An artificial intelligence-based microemulsion oil displacement agent formula generation and optimization method

Abstract

The present application relates to a kind of microemulsion oil displacement agent formula generation and optimization method based on artificial intelligence, method includes: constructing the formula performance database comprising input feature and output feature;Machine learning prediction model is constructed and trained;According to target performance parameter, candidate oil displacement agent formula is generated using generative artificial intelligence model, and performance evaluation and screening are carried out using trained prediction model;The digital twin of microemulsion oil displacement process is built, and the candidate formula screened is injected into digital twin and carries out virtual oil displacement experiment, and according to the verification result, the prediction model and / or generative model are optimized;Output optimal formula.The present application combines generative model, prediction model and digital twin, forms the closed-loop system of generation, prediction, screening, verification, optimization, greatly shortens the development cycle of microemulsion oil displacement agent, and reduces the development cost.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development technology, and in particular to a method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence. Background Technology

[0002] Microemulsion flooding is a highly efficient tertiary oil recovery technology. Its core mechanism lies in injecting a microemulsion system composed of water, oil, surfactants, and co-surfactants to create an ultra-low interfacial tension (up to 10) at the oil-water interface. - (on the order of 3 mN / m), thereby effectively initiating and displacing residual oil in the micropores of the reservoir, significantly improving crude oil recovery.

[0003] However, traditional microemulsion displacement agent design methods heavily rely on researchers' experience and extensive trial-and-error experiments. To find the optimal formulation that forms a stable intermediate-phase or single-phase microemulsion under specific reservoir conditions (such as temperature, salinity, and crude oil properties), hundreds or even thousands of experimental screenings are typically required. This approach suffers from long development cycles, high costs, and low formulation screening efficiency, making it difficult to quickly respond to the complex, variable, and demanding reservoir development needs, especially for the exploitation of unconventional oil and gas resources such as low-permeability reservoirs and heavy oil reservoirs.

[0004] In recent years, artificial intelligence technology has made significant progress in the application of chemical formulation design. A patent for an automated chemical formulation device and method (publication number US20220397886A1) discloses a technical solution for formulation prediction based on a machine learning prediction model. This model generates a prediction model based on target properties by training a first machine learning model, and then generates a new input dataset within boundary conditions for prediction. Furthermore, a patent for a method for predicting compound properties based on a multi-task deep learning model (publication number CN114566230A) discloses a technical solution for predicting compound properties using a multi-task deep learning model. A patent for a method for training a compound property prediction model (application number CN202111482319.9) discloses a training framework for a compound property prediction model based on multi-stage pre-training and knowledge transfer.

[0005] However, the existing technologies mentioned above still have the following shortcomings: (1) Most existing AI formulation design methods are open-loop systems, lacking efficient closed-loop feedback optimization mechanisms, making it difficult to achieve automatic iterative optimization of formulations; (2) Existing methods lack a technical solution that systematically integrates generative exploration, predictive evaluation and virtual verification, and formulation verification still relies on physical experiments, which limits the further improvement of R&D efficiency; (3) For the specific complex system of microemulsion oil displacement agents, existing methods have not solved technical problems such as synergistic prediction of multiple performance indicators and in-depth characterization of surfactant molecular structure. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an artificial intelligence-based method for generating and optimizing microemulsion oil displacement agent formulations, which solves the problems of long development cycle, high cost and low efficiency of existing microemulsion oil displacement agent design methods, as well as the lack of closed-loop feedback optimization and the reliance on physical experiments for verification in existing AI formulation design methods.

[0007] The technical solution adopted by this invention to solve its technical problem is: The beneficial effect of this invention is that it solves the defects existing in the prior art. I. This invention combines generative artificial intelligence models with machine learning predictive models and introduces digital twins for virtual verification, forming a closed-loop system of generation, prediction, screening, verification, and optimization. The generative model is responsible for exploring the formula space, the predictive model for rapid evaluation, and the digital twin for high-fidelity virtual verification. These three work collaboratively, significantly improving the efficiency of formula optimization. Compared to traditional methods that rely on extensive trial and error, this invention shortens the development cycle from months or even years to days or weeks, reducing development costs by approximately 90%.

