encoded vector to a 12-dimensional feature vector, which is used as the input for the subsequent model.

[0044] In some other preferred embodiments, in step S3, the kernel function includes: radial basis function kernel (RBF kernel), radial basis function-periodic composite kernel, rational quadratic kernel (RQ kernel), Matern-3 / 2 kernel, and pairwise kernel.

[0045] More preferably, the step of fusing multiple trained single models through hyperparameter optimization and Bayesian averaging to obtain a predictive model for the molecular properties of the mixing agent specifically includes: For each Gaussian process regression single model, the optimization objective is to maximize the logarithmic marginal likelihood function of the training set data. The maximum likelihood estimation method is used to automatically optimize the hyperparameters of the model kernel function, obtaining the optimal hyperparameter combination for each single model. Then, based on the Bayesian model averaging framework, the normalized fusion weights of each single model are calculated using the Bayesian formula based on the logarithmic marginal likelihood values ​​of each optimized single model. The predicted values ​​of MMP reduction magnitude of multiple single models are then weighted and summed according to the obtained weights, finally obtaining the molecular performance prediction model of the mixing agent.

[0046] In this embodiment of the invention, by employing the kernel function, hyperparameter optimization method and Bayesian averaging fusion described above, the prediction accuracy of the final obtained molecular performance prediction model for the additive can be further improved.

[0047] The present invention will be described in detail below through embodiments, but the scope of protection of the present invention is not limited thereto.

[0048] Example 1 Adopting such Figures 1-2 The method for screening molecular additives that reduce minimum miscibility pressure in carbon dioxide flooding is shown. Specifically, this method includes: S1. Obtain the molecular dataset and filter the molecules in the molecular dataset to obtain the first candidate molecule set; The molecular dataset was obtained from a public chemical database, specifically the PubChem public chemical database. The specific steps of screening molecules within the molecular dataset include: The molecules in the aforementioned molecular dataset are screened for molecular weight, functional groups, and unwanted atoms. Molecular weight screening retains molecules with a molecular weight below 500. Functional group screening removes molecules containing one or more of the following functional groups: cyano, nitro, organophosphorus, thiocyanate, aldehyde, hydrazine, nitrosamine, peroxide, azide, epoxy three-membered ring, azacyclopropane, anhydride, acyl chloride, isocyanate, diazo, imine, carboxylic acid, carboxylate, β-carbonyl ketone, β-carbonyl ester, heterocyclic structures containing nitrogen / oxygen / sulfur / phosphorus atoms, and primary and secondary amino groups. Undesirable atom screening removes molecules containing one or more of the following atoms: mercury, lead, cadmium, arsenic, selenium, chromium, titanium, silver, nickel, cobalt, copper, thallium, barium, antimony, gold, chlorine, bromine, iodine, and fluorine. S2. Select molecules with broad structural coverage and high diversity from the first candidate molecule set to construct a syntheticity scoring model training molecule set. Then, obtain the syntheticity information of molecules in the syntheticity scoring model training molecule set through retrosynthetic analysis, construct a syntheticity label dataset, and then train a molecular syntheticity scoring model based on the syntheticity label dataset. Finally, use the molecular syntheticity scoring model to predict the molecules in the first candidate molecule set and filter out molecules that do not meet the syntheticity prediction criteria to obtain a second candidate molecule set. The step of selecting molecules with broad structural coverage and high diversity within the first candidate molecule set to construct a syntheticity scoring model training molecule set specifically includes: First, based on the functional group category, the first candidate molecule set is divided into eleven major categories: aromatic ring, carbonyl, ether, alcohol hydroxyl, carbon-carbon double bond, ester, sulfur-containing functional group, amine, amide, siloxane and others. Then, based on the intermolecular similarity to characterize the degree of molecular structure similarity, the Butina clustering method is used to screen the structure of molecules in the eleven major categories, retaining molecules with an intermolecular similarity of less than 0.2, and finally removing duplicate molecules to construct a syntheticity scoring model to train the molecule set. The synthesizability information of molecules in the training set of the synthesizability scoring model obtained by retrosynthetic analysis specifically includes: whether there is an achievable retrosynthetic path, the total number of synthetic paths, and the number of synthetic steps corresponding to each synthetic path. The construction of the syntheticity label dataset specifically includes: The molecules in the training set of the syntheticity scoring model are encoded using extended connectivity fingerprints to obtain the molecular coding features of each molecule, which are then used as input to the piecewise function. The syntheticity score of each molecule is then calculated using the piecewise function. Finally, the molecular coding features of each molecule are mapped one-to-one with its syntheticity score to construct a syntheticity label dataset. The formula for calculating the piecewise function is:

[0049] in, The score represents the syntheticity score for a single molecule; This represents the total number of synthetic pathways for a single molecule. The maximum number of synthetic pathways for all molecules in the molecular set is trained for the syntheticity scoring model. This represents the minimum number of synthesis steps required for a single molecule. The maximum number of synthesis steps for all molecules in the molecular set is trained to create a syntheticity scoring model. The minimum number of synthesis steps required to train the syntheticity scoring model for all molecules in the molecular set; The molecular synthesizability scoring model trained based on the synthesizability label dataset specifically includes: Using the molecular coding features of each molecule in the syntheticity label dataset as input and its corresponding syntheticity score as output, a molecular syntheticity scoring model is obtained by training with a random forest algorithm. Specifically, the screening out of molecules that fail to meet the syntheticity prediction criteria involves screening out molecules whose syntheticity score predicted by the syntheticity scoring model is below 0.8. S3. Molecular structure encoding is performed on existing blending agent molecules and molecules in the second candidate molecule set using ECFP4 extended connectivity fingerprinting, generating a 1024-dimensional vector. Then, principal component analysis (PCA) is used to reduce the dimensionality of the encoded vector to 12-dimensional feature vectors, which are used as inputs for subsequent models. The true values ​​of the MMP reduction magnitude of existing blending agent molecules are obtained through indoor experiments, and after mapping them one-to-one with existing blending agent molecules, an initial label dataset is constructed. Then, multiple Gaussian process regression single models with different kernel functions are constructed based on the active learning framework, and the initial label dataset is input into each single model for training. Finally, the trained single models are fused through hyperparameter optimization and Bayesian averaging to obtain the blending agent molecule performance prediction model. The true value of the MMP reduction of existing mixing agent molecules was obtained through indoor experiments, specifically including: a. Sample physical property testing: Determine the density and viscosity of shale oil in the target reservoir under different formation temperature and pressure conditions to provide basic physical property parameters for interfacial tension calculation; b. Interfacial tension test: Using an interfacial tensiometer, the equilibrium interfacial tension of the pure CO2-shale oil blank system and the CO2-shale oil experimental system with added mixing agent were determined by the pendant drop method under different pressure conditions. The test was conducted with the actual formation temperature of the target reservoir as a constant temperature, and a pressure gradient covering the CO2 critical pressure to the target formation pressure was set. After the system was kept at a constant temperature at each pressure point until it was completely in equilibrium, the morphological data of 3-5 parallel droplets were collected, and the average value was taken as the interfacial tension value at that pressure point. c. MMP Calculation and Performance Evaluation: Using the test pressure as the abscissa and the equilibrium interfacial tension as the ordinate, the linear relationship between the pressure and interfacial tension in the high-pressure section is fitted. The linear extrapolation is performed to the pressure corresponding to the interfacial tension being 0. Then, using the MMP of the pure CO2-shale oil blank system as the benchmark, the relative percentage reduction of MMP in the experimental system of the additive is calculated to obtain the true value of the MMP reduction corresponding to the additive molecule. The kernel functions include: radial basis function kernels, radial basis function-periodic composite kernels, rational quadratic kernels, Matern-3 / 2 kernels, and paired kernels; The hyperparameter optimization specifically includes: for each Gaussian process regression single model, taking the maximization of the logarithmic marginal likelihood function of the training set data as the optimization objective, and using the maximum likelihood estimation method to complete the automatic optimization of the hyperparameters of the model kernel function; The Bayesian averaging fusion specifically includes: based on the Bayesian model averaging framework, using the log marginal likelihood value of each optimized single model as a basis, calculating the fusion weight corresponding to each single model through the Bayesian formula, and weighting and summing the predicted MMP reduction magnitude of each single model according to the obtained weights to obtain the fusion output result of the admixture performance prediction model. S4. The molecular performance prediction model of the blending agent is used to predict the molecules in the second candidate molecule set, and the predicted value and prediction variance of the MMP reduction of each molecule in the second candidate molecule set are obtained. Then, the expected improvement value of each molecule is calculated, and the molecules are sorted from high to low according to the expected improvement value. The molecules with the highest expected improvement value are selected and the functional group distribution statistical analysis is performed. Then, combined with the molecular syntheticity score, raw material acquisition cost and industrial application feasibility, several high-performance blending agent molecules are obtained. Step S5: Referring to step S3, conduct indoor experiments on multiple high-performance admixture molecules, and after matching them one-to-one with the true values ​​of MMP reduction obtained from the indoor experiments, add them to the initial label dataset. Then, based on the expanded initial label dataset, regenerate the admixture molecule performance prediction model, and finally return to step S4 to obtain multiple high-performance admixture molecules again.

