A method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs

Abstract

This invention provides a method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs, belonging to the technical field of determining the carbon dioxide huff and puff mechanism in fractured reservoirs. This invention establishes a unified digital core by selecting outcrop cores to create cast thin sections and employing multi-scale fusion reconstruction technology. Topological connectivity is evaluated based on a seepage topological invariant fracture network connectivity evaluation algorithm. During well shut-in, a diffusion-convection hybrid encoding / decoding model is used to predict migration in real time, and the model learning rate is adjusted by a dynamic learning rate adjustment function. After well shut-in, a feedforward-feedback composite control strategy is used to gradually reduce back pressure to complete the production stage. The experiment ends after cleaning the chip with anhydrous ethanol. Finally, the degree of crude oil recovery is statistically analyzed using image processing software. This invention solves the technical problem of not being able to predict the migration process in fractures and matrix of fractured reservoirs in real time during well shut-in.

Description

Technical Field

[0001] This invention belongs to the technical field of determining the carbon dioxide huff and puff mechanism in fractured reservoirs, and specifically relates to a method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs. Background Technology

[0002] Fractured reservoirs, due to the highly heterogeneous nature of their fracture-matrix coupling structure, make... Huff and puff (Huff and Puff) has become an important means of enhancing oil recovery. In current microfluidic visualization experiments, researchers typically use high-speed cameras to record fluid distribution during the injection and production phases, and combine this with image processing methods to statistically analyze the degree of oil recovery. However, during the well-clogging phase of the huff and puff process, the high-speed camera cannot directly capture the fluid distribution because the chip is in a closed state. The diffusion and convection transport processes in the fractures and matrix result in a lack of information on the evolution of the phase interface during well shut-in.

[0003] While traditional numerical simulation methods, such as the lattice Boltzmann method, can simulate two-phase flow, their high computational complexity makes it difficult to perform real-time frame-by-frame inference during experiments, thus failing to meet the requirements of the well-closing stage. There is a need for dynamic monitoring of transport. When the training samples are limited, purely data-driven neural network models are prone to deviations that violate physical laws, making it difficult to guarantee the physical rationality of phase interface predictions.

[0004] In other words, existing technologies have limitations in enabling the monitoring of the well during the well-sealing process. The technical problem of real-time prediction of migration processes in fractures and matrix of fractured reservoirs. Summary of the Invention

[0005] In view of this, the present invention provides a method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs, which can solve the problem in the prior art that it is impossible to determine the mechanism during well shut-in. The technical problem of real-time prediction of migration processes in fractures and matrix of fractured reservoirs.

[0006] This invention is achieved as follows: This invention provides a method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs, comprising the following steps:

[0007] Select outcrop cores from the target block to produce cast thin sections, use multi-scale fusion reconstruction technology to establish a unified digital core, and fabricate a fractured reservoir microfluidic chip based on the unified digital core.

[0008] Simulated oil was prepared and dyed. Carbon dioxide was pressurized using a gas booster pump. Simulated oil, carbon dioxide, and anhydrous ethanol were placed in three intermediate containers respectively. After cleaning the pipeline, the microfluidic chip of the fractured reservoir was installed into the microfluidic chip holder to establish back pressure, confining pressure, and reservoir temperature.

[0009] The displacement pump saturates the simulated oil into the microfluidic chip of the fractured reservoir through an intermediate container containing simulated oil. After saturation, the simulated oil is aged in the microfluidic chip of the fractured reservoir.

[0010] Open the outlet three-way switching valve and close the inlet three-way switching valve to inject carbon dioxide into the fractured reservoir microfluidic chip through the intermediate container containing carbon dioxide. Use the displacement pump to displace the simulated oil in the fractured reservoir microfluidic chip with low pressure difference. After the simulated oil stops flowing or basically stops flowing, use a high-speed camera to record the distribution of the remaining oil during the injection stage. Calculate the topological connectivity index based on the fracture network connectivity evaluation algorithm based on the seepage topological invariant. Determine the number of effective seepage paths and the carbon dioxide breakthrough time based on the topological connectivity index.

[0011] Close the outlet three-way switching valve and displacement pump, and start well shut-in. During well shut-in, a diffusion-convection hybrid encoding and decoding model is used to predict the transport of carbon dioxide in fractures and matrix in real time. The learning rate parameter of the diffusion-convection hybrid encoding and decoding model is adjusted according to the dynamic learning rate adjustment function.

[0012] After the well is shut down, the distribution of remaining oil during the shut-down phase is recorded by a high-speed camera. The inlet three-way switching valve is opened, and the outlet back pressure is gradually reduced to the specified pressure based on the feedforward-feedback composite control strategy. After the crude oil in the microfluidic chip of the fractured reservoir stabilizes, the inlet three-way switching valve is closed, and the distribution of remaining oil during the production phase is recorded by a high-speed camera.

[0013] Open the inlet three-way switching valve and the outlet three-way switching valve, gradually reduce the back pressure at the outlet to release the pressure to atmospheric pressure, and clean the pipeline and the microfluidic chip of the fractured reservoir with anhydrous ethanol. The experiment ends.

