A method and apparatus for designing a carbon dioxide thickened polymer structure

Abstract

The application discloses a design method and device for carbon dioxide thickening polymer structure. The method comprises the following steps: obtaining training input data generated by high-throughput molecular dynamics simulation and completing preprocessing; constructing a double initial model of basic performance and accurate performance, and obtaining a corresponding prediction model through training and optimization; refining structure optimization design rules through explainability analysis of the accurate performance prediction model; combining the structure optimization design rules with polymer structure chemical grammar rules and high-performance structure features to generate candidate polymer structures; completing reinforcement learning iteration optimization through the basic performance prediction model preliminary screening, a multi-objective reward function and the accurate performance prediction model, and obtaining high-potential candidate structures; and obtaining a target carbon dioxide thickening polymer structure. The application realizes high-throughput screening and intelligent design of carbon dioxide thickening polymer, and improves design efficiency and performance prediction accuracy.

Description

Technical Field

[0001] This application relates to the fields of oil and gas development and polymer material design technology, and in particular to a design method and apparatus for a carbon dioxide thickening polymer structure. Background Technology

[0002] As oil and gas field development moves towards deeper and ultra-deep formations, carbon dioxide, as a highly efficient fracturing fluid, is widely used in unconventional oil and gas resource extraction, offering significant advantages such as water conservation, no reservoir damage, and high flowback efficiency. However, the extremely low viscosity of pure liquid or supercritical carbon dioxide results in poor proppant carrying capacity and low fracture-creating efficiency, severely impacting fracturing effectiveness. To address this issue, adding a small amount of thickening polymer to carbon dioxide has become a key technology. Thickening polymers dissolve and extend in carbon dioxide to form a network structure, significantly increasing viscosity.

[0003] Currently, the design methods for carbon dioxide thickening polymers mainly include experimental screening and traditional molecular dynamics simulation screening. Experimental screening relies on researchers' experience to adjust structural parameters such as functional groups and chain lengths of polymers, verifying performance through extensive synthesis and experimental testing. However, this method suffers from long development cycles, high costs, and difficulty in revealing structure-activity relationships at the molecular level. Traditional molecular dynamics simulation screening virtually screens polymers by simulating their behavior in a carbon dioxide environment at the atomic scale. However, it is limited by time and spatial scales, resulting in extremely time-consuming calculations and preventing large-scale screening. Furthermore, it lacks accuracy in quantifying the weak interactions between polymers with special functional groups such as fluorine and silicon and carbon dioxide, leading to limited prediction precision. In addition, existing design processes are characterized by a unidirectional open-loop nature, with experimental verification, simulation calculations, and model analysis operating independently. The lack of automated data feedback and iteration mechanisms makes it difficult to collaboratively optimize multiple objectives such as thickening efficiency, solubility, stability, and synthetic feasibility.

[0004] Therefore, how to achieve high-throughput screening and high-precision prediction of carbon dioxide thickening polymers, shorten the R&D cycle, and reduce R&D costs is an urgent problem to be solved in the field. Summary of the Invention

[0005] In view of this, the design method and apparatus for carbon dioxide thickening polymer structures provided in this application can integrate the advantages of machine learning and molecular dynamics to achieve high-throughput screening and intelligent iterative optimization, thereby improving design efficiency and prediction accuracy and reducing R&D costs. The design method and apparatus for carbon dioxide thickening polymer structures provided in this application are implemented as follows: This application provides a method for designing a carbon dioxide thickening polymer structure, including: Acquire training input data, preprocess the training input data to obtain processed training input data, the training input data being generated based on high-throughput molecular dynamics simulation; Construct a basic performance initial model and a precise performance initial model, and input the training input data into the basic performance initial model and the precise performance initial model respectively for model training and optimization to obtain a basic performance prediction model and a precise performance prediction model. The accurate performance prediction model is analyzed and processed to obtain structural optimization design rules; Acquire polymer structural chemical syntax rules and high-performance polymer structural features, and learn and generate the polymer structural chemical syntax rules, high-performance polymer structural features and structural optimization design rules to obtain candidate polymer structures; The candidate polymer structures are input into the basic performance prediction model for rapid initial screening to obtain the initial screening candidate structures. Obtain a multi-objective reward function that is compatible with the accurate performance prediction model, and perform reinforcement learning iterative optimization on the initial screening candidate structure based on the multi-objective reward function and the accurate performance prediction model to obtain a high-potential candidate structure. The high-potential candidate structures were simulated and verified to obtain high-performance polymer structures; The structure-performance data corresponding to the high-performance polymer structure is added to the training input data. The training input data is then input into the basic performance prediction model and the accurate performance prediction model for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

[0006] In some embodiments, the training input data includes a training set, a validation set, and a structured feature matrix. The structured feature matrix is ​​a fixed set of structured features constructed by integrating polymer structure data generated based on high-throughput molecular dynamics simulation in the preprocessing stage after structure encoding and feature extraction. It is used to characterize polymer structure information related to carbon dioxide thickening performance. The construction of the basic performance initial model and the accurate performance initial model involves inputting the training input data into the basic performance initial model and the accurate performance initial model, respectively, for model training and optimization to obtain the basic performance prediction model and the accurate performance prediction model, including: Construct an initial basic performance model, which includes decision tree configuration parameters and feature splitting criteria; Using the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, the basic performance initial model is trained based on the decision tree configuration parameters and the feature splitting criteria to obtain the basic performance initial prediction model. Based on the validation set, the basic performance prediction model is evaluated and hyperparameters are tuned to obtain the basic performance prediction model. A precise performance initial model based on graph neural networks is constructed, which includes a network architecture adapted for polymer molecular structure feature extraction and initial parameters for model training. Using the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, the accurate performance initial model is trained to obtain the accurate performance initial prediction model. Based on the validation set, the performance of the initial accurate performance prediction model is evaluated and the parameters are optimized to obtain the accurate performance prediction model.

[0007] In some embodiments, the step of obtaining polymer structural chemical syntax rules and high-performance polymer structural features, and learning and generating the polymer structural chemical syntax rules, high-performance polymer structural features, and structural optimization design rules to obtain candidate polymer structures includes: Multiple sets of initial polymer structures, coding sequences, and high-performance polymer structural features are obtained. The polymer structure chemical syntax rules are composed of the chemical bonding rules of multiple sets of initial polymer structures and coding sequences. The high-performance polymer structural features are polymer structural features that meet the preset thickening performance standards. A structure generation model is obtained by using the encoding sequences corresponding to the multiple initial polymer structures as pre-training corpus to pre-train the structure generation model, thereby obtaining the processed structure generation model. Using the structural features of the high-performance polymer and the structural optimization design rules as fine-tuning data, the processed structure generation model is fine-tuned to obtain the target structure generation model. The target structure generation model is used to learn the chemical syntax rules of polymer structure and characterize structural features that have a positive contribution to carbon dioxide thickening performance. Specific operating condition parameters and functional group requirements are obtained, and the specific operating condition parameters and functional group requirements are input into the target structure generation model for sequence generation processing to obtain multiple sets of candidate polymer coding sequences. The candidate polymer coding sequences are decoded to obtain the candidate polymer structures.