[0050] Specifically, in step S1, 150 million SDF structure files of molecules are first downloaded from the PubChem public chemical database. Then, a Python script is written based on the RDKit toolkit to read the SDF files in batches, convert the molecular structures in MolBlock form into SMILES expressions, and construct a molecular dataset. Finally, based on the field application requirements of the target reservoir, the molecular dataset is subjected to three levels of automated initial screening. After molecular weight screening, about 100 million molecules remain in the molecular dataset. After functional group screening and undesirable atom screening, about 10 million molecules remain in the molecular dataset, which are the first candidate molecule set.

[0051] In step S2, the first candidate molecule set obtained in step S1 is first divided into eleven categories according to functional group type: aromatic ring, carbonyl, ether, alcohol hydroxyl, carbon-carbon double bond, ester, sulfur-containing functional group, amino, amide, siloxane and others. Then, based on the intermolecular similarity, the Butina clustering method is used to screen the structure of molecules in each category. Molecules with an intermolecular similarity of less than 0.2 are retained, and duplicate structures are removed. A syntheticity scoring model training molecule set containing 50,000 molecules is then constructed. The AiZynthfinder retrosynthetic analysis tool was used, with a search time of 4 minutes for a single molecule and other parameters set to default. The synthesizability information of each molecule in the training set of the synthesizability scoring model was obtained in batches, including the existence of feasible retrosynthetic paths, the total number of synthetic paths, and the number of synthetic steps corresponding to each synthetic path. The molecular structures in the training set of the synthesizability scoring model were encoded into 1024-dimensional vectors using ECFP4 extended connectivity fingerprints and used as input features of the model. The synthesizability score of each molecule was calculated using a piecewise function. The molecular encoding features of each molecule were mapped one-to-one with their synthesizability scores to construct a synthesizability label dataset. Before training the molecular syntheticity scoring model, molecules in the syntheticity label dataset were randomly divided into training and test sets at a 9:1 ratio. The training set contained 45,000 molecular data points, and the test set contained 5,000 molecular data points. A 5-fold cross-validation combined with a grid search method was used to automatically optimize the core hyperparameters of the random forest regression model, using the cross-validation coefficient of determination R0. 2 Maximizing this as the optimization objective yields the optimal combination of hyperparameters, which in turn leads to the optimal molecular syntheticity scoring model. The model is then validated on a test set, and the final model's test set determination coefficient R0 is [value missing]. 2 =0.90, mean absolute error (MAE)=0.07, which meets the batch prediction accuracy requirements; Using the trained molecular syntheticity scoring model, the syntheticity of 10 million molecules in the first candidate molecular set is predicted in batches. Molecules with syntheticity scores below 0.8 are screened out, and finally a second candidate molecular set containing 1 million molecules is obtained.