[0008] II. This invention employs a multi-task learning neural network model based on an attention mechanism, which can simultaneously predict multiple key performance indicators such as interfacial tension, phase state, and adsorption capacity. Furthermore, it utilizes the attention mechanism to automatically identify the key features that have the greatest impact on different tasks, significantly improving the prediction accuracy for the complex performance of microemulsion systems. Experimental verification shows that the model achieves an R² of over 0.94 on the interfacial tension prediction task and a phase state classification accuracy of over 92%.

[0009] Third, this invention introduces a surfactant molecule encoder based on graph neural networks, which can directly learn molecular topological features from molecular structure diagrams without relying on manually designed molecular descriptors. This achieves end-to-end modeling from molecular structure to macroscopic properties, and has stronger generalization ability for the development of novel surfactants. Experimental verification shows that the prediction accuracy (R²) of this GNN encoder for novel surfactants not found in the training set can reach over 0.85, significantly better than models based on traditional descriptors (R² approximately 0.72).

[0010] IV. This invention constructs a digital twin of the microemulsion flooding process integrating a multiphase flow simulator and a machine learning prediction model. This enables efficient formulation performance verification in a virtual environment, avoiding extensive physical core displacement experiments and significantly reducing R&D and time costs. The digital twin adopts a master-slave collaborative mode, reducing the overall verification time to less than 20% of the traditional full simulation scheme while ensuring verification accuracy. Furthermore, the relative error between the virtual verification results and the measured values ​​from physical experiments can be controlled within 2%.

[0011] Fifth, this invention covers the entire process of oil displacement agent research and development, from data construction, model training, formula generation, virtual verification to injection process optimization. It can design the most suitable formula and injection scheme according to different reservoir conditions, providing key technical support for the intelligent transformation of oil extraction. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0013] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0014] like Figure 1 The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence is shown. By systematically integrating generative models, multi-task prediction models, digital twins, and incremental learning, a closed loop of generation, prediction, screening, verification, and optimization is constructed. The synergistic effect of each link results in an overall effect that is significantly better than the simple superposition of individual modules. The specific implementation method is as follows: Step S1: Data Construction Historical formulation data and corresponding performance test data of microemulsion oil displacement agents were collected and integrated to construct a formulation performance database that includes input and output characteristics.

[0015] Specifically, the input features include: crude oil composition (such as the content of saturated hydrocarbons, aromatic hydrocarbons, resins, and asphaltene), formation water salinity, and ionic composition (such as Na). + Ca² + Mg² + Cl - Plasma concentration), reservoir temperature, surfactant molecular structure descriptors (including molecular weight, hydrophilic-lipophilic balance value (HLB), critical micelle concentration, etc.), co-surfactant type and concentration, salinity, etc.

[0016] Output characteristics include: interfacial tension (oil-water interfacial tension value), optimal salinity (salinity when the microemulsion forms the middle phase), solubilization parameters (amount of oil or water solubilized by a unit concentration of surfactant), phase type (Winsor type I, II, III), adsorption amount (adsorption loss of surfactant on the rock surface), and core displacement efficiency (enhanced oil recovery value from oil displacement experiments).

[0017] Historical data is derived from laboratory tests, literature data, or field test data in oilfields. After data collection, data cleaning (removing outliers and handling missing values), feature standardization, and normalization are performed to construct a structured formulation performance database. To ensure data quality, multiple imputation is used for missing values, and outliers are detected and corrected based on local outlier factors. The initial database size is recommended to be no less than 1000 valid samples, covering different reservoir types (medium-high permeability, low permeability, and heavy oil) and different surfactant systems.

[0018] Step S2: Model Building and Training

[0019] A machine learning prediction model is constructed and trained, and the model is trained using the database to establish a mapping relationship from the input features to the output features.

[0020] The machine learning prediction model is a multi-task learning neural network model based on an attention mechanism. The model includes: an input layer that receives the aforementioned input feature vectors; a shared feature extraction layer, composed of multiple fully connected layers and activation functions, used to extract shared representations of the input features; and multiple task-specific output layers, each corresponding to a different prediction task (such as interfacial tension prediction, phase classification, adsorption prediction, etc.), with each task-specific output layer containing an independent fully connected layer and output unit. The attention mechanism is used to dynamically identify the key feature combinations in the input features that contribute the most to different prediction tasks, thereby improving the model's prediction accuracy and generalization ability in complex microemulsion systems.