[0014] The remaining oil distribution during the injection stage, the simmering stage, and the production stage, as recorded by high-speed cameras, were loaded into image processing software. The degree of crude oil recovery and the type of remaining oil were statistically analyzed. An equivalent mapping framework from two-dimensional experimental results to three-dimensional reservoirs was established. The carbon dioxide huff and puff mechanism and the characteristics of remaining oil distribution were analyzed. The displacement pump displacement pressure differential, outlet back pressure, and simmering time were changed. The aforementioned steps were repeated to simulate the impact of different huff and puff parameters on crude oil utilization in fractured reservoirs.

[0015] Specifically, the multi-scale fusion reconstruction technology uses macroscopic CT scanning to obtain the spatial topology of fractures in the fractured body, and uses focused ion beam scanning electron microscopy to reconstruct the three-dimensional structure of matrix pores in the fracture zone and matrix zone respectively. Through a scale upgrade algorithm, the pore permeability tensor at the focused ion beam scanning electron microscopy scale is mapped to the grid nodes at the macroscopic CT scanning scale to form a unified digital core across scales.

[0016] Specifically, the scaling algorithm involves performing a volume-weighted average of the permeability tensors of all focused ion beam scanning electron microscope daughter voxels within each macroscopic CT scan grid cell to obtain an equivalent permeability tensor representing that macroscopic CT scan grid cell.

[0017] Specifically, the seepage topological invariant crack network connectivity evaluation algorithm takes a binarized image of the crack network as input, constructs a simple complex sequence at different scales using Vietoris-Rips complexes, calculates the Betti number at each scale, generates a persistent barcode image, and calculates the topological connectivity index.

[0018] The overall architecture of the diffusion-convection hybrid encoding / decoding model is a spatiotemporal convolutional autoencoder. The encoder consists of multiple improved 3D convolutional modules connected in series. A physical prior module is embedded in the bottleneck layer. The diffusion operator discretized from Fick's second law and the convection operator discretized from the convection equation are respectively constructed as convolutional kernels with fixed weights. The decoder consists of multiple deconvolutional modules, and the final output... Phase probability distribution diagram and oil phase probability distribution diagram.

[0019] Specifically, the training dataset of the diffusion-convection hybrid encoding and decoding model is generated by numerical simulation on a unified digital core using the lattice Boltzmann method, and then mixed with the microfluidic image sequence collected in the pre-experiment according to the training ratio. Each sample consists of multiple consecutive frames of images, and the label is the two-phase distribution map of the next frame.

[0020] Specifically, the dynamic learning rate adjustment function is defined as a comprehensive evaluation index. The loss is calculated by weighting three factors: physical residual loss, validation set pixel-level cross-entropy loss, and loss reduction rate across consecutive rounds, based on a comprehensive evaluation index. The range of values ​​is adjusted by using different learning rate multiplication factors to adjust the current learning rate.

[0021] Specifically, the feedforward-feedback composite control strategy is pre-calculated based on thermodynamic equations. When the back pressure at the outlet drops to a threshold distance ahead of the bubble nucleation pressure node, the system switches to a small-step pressure reduction mode. At the same time, the fast-response electromagnetic proportional valve performs closed-loop correction of the pressure reduction rate based on real-time feedback from the pressure sensor.

[0022] Specifically, the equivalent mapping framework from the two-dimensional experimental results to the three-dimensional reservoir is established by using the lattice Boltzmann method to perform benchmark simulation on a unified digital core, and using the dimensionless capillary number and dimensionless Bond number as similarity criteria to establish the conversion relationship between two-dimensional microfluidic observations and three-dimensional equivalent parameters.

[0023] Specifically, the fabrication of the cast thin section and the fabrication of the microfluidic chip for the fractured reservoir involve selecting a long 8 outcrop core that meets the porosity and permeability standard to fabricate a cast thin section, performing binarization and smoothing on the result image to obtain a base map, overlaying the base map to process details to obtain a matrix map of the fractured reservoir, and then overlaying the main fracture and several micro fractures according to the zoning method to fabricate the microfluidic chip for the fractured reservoir using a masking method.

[0024] The physical residual loss refers to the mean square value of the residual obtained by substituting the time evolution results of the latent variables of the diffusion-convection hybrid encoding and decoding model into the discretization equation of Fick's second law.

[0025] Among them, the The bubble nucleation pressure node refers to the pressure at which, under the current reservoir temperature and fluid composition conditions, The critical pressure value corresponding to the transition from the dissolved state to the bubble state is calculated by combining the Peng-Robinson equation of state with fluid composition data.

[0026] The process includes, prior to the saturated simulated oil step, a step of cleaning the pipeline with a displacement pump through an intermediate container filled with anhydrous ethanol, turning off the displacement pump after the liquid flowing out of the microfluidic chip holder and the outlet end is pure anhydrous ethanol, and then loading the microfluidic chip of the fractured reservoir into the microfluidic chip holder after the anhydrous ethanol has evaporated.