[0052] In step S3, the molecular structure of the 12 existing co-mixing agent molecules with CO2-driven mixing and the molecules in the second candidate molecule set is encoded using ECFP4 extended connectivity fingerprint to generate a 1024-dimensional vector; the dimensionality of the encoded vector is reduced by principal component analysis (PCA) to obtain a 12-dimensional feature vector, which is used as the input feature of the subsequent model. Using the actual formation temperature of the target reservoir (110℃) as a constant test temperature, a DSA100 high-temperature and high-pressure interfacial tensiometer from KRUSS (Germany) was employed to determine the MMP reduction performance data of 12 existing blending agents using the pendant drop method. The test included a blank control group and an experimental group containing the blending agent. The control group consisted of a pure CO2-shale oil blank system without any added blending agent, while the experimental group consisted of a CO2-shale oil system with 0.1% (by mass) of the target blending agent. The temperature, pressure gradient, equilibrium time, and operating procedures were completely identical for both groups. The test pressure... The gradient was set to 2 MPa / point, and the pressure range was 2-25 MPa. After the system was kept at a constant temperature for 4 hours at each pressure point until it was completely in equilibrium, the morphology data of 3-5 parallel droplets were collected, and the equilibrium interfacial tension at the corresponding pressure was calculated. A linear fit was performed with the test pressure as the abscissa and the equilibrium interfacial tension as the ordinate, and the pressure corresponding to the interfacial tension of 0 was extrapolated to obtain the MMP of the system. With the MMP of the blank control group as the benchmark, the relative percentage reduction of MMP of each mixing agent molecule was calculated to obtain the true value of the MMP reduction of the corresponding mixing agent molecule. The 12-dimensional reduced feature vectors of the 12 blending agent molecules were matched one by one with the corresponding true values ​​of MMP reduction magnitude to construct the initial label dataset. Based on an active learning framework, four Gaussian process regression single models with different kernel functions were constructed: the radial basis function kernel (RBF kernel), the rational quadratic kernel (RQ kernel), the Matern-3 / 2 kernel, and the pairwise kernel. For each single model, the optimization objective was to maximize the logarithmic marginal likelihood function of the training set data. The maximum likelihood estimation method was used to automatically optimize the hyperparameters of the kernel function, obtaining the optimal hyperparameter combination for each single model. Based on the Bayesian model averaging framework, the normalized fusion weights of each single model were calculated using the Bayesian formula, based on the logarithmic marginal likelihood values ​​of each optimized single model. The weights were negatively correlated with the prediction variance of the single model. The predicted MMP reduction values ​​of the four single models were weighted and summed according to the obtained weights to obtain the comprehensive prediction model for the performance of the mixing agent.

[0053] In step S4, the trained molecule performance prediction model for the mixing aid is used to predict the MMP reduction magnitude of 1 million molecules in the second candidate molecule set. Based on the predicted MMP reduction magnitude and prediction variance output by the model, the expected improvement value of each high-potential candidate molecule is calculated. The molecules are then sorted from highest to lowest expected improvement value, and the top 5000 molecules are selected. The functional group distribution of these 5000 molecules is statistically analyzed, and the results are attached. Figure 2 As shown; then, considering the molecular syntheticity score, raw material acquisition cost, and feasibility of industrial application, 15 high-performance mixing agent molecules were selected.

[0054] In step S5, the true values ​​of the MMP reduction magnitude of the 15 molecules obtained from the indoor experiment are added to the initial label dataset to complete the incremental optimization of the molecular performance prediction model of the mixing agent.

[0055] This invention employs a multi-round active learning closed-loop iteration to gradually improve model prediction accuracy and further screen for superior mixing aid molecules. However, due to time constraints, only one round of active learning screening was completed in this embodiment. Nevertheless, it has already achieved screening results far exceeding expectations: the dominant functional group type with the highest proportion and most concentrated distribution among the top 5000 high-potential molecules in the first round of active learning screening highly matches the functional groups contained in the molecules with the best mixing aid effect in indoor experimental verification. This indicates that this invention, through a fusion model of active learning and Gaussian process regression Bayesian model, can accurately capture characteristic structures beneficial to reducing the minimum miscibility pressure of CO2-shale oil, and can directionally enrich highly active functional groups, significantly narrowing down the range of preferred molecules. Even with only one round of active learning, candidate molecules with excellent demixing potential have been efficiently identified, fully verifying the efficiency and targeting of the screening method of this invention. If multiple rounds of active learning iteration are carried out without considering time and experimental costs, the model accuracy can be further improved, and mixing aid molecules with even better demixing performance can be screened.