[0021] The training process employs a multi-task joint loss function: ; Among them, L i Let w be the loss function for the i-th prediction task (mean squared error for regression tasks, cross-entropy for classification tasks). i Task weights are assigned; model parameters are optimized using the backpropagation algorithm; early stopping and learning rate decay strategies are employed during training to prevent overfitting.

[0022] To enhance the characterization ability of surfactant molecular structures, the multi-task learning neural network model also includes a surfactant molecular encoder (GNN encoder) based on a graph neural network. This encoder maps the molecular structure diagram of surfactant molecules (with atoms as nodes, chemical bonds as edges, and including attributes such as atom type, bond order, and charge) into a fixed-dimensional molecular structure feature vector through a multi-layer graph convolutional network, and inputs it to the shared feature extraction layer. In this way, the model can automatically learn the deep correlation between the topological structure of surfactant molecules and their macroscopic properties, avoiding the limitations of traditional methods that rely on manual descriptors. Experimental results show that the GNN encoder achieves a prediction accuracy (R²) of over 0.85 for novel surfactants not found in the training set (such as derivatives with different alkyl chain lengths and different hydrophilic head groups), significantly outperforming models based on traditional descriptors (R² approximately 0.72), demonstrating its good generalization ability and supporting the rapid screening of novel surfactants.

[0023] Step S3: Recipe Generation

[0024] Based on the target performance parameters, a generative artificial intelligence model is used to generate candidate oil displacement agent formulations, and the trained machine learning prediction model is used to evaluate and screen the generated candidate formulations.

[0025] In this embodiment, the generative artificial intelligence model employs a conditional variational autoencoder (CVAE). Target performance parameters (such as target interfacial tension ≤ 0.01 mN / m, target oil displacement efficiency ≥ 80%) are input into the generative artificial intelligence model as conditional variables to constrain the performance range of the generated formulation, ensuring that the generated formulation remains within the target performance range. The latent space dimension of the CVAE is set to 16, and the generative network adopts a three-layer fully connected structure, with batch normalization and Dropout layers added to enhance the diversity of generation. The CVAE decoder output is a normalized continuous value vector of each component concentration, which, after inverse normalization, yields the actual formulation composition.

[0026] For each generated candidate formulation (500–1000 formulations generated at a time), the machine learning prediction model trained in step S2 is used to quickly predict its performance indicators such as interfacial tension, adsorption capacity, and oil displacement efficiency. Based on the prediction results, the top K candidate formulations (K=20–50) that meet the target performance parameters or have high overall scores are retained for subsequent validation. This collaborative mechanism of generation, prediction, and selection can efficiently locate the optimal solution region within a vast formulation space.

[0027] Step S3 also includes interpretability analysis: Based on the trained machine learning prediction model, feature importance analysis methods (such as SHAP, LIME, etc., this embodiment uses SHAP values) are used to output the ranking of key features that contribute most to the performance of candidate formulations (for example, when predicting interfacial tension, salinity and surfactant HLB values ​​often rank in the top two). This ranking of key features is used as feedback information and passed to the constraints of the generative model; specifically, the top 5 key features and their expected direction of change (e.g., reducing salinity is beneficial for reducing interfacial tension) are transformed into additional weighting factors in the CVAE condition input, allowing the generator to prioritize adjusting these high-impact features in subsequent iterations, thereby accelerating convergence. This step improves the transparency of the formulation decision-making process, making it easier for researchers to understand the model's judgment criteria and perform targeted optimization accordingly.

[0028] Meanwhile, this step also implements feature linkage constraints: the key feature combination identified by the attention mechanism in step S2 (for example, for the interfacial tension prediction task, the feature combination with the highest attention weight is salinity-surfactant HLB value-co-surfactant concentration) is directly used as part of the CVAE condition constraints in step S3. That is, the joint distribution constraint of this combination feature is added to the generation conditions to ensure that the generated formulation meets the physicochemical compatibility in the key feature combination region, avoids generating unreasonable extreme combinations, and thus improves the actual feasibility of the generated formulation.

[0029] Step S4: Formulation Optimization and Validation

[0030] A digital twin of the microemulsion oil displacement process is constructed, and the screened candidate formulations are injected into the digital twin to conduct virtual oil displacement experiments. The virtual experimental results are obtained, the candidate oil displacement agent formulations are verified based on the virtual experimental results, and the machine learning prediction model and / or the generative artificial intelligence model are optimized based on the verification results.