[0027] Specifically, the well-sealing time refers to a series of well-sealing pre-experiments conducted within a well-sealing time range threshold, predicted by a diffusion-convection hybrid encoding / decoding model. The time required for the diffusion front to reach the far end of the matrix zone is a theoretical reference. Combined with the measured data of crude oil recovery under various well-closing times, the shortest well-closing time corresponding to the time when the increase in crude oil recovery tends to be stable is taken.

[0028] The back pressure was set to 10 MPa, the confining pressure to be greater than the injection pressure by 2 MPa, the reservoir temperature to 55℃, the displacement pump flow rate to 0.001 mL / min, the low pressure differential to 0.2 MPa, the well shut-in time to 1 hour, the outlet back pressure to be gradually reduced from 10 MPa to 6 MPa, the advance switching distance threshold to 0.5 MPa, the small step pressure reduction step size not exceeding 0.1 MPa, and the pressure control accuracy to be maintained within ±10 kPa.

[0029] This invention constructs a diffusion-convection hybrid encoding and decoding model, embedding physical prior modules into the model's latent space. It embeds the discretized diffusion operator (derived from Fick's second law) and the discretized convection operator (derived from the convection equation) as non-trainable differentiable operators, forcing the temporal evolution trajectory of latent variables to simultaneously satisfy the physical constraints of both diffusion and convection transport. This solves the problem of the inability to predict in real time during well stagnation. Technical issues concerning the migration process within fractures and matrix of fractured reservoirs.

[0030] This invention utilizes a spatiotemporal convolutional autoencoder structure to simultaneously extract pore geometric features and fluid motion dynamics, enabling the model to output physically reasonable values ​​even with limited training samples. The phase probability distribution map and the oil phase probability distribution map overcome the shortcomings of pure data-driven models that produce biases during extrapolation.

[0031] In summary, the present invention solves the problem mentioned in the background art of the inability to control the well during the well-sealing period. The technical problem of real-time prediction of migration processes in fractures and matrix of fractured reservoirs. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method of the present invention.

[0033] Figure 2 This is a schematic diagram of the overall structure of the experimental apparatus of the present invention.

[0034] Figure 3 This is a schematic diagram of a thin section of a core casting of a fractured reservoir according to the present invention.

[0035] Figure 4 This is a thin section processing diagram of the core casting of the fractured reservoir according to the present invention.

[0036] Figure 5 This is a design diagram of the microfluidic chip for fractured reservoirs according to the present invention.

[0037] Figure 6 This is a physical image of the microfluidic chip for fractured reservoirs according to the present invention.

[0038] Figure 7 This is a distribution map of CO2 huff-and-puff injection-well shut-off-residual oil in a fractured reservoir according to the present invention.

[0039] Figure 8 This is a bar chart showing the degree of crude oil recovery at different production stages of this invention.

[0040] Figure 9 This is a bar chart showing the saturation of different types of residual oil after the injection stage of this invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0042] like Figure 1 The diagram shown is a flowchart of a method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs provided by this invention. This method includes the following steps:

[0043] S01. Select outcrop cores from the target block to produce cast thin sections, use multi-scale fusion reconstruction technology to establish a unified digital core, and fabricate a fractured reservoir microfluidic chip based on the unified digital core.

[0044] S02. Prepare simulated oil and stain it with Oil Red O dye. Use a gas booster pump to pressurize carbon dioxide. Place the simulated oil, carbon dioxide and anhydrous ethanol in three intermediate containers respectively. After cleaning the pipeline, install the microfluidic chip of the fractured reservoir into the microfluidic chip holder and establish back pressure, confining pressure and reservoir temperature.

[0045] S03. The simulated oil is saturated into the microfluidic chip of the fractured reservoir through an intermediate container containing simulated oil using a displacement pump. After saturation, the simulated oil is aged in the microfluidic chip of the fractured reservoir for 24 hours.

[0046] S04. Open the outlet three-way switching valve and close the inlet three-way switching valve. Use the displacement pump to inject carbon dioxide into the microfluidic chip of the fractured reservoir through the intermediate container containing carbon dioxide. Use the low pressure difference to displace the simulated oil in the microfluidic chip of the fractured reservoir. After the simulated oil basically stops flowing, use a high-speed camera to record the distribution of the remaining oil during the injection stage. Calculate the topological connectivity index based on the fracture network connectivity evaluation algorithm of seepage topological invariant. Determine the number of effective seepage paths and the carbon dioxide breakthrough time based on the topological connectivity index.

[0047] S05. Close the outlet three-way switching valve and displacement pump, and start well shut-in. During well shut-in, the diffusion-convection hybrid encoding and decoding model is used to predict the transport of carbon dioxide in the fracture and matrix in real time. The learning rate parameter of the diffusion-convection hybrid encoding and decoding model is adjusted according to the dynamic learning rate adjustment function.

[0048] S06. After the well is shut down, use a high-speed camera to record the distribution of remaining oil during the shut-down stage. Open the three-way switching valve at the inlet end and gradually reduce the back pressure at the outlet end to the specified pressure based on the feedforward-feedback composite control strategy. After the crude oil in the microfluidic chip of the fractured reservoir stabilizes, close the three-way switching valve at the inlet end and use a high-speed camera to record the distribution of remaining oil during the production stage.