[0008] In some embodiments, obtaining a multi-objective reward function adapted to the accurate performance prediction model, and performing reinforcement learning iterative optimization on the initially screened candidate structures based on the multi-objective reward function and the accurate performance prediction model to obtain high-potential candidate structures, includes: The core performance indicators, synthesis feasibility, and stability of the carbon dioxide thickening polymer are obtained. These indicators are then weighted and combined to construct a multi-objective reward function adapted to the precise performance prediction model. The expression of the multi-objective reward function is: R = 0.4 × R1 + 0.3 × R2 + 0.2 × R3 + 0.1 × R4, where R1 is the viscosity enhancement factor among the core performance indicators, R2 is the carbon dioxide solubility among the core performance indicators, R3 is the synthesis feasibility, and R4 is the stability. Based on the aforementioned structural optimization design rules, the candidate structures after initial screening are subjected to structural modification processing that conforms to chemical rationality, resulting in modified candidate structures. The modified candidate structure is input into the precise performance prediction model for performance evaluation, and the core performance evaluation results corresponding to each modified candidate structure are obtained. The core performance evaluation results and the multi-objective reward function are quantitatively calculated to obtain the reward function value of the corresponding candidate structure; Based on the reward function value, the parameters of the reinforcement learning optimization model are updated to obtain the processed reinforcement learning optimization model. The processed reinforcement learning optimization model performs a preset number of iterations on the initially screened candidate structures. After the iterations are completed, candidate structures whose reward function values ​​meet the preset requirements are selected to obtain high-potential candidate structures.

[0009] In some embodiments, the analysis and processing of the accurate performance prediction model to obtain structural optimization design rules includes: An interpretability analysis tool is obtained, and the accurate performance prediction model is associated with the interpretability analysis tool to obtain a correspondence. The correspondence is used to characterize the correspondence between the model prediction logic and the contribution analysis of polymer structural features related to carbon dioxide thickening performance. The accurate performance prediction model is correlated with the core performance indicators in the training input data to obtain the contribution correlation results. The contribution correlation results are quantified to obtain the feature contribution results; The feature contribution results are analyzed and processed to obtain the structural feature contribution ranking results and correlation relationships. The correlation relationships are used to characterize the relationship between the structural feature contribution ranking results and the core performance indicators. The feature contribution results, structural feature contribution ranking results, and correlation relationships are processed by rule extraction to obtain structural optimization design rules.

[0010] In some embodiments, the simulation verification process of the high-potential candidate structure to obtain a high-performance polymer structure includes: The appropriate parameters and environmental conditions for molecular dynamics simulation are obtained. These parameters and environmental conditions are fixed simulation parameters and operating conditions preset based on the actual application scenario of carbon dioxide thickening polymers. They are used to simulate and verify the performance of high-potential candidate structures under a unified benchmark. The high-potential candidate structure, the adaptation parameters, and the environmental conditions are simulated to obtain simulation results, which include structural conformation, interaction characteristics, and core performance data. The core performance data is extracted and analyzed to obtain the core performance verification results; Obtain a preset performance verification standard, compare the core performance verification results with the preset performance verification standard, and obtain candidate structures that meet the standard; The candidate structures that meet the standards are subjected to stability and feasibility verification to obtain high-performance polymer structures.

[0011] This application provides a carbon dioxide thickening polymer structure device, comprising: The acquisition module is used to acquire training input data, preprocess the training input data, and obtain processed training input data, wherein the training input data is generated based on high-throughput molecular dynamics simulation. A construction module is used to construct a basic performance initial model and a precise performance initial model. The training input data is input into the basic performance initial model and the precise performance initial model respectively for model training and optimization to obtain a basic performance prediction model and a precise performance prediction model. The processing module is used to analyze and process the accurate performance prediction model to obtain structural optimization design rules; The acquisition module is also used to acquire polymer structure chemical syntax rules and high-performance polymer structure features, and to learn and generate the polymer structure chemical syntax rules, high-performance polymer structure features and structure optimization design rules to obtain candidate polymer structures. The processing module is also used to input the candidate polymer structure into the basic performance prediction model for rapid initial screening to obtain the initial screening candidate structure. The acquisition module is further configured to acquire a multi-objective reward function adapted to the accurate performance prediction model, and perform reinforcement learning iterative optimization on the initial screening candidate structure based on the multi-objective reward function and the accurate performance prediction model to obtain a high-potential candidate structure. The processing module is also used to perform simulation verification on the high-potential candidate structure to obtain a high-performance polymer structure. The processing module is further configured to add the structure-performance data corresponding to the high-performance polymer structure to the training input data, and input the training input data into the basic performance prediction model and the accurate performance prediction model respectively for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

[0012] The computer device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0013] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method described in this application embodiment.

[0014] The present application provides a method and apparatus for designing carbon dioxide thickening polymer structures. This method involves acquiring and preprocessing training input data generated from high-throughput molecular dynamics simulations; constructing dual initial models for basic and precise performance, and optimizing them through training to obtain corresponding prediction models; extracting structural optimization design rules through interpretability analysis of the precise performance prediction model; generating candidate polymer structures by combining these rules with polymer structural chemistry rules and high-performance structural features; and obtaining high-potential candidate structures through initial screening using the basic performance prediction model and iterative optimization via reinforcement learning in collaboration with a multi-objective reward function and the precise performance prediction model. This yields the target carbon dioxide thickening polymer structure. This approach integrates the advantages of machine learning and molecular dynamics, enabling high-throughput screening and intelligent iterative optimization, improving design efficiency and prediction accuracy, reducing R&D costs, and addressing the technical problems mentioned in the background section. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram illustrating the implementation process of a design method for a carbon dioxide thickening polymer structure provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the implementation process for obtaining candidate polymer structures, provided in an embodiment of this application. Figure 3 This is a schematic diagram of a carbon dioxide thickening polymer structure device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0019] Figure 1 This is a schematic flowchart illustrating the implementation of a carbon dioxide thickening polymer structure design method provided in this application embodiment, including steps 101 to 108. Wherein, Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order of a method for designing carbon dioxide thickening polymer structures. Where the final result can be achieved, Figure 1 The steps shown can be performed in parallel or in reverse order.

[0020] Step 101: Obtain training input data, preprocess the training input data, and obtain processed training input data.

[0021] In this embodiment, multiple sets of polymer structure-property data generated based on high-throughput molecular dynamics simulations are first acquired as training input data. The training input data is then preprocessed.

[0022] The polymer molecular structure is encoded by clearly marking information such as the boundaries of repeating units, functional group types, and topological structure, and the polymer molecular structure is transformed into recognizable sequence data. At the same time, molecular descriptors such as molecular weight, number of functional groups, number of hydrogen bond donors, and number of hydrogen bond acceptors are extracted and integrated to form a structured feature matrix.

[0023] For the three core performance indicators of viscosity enhancement factor, carbon dioxide solubility and thermal stability, the Min-Max normalization method is used to process the data, mapping the original data to the [0,1] interval to eliminate the influence of dimensional differences on model training.

[0024] The preprocessed structured feature matrix and standardized performance data are divided into training, validation, and test sets in a 7:2:1 ratio. The training, validation, and test sets, along with the structured feature matrix and standardized performance data, are then integrated to obtain the processed training input data.

[0025] Step 102: Construct the basic performance initial model and the accurate performance initial model. Input the training input data into the basic performance initial model and the accurate performance initial model respectively for model training and optimization to obtain the basic performance prediction model and the accurate performance prediction model.

[0026] In this embodiment of the application, a basic performance initial model is constructed. The basic performance initial model adopts the random forest algorithm, sets the number of decision trees to 100, does not limit the maximum tree depth, sets the minimum number of samples for node splitting to 2, sets the minimum number of samples for leaf nodes to 1, sets the random seed to 42, and uses the Gini coefficient as the feature splitting criterion.

[0027] Based on decision tree configuration parameters and feature splitting criteria, the structured feature matrix in the processed training input data is used as input, and the three standardized core performance indicators are used as output. The initial basic performance model is trained using the training set: 100 decision trees are constructed, and each decision tree learns parameters based on a randomly selected feature subset. Finally, the prediction results of each decision tree are fused by voting to obtain the initial basic performance prediction model.