[0031] This step specifically includes: (1) A digital twin of the microemulsion oil displacement process is constructed, which is used to simulate the transport, adsorption, and oil displacement efficiency of candidate oil displacement agent formulations in porous media. This digital twin integrates a multiphase flow simulator and the machine learning prediction model from step S2. The specific construction method is as follows: Multiphase Flow Simulator: A core-scale multiphase, multi-component simulator based on the finite volume method is used to solve the mass conservation equations and interphase mass transfer equations for the three phases including oil, water, and microemulsions. Spatial discretization employs an orthogonal grid, and time progression uses a fully implicit scheme. The simulator can calculate the pressure field, saturation field, component concentration field, and interfacial tension field, and output the recovery curve. To balance computational efficiency and accuracy, the grid number is set to 50×50×10, and the time step is adaptively adjusted.

[0032] Machine Learning Surrogate Model: Using simulation data generated by a multiphase flow simulator in a representative parameter space (covering different formulations and reservoir conditions) as the training set, a lightweight neural network (3 hidden layers, 64 neurons per layer) is trained as a surrogate model to quickly predict key indicators of oil displacement efficiency (such as ultimate recovery and breakthrough time). The surrogate model's inputs are formulation parameters and reservoir parameters, and its output is the predicted recovery value. After training, the prediction error of the surrogate model (compared to the high-fidelity simulator) is controlled within 3%, while the computation speed is improved by approximately 1000 times.

[0033] Data Interaction and Fusion Strategy: The digital twin adopts a master-slave collaborative model. The multiphase flow simulator serves as the master module, responsible for accurately simulating physical transport processes (such as phase changes, adsorption and diffusion), with each complete simulation taking several hours. The machine learning proxy model serves as the slave module, providing rapid predictions at the beginning of the simulator's operation to help determine whether to continue running the high-fidelity simulation. Specifically, for each candidate formulation, the proxy model first provides a preliminary recovery rate prediction within 1 second. If the predicted value is lower than a preset baseline (e.g., below 50%), the formulation is directly eliminated, and the multiphase flow simulator is no longer invoked. If the predicted value is higher than the baseline, the multiphase flow simulator is invoked for accurate verification. This fusion strategy reduces the overall verification time to less than 20% of the traditional full simulation scheme while ensuring verification accuracy.

[0034] (2) The selected candidate formulations were injected into the digital twin to conduct virtual oil displacement experiments and obtain virtual experimental results, including recovery curves, pressure distribution, adsorption distribution, etc. Each formulation was run three times in the digital twin under the same initial conditions and the average value was taken to eliminate numerical fluctuations.

[0035] (3) Calculate the overall fitness score based on the virtual experiment results. The overall fitness score can be defined as the weighted sum of multiple performance indicators. Each indicator is normalized and then weighted and summed. ; Where EOR is the recovery rate (%), τ is the interfacial tension (mN / m), Ads is the adsorption capacity (mg / g), and α, β, and γ are weighting coefficients (in this embodiment, α=0.6, β=0.3, and γ=0.1, after normalization).

[0036] (4) Iterative and incremental learning: Triggering condition: If the overall fitness score is lower than the preset threshold (the threshold is set to 0.75 in this embodiment, and the full score is 1.0), the candidate formula and its virtual experiment results are added to the database of step S1 as incremental samples.

[0037] Incremental Update Algorithm: When the cumulative number of incremental samples reaches the preset batch size (20 in this embodiment), the incremental learning update of the machine learning prediction model in step S2 is triggered. The incremental update adopts the Elastic Weight Consolidation (EWC) algorithm, which applies a secondary penalty to important parameters of the old task while fine-tuning the model parameters to prevent catastrophic forgetting. After the update is completed, step S3 is re-executed using the updated machine learning prediction model to generate a new round of candidate recipes.

[0038] Convergence criteria: If the change in the highest overall fitness score in two consecutive iterations is less than 0.01, or the total number of iterations reaches the preset upper limit (e.g., 20 iterations), or a recipe that meets the fitness threshold (≥0.75) has been obtained, then the iteration is terminated; otherwise, return to step S3 to continue the loop.

[0039] Through the above closed-loop iteration, the present invention can quickly approximate the globally optimal formulation under a limited experimental / simulation budget.