[0049] S07. Open the inlet three-way switching valve and the outlet three-way switching valve, gradually reduce the outlet back pressure to release the pressure to atmospheric pressure, and clean the pipeline and the microfluidic chip of the fractured reservoir with anhydrous ethanol. The experiment ends.

[0050] S08. Load the remaining oil distribution during the injection stage, the remaining oil distribution during the well-closing stage, and the remaining oil distribution during the production stage recorded by the high-speed camera into the image processing software, statistically analyze the crude oil recovery degree and the type of remaining oil, establish an equivalent mapping framework from the two-dimensional experimental results to the three-dimensional reservoir, analyze the carbon dioxide huff and puff mechanism and the characteristics of the remaining oil distribution, change the displacement pump displacement pressure difference, the outlet back pressure and the well-closing time, repeat S01 to S07, and simulate the influence of different huff and puff parameters on the crude oil utilization of the fractured reservoir.

[0051] The specific steps of the multi-scale fusion reconstruction technology are as follows: Macroscopic CT scanning is used to obtain the spatial topology of fractures in the fracture zone, with a scanning spatial resolution of 1–10 μm; representative voxels are selected for the fracture zone and matrix zone respectively, and the three-dimensional structure of matrix pores is reconstructed using focused ion beam scanning electron microscopy at a resolution of 50–100 nm; a scale-up algorithm is used to map the pore permeability tensor at the focused ion beam scanning electron microscopy scale to the grid nodes at the macroscopic CT scanning scale, forming a unified digital core covering a scale spanning from 50 nm to 500 μm. This multi-scale fusion reconstruction technology, by coupling information from macroscopic CT scanning and focused ion beam scanning electron microscopy within a unified framework, solves the problem that single-resolution imaging cannot simultaneously capture two types of structural details when the fracture aperture (50–500 μm) and matrix pore throat radius (5–50 μm) span two orders of magnitude. This allows subsequent numerical simulations to accurately reflect fluid seepage behavior at different scales, thereby improving the accuracy of the correspondence between experimental results and actual reservoirs.

[0052] The scale upgrade algorithm refers to a mathematical method that maps the permeability tensor of the reconstructed scale of focused ion beam scanning electron microscopy to the nodes of macroscopic CT scanning grids using a volume averaging method. Specifically, it involves performing a volume-weighted average of the permeability tensors of all focused ion beam scanning electron microscopy daughter voxels within each macroscopic CT scanning grid cell to obtain an equivalent permeability tensor representing the equivalent flow capacity of that macroscopic CT scanning grid cell, thereby achieving a unified expression of information from the two imaging scales.

[0053] The steps for fabricating the cast thin section and the microfluidic chip for the fractured reservoir are as follows: Select a long 8 outcrop core that meets the porosity and permeability standard, fabricate a cast thin section, select a suitable result image and perform binarization-smoothing processing to obtain a base image, overlay the base image and process the details to obtain a matrix image of the fractured reservoir, overlay the main fracture and several micro fractures according to the partition to obtain an effect image of the microfluidic chip for the fractured reservoir, determine the etching depth of 15μm and the etching area size of 15mm×10mm, and use the mask method to fabricate the microfluidic chip for the fractured reservoir.

[0054] In S02, the back pressure is set to 10 MPa, the confining pressure is set to be greater than the injection pressure by 2 MPa, and the reservoir temperature is set to 55℃. The back pressure, confining pressure, and reservoir temperature values ​​are obtained by: based on the measured formation pressure and geothermal gradient data of the target block, combined with the fluid seepage stability data measured in multiple rounds of pre-experiments, and by comparing the saturated oil effect under different pressure-temperature combinations, the optimal parameter combination is determined iteratively.

[0055] In S03, the displacement pump flow rate was set to 0.001 mL / min. The displacement pump flow rate was obtained by adjusting the flow rate stepwise within the range of 0.0001 to 0.01 mL / min through a series of preliminary experiments. The standard for judgment was to observe a continuous and stable oil flow without bubble entrainment at the outlet of the microfluidic chip in the fractured reservoir. The lowest displacement pump flow rate that met the standard was taken as the displacement pump flow rate for the formal experiment.

[0056] In S04, the low pressure differential is 0.2 MPa. The low pressure differential is obtained by conducting a series of displacement pre-experiments in the range of 0.05 to 1 MPa, recording the stability of the displacement front and the degree of fracture-matrix crossflow under each pressure differential condition, and using the uniform advancement of the displacement front and the effective mobilization of simulated oil in the matrix as the judgment criteria. The optimal displacement pressure differential is determined through iterative analysis of experimental data.