[0028] The initial prediction model for basic performance is evaluated based on the validation set, and evaluation metrics such as mean absolute error and coefficient of determination are calculated. If the coefficient of determination does not reach 0.85, hyperparameters such as the number of decision trees and the maximum tree depth are adjusted by grid search (e.g., the number of decision trees is adjusted to 150 and the maximum tree depth is adjusted to 30), and the training process is repeated until the prediction error on the test set is controlled within the preset error range, thus obtaining the basic performance prediction model.

[0029] An initial model for accurate performance was constructed using a graph neural network algorithm. The initial parameters included the network structure, optimizer, loss function, and regularization strategy: The network structure consisted of 3 graph convolutional layers, 1 global pooling layer, and 2 fully connected layers. The output dimensions of the 3 graph convolutional layers were 64, 128, and 256, respectively, used to extract local and global features from the molecular graph. The global pooling layer used mean pooling to transform the variable-length molecular graph features into fixed-dimensional feature vectors. The output dimensions of the 2 fully connected layers were 128 and 3, respectively, ultimately outputting the predicted values ​​of the three core performance metrics. The optimizer used Adam with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The loss function was mean squared error, and the regularization strategy was L2 regularization with a regularization coefficient of 0.0001.

[0030] Using the training set and encoded polymer molecular graph data from the processed training input data as input, the initial model for accurate performance is trained and the network weight parameters are iteratively updated to obtain the initial prediction model for accurate performance.

[0031] The initial accurate performance prediction model is evaluated based on the validation set. If the mean square error of the validation set does not decrease for 10 consecutive rounds, the training is stopped early and the optimal parameters are saved. The saved optimal parameter model is then validated using the test set. When the coefficient of determination is ≥0.92, the accurate performance prediction model is obtained.

[0032] Step 103: Analyze and process the accurate performance prediction model to obtain structural optimization design rules.

[0033] In this embodiment of the application, an interpretability analysis tool adapted to the accurate performance prediction model is obtained, and the accurate performance prediction model and the interpretability analysis tool are associated to establish a correspondence between the model prediction logic and the structural feature contribution analysis.

[0034] Using interpretability analysis tools, the polymer structural features learned by the accurate performance prediction model are correlated with the core performance indicators in the processed training input data to obtain the contribution correlation results between each structural feature and the core performance indicators. The core performance indicators are the viscosity enhancement factor, carbon dioxide solubility, and thermal stability, which have been standardized in the training input data preprocessing stage.

[0035] The contribution correlation results are quantified to obtain the feature contribution results of each structural feature to each core performance indicator. A positive feature contribution result indicates that the corresponding structural feature improves the core performance indicator, while a negative feature indicates that the corresponding structural feature inhibits the core performance indicator. The absolute value reflects the degree of contribution of the structural feature to the core performance indicator.

[0036] The feature contribution results are comprehensively analyzed and processed to obtain the correlation between the structural feature contribution ranking results and single structural features and core performance indicators. The feature contribution results, structural feature contribution ranking results and correlation are processed by rule extraction to obtain the influence rules of different structural features (such as functional group type, chain length, topology, etc.) on carbon dioxide thickening performance, that is, structural optimization design rules.

[0037] Step 104: Obtain the polymer structure chemical syntax rules and high-performance polymer structure features. Learn and generate the polymer structure chemical syntax rules, high-performance polymer structure features, and structure optimization design rules to obtain candidate polymer structures.

[0038] In the embodiments of this application, multiple sets of initial polymer structures, coding sequences of various known carbon dioxide thickening polymers, and high-performance polymer structural features are obtained. The polymer structure chemical syntax rules are composed of multiple sets of initial polymer structures and coding sequences of known carbon dioxide thickening polymers. The high-performance polymer structural features are polymer structural features with a viscosity increase factor ≥ 5.0.

[0039] A structure generation model was obtained. Multiple sets of initial polymer structures and the encoding sequences of known carbon dioxide thickening polymers were used as pre-training data to pre-train the structure generation model. The number of training rounds was set to 50, the learning rate to 0.0001, and the batch size to 32. This allowed the structure generation model to learn the chemical grammar rules of polymer structures (such as the connection of repeating units and the substitution rules of functional groups), resulting in the processed structure generation model.

[0040] Using the structural features of high-performance polymers and structural optimization design rules as fine-tuning data, the processed structure generation model is fine-tuned to focus on structural features that positively contribute to the thickening performance of carbon dioxide, thus obtaining the target structure generation model.

[0041] The specific operating parameters and functional group requirements corresponding to the actual application scenarios of carbon dioxide thickening polymers are obtained. The specific operating parameters and functional group requirements are input into the target structure generation model for sequence generation processing. The maximum sequence length is set to 100 and the sampling temperature is 0.7 to obtain multiple sets of candidate polymer coding sequences.

[0042] The coding sequences of multiple candidate polymers are decoded to obtain the corresponding molecular structures. Monte Carlo simulation is used to conduct preliminary screening of synthesis feasibility, filtering out thermodynamically unstable and unfeasible structures, and retaining 30-50 sets of molecular structures that meet the requirements, i.e., candidate polymer structures.

[0043] Step 105: Input the candidate polymer structure into the basic performance prediction model for rapid initial screening to obtain the candidate structure after initial screening.

[0044] In this embodiment, the generated candidate polymer structures are input into the basic performance prediction model for rapid initial screening. The basic performance prediction model outputs the core performance prediction results corresponding to each candidate polymer structure at a speed of milliseconds. Candidate structures whose core performance prediction results meet the preset initial screening criteria are selected to obtain the initial screening candidate structures.

[0045] Step 106: Obtain the multi-objective reward function that is adapted to the accurate performance prediction model. Based on the multi-objective reward function and the accurate performance prediction model, perform reinforcement learning iterative optimization on the candidate structures after the initial screening to obtain high-potential candidate structures.

[0046] In this embodiment of the application, a multi-objective reward function is obtained. This multi-objective reward function is a composite reward function, and its expression is R=0.4×R1+0.3×R2+0.2×R3+0.1×R4, where R1 is the standardized value of viscosity enhancement factor, R2 is the standardized value of carbon dioxide solubility, R3 is the synthesis feasibility score (within the range of 0-1), and R4 is the standardized value of thermal stability.

[0047] The reinforcement learning optimization model is obtained by taking the candidate structures after initial screening as the initial input. The agent of the reinforcement learning optimization model modifies the structure based on the sequence of candidate structures (such as replacing functional groups and adjusting chain lengths) to obtain the modified candidate structures. The modified candidate structures are then input into the accurate performance prediction model for performance evaluation to obtain the core performance evaluation results corresponding to each modified candidate structure. Based on the core performance evaluation results and the multi-objective reward function, quantitative calculation is performed to obtain the corresponding reward function values.

[0048] The reinforcement learning optimization model and reward function value are updated with parameters. The agent learns the structure modification policy corresponding to high reward through the policy gradient algorithm. The reinforcement learning optimization model is iterated and optimized in multiple steps, and the optimal model is saved every 1000 steps. After the iteration is completed, the 10 candidate structures with the highest reward function value are selected to obtain high potential candidate structures.

[0049] Step 107: Perform simulation verification on the high-potential candidate structures to obtain high-performance polymer structures.

[0050] In this embodiment, the high-potential candidate structure is subjected to simulation verification, specifically as follows: The appropriate parameters and environmental conditions required for molecular dynamics simulation are obtained (determined based on the actual application scenario of the carbon dioxide thickening polymer); the high-potential candidate structure, appropriate parameters, and environmental conditions are input into the molecular dynamics simulation system for simulation processing to obtain simulation results containing structural conformation, interaction characteristics, and core performance-related data; the core performance-related data in the simulation results are extracted and analyzed to obtain core performance verification results; a preset performance verification standard is obtained; the core performance verification results are compared with the preset performance verification standard to screen out candidate structures that meet the standard; the stability and feasibility of the candidate structures that meet the standard are verified to obtain polymer structures that pass the verification, i.e., high-performance polymer structures.