[0040] Step S5: Output

[0041] Output the validated and optimized optimal formulation, including the type and concentration of surfactants, the type and concentration of co-surfactants, and the proportions of key components such as salinity; at the same time, output the predicted performance (interfacial tension, adsorption capacity, oil displacement efficiency) and digital twin validation results of the formulation.

[0042] Injection process parameter optimization steps

[0043] As a further optimization scheme, the method also includes an injection process parameter optimization step: the optimized displacement agent formulation and reservoir geological characteristic parameters (such as permeability, porosity, heterogeneity, etc.) are input into a reservoir numerical simulator (such as the commercial software ECLIPSE), and the particle swarm optimization algorithm (PSO) is used to simultaneously optimize the displacement agent injection concentration (0.1–0.5 PV), injection slug size (0.3–1.0 PV), and injection rate (0.5–5 m / d), with the goal of maximizing net present value (NPV), to generate an injection process scheme adapted to the specific reservoir type. This step combines formulation design with field application technology to ensure that the design scheme has engineering practicality.

[0044] Experimental Examples and Effect Verification

[0045] To verify the actual effect of the method of the present invention, a complete formulation optimization process was carried out in a medium-high permeability sandstone oil reservoir (permeability 120 mD, temperature 65℃, formation water salinity 35000 ppm, crude oil viscosity 8.5 mPa·s) as the target block.

[0046] 1. Experimental conditions and initial setup

[0047] The initial database contains 2,150 valid formulation-performance data points from historical literature and laboratory tests, covering 12 surfactant systems, 5 co-surfactants, and 3 salinity ranges. Based on this, 120 additional historical formulation and performance data points for this block were added, specifically targeting the characteristics of medium-to-high permeability reservoirs, bringing the total database sample size to 2,270.

[0048] The target performance parameters are set as follows: interfacial tension ≤ 0.005 mN / m, oil displacement efficiency ≥ 85%, and surfactant adsorption ≤ 1.5 mg / g.

[0049] 2. Model Training and Recipe Generation

[0050] The expanded database was used to train a multi-task learning neural network model based on the attention mechanism. The training set, validation set, and test set were divided into 7:1.5:1.5. After 100 rounds of training iterations, the model achieved an R² of 0.941 for interfacial tension prediction, a phase classification accuracy of 92.7%, and an R² of 0.896 for adsorption amount prediction on the test set.

[0051] Using the target performance parameters as conditions, 800 candidate formulations are generated using a conditional variational autoencoder (CVAE). After rapid screening by a machine learning prediction model, the top 40 formulations with the highest comprehensive scores are retained to enter the digital twin verification stage.

[0052] 3. Digital Twin Validation and Iterative Optimization

[0053] Forty candidate formulations were sequentially injected into a digital twin for virtual oil displacement experiments. The digital twin used a 50×50×10 orthogonal grid, with a core size of 30 cm×4.5 cm×4.5 cm, an injection rate of 0.8 m / d, and an injection slug size of 0.5 PV. A multiphase flow simulator was used to solve the mass conservation equations and interphase mass transfer equations for the oil, water, and microemulsion phases. A machine learning surrogate model (3 hidden layers, 64 neurons per layer, prediction error <3%, and approximately 1000 times faster computation) assisted in rapid prediction. Each formulation was run three times, and the average value was taken to eliminate numerical fluctuations.

[0054] After the first round of screening, the formula with the highest overall fitness score (number F-017) scored 0.68, which is lower than the preset threshold of 0.75. This formula and the virtual experiment results were added to the database as incremental samples, triggering incremental learning updates (using an elastic weight consolidation algorithm with a batch size of 20). After four rounds of closed-loop iterations, a total of 3200 candidate formulas were generated, of which 152 were verified by digital twins, reducing the single verification time to less than 20% of the traditional full simulation scheme.

[0055] 4. Optimal formulation and performance

[0056] In the fourth iteration, the optimal formulation (number F-239) that met the fitness threshold was obtained: surfactant C concentration 0.25 wt%, co-surfactant D concentration 0.12 wt%, and NaCl concentration 1.8 wt%. The key performance indicators of this formulation in the digital twin are as follows:

[0057] 5. Physical Experiment Verification

[0058] To verify the reliability of the digital twin prediction results, an oil displacement agent system was prepared according to the optimal formula F-239, and an indoor physical simulation oil displacement experiment was conducted. The experiment used artificial homogeneous sandstone cores (permeability 125 mD, porosity 28.5%), the experimental temperature was 65℃, the injection rate was 0.8 m / d, the injection slug was 0.5 PV, and subsequent water flooding was completed until the water cut reached 98%.