[0057] The principle and implementation of the seepage topological invariant fracture network connectivity evaluation algorithm are as follows: Using a binarized image of the fracture network as input, a simple complex sequence is constructed at different scales using Vietoris-Rips complexes, and the Betti number at each scale is calculated. Indicates the number of connected components. The topological connectivity index represents the number of loops and generates a persistent barcode map to quantify the multi-scale connectivity characteristics of the fracture network. The formula for calculating the topological connectivity index is as follows:

[0058] ;

[0059] in This is a scale parameter, with units of μm. The weighting function is positively correlated with the duration length and is dimensionless. In order to scale The number of loops below, dimensionless The topological connectivity index is dimensionless. and The lower and upper limits of the scale integral are determined by the range of crack aperture distribution. The topological connectivity index is then compared with... Breaking through time to establish regression relationships, achieving Topological prediction of throughput leading edge arrival time has a computational complexity of O(n log n). ,in The number of nodes in the fracture network is given. The aforementioned seepage topological invariant fracture network connectivity evaluation algorithm quantifies the connectivity and loop structure of the fracture network at different observation scales using algebraic topology methods. The resulting topological connectivity index can directly extract the number of effective seepage paths from the binarized image without relying on a complete pore network model, thereby enabling... A rapid assessment of the seepage capacity of the fracture network before injection provides a quantitative basis for selecting the injection pressure, avoiding low injection efficiency due to insufficient seepage paths. and The method of obtaining the data is as follows: based on the statistical distribution of crack aperture in macroscopic CT scan images, the 5th percentile and 95th percentile values ​​are taken as... and The coverage integrity of the range to the calculation results of the topological connectivity index was verified through multiple rounds of experiments.

[0060] The Betti number is an integer invariant in algebraic topology that describes the connectivity of a topological space. Indicates the number of independent connected components. The number of independent loops is represented, and together they characterize the topological features of the crack network at a specified scale.

[0061] The persistence barcode graph is a visualization output of persistent cohomology computation. The horizontal axis represents the scale parameter, and the vertical axis represents the birth and death events of topological features. The barcode length, i.e. persistence length, reflects the stability of the corresponding topological feature in multi-scale changes. Features with longer persistence lengths correspond to large-scale connected loops that actually exist in the crack network, while features with shorter persistence lengths correspond to noise or small pore disturbances.

[0062] The Vietoris-Rips complex is a method for constructing a simple complex sequence from point cloud data, using crack skeleton nodes as the point set, where the distance between two nodes does not exceed the current scale parameter. When an edge is established between two nodes, then... By gradually forming high-dimensional simplexes such as edges, triangular faces, and tetrahedrons, the topological evolution process of the crack network can be captured at different resolutions.

[0063] The specific structure of the diffusion-convection hybrid encoding / decoding model is as follows: The model takes microfluidic image sequences as input, and its overall architecture is a spatiotemporal convolutional autoencoder. The encoder consists of four improved 3D convolutional modules connected in series. Each improved 3D convolutional module contains a 3D convolutional layer with a kernel size of 3×3×3, a batch normalization layer, and a linear rectified activation function. It extracts local texture features of the pore structure in the spatial dimension and extracts the inter-frame evolution trend of the phase interface in the temporal dimension. Each improved 3D convolutional module is followed by a max-pooling layer with a stride of 2 to perform downsampling, reducing the feature map resolution sequentially to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original input. The bottleneck layer embeds a physical prior module, constructing fixed-weight convolutional kernels from the discretized diffusion operator (Fick's second law) and the discretized convection operator (convection equation), respectively. These kernels are embedded in the latent space as non-trainable differentiable operators, forcing the temporal evolution trajectory of the latent variables to simultaneously satisfy the physical constraints of diffusion and convection transport. The dimension of the latent variables is 512. The decoder consists of four deconvolutional modules. Each deconvolutional module contains a transposed convolutional layer, a batch normalization layer, and a linear rectified activation function, progressively upsampling the latent variables to the same spatial resolution as the input image. The final output has two channels, corresponding to... Phase probability distribution maps and oil phase probability distribution maps are presented. An internal iterative mechanism is introduced into the decoder, performing self-correction three times during the decoding stage. In each iteration, the previous prediction result is added to the current decoded feature map via residual connections, and the prediction error information is fused before continuing decoding. Skip connections are set between corresponding layers of the encoder and decoder, splicing the feature maps of each layer of the encoder with the corresponding layer of the decoder to retain fine-grained pore structure information. The diffusion-convection hybrid encoding and decoding model embeds a physical prior module in the latent space, constraining the model's feature evolution process by the diffusion and convection equations. This ensures physically reasonable phase interface prediction results even with limited training samples, avoiding prediction biases that violate physical laws during extrapolation in purely data-driven models. Simultaneously, the spatiotemporal convolution structure can simultaneously capture pore geometry and fluid movement dynamics, enabling real-time prediction of the diffusion field, which is difficult to observe directly during well stagnation, providing a basis for quantitative optimization of well stagnation time.

[0064] The steps for establishing the training dataset for the diffusion-convection hybrid encoding / decoding model specifically include: using the lattice Boltzmann method on a unified digital core under different injection pressures and well-drainage times. - Numerical simulation of two-phase flow of oil was performed, and the time series of two-phase distribution obtained from the simulation was projected onto a two-dimensional cross section to generate a synthetic image sequence. At the same time, microfluidic image sequences acquired by a high-speed camera in the pre-experiment were collected. The synthetic image sequence and the microfluidic image sequence were mixed at a ratio of 8:2 to form a training dataset. Each sample consists of 10 consecutive frames of images, labeled as the two-phase distribution map of the 11th frame.