[0056] Example 2 This invention demonstrates how to obtain blending agent molecules capable of achieving near-miscible flow at 20 MPa using the blending agent molecule screening method of the present invention, targeting the low-permeability shale oil reservoir in the Gulong area of ​​Daqing. Following the example in Example 1, except that steps S1-S2 were performed to obtain a second candidate molecule set (approximately 460,000 molecules); in step S3, an initial labeled dataset was constructed specifically for the crude oil components of this shale oil reservoir (containing a large amount of long-chain alkanes and colloids). Twelve blending agent molecules containing different functional groups (including esters, ethers, and ketones) were selected, and their actual MMP reduction values ​​were measured at 110°C. The MMP of the control group pure CO2-shale oil system was 31.22 MPa, while the best-performing blending agent molecule in the experimental group (a polyethylene glycol monomethyl ether) reduced the MMP to 23.97. The MMP was reduced by 29.63%. Finally, using the constructed initial labeled dataset, multiple Gaussian process regression models with different kernel functions were built based on an active learning framework to obtain a molecular performance prediction model for the additive. In step S4, after predicting 460,000 molecules, the molecular performance prediction model paid special attention to molecules (approximately 1200) with predicted MMP reductions >20%. Further considering the field constraints (environmentally friendly, non-bioaccumulative, low cost), 10 molecules were selected from the top 500 molecules with the highest expected improvement. Experimental verification showed that one molecule containing two ether bonds and one carbonyl group (named SD-07), at an addition concentration of 1.5 wt%, significantly reduced the MMP of CO2-shale oil from 31.22 MPa to 20.65 MPa, a reduction of 33.86%. Under simulated formation pressure of 20... Under MPa conditions, CO2 displacement experiments assisted by SD-07 showed that the oil displacement efficiency increased from 18.54% in immiscible flooding to 29.87%, approaching the level of miscible flooding. Subsequently, the experimental data of SD-07 was added to the initial labeled dataset to regenerate a new predictive model for the performance of blending agents. After the model update, the second round of active learning predicted three more molecules with similar structures but lower synthesis costs. Experimental verification showed that one of them reduced the MMP to 20.18 MPa, with a cost of only 60% of that of SD-07. This again demonstrates that the blending agent molecule screening method of the present invention can quickly and efficiently screen blending agent molecules that can significantly reduce minimum miscibility pressure from a large number of compounds while ensuring molecular syntheticity, thereby improving the efficiency of blending agent development, reducing experimental costs, and providing reliable technical support for improving oil recovery through supercritical carbon dioxide flooding.

[0057] Example 3 The implementation follows the same procedure as Example 1, but with a key difference: this example focuses on demonstrating the impact of syntheticity screening in step S2 on the engineering conversion rate of the final screening results, and compares the experimental verification success rate with and without syntheticity constraints.

[0058] Specifically, step S1 yields a first candidate molecule set (containing 2.1 million molecules), from which 100,000 molecules are randomly selected and divided into two groups: Group A: Based solely on the molecular performance prediction model of the mixing aids in step S3 (without going through the syntheticity screening in step S2), the true value of the reduction in MMP was directly predicted, and the top 50 molecules in the prediction were selected for experimental verification.

[0059] Group B: First, the molecular syntheticity scoring model in step S2 is used to screen the molecules (threshold 0.8) to obtain a second set of candidate molecules that meet the syntheticity standard (approximately 460,000). Then, the top 50 molecules are predicted and selected in steps S3-S4 for experimental verification.

[0060] Group A Results: Of the 50 molecules, 38 were synthesized using commercial reagents or through in-house laboratory attempts. Of these, 12 could not be synthesized within a reasonable number of steps (≤6 steps), resulting in 26 actually achievable molecules. Experimental results showed that 5 of these 26 molecules reduced MMP by more than 15% (success rate approximately 19%), with the best reduction being 23.1%. However, the synthetic routes for 2 of these 5 effective molecules required expensive palladium-catalyzed coupling reactions, making large-scale application difficult.

[0061] Group B Results: All 50 molecules had a synthesizability score ≥0.8, and 45 of them could be synthesized within 4 steps, using commercially available compounds as raw materials. Experimental verification showed that 18 of the 45 molecules reduced the MMP by more than 15% (40% success rate), with the best molecule (a hexanediol derivative with a diether structure) achieving a reduction of 29.46%. Furthermore, 16 of the 18 effective molecules had synthetic routes of no more than 4 steps, demonstrating good potential for industrial scale-up.

[0062] This embodiment fully demonstrates that the syntheticity screening step (S2) introduced in the molecule screening method for mixing aids described in this invention not only improves the success rate of experimental verification (from about 19% to 40%), but also ensures that the screened high-potential molecules have practical engineering application value, avoiding the technical dilemma of "seeing but not being able to synthesize".

[0063] The method for screening mixing agents that reduce minimum miscibility pressure for carbon dioxide flooding provided by this invention can rapidly and efficiently screen mixing agent molecules that can significantly reduce minimum miscibility pressure from a large number of compounds while ensuring molecular syntheticability. This improves the efficiency of mixing agent development, reduces experimental costs, and provides reliable technical support for enhancing oil recovery in supercritical carbon dioxide flooding.