[0051] Step 108: Add the structure-performance data corresponding to the high-performance polymer structure to the training input data, and input the training input data into the basic performance prediction model and the accurate performance prediction model respectively for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

[0052] In this embodiment, the structure-performance data corresponding to the high-performance polymer structure is supplemented into the training input data, and the basic performance prediction model and the accurate performance prediction model are updated and trained to optimize the prediction accuracy and generalization ability of the model. After the model update is completed, the high-performance polymer structure is output, which is the target carbon dioxide thickening polymer structure.

[0053] This application's embodiments achieve a tiered screening process—from rapid initial screening of a large number of candidate structures to high-precision performance evaluation—through the synergistic application of a basic performance prediction model and a precise performance prediction model. This addresses the pain points of low efficiency and high computational cost in traditional simulation screening, significantly increasing screening throughput. By integrating structural optimization design rules with reinforcement learning multi-objective optimization, it achieves synergistic optimization of thickening performance, solubility, stability, and synthetic feasibility, overcoming the limitations of single-objective design in existing technologies and enhancing the comprehensive application value of the target polymer. The model update mechanism forms a self-evolving cycle of design-verification-learning-redesign, continuously improving model prediction accuracy and structural design rationality, reducing reliance on human experience, shortening the R&D cycle, and lowering R&D costs.

[0054] In the above Figure 1 Based on the above, embodiments of this application also provide a schematic diagram of the implementation process for obtaining candidate polymer structures. For example... Figure 2 As shown, steps 201 to 205 are included: Step 201: Obtain multiple sets of initial polymer structures, coding sequences, and high-performance polymer structural features.

[0055] In this embodiment, multiple sets of initial polymer structures, coding sequences, and high-performance polymer structural features are first obtained. The coding sequences adopt the P-SMILES format, containing P-SMILES sequences of 353 known carbon dioxide thickening polymers, with over 100,000 initial polymer structures. The polymer structural chemical syntax rules are composed of the implicit rules between the multiple sets of initial polymer structures and the P-SMILES coding sequences, specifically including chemical rationality rules such as the connection mode of polymer repeating units and the substitution rules of functional groups. High-performance polymer structural features are defined as structural features that meet a preset thickening performance standard, which is set as a viscosity increase factor ≥ 5.0, corresponding to the common features (such as specific functional group types, topological characteristics, etc.) inherent in polymer structures with this performance level in the high-throughput molecular dynamics simulation dataset.

[0056] Step 202: Obtain the structure generation model. Using the encoding sequences corresponding to multiple initial polymer structures as pre-training corpus, pre-train the structure generation model to obtain the processed structure generation model.

[0057] In this embodiment, a structure generation model is obtained. Multiple sets of initial polymer structures and P-SMILES encoding sequences of 353 known carbon dioxide thickening polymers are used as pre-training data to pre-train the structure generation model. The parameters for the pre-training process are set as follows: 50 training epochs, a learning rate of 0.0001, and a batch size of 32. Through this pre-training, the structure generation model fully learns the chemical syntax rules of polymer structures, accurately captures the periodic characteristics and structural diversity of polymer chains, and ensures that the generated polymer structures are logically sound in terms of chemical connections. After pre-training, the processed structure generation model is obtained.

[0058] Step 203: Using the structural characteristics of high-performance polymers and structural optimization design rules as fine-tuning data, the processed structure generation model is fine-tuned to obtain the target structure generation model.

[0059] In this embodiment, high-performance polymer structural features and structural optimization design rules are used as fine-tuning data to fine-tune the processed structure generation model. Through the fine-tuning process, the structure generation model further learns and strengthens its ability to characterize structural features that positively contribute to carbon dioxide thickening performance, while also taking into account the constraints of polymer structural chemical syntax rules, ultimately obtaining the target structure generation model.

[0060] Step 204: Obtain specific operating parameters and functional group requirements, input the specific operating parameters and functional group requirements into the target structure generation model for sequence generation processing, and obtain multiple sets of candidate polymer coding sequences.

[0061] In this embodiment, specific operating parameters and functional group requirements corresponding to actual application scenarios of carbon dioxide thickening polymers are obtained. The specific operating parameters refer to application conditions in actual oil and gas development, such as common supercritical carbon dioxide application conditions like a temperature of 377K and a pressure of 55MPa. Functional group requirements are determined based on the thickening performance improvement target, such as functional groups containing ether groups or aromatic rings that positively contribute to carbon dioxide thickening performance. The specific operating parameters and functional group requirements are used as input conditions and fed into the target structure generation model for sequence generation. The parameters for the generation process are set as follows: maximum sequence length is 100, and sampling temperature is 0.7℃. Through the generation process, the model outputs 100-200 sets of candidate polymer P-SMILES encoding sequences that meet the requirements of operating condition adaptability and chemical rationality.

[0062] Step 205: Decode the multiple candidate polymer coding sequences to obtain the candidate polymer structures.

[0063] In this embodiment, the generated 100-200 sets of candidate polymer P-SMILES encoding sequences are decoded to convert the sequence data into corresponding polymer molecular structures. Subsequently, Monte Carlo simulation is used to preliminarily screen the synthetic feasibility of the decoded molecular structures. The screening criteria are thermodynamic stability and polymerization reaction feasibility: molecular structures containing easily broken chemical bonds, thermodynamically unstable structures, and structures that are difficult to polymerize under existing synthetic process conditions are filtered out. Finally, 30-50 sets of molecular structures that meet the screening requirements are retained as candidate polymer structures.

[0064] This application's embodiments explicitly define the training input data as including a training set, a validation set, and a structured feature matrix. Through a scientific dataset partitioning and model training process, it ensures the comprehensiveness and generalization ability of the model's learning, addressing the problems of low training data utilization and poor prediction stability in traditional models. The initial model for basic performance employs the random forest algorithm, whose robustness and high training efficiency enable rapid initial screening of candidate structures, significantly improving screening efficiency. The initial model for precise performance uses a graph neural network algorithm to accurately capture the intrinsic relationship between molecular structure and performance. Further optimization techniques, such as decision tree configuration parameter tuning and early stopping strategies, enhance the prediction reliability and stability of the two models.

[0065] In some embodiments, the training input data includes a training set, a validation set, and a structured feature matrix. The structured feature matrix is ​​a fixed set of structured features constructed by integrating polymer structure data generated based on high-throughput molecular dynamics simulations during the preprocessing stage after structure encoding and feature extraction. It is used to characterize polymer structure information related to carbon dioxide thickening performance. The basic performance initial model and the precise performance initial model are constructed by inputting the training input data into the basic performance initial model and the precise performance initial model, respectively, for model training and optimization to obtain the basic performance prediction model and the precise performance prediction model. This includes: constructing the basic performance initial model, which includes decision tree configuration parameters and feature splitting criteria.

[0066] Specifically, a basic performance initial model is constructed, which adopts the random forest algorithm. The specific settings of the decision tree configuration parameters and feature splitting criteria are as follows: the number of decision trees is set to 100, the maximum tree depth is not manually limited and is automatically determined by the characteristics of the training data itself, the minimum number of samples for node splitting is 2, the minimum number of samples for leaf nodes is 1, the random seed is set to 42, and the feature splitting criterion adopts the Gini coefficient. The effectiveness of feature splitting is judged by the criterion to ensure that the model can effectively capture the correlation between polymer structural features and thickening performance.

[0067] Furthermore, using the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, the initial basic performance model is trained based on the decision tree configuration parameters and feature splitting criteria to obtain the initial basic performance prediction model.