[0065] The specific steps for training the diffusion-convection hybrid encoding / decoding model include: using the weighted sum of pixel-level cross-entropy loss and physical residual loss as the total loss function, where the physical residual loss is calculated by substituting the latent variable evolution result into the residual after discretizing the equation according to Fick's second law; the initial learning rate is determined by a dynamic learning rate adjustment function; backpropagation is performed using the Adam optimizer, with a batch size of 8 and 200 training epochs, evaluating the prediction accuracy and saving the optimal weights every 20 epochs on the validation set; after training, the diffusion-convection hybrid encoding / decoding model is deployed to an image acquisition system, and frame-by-frame inference is performed on the real-time microfluidic image sequence acquired by the high-speed camera during the well-closing phase, outputting... Phase probability distribution diagram and oil phase probability distribution diagram.

[0066] The principle and implementation of the dynamic learning rate adjustment function are as follows: Define the comprehensive evaluation index. Comprehensive evaluation index Loss due to physical residuals Pixel-level cross-entropy loss on validation set and the rate of decrease in losses across consecutive rounds The formula is obtained by weighting the three factors and is expressed as follows: ;in This is the initial reference value for physical residual loss. To verify the initial reference value of pixel-level cross-entropy loss, This serves as an initial reference value for the rate of decrease in losses across consecutive rounds. , , For the weighting coefficients, satisfying , This is a dimensionless comprehensive evaluation index. Based on the comprehensive evaluation index... The range of values ​​is adjusted using different learning rate strategies: when When the current learning rate is multiplied by 0.5, when... When, maintain the current learning rate; when When the current learning rate is multiplied by 1.2, then... At that time, the current learning rate is multiplied by 2.0. The thresholds of 0.5, 0.8, and 1.2 are obtained by using a comprehensive evaluation index on 200 rounds of training data from pre-experimental data. The value distribution was statistically analyzed, and a comprehensive evaluation index was obtained. The 25th, 60th, and 90th quantiles of the distribution were used as three threshold values. The final values ​​were determined after three independent training experiments to verify the model's convergence stability under these threshold divisions. The weight coefficients... , , The method for obtaining the results is as follows: in the pre-experiment training, sensitivity analysis is performed on three items: physical residual loss, validation set pixel-level cross-entropy loss, and loss reduction rate in consecutive rounds. Based on the degree of response of the validation set prediction accuracy to the changes in each loss, the optimal combination of weight coefficients is determined in the parameter space with a 0.1 interval by using the grid search method.

[0067] In S05, the well-sealing time is 1 hour. The well-sealing time is obtained by conducting a series of pre-well-sealing experiments within the range of 0.5 to 6 hours, using a diffusion-convection hybrid encoding / decoding model for prediction. The time required for the diffusion front to reach the far end of the matrix zone is used as a theoretical reference. Combined with the measured data of crude oil recovery under various well-closing times, the shortest well-closing time corresponding to the time when the increase in crude oil recovery tends to be stable is taken as the formal experimental well-closing time.

[0068] Specifically, the feedforward-feedback composite control strategy is implemented by pre-calculating the current reservoir temperature and fluid composition conditions based on thermodynamic equations. The pressure node for bubble nucleation is reached when the back pressure at the outlet drops to a distance. When the bubble nucleation pressure node reaches 0.5 MPa, the system actively switches to a small-step pressure reduction mode with a step size not exceeding 0.1 MPa. Simultaneously, a fast-response electromagnetic proportional valve performs closed-loop correction of the pressure reduction rate based on real-time feedback from the pressure sensor, maintaining pressure control accuracy within ±10 kPa. The 0.5 MPa advance switching distance and 0.1 MPa small step size are obtained by conducting a series of pressure reduction experiments under target reservoir temperature and pressure conditions, recording the pressure overshoot under different combinations of advance switching distance and pressure reduction step size. A pressure overshoot below ±10 kPa is considered acceptable. The optimal advance switching distance and pressure reduction step size are determined iteratively through experimental data analysis. This feedforward-feedback composite control strategy predicts in advance based on thermodynamic equations. Bubble nucleation pressure node, near When the pressure node of bubble nucleation occurs, the system actively switches to a small-step pressure reduction mode. Combined with the closed-loop correction of the fast-response electromagnetic proportional valve, it solves the problem of severe overshoot in traditional proportional-integral-derivative control when the fluid phase changes abruptly. This enables high-precision synchronous control of the pressure transient during the pressure reduction production stage in multi-round throughput experiments.

[0069] In S06, the outlet back pressure gradually decreases from 10 MPa to 6 MPa. The target back pressure of 6 MPa is obtained by: based on the measured data of the target block's crude oil bubble point pressure and... Based on the calculated miscibility pressure, and under the premise of ensuring that the fluid remains in a miscible state, the optimal termination pressure is determined iteratively by comparing the degree of crude oil recovery under different termination back pressures through a series of pressure reduction experiments.