[0059] Physical experiments showed that the core oil displacement efficiency was 86.2%, the interfacial tension was 0.0048 mN / m, and the adsorption capacity was 1.3 mg / g. The relative errors between the digital twin predictions and the physical experimental measurements were 1.6% (oil displacement efficiency), 12.5% ​​(interfacial tension, with an absolute error of only 0.0006 mN / m), and 15.4% (adsorption capacity), respectively. The prediction errors for all indicators were controlled within acceptable ranges, verifying the accuracy of the digital twin and the reliability of the method of this invention.

[0060] 6. Comparative Experiment

[0061] To verify the superiority of the method of the present invention, three sets of comparative experiments were set up: (1) Traditional empirical method: The formulation was optimized by experienced formulation engineers based on phase diagram method and bottle test method. A total of 85 bottle tests and 12 core displacement experiments were carried out, which took about 11 months. The best formulation obtained had an oil displacement efficiency of 79.3% and an interfacial tension of 0.011 mN / m.

[0062] (2) Predictive model only (no generative model): No generative artificial intelligence model is used; the preset formula library (containing 500 predefined formulas) is screened and optimized solely by machine learning predictive model. After 5 rounds of iteration, the optimal formula has an oil displacement efficiency of 81.7% and an interfacial tension of 0.009 mN / m.

[0063] (3) Generative model only (no predictive model): Generative artificial intelligence model is used to generate formulas, but without screening by machine learning predictive model, digital twin verification is performed directly. Due to limited verification resources (only 80 formulas can be verified), the optimal formula oil displacement efficiency is 80.5% and the interfacial tension is 0.010 mN / m.

[0064] 7. Result Comparison

[0065] 8. Experimental Conclusions

[0066] This experimental example shows that: (1) Under the conditions of medium-to-high permeability sandstone reservoirs, the optimal formulation of the method of the present invention achieves an oil displacement efficiency of 87.6% and an interfacial tension of 4.2 × 10⁻⁶. - ³ mN / m, with overall performance significantly better than the traditional empirical method (oil displacement efficiency increased by 8.3 percentage points, interfacial tension reduced by 61.8%). (2) The prediction results of the digital twin are in good agreement with the measured values ​​of the physical experiment (oil displacement efficiency error <2%), which verifies the reliability and engineering applicability of the virtual verification. (3) Generative and predictive models are indispensable. Their synergistic effect improves oil displacement efficiency by 6.0% and 7.1% respectively compared to using them alone, proving the core value of the generative-predictive synergistic mechanism. (4) The method of the present invention shortens the research and development cycle from the traditional 11 months to about 10 days, improves efficiency by about 33 times, and reduces research and development costs (experimental consumables and manpower) by about 90%, which verifies the universality and significant economic benefits of the method in different reservoir types.

[0067] Synergistic effect analysis

[0068] The closed-loop system of this invention is not a simple stacking of modules, but a deeply coupled system designed to meet the special needs of high-dimensional, highly nonlinear, and multi-objective coupling in the formulation design of microemulsion oil displacement agents. The overall effect is significantly better than the sum of the independent effects of each module.

[0069] Generative and predictive models form an efficient positive feedback loop: the generative model explores the recipe space, while the predictive model rapidly evaluates performance, ensuring each generation has a clear direction. Without either, generation becomes aimless or lacks innovation. Digital twins and predictive models work together to achieve a balance between accuracy and efficiency: the surrogate model quickly eliminates inferior recipes, and the multiphase flow simulator accurately verifies high-potential candidates, significantly reducing computational costs and overcoming the traditional notion that accuracy and efficiency are mutually exclusive. Incremental learning continuously feeds verification results back to both the predictive and generative models, increasing model accuracy round by round, with a convergence speed significantly better than the control group without incremental learning. Interpretable analysis and feature linkage directly transmit key feature combinations to the constraints of the generative model, guiding the generation process with physicochemical coordination and avoiding aimless exploration.