[0068] Specifically, model training is conducted based on decision tree configuration parameters and feature splitting criteria. The structured feature matrix from the training input data is used as input data. This matrix contains molecular descriptor information such as the polymer's molecular weight, number of functional groups, and number of hydrogen bond donors / acceptors. Three core performance indicators—preprocessed and standardized viscosity enhancement factor, carbon dioxide solubility, and thermal stability—are used as output targets. The initial basic performance model is trained using the training set. During training, samples are drawn from the training set to construct 100 independent decision trees. Each decision tree learns parameters based on a randomly selected subset of features, without relying on all feature dimensions, effectively reducing the risk of overfitting. After training, the prediction results of the 100 decision trees are merged using a voting method, and a comprehensive multi-objective performance prediction value is output to obtain the initial basic performance prediction model.

[0069] Furthermore, based on the validation set, the initial prediction model for basic performance is evaluated and its hyperparameters are tuned to obtain the basic performance prediction model.

[0070] Specifically, the initial prediction model for basic performance is evaluated based on the validation set in the training input data. Core evaluation metrics such as mean absolute error and coefficient of determination are calculated. The coefficient of determination measures the degree of fit between the model's predicted values ​​and the actual values; a value closer to 1 indicates higher prediction accuracy. If the evaluation results show that the coefficient of determination does not reach the preset threshold of 0.85, the decision tree configuration parameters are optimized using a grid search method. For example, the number of decision trees is adjusted to 150, and the maximum tree depth is set to 30. The above training process is repeated until the model's prediction error on the test set is controlled within the preset error range. At this point, the model has a stable and reliable basic performance prediction capability and is determined to be a basic performance prediction model.

[0071] Furthermore, a precise performance initial model based on graph neural networks is constructed, which includes a network architecture adapted for polymer molecular structure feature extraction and initial parameters for model training.

[0072] Specifically, an initial model for accurate performance is constructed using a graph neural network algorithm. The network structure employs a combination of graph convolutional layers, global pooling layers, and fully connected layers. Three graph convolutional layers, each with output dimensions of 64, 128, and 256, are used to extract local and global features from the polymer molecular graph layer by layer. The global pooling layers use mean pooling to transform the variable-length molecular graph features into fixed-dimensional feature vectors, ensuring stable processing by subsequent fully connected layers. Two fully connected layers, with output dimensions of 128 and 3, ultimately output the predicted values ​​of the three core performance indicators. The optimizer uses the Adam optimizer with a learning rate of 0.001 to adaptively adjust the parameter update step size during model training. The loss function uses mean squared error to quantify the deviation between the model's predicted values ​​and the true values, guiding iterative optimization of the model parameters. L2 regularization is used with a regularization coefficient of 0.0001 to penalize the model parameters, suppressing overfitting and improving generalization ability.

[0073] Furthermore, the initial accurate performance model is trained by taking the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, thus obtaining the initial accurate performance prediction model.

[0074] Specifically, using the training set from the training input data as the data foundation, the encoded polymer molecular diagram data is combined with the structured feature matrix as input data for the initial accurate performance model, and model training is carried out. During training, the training data is divided into batches of 64, and the number of training epochs is set to 100. The prediction results are calculated through forward propagation, the bias is calculated using the loss function, and then the network weight parameters are iteratively updated through backpropagation to continuously optimize the model's prediction accuracy. After training, the initial accurate performance prediction model is obtained.

[0075] Furthermore, based on the validation set, the performance of the initial accurate performance prediction model is evaluated and the parameters are optimized to obtain the accurate performance prediction model.

[0076] Specifically, the performance of the initial accurate performance prediction model is evaluated based on the validation set in the training input data, with a focus on monitoring the trend of the mean squared error (MSE) of the validation set. If the MSE of the validation set does not decrease for 10 consecutive rounds, it indicates that the model has reached its optimal training state. At this point, an early stopping strategy is adopted to stop training, avoiding overfitting due to overtraining, and the current optimal model parameters are saved. The saved optimal parameter model is then finally validated using a test set. When the coefficient of determination on the test set is ≥0.92, it indicates that the model has high-precision performance prediction capabilities and is identified as an accurate performance prediction model.

[0077] This application's embodiments, based on multiple initial polymer structures and known thickening polymer coding sequences, enable the structure generation model to fully learn polymer chemical syntax rules, ensuring the chemical rationality of the generated structures and addressing the problems of insufficient diversity and difficulty in guaranteeing chemical feasibility in traditional artificially designed structures. By combining high-performance polymer structural characteristics and structural optimization design rules, the model is fine-tuned, focusing the generated candidate structures on performance-oriented features, reducing the generation of invalid structures, improving the quality rate of candidate structures, and lowering subsequent screening costs. Specific operating parameters and functional group requirements are introduced as generation constraints, making the candidate structures adaptable to actual application scenarios, solving the problem of mismatch between traditional designed structures and operating conditions and poor practical application results. Furthermore, through decoding and synthesis feasibility screening, the industrialization potential of the structures is further guaranteed.

[0078] In some embodiments, a multi-objective reward function adapted to the accurate performance prediction model is obtained, and the candidate structures after initial screening are subjected to reinforcement learning iterative optimization based on the multi-objective reward function and the accurate performance prediction model to obtain high-potential candidate structures. This includes: obtaining the core performance indicators, synthesis feasibility and stability of carbon dioxide thickening polymers, performing weighted combination processing on the core performance indicators, synthesis feasibility and stability, and constructing a multi-objective reward function adapted to the accurate performance prediction model.

[0079] Specifically, the core performance indicators are viscosity increase factor and carbon dioxide solubility, and stability specifically refers to thermal stability. All three are key performance evaluation dimensions for carbon dioxide thickening polymers. Synthesis feasibility is a feasibility score based on existing polymerization process conditions (within the range of 0-1). The higher the score, the easier it is to achieve industrial synthesis.

[0080] The four evaluation dimensions mentioned above are weighted and combined to construct a composite multi-objective reward function, with the function expression R = 0.4 × R1 + 0.3 × R2 + 0.2 × R3 + 0.1 × R4. Here, R1 is the standardized value of the viscosity improvement factor, R2 is the standardized value of carbon dioxide solubility, R3 is the synthesis feasibility score, and R4 is the standardized value of thermal stability. All standardized values ​​are mapped to the [0,1] interval through Min-Max standardization to ensure the rationality of the weighted calculation. The final reward function value ranges from [0,1], with higher values ​​indicating better overall performance of the candidate structure.

[0081] Furthermore, based on structural optimization design rules, the candidate structures after initial screening are modified to conform to chemical rationality, resulting in modified candidate structures.

[0082] Specifically, candidate structures obtained from the initial screening using a basic performance prediction model are acquired. These candidate structures have passed preliminary screening for synthetic feasibility and possess basic chemical rationality and performance potential. A reinforcement learning agent is then used to modify the candidate structures. Modifications are based on the SMILES encoding sequence of the candidate structures and include functional group replacement, chain length adjustment, and topology optimization. All modifications adhere to the rules of polymer structural chemistry syntax to ensure that the modified structures possess thermodynamic stability and chemical rationality, resulting in modified candidate structures.

[0083] Furthermore, the modified candidate structures are input into the accurate performance prediction model for performance evaluation, and the core performance evaluation results corresponding to each modified candidate structure are obtained.

[0084] Specifically, the modified candidate structure is input into the accurate performance prediction model for performance evaluation. The accurate performance prediction model extracts the local and global molecular features of the modified candidate structure through graph convolution operations, accurately predicts its corresponding viscosity enhancement factor, carbon dioxide solubility, and thermal stability, and outputs the standardized prediction results (mapped to the [0,1] interval). These prediction results are the performance evaluation results.

[0085] Furthermore, the core performance evaluation results and the multi-objective reward function are quantitatively calculated to obtain the reward function values ​​of the corresponding candidate structures.