[0070] The specific implementation of the equivalent mapping framework from two-dimensional experimental results to three-dimensional reservoirs involves: using the lattice Boltzmann method for benchmarking simulation on a unified digital core; establishing the conversion relationship between two-dimensional microfluidic observations and three-dimensional equivalent parameters using dimensionless capillary number and dimensionless Bond number as similarity criteria; and eliminating systematic biases in the two-dimensional fractured reservoir microfluidic chip experiment regarding geometric topological differences, missing gravity effects, and capillary force anisotropy. The formula for the dimensionless capillary number is as follows: The formula for the dimensionless Bond number is as follows: ;in The fluid seepage velocity is expressed in m / s. For reference seepage velocity, the unit is m / s Fluid viscosity, in units of Reference fluid viscosity, in units of Interfacial tension, unit: For reference interface tension, the unit is... The density difference between the two phases, in units of For reference, the density difference between the two phases is expressed in units of... This is the acceleration due to gravity, in units of 1. For reference to gravitational acceleration, the unit is . The characteristic length is expressed in meters (m). The reference characteristic length is in meters; each reference value is taken from the corresponding physical property parameter value under the measured stratigraphic conditions of the target block.

[0071] The Oil Red O staining agent is used to dye simulated oil red to enhance its contrast with microfluidic images. The color contrast of phases and anhydrous ethanol facilitates two-phase segmentation and crude oil recovery statistics in subsequent image processing.

[0072] Among them, the The bubble nucleation pressure node refers to the pressure at which, under the current reservoir temperature and fluid composition conditions, the pressure at which the bubble nucleates is at a given pressure. The critical pressure value corresponding to the transition from the dissolved state to the bubble state is calculated by combining the Peng-Robinson equation of state with fluid composition data.

[0073] The physical residual loss refers to the mean square value of the residual obtained after substituting the time evolution results of the latent variables of the diffusion-convection hybrid encoding and decoding model into the discretized equation of Fick's second law, which is used to constrain the model prediction results to meet the physical laws of diffusion and transport.

[0074] This invention also provides an experimental apparatus for determining the carbon dioxide huff and puff mechanism of fractured reservoirs. The experimental apparatus consists of six modules, including a microfluidic chip model, an injection system, a huff and puff control system, a confining pressure and temperature control system, a visualization monitoring system, and a data acquisition system. The microfluidic chip model consists of two square, 76mm square, heat- and pressure-resistant glass plates, each 2mm thick. Each plate is divided into an upper capping layer and a lower etched layer, which are bonded together via plasma reaction cleaning and vacuum bonding. Through-holes are made at both ends of the upper capping layer, connecting it to the lower etched layer. A wet etching technique is used to form an etched region with a depth of 15μm and a size of 15mm × 10mm in the lower etched layer. The fragmentation zone is located on the far left of the etched region, measuring 3mm × 10mm; the crack zone is located in the middle of the etched region, measuring 3mm × 10mm; and the matrix zone is located on the far right of the etched region, measuring 9mm × 10mm. The fragmentation zone contains two main cracks and several microcracks in addition to the matrix, with crack apertures ranging from 50 to 500μm. The crack zone contains several microcracks in addition to the matrix, with the microcrack density increasing closer to the fragmentation zone and decreasing closer to the matrix zone. The matrix pore throat radius ranges from 5 to 50μm. The injection system includes a constant speed and constant pressure pump, a gas booster pump, an intermediate container, and... Gas cylinder. The throughput control system includes a three-way switching valve. Opening the outlet three-way switching valve allows fluid to flow from the inlet to the outlet. Closing the outlet three-way switching valve and opening the inlet three-way switching valve connects the inlet directly to the outlet, completing the throughput process. The confining pressure and temperature control system includes a high-temperature and high-pressure chamber, a microfluidic chip holder, a confining pressure tracking pump, a pressure sensor, a temperature control system, and a temperature sensor. The visual monitoring system includes a high-speed camera and an image acquisition system. The acquisition system includes a collection device and a back pressure tracking pump.

[0075] The specific implementation of step S01 is as follows: Select an outcrop core from the target block, fabricate a thin section, and binarize and smooth the resulting image to obtain a base map reflecting the matrix pore structure. Overlay the base map and process details to obtain a matrix map of the fractured reservoir. Based on the zoning relationship between the fracture zone, fracture zone, and matrix zone, overlay the main fracture and several microfractures to obtain a microfluidic chip rendering of the fractured reservoir. In the multi-scale fusion reconstruction stage, macroscopic CT scanning is used to obtain the spatial topology of the fractures in the fractured body at a spatial resolution of 1–10 μm. Representative voxels are selected for the fracture zone and matrix zone respectively, and focused ion beam scanning electron microscopy is used to reconstruct the three-dimensional structure of the matrix pores at a resolution of 50–100 nm. Subsequently, through a scale-up algorithm, the permeability tensor of all focused ion beam scanning electron microscopy daughter voxels within each macroscopic CT scanning grid unit is volume-weighted and mapped to the macroscopic CT scanning scale grid nodes, forming a unified digital core covering a scale from 50 nm to 500 μm. The etching depth of 15 μm and the etching area size of 15 mm × 10 mm were determined. The microfluidic chip of the fractured reservoir was fabricated using the mask method. The fracture zone size was 3 mm × 10 mm, the crack zone size was 3 mm × 10 mm, the matrix zone size was 9 mm × 10 mm, the crack aperture was 50–500 μm, and the matrix pore throat radius was 5–50 μm.