[0070] In summary, this invention, through the systematic integration of generative models, predictive models, digital twins, and incremental learning, complements and reinforces each other in a closed loop. This not only achieves a breakthrough improvement in the efficiency of formula optimization, but also demonstrates significantly better performance than traditional solutions in terms of multi-objective trade-offs, generalization ability, and cost savings. It has outstanding substantive features and significant progress.

[0071] The above description is only a specific embodiment of the present invention. Various examples and illustrations do not constitute a limitation on the substantive content of the present invention. Those skilled in the art can make modifications or variations to the above-described specific embodiments after reading the specification without departing from the substance and scope of the invention.

Claims

1. A method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence, characterized in that: Includes the following steps: S1. Data Construction: Collect and integrate historical formulation data and corresponding performance test data of microemulsion oil displacement agents to construct a formulation performance database containing input and output features; S2. Model building and training: Build and train a machine learning prediction model, use the database to train the machine learning prediction model, and establish a mapping relationship from the input features to the output features; S3. Formulation generation: Based on the target performance parameters, a generative artificial intelligence model is used to generate candidate oil displacement agent formulations, and the trained machine learning prediction model is used to evaluate and screen the generated candidate formulations. S4. Formulation optimization and verification: Construct a digital twin of the microemulsion oil displacement process, inject the screened candidate formulations into the digital twin to conduct virtual oil displacement experiments, obtain virtual experimental results, verify the candidate oil displacement agent formulations based on the virtual experimental results, and optimize the machine learning prediction model and / or the generative artificial intelligence model based on the verification results. S5. Output: Output the verified and optimized optimal formula.

2. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: In step S1, the input features include at least one of the following: crude oil composition, formation water salinity and ionic composition, reservoir temperature, surfactant molecular structure descriptor, co-surfactant type and concentration, and salinity. The output characteristics include at least one of interfacial tension, optimal salinity, solubilization parameters, phase type, adsorption capacity, and core displacement efficiency.

3. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: The machine learning prediction model in step S2 is a multi-task learning neural network model based on an attention mechanism; the multi-task learning neural network model includes a shared feature extraction layer and multiple task-specific output layers; the attention mechanism is used to dynamically identify the key feature combination that contributes the most to different prediction tasks among the input features.

4. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 3, characterized in that: The multi-task learning neural network model also includes a surfactant molecule encoder based on a graph neural network. The surfactant molecule encoder is used to map the molecular structure diagram of surfactant molecules into a fixed-dimensional molecular structure feature vector and input it into the shared feature extraction layer.

5. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: The generative artificial intelligence model in step S3 is a conditional variational autoencoder or a conditional generative adversarial network; the target performance parameters are input into the generative artificial intelligence model in the form of conditional variables to constrain the performance range of the generated recipe.

6. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: In step S4, the digital twin integrates a multiphase flow simulator and the machine learning prediction model from step S2; the multiphase flow simulator is used to simulate the physical transport process of microemulsions in porous media, and the machine learning prediction model is used to accelerate the calculation of key indicators of oil displacement efficiency.

7. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 6, characterized in that: The digital twin uses a master-slave collaborative mode for data interaction: the machine learning prediction model acts as a slave module to quickly pre-evaluate candidate formulations and eliminate formulations that are below a preset baseline; the multiphase flow simulator acts as a master module to accurately simulate and verify candidate formulations that have passed the pre-evaluation.

8. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: In step S4, the feedback optimization specifically includes: calculating a comprehensive fitness score based on the virtual experiment results; if the comprehensive fitness score is lower than a preset threshold, adding the candidate formula and its virtual experiment results as incremental samples to the database of step S1, and triggering the incremental learning update of the machine learning prediction model; and re-executing step S3 using the updated machine learning prediction model until an optimized formula that meets the preset fitness requirements is obtained.

9. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: Step S3 further includes an interpretability analysis step: based on the trained machine learning prediction model, the key features that contribute the most to the performance of the candidate formulation are ranked using the feature importance analysis method, and the ranking of the key features is used as the basis for formulation adjustment.

10. The method for generating and optimizing microemulsion oil displacement agent formulations based on artificial intelligence as described in claim 1, characterized in that: It also includes an injection process parameter optimization step: inputting the optimized oil displacement agent formula and reservoir geological characteristic parameters into the reservoir numerical simulator, optimizing the oil displacement agent injection concentration, injection slug size and injection speed, and generating an injection process scheme adapted to a specific reservoir type.

Images