[0086] Specifically, based on the performance evaluation results output by the accurate performance prediction model, and combined with the synthesis feasibility score, the performance evaluation results and the multi-objective reward function are quantitatively calculated. The specific calculation process is as follows: the predicted results of R1 (standardized value of viscosity improvement factor), R2 (standardized value of carbon dioxide solubility), and R4 (standardized value of thermal stability), along with the pre-determined R3 (synthesis feasibility score), are substituted into the multi-objective reward function formula. The reward function value corresponding to each modified candidate structure is obtained by weighted summation. The reward function value directly reflects the comprehensive performance of the candidate structure.

[0087] Furthermore, the reinforcement learning optimization model is updated with parameters based on the reward function value to obtain the processed reinforcement learning optimization model.

[0088] Specifically, a reinforcement learning optimization model is obtained. The core architecture of the reinforcement learning optimization model is an agent-environment-reward interaction mechanism. The calculated reward function value is used as the core basis for updating the model parameters. The agent parameters of the reinforcement learning optimization model are updated through the policy gradient algorithm, so that the agent gradually learns the structural modification strategies that can obtain high reward function values. That is, it identifies which functional group replacements, chain length adjustments, or topology optimization methods are more likely to improve the overall performance, resulting in the processed reinforcement learning optimization model.

[0089] Furthermore, the processed reinforcement learning optimization model is used to perform a preset number of iterations on the candidate structures after the initial screening. After the iteration is completed, candidate structures whose reward function values ​​meet the preset requirements are selected to obtain high-potential candidate structures.

[0090] Specifically, the reinforcement learning optimization model is iteratively optimized based on the reward function value. The total number of iterations is set to 10,000 steps, and the current optimal model parameters are saved every 1,000 steps to avoid overfitting or performance degradation due to excessive iteration. During the iteration process, the agent continuously modifies, evaluates, and updates the parameters of candidate structures based on the learned optimization strategy, constantly improving the overall performance of the generated structures. After iteration, the 10 candidate structures with the highest reward function values ​​are selected. These 10 structures demonstrate the best performance in terms of thickening effect, solubility, stability, and synthetic feasibility, and are thus considered high-potential candidate structures.

[0091] This application's embodiments utilize a multi-objective reward function to achieve synergistic optimization of multiple design objectives through a weighted combination of core performance indicators, synthetic feasibility, and stability. This addresses the problem of existing technologies' inability to systematically balance multiple objective requirements, thereby improving the overall performance and application value of candidate structures. The reinforcement learning optimization model, through an iterative mechanism of structure modification-performance evaluation-parameter update, gradually learns the structural optimization strategies corresponding to high rewards, effectively exploring the chemical molecular space, increasing the probability of discovering novel, high-performance polymer structures, and avoiding getting trapped in local optima. The reward function value is calculated based on the evaluation results of the accurate performance prediction model, ensuring the accuracy of the optimization direction. Simultaneously, multiple iterations and phased model saving ensure the stability and traceability of the optimization process, improving the reliability of research results.

[0092] In some embodiments, the accurate performance prediction model is analyzed and processed to obtain structural optimization design rules, including: obtaining an interpretability analysis tool, associating the accurate performance prediction model with the interpretability analysis tool to obtain a correspondence, and the correspondence is used to characterize the correspondence between the model prediction logic and the contribution analysis of polymer structural features related to carbon dioxide thickening performance.

[0093] Specifically, the interpretability analysis tool is first acquired. This tool employs the SHAP (Shapley Additive ex Planations) algorithm, adapted to the graph structure characteristics of the accurate performance prediction model, and uses the TreeExplainer interpreter as the specific analysis component. The accurate performance prediction model is then correlated with this interpretability analysis tool. Since the SHAP algorithm supports interpretability analysis of mainstream deep learning models, the prediction logic and structural feature contribution information of the graph neural network algorithm model can be directly extracted. By establishing the correlation between the model's prediction results and the structural feature contribution analysis, this correlation clearly characterizes the influence path and logic of each polymer structural feature on the prediction results when the model outputs performance prediction values.

[0094] Furthermore, the core performance indicators in the training input data are correlated with the accurate performance prediction model to obtain contribution correlation results.

[0095] Specifically, the core performance indicators in the training input data of the accurate performance prediction model are correlated. These core performance indicators are defined as three key metrics standardized during the preprocessing stage of the training input data: viscosity enhancement factor, carbon dioxide solubility, and thermal stability. These three metrics directly reflect the core application performance of carbon dioxide thickening polymers. During the correlation process, based on the established model-tool correspondence, the polymer structural features learned by the accurate performance prediction model (such as functional group type, chain length, topology, and number of functional groups) are extracted. Each structural feature is then matched with one of the three core performance indicators to clarify the potential correlation between different structural features and each performance indicator. Finally, the contribution correlation results between each structural feature and the core performance indicators are obtained, i.e., which structural features may initially affect one or more performance indicators.

[0096] Furthermore, the contribution correlation results are quantified to obtain the characteristic contribution results.

[0097] Specifically, the contribution correlation results are quantified. The SHAP algorithm interpreter calculates the feature contribution of each structural feature to each core performance indicator, which is represented by the SHAP value. The core logic of the quantification calculation is to decompose the model prediction result into the sum of the independent contributions of each structural feature. A positive SHAP value indicates that the corresponding structural feature has a positive effect on the core performance indicator, while a negative SHAP value indicates that the corresponding structural feature has an inhibitory effect on the core performance indicator. The absolute value of the SHAP value directly reflects the degree of contribution of the structural feature to the performance indicator; the larger the absolute value, the stronger the contribution.

[0098] Furthermore, the feature contribution results are analyzed and processed to obtain the ranking results of structural feature contribution and the correlation relationship. The correlation relationship is used to characterize the relationship between the ranking results of structural feature contribution and the core performance indicators.

[0099] Specifically, the quantified feature contribution results are analyzed and processed. On the one hand, the structural features are sorted according to the absolute value of their SHAP values ​​to obtain the ranking results of structural feature contribution, thus identifying the key structural features that have the most significant impact on core performance indicators. On the other hand, the correlation pattern between a single structural feature and various core performance indicators is analyzed through the dependency graph of the SHAP algorithm to obtain the correlation between the ranking results of structural feature contribution and the core performance indicators. The correlation is specifically a non-linear correlation, such as the positive non-linear correlation between the number of aromatic rings and the viscosity increase factor, that is, when the number of aromatic rings increases within a certain range, the viscosity increase factor shows a non-linear growth trend.

[0100] Furthermore, rule extraction is performed on the feature contribution results, structural feature contribution ranking results, and correlation relationships to obtain structural optimization design rules.

[0101] Specifically, a comprehensive rule extraction process is performed on the feature contribution results, the ranking results of structural feature contributions, and the nonlinear correlation relationships. Combining the positive and negative values ​​of the SHAP values ​​of various structural features, their contribution magnitude, and the nonlinear correlation patterns with performance indicators, rules are extracted to determine the impact of different structural features on carbon dioxide thickening performance. For example, based on the positive SHAP values ​​and high contribution ranking results corresponding to ether groups, a rule is extracted that introducing ether groups can increase the carbon dioxide solubility and viscosity of the polymer; based on the negative SHAP values ​​corresponding to easily broken chemical bonds, a rule is extracted that avoiding easily broken structures such as SS bonds can improve the thermal stability of the polymer; and combining the nonlinear correlation between chain length and performance indicators, a rule is extracted that the thickening performance is optimal within a specific chain length range, and performance decreases beyond this range. The rules obtained from the above comprehensive extraction are the structural optimization design rules.

[0102] This application's embodiments establish a correspondence between model prediction logic and structural feature contributions using interpretability analysis tools, overcoming the black-box problem of traditional machine learning models. This visualizes and quantifies the structure-performance relationship, providing a clear basis for structural optimization. By quantifying and ranking feature contributions, key structural features (such as silicon-containing groups and easily broken chemical bonds) that have a positive / negative impact on thickening performance are accurately identified, avoiding the blindness of traditional design and improving the targeting and efficiency of structural optimization. The extracted structural optimization design rules can directly guide the targeted generation and modification of candidate structures, achieving rule-driven intelligent design, reducing invalid attempts, and further shortening the R&D cycle and reducing R&D costs.