[0076] The specific implementation of step S02 is as follows: Simulated oil with viscosity and density matching reservoir conditions is prepared by mixing white oil and formation crude oil; Oil Red O dye is added to dye the simulated oil red to enhance its similarity to oil. The color contrast of the phase and anhydrous ethanol was measured, and the residue was filtered out for later use. A gas booster pump was used to... Pressurize to the required experimental pressure, then add the simulated oil... Anhydrous ethanol was separately loaded into three intermediate containers. The pipeline was cleaned by a displacement pump through the intermediate containers containing anhydrous ethanol. After the liquid flowing out of the microfluidic chip holder and the outlet was pure anhydrous ethanol, the displacement pump was turned off. After the anhydrous ethanol had fully evaporated, the microfluidic chip for the fractured reservoir was loaded into the microfluidic chip holder and the connection was completed. The back pressure tracking pump, confining pressure tracking pump and temperature control system were turned on. The back pressure was set to 10 MPa, the confining pressure was set to be 2 MPa greater than the injection pressure, and the reservoir temperature was set to 55℃. Each parameter was determined iteratively based on the measured formation pressure and geothermal gradient data of the target block combined with the preliminary experimental data.

[0077] The specific implementation of step S03 is as follows: Open the displacement pump and the intermediate container containing the simulated oil. Set the flow rate of the displacement pump to 0.001 mL / min to saturate the simulated oil into the microfluidic chip of the fractured reservoir. This flow rate is determined by adjusting it stepwise within the range of 0.0001 to 0.01 mL / min, using the observation of a continuous and stable oil flow without entrainment of air bubbles at the outlet as the criterion. After saturation, close the displacement pump and the intermediate container containing the simulated oil, allowing the simulated oil to age in the chip for 24 hours to simulate the reservoir state where crude oil is in full contact with the pore surface.

[0078] The specific implementation of step S04 is as follows: Open the outlet three-way switching valve and close the inlet three-way switching valve to drive the pump through the valve equipped with... The intermediate container injects microfluidic chip into the fractured reservoir. Simulated oil was displaced at a low pressure differential of 0.2 MPa. This pressure differential was determined iteratively through a series of pre-displacement experiments within the range of 0.05–1 MPa, with uniform advancement at the displacement front and effective mobilization of simulated oil in the matrix as the criteria. After the simulated oil essentially ceased to flow, the distribution of remaining oil during the injection phase was recorded using a high-speed camera. Subsequently, a seepage topological invariant fracture network connectivity evaluation algorithm was used to calculate the topological connectivity index. This algorithm takes a binary image of the fracture network as input, constructs a simple complex sequence at different scales using Vietoris-Rips complexes, and calculates the Betti number. and Generate a persistent barcode image using the following formula:

[0079] ;

[0080] Calculate the topological connectivity index, where and Taking the 5th and 95th quantiles of the statistical distribution of crack aperture respectively, the computational complexity is... The number of effective seepage paths is determined based on the topological connectivity index. Breaking through time.

[0081] The specific implementation of step S05 is as follows: Close the outlet three-way switching valve and the displacement pump, and begin well shut-in. The shut-in time is set to 1 hour. This time is determined by conducting a series of well shut-in pre-experiments within the range of 0.5 to 6 hours and taking the shortest shut-in time corresponding to when the increase in crude oil production tends to stabilize. During well shut-in, a diffusion-convection hybrid encoding and decoding model is used to... This model enables real-time prediction of transport within cracks and the matrix. Its overall architecture is a spatiotemporal convolutional autoencoder. The encoder consists of four cascaded improved 3D convolutional modules, with a bottleneck layer embedding a physical prior module. The diffusion operator (discreteized from Fick's second law) and the convection operator (discreteized from the convection equation) are constructed as fixed-weight differentiable operators embedded in the latent space, with a latent variable dimension of 512. The decoder comprises four deconvolutional modules and incorporates an internal iterative mechanism for three rounds of self-calibration. The final output... Phase probability distribution diagrams and oil phase probability distribution diagrams. The training dataset consists of a synthetic image sequence generated by numerical simulation using the lattice Boltzmann method and a pre-experimental microfluidic image sequence mixed in an 8:2 ratio. The total loss function is a weighted sum of pixel-level cross-entropy loss and physical residual loss. The physical residual loss is the mean square value of the residuals obtained by substituting the latent variable evolution into the discretized equation of Fick's second law. The dynamic learning rate adjustment function defines the comprehensive evaluation index. ,when Multiply the learning rate by 0.5 when Time remains unchanged, when Multiply by 1.2 when Multiply by 2.0, and the thresholds of 0.5, 0.8, and 1.2 correspond to the 25th, 60th, and 90th percentile values ​​of the comprehensive evaluation index distribution of the pre-experimental data, respectively.

[0082] The specific implementation of step S06 is as follows: After the well is shut down, a high-speed camera records the distribution of remaining oil during the shut-down