[0064] The preferred embodiments of the present invention have been described in detail above; however, the present invention is not limited thereto. Within the scope of the inventive concept, various simple modifications can be made to the technical solutions of the present invention. To avoid unnecessary repetition, the present invention will not describe all possible combinations separately. However, these simple modifications and combinations should also be considered as part of the content disclosed in this invention and are all within the protection scope of this invention.

Claims

1. A method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding, characterized in that, The molecular screening method for mixing aids used in carbon dioxide flooding to reduce minimum miscibility pressure includes: S1. Obtain the molecular dataset and filter the molecules in the molecular dataset to obtain the first candidate molecule set; S2. Select molecules with broad structural coverage and high diversity from the first candidate molecule set to construct a syntheticity scoring model training molecule set. Then, obtain the syntheticity information of molecules in the syntheticity scoring model training molecule set through retrosynthetic analysis, construct a syntheticity label dataset, and then train the molecular syntheticity scoring model based on the syntheticity label dataset. Finally, use the molecular syntheticity scoring model to predict the molecules in the first candidate molecule set and filter out molecules that do not meet the syntheticity prediction criteria to obtain the second candidate molecule set. S3. Obtain the true value of the MMP reduction of existing mixing agent molecules through indoor experiments, and after mapping them one by one with existing mixing agent molecules, construct an initial label dataset. Then, construct multiple Gaussian process regression single models with different kernel functions based on the active learning framework, and input the initial label dataset into each single model for training. Finally, the trained single models are fused by hyperparameter optimization and Bayesian averaging to obtain the mixing agent molecule performance prediction model. S4. The molecular performance prediction model of the blending agent is used to predict the molecules in the second candidate molecule set, and the predicted value and prediction variance of the MMP reduction of each molecule in the second candidate molecule set are obtained. Then, the expected improvement value of each molecule is calculated, and the molecules are sorted from high to low according to the expected improvement value. The molecules with the highest expected improvement value are selected for functional group distribution statistical analysis. Then, combined with the molecular syntheticability score, raw material acquisition cost and industrial application feasibility, several high-performance blending agent molecules are obtained.

2. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1, characterized in that, It also includes: step S5, referring to step S3, conducting indoor experiments on multiple high-performance admixture molecules, and then adding them to the initial label dataset after corresponding them one by one with the true values ​​of MMP reduction obtained from the indoor experiments. Then, based on the expanded initial label dataset, a new admixture molecule performance prediction model is generated, and finally, step S4 is returned to obtain multiple high-performance admixture molecules again.

3. The method for screening molecule-based mixing aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1, characterized in that, In step S1, the data source for the molecular dataset is a public chemical database, specifically the PubChem public chemical database.

4. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1 or 3, characterized in that, Step S1, specifically screening the molecules within the molecular dataset, includes: Molecules within the aforementioned molecular dataset are screened for molecular weight, functional groups, and unwanted atoms. Molecular weight screening retains molecules with a molecular weight below 500. Functional group screening removes molecules containing one or more of the following functional groups: cyano, nitro, organophosphorus, thiocyanate, aldehyde, hydrazine, nitrosamine, peroxide, azide, epoxy three-membered ring, azacyclopropane, anhydride, acyl chloride, isocyanate, diazo, imine, carboxylic acid, carboxylate, β-carbonyl ketone, β-carbonyl ester, heterocyclic structures containing nitrogen / oxygen / sulfur / phosphorus atoms, and primary and secondary amino groups. Undesirable atom screening removes molecules containing one or more of the following atoms: mercury, lead, cadmium, arsenic, selenium, chromium, titanium, silver, nickel, cobalt, copper, thallium, barium, antimony, gold, chlorine, bromine, iodine, and fluorine.

5. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1, characterized in that, In step S2, the step of selecting molecules with broad structural coverage and high diversity within the first candidate molecule set to construct a syntheticity scoring model training molecule set specifically includes: First, based on the functional group category, the first candidate molecule set is divided into eleven major categories: aromatic ring, carbonyl, ether, alcohol hydroxyl, carbon-carbon double bond, ester, sulfur-containing functional group, amine, amide, siloxane and others. Then, based on the intermolecular similarity to characterize the degree of molecular structural similarity, a clustering method is used to screen the molecules in the eleven major categories, retaining molecules with an intermolecular similarity of less than 0.3, and finally removing duplicate molecules to construct a syntheticity scoring model training molecule set.

6. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1, characterized in that, In step S2, the synthesizability information of molecules in the syntheticity scoring model training molecule set obtained by retrosynthetic analysis specifically includes: whether there is an achievable retrosynthetic path, the total number of synthetic paths, and the number of synthetic steps corresponding to each synthetic path.

7. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to any one of claims 1, 5, and 6, characterized in that, In step S2, constructing the syntheticity label dataset specifically includes: The molecules in the training set of the syntheticity scoring model are encoded using extended connectivity fingerprints to obtain the molecular coding features of each molecule, which are then used as input to the piecewise function. The syntheticity score of each molecule is then calculated using the piecewise function. Finally, the molecular coding features of each molecule are mapped one-to-one with its syntheticity score to construct a syntheticity label dataset. The formula for calculating the piecewise function is: in, The score represents the syntheticity score for a single molecule; This represents the total number of synthetic pathways for a single molecule. The maximum number of synthetic pathways for all molecules in the molecular set is trained for the syntheticity scoring model. This represents the minimum number of synthesis steps required for a single molecule. The maximum number of synthesis steps for all molecules in the molecular set is trained to create a syntheticity scoring model. The minimum number of synthesis steps required to train the syntheticity scoring model for all molecules in the set.

8. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 7, characterized in that, In step S2, the process of training a molecular synthesizability scoring model based on the synthesizability label dataset specifically includes: Using the molecular coding features of each molecule in the syntheticity label dataset as input and its corresponding syntheticity score as output, a molecular syntheticity scoring model is obtained by training with a random forest algorithm. Specifically, the screening of molecules that fail to meet the syntheticity prediction criteria involves screening out molecules whose syntheticity value predicted by the syntheticity scoring model is below 0.

8.

9. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1, characterized in that, In step S3, the true value of the MMP reduction of the existing mixing agent molecules is obtained through indoor experiments, specifically including: a. Sample physical property testing: Determine the density and viscosity of shale oil in the target reservoir under different formation temperature and pressure conditions to provide basic physical property parameters for interfacial tension calculation; b. Interfacial tension test: Using an interfacial tensiometer, the equilibrium interfacial tension of the pure CO2-shale oil blank system and the CO2-shale oil experimental system with added mixing agent were determined by the pendant drop method under different pressure conditions. The test was conducted with the actual formation temperature of the target reservoir as a constant temperature, and a pressure gradient covering the CO2 critical pressure to the target formation pressure was set. After the system was kept at a constant temperature at each pressure point until it was completely in equilibrium, the morphological data of 3-5 parallel droplets were collected, and the average value was taken as the interfacial tension value at that pressure point. c. MMP Calculation and Performance Evaluation: Using the test pressure as the abscissa and the equilibrium interfacial tension as the ordinate, a linear relationship between the pressure and interfacial tension in the high-pressure section is fitted. The linear extrapolation is then performed to the pressure corresponding to the interfacial tension being 0. Using the MMP of the pure CO2-shale oil blank system as a benchmark, the relative percentage reduction of MMP in the experimental system of the additive is calculated to obtain the true value of the MMP reduction corresponding to the additive molecule.

10. The method for screening molecular mixture aids to reduce minimum miscibility pressure for carbon dioxide flooding according to claim 1 or 9, characterized in that, In step S3, the kernel functions include: radial basis function kernels, radial basis function-periodic composite kernels, rational quadratic kernels, Matern-3 / 2 kernels, and paired kernels; The process of fusing multiple trained single models through hyperparameter optimization and Bayesian averaging to obtain a predictive model for the molecular properties of the additives specifically includes: For each Gaussian process regression single model, the optimization objective is to maximize the logarithmic marginal likelihood function of the training set data. The maximum likelihood estimation method is used to automatically optimize the hyperparameters of the model kernel function, obtaining the optimal hyperparameter combination for each single model. Then, based on the Bayesian model averaging framework, the normalized fusion weights of each single model are calculated using the Bayesian formula based on the logarithmic marginal likelihood values ​​of each optimized single model. The predicted values ​​of MMP reduction magnitude of multiple single models are then weighted and summed according to the obtained weights, finally obtaining the molecular performance prediction model of the mixing agent.

Images