[0103] In some embodiments, high-potential candidate structures are simulated and verified to obtain high-performance polymer structures, including: obtaining adaptation parameters and environmental conditions for molecular dynamics simulation, wherein the adaptation parameters and environmental conditions are fixed simulation parameters and operating conditions preset based on the actual application scenario of carbon dioxide thickening polymers, which are used to simulate and verify the performance of high-potential candidate structures under a unified benchmark.

[0104] Specifically, the required adaptation parameters and environmental conditions for molecular dynamics simulation are obtained. The adaptation parameters are the core calculation parameters for the simulation, including: force field parameters, a simulation step size of 1 fs, a total simulation duration of 10 ns, and a sampling interval of 10 ps. The environmental conditions are determined based on the actual application scenario of the carbon dioxide thickening polymer, specifically supercritical carbon dioxide conditions, including a temperature of 377 K, a pressure of 55 MPa, a supercritical carbon dioxide phase, and a polymer mass fraction of 0.1%–2.0 wt% in the simulation system.

[0105] Furthermore, high-potential candidate structures, adaptation parameters, and environmental conditions are simulated to obtain simulation results, including structural conformation, interaction characteristics, and core performance data.

[0106] Specifically, molecular dynamics simulations are employed to synergistically simulate high-potential candidate structures, adaptation parameters, and environmental conditions. During the simulation, a simulation system comprising high-potential candidate polymer molecules and supercritical carbon dioxide molecules is first constructed. The initial configuration of the system is optimized based on predetermined force field parameters to eliminate unreasonable intermolecular overlap. Subsequently, under specified temperature and pressure conditions, the system is equilibrated sequentially through canonical and isothermal-isobaric ensembles, followed by molecular dynamics trajectory simulations. The simulation results are output, including structural conformation, interaction characteristics, and core performance data.

[0107] Furthermore, the core performance data is extracted and analyzed to obtain the core performance verification results.

[0108] Specifically, the core performance data from the simulation results were extracted and analyzed in a targeted manner. The core performance data focused on three key indicators: viscosity enhancement factor, carbon dioxide solubility, and thermal stability. During extraction, invalid data from the initial equilibrium stage of the simulation were removed, retaining only the valid trajectory data after the system stabilized. In the analysis, statistical mechanics methods were used to transform the simulation data into actual performance index values. For example, the viscosity enhancement factor was calculated based on the correlation formula between the diffusion coefficient and viscosity; carbon dioxide solubility was determined based on the polymer's equilibrium concentration in carbon dioxide; and thermal stability was assessed based on the integrity of the polymer structure and bond breakage during the simulation period. Finally, the analysis results of the three indicators were integrated to obtain the core performance verification results.

[0109] Furthermore, a preset performance verification standard is obtained, and the core performance verification results are compared with the preset performance verification standard to obtain a candidate structure that meets the standard.

[0110] Specifically, preset performance verification standards are obtained. These standards are based on the industrial application requirements and high-performance definition of carbon dioxide thickening polymers, and are specifically: viscosity increase factor ≥ 5.0, carbon dioxide solubility reaching a preset dissolution threshold, and thermal stability meeting the requirement of ≥ 95% structural retention rate during simulation at 377K. The core performance verification results are compared with these preset performance verification standards one by one. Candidate structures that meet all three indicators are selected as compliant candidate structures. If any indicator fails to meet the standard, the candidate structure is directly eliminated and does not proceed to the next step.

[0111] Furthermore, the stability and feasibility of the candidate structures that meet the standards are verified to obtain high-performance polymer structures.

[0112] Specifically, candidate structures that meet the standards undergo stability and feasibility verification. Stability verification involves extending the simulation time to 20 ns to monitor structural changes over a longer timescale, ensuring stable conformation and performance under continuous operating conditions in practical applications, without significant degradation or performance decline. Feasibility verification, based on existing polymerization process conditions and combined with Monte Carlo simulation analysis, assesses the synthetic feasibility of candidate structures, specifically evaluating functional group reactivity, chain segment connection efficiency, and the difficulty of branching structure formation, ensuring that candidate structures can be industrialized through conventional polymerization reactions, while excluding structures that are thermodynamically unstable or have excessively high polymerization barriers. Candidate structures that pass both stability and feasibility verifications are considered high-performance polymer structures.

[0113] This application's embodiments precisely verify the core performance of high-potential candidate structures through molecular dynamics simulations, ensuring the stability and effectiveness of the structures under target operating conditions and solving the problem of the disconnect between traditional design and practical application. By extracting core performance data and comparing it with preset standards, combined with stability and feasibility verification, high-performance polymer structures that meet industrial application requirements are screened layer by layer, filtering out thermodynamically unstable and synthetically infeasible invalid structures, thus ensuring the reliability of the output structures.

[0114] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0115] like Figure 3 As shown in the embodiments of this application, a carbon dioxide thickening polymer structure device 300 is also provided. The device includes: The acquisition module 301 is used to acquire training input data, preprocess the training input data, and obtain processed training input data. The training input data is generated based on high-throughput molecular dynamics simulation.

[0116] Module 302 is used to build a basic performance initial model and a precise performance initial model. The training input data is input into the basic performance initial model and the precise performance initial model respectively for model training and optimization, so as to obtain a basic performance prediction model and a precise performance prediction model.

[0117] The processing module 303 is used to analyze and process the accurate performance prediction model to obtain structural optimization design rules.

[0118] The acquisition module 301 is also used to acquire polymer structure chemical syntax rules and high-performance polymer structure features, and to learn and generate polymer structure chemical syntax rules, high-performance polymer structure features and structure optimization design rules to obtain candidate polymer structures.

[0119] The processing module 303 is also used to input the candidate polymer structure into the basic performance prediction model for rapid initial screening to obtain the initial screening candidate structure.

[0120] The acquisition module 301 is also used to acquire a multi-objective reward function that is adapted to the accurate performance prediction model, and to perform reinforcement learning iterative optimization on the candidate structures after the initial screening based on the multi-objective reward function and the accurate performance prediction model to obtain high-potential candidate structures.

[0121] The processing module 303 is also used to perform simulation verification of high-potential candidate structures to obtain high-performance polymer structures.

[0122] The processing module 303 is also used to add the structure-performance data corresponding to the high-performance polymer structure to the training input data, and input the training input data into the basic performance prediction model and the accurate performance prediction model respectively for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

[0123] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0124] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0125] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, such as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of a memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the device used to implement various functions can be viewed as either a software module that implements the method or a structure within a hardware component.

[0126] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0127] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0128] Furthermore, in the various embodiments of the present invention, each functional module can be integrated into a processing module, or each module can exist independently, or two or more modules can be integrated into a single module.

[0129] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0130] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0131] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0132] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A method for designing a carbon dioxide thickening polymer structure, characterized in that, include: Acquire training input data, preprocess the training input data to obtain processed training input data, the training input data being generated based on high-throughput molecular dynamics simulation; Construct a basic performance initial model and a precise performance initial model, and input the training input data into the basic performance initial model and the precise performance initial model respectively for model training and optimization to obtain a basic performance prediction model and a precise performance prediction model. The accurate performance prediction model is analyzed and processed to obtain structural optimization design rules; Acquire polymer structural chemical syntax rules and high-performance polymer structural features, and learn and generate the polymer structural chemical syntax rules, high-performance polymer structural features and structural optimization design rules to obtain candidate polymer structures; The candidate polymer structures are input into the basic performance prediction model for rapid initial screening to obtain the initial screening candidate structures. Obtain a multi-objective reward function that is compatible with the accurate performance prediction model, and perform reinforcement learning iterative optimization on the initial screening candidate structure based on the multi-objective reward function and the accurate performance prediction model to obtain a high-potential candidate structure. The high-potential candidate structures were simulated and verified to obtain high-performance polymer structures; The structure-performance data corresponding to the high-performance polymer structure is added to the training input data. The training input data is then input into the basic performance prediction model and the accurate performance prediction model for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

2. The method according to claim 1, characterized in that, The training input data includes a training set, a validation set, and a structured feature matrix. The structured feature matrix is ​​a fixed set of structured features constructed by integrating polymer structure data generated based on high-throughput molecular dynamics simulation in the preprocessing stage after structure encoding and feature extraction. It is used to characterize polymer structure information related to carbon dioxide thickening performance. The construction of the basic performance initial model and the accurate performance initial model involves inputting the training input data into the basic performance initial model and the accurate performance initial model, respectively, for model training and optimization to obtain the basic performance prediction model and the accurate performance prediction model, including: Construct an initial basic performance model, which includes decision tree configuration parameters and feature splitting criteria; Using the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, the basic performance initial model is trained based on the decision tree configuration parameters and the feature splitting criteria to obtain the basic performance initial prediction model. Based on the validation set, the basic performance prediction model is evaluated and hyperparameters are tuned to obtain the basic performance prediction model. A precise performance initial model based on graph neural networks is constructed, which includes a network architecture adapted for polymer molecular structure feature extraction and initial parameters for model training. Using the structured feature subset corresponding to the training set as input and the core performance index subset corresponding to the training set as output, the accurate performance initial model is trained to obtain the accurate performance initial prediction model. Based on the validation set, the performance of the initial accurate performance prediction model is evaluated and the parameters are optimized to obtain the accurate performance prediction model.

3. The method according to claim 1, characterized in that, The process involves acquiring polymer structural chemical syntax rules and high-performance polymer structural features, and then learning and generating these rules, along with structural optimization design rules, to obtain candidate polymer structures, including: Multiple sets of initial polymer structures, coding sequences, and high-performance polymer structural features are obtained. The polymer structure chemical syntax rules are composed of the chemical bonding rules of multiple sets of initial polymer structures and coding sequences. The high-performance polymer structural features are polymer structural features that meet the preset thickening performance standards. A structure generation model is obtained by using the encoding sequences corresponding to the multiple initial polymer structures as pre-training corpus to pre-train the structure generation model, thereby obtaining the processed structure generation model. Using the structural features of the high-performance polymer and the structural optimization design rules as fine-tuning data, the processed structure generation model is fine-tuned to obtain the target structure generation model. The target structure generation model is used to learn the chemical syntax rules of polymer structure and characterize structural features that have a positive contribution to carbon dioxide thickening performance. Specific operating condition parameters and functional group requirements are obtained, and the specific operating condition parameters and functional group requirements are input into the target structure generation model for sequence generation processing to obtain multiple sets of candidate polymer coding sequences. The candidate polymer coding sequences are decoded to obtain the candidate polymer structures.

4. The method according to claim 1, characterized in that, The process involves obtaining a multi-objective reward function adapted to the accurate performance prediction model, and then performing reinforcement learning iterative optimization on the initially screened candidate structures based on the multi-objective reward function and the accurate performance prediction model to obtain high-potential candidate structures, including: The core performance indicators, synthesis feasibility, and stability of the carbon dioxide thickening polymer are obtained. These indicators are then weighted and combined to construct a multi-objective reward function adapted to the precise performance prediction model. The expression of the multi-objective reward function is: R = 0.4 × R1 + 0.3 × R2 + 0.2 × R3 + 0.1 × R4, where R1 is the viscosity enhancement factor among the core performance indicators, R2 is the carbon dioxide solubility among the core performance indicators, R3 is the synthesis feasibility, and R4 is the stability. Based on the aforementioned structural optimization design rules, the candidate structures after initial screening are subjected to structural modification processing that conforms to chemical rationality, resulting in modified candidate structures. The modified candidate structure is input into the precise performance prediction model for performance evaluation, and the core performance evaluation results corresponding to each modified candidate structure are obtained. The core performance evaluation results and the multi-objective reward function are quantitatively calculated to obtain the reward function value of the corresponding candidate structure; Based on the reward function value, the parameters of the reinforcement learning optimization model are updated to obtain the processed reinforcement learning optimization model. The processed reinforcement learning optimization model performs a preset number of iterations on the initially screened candidate structures. After the iterations are completed, candidate structures whose reward function values ​​meet the preset requirements are selected to obtain high-potential candidate structures.

5. The method according to claim 1, characterized in that, The analysis and processing of the accurate performance prediction model yields structural optimization design rules, including: An interpretability analysis tool is obtained, and the accurate performance prediction model is associated with the interpretability analysis tool to obtain a correspondence. The correspondence is used to characterize the correspondence between the model prediction logic and the contribution analysis of polymer structural features related to carbon dioxide thickening performance. The accurate performance prediction model is correlated with the core performance indicators in the training input data to obtain the contribution correlation results. The contribution correlation results are quantified to obtain the feature contribution results; The feature contribution results are analyzed and processed to obtain the structural feature contribution ranking results and correlation relationships. The correlation relationships are used to characterize the relationship between the structural feature contribution ranking results and the core performance indicators. The feature contribution results, structural feature contribution ranking results, and correlation relationships are processed by rule extraction to obtain structural optimization design rules.

6. The method according to claim 1, characterized in that, The simulation verification process for the high-potential candidate structures to obtain high-performance polymer structures includes: The appropriate parameters and environmental conditions for molecular dynamics simulation are obtained. These parameters and environmental conditions are fixed simulation parameters and operating conditions preset based on the actual application scenario of carbon dioxide thickening polymers. They are used to simulate and verify the performance of high-potential candidate structures under a unified benchmark. The high-potential candidate structure, the adaptation parameters, and the environmental conditions are simulated to obtain simulation results, which include structural conformation, interaction characteristics, and core performance data. The core performance data is extracted and analyzed to obtain the core performance verification results; Obtain a preset performance verification standard, compare the core performance verification results with the preset performance verification standard, and obtain candidate structures that meet the standard; The candidate structures that meet the standards are subjected to stability and feasibility verification to obtain high-performance polymer structures.

7. A carbon dioxide thickening polymer structure device, characterized in that, include: The acquisition module is used to acquire training input data, preprocess the training input data, and obtain processed training input data, wherein the training input data is generated based on high-throughput molecular dynamics simulation. A construction module is used to construct a basic performance initial model and a precise performance initial model. The training input data is input into the basic performance initial model and the precise performance initial model respectively for model training and optimization to obtain a basic performance prediction model and a precise performance prediction model. The processing module is used to analyze and process the accurate performance prediction model to obtain structural optimization design rules; The acquisition module is also used to acquire polymer structure chemical syntax rules and high-performance polymer structure features, and to learn and generate the polymer structure chemical syntax rules, high-performance polymer structure features and structure optimization design rules to obtain candidate polymer structures. The processing module is also used to input the candidate polymer structure into the basic performance prediction model for rapid initial screening to obtain the initial screening candidate structure. The acquisition module is further configured to acquire a multi-objective reward function adapted to the accurate performance prediction model, and perform reinforcement learning iterative optimization on the initial screening candidate structure based on the multi-objective reward function and the accurate performance prediction model to obtain a high-potential candidate structure. The processing module is also used to perform simulation verification on the high-potential candidate structure to obtain a high-performance polymer structure. The processing module is further configured to add the structure-performance data corresponding to the high-performance polymer structure to the training input data, and input the training input data into the basic performance prediction model and the accurate performance prediction model respectively for iterative updates and structure generation processing to obtain the target carbon dioxide thickening polymer structure.

8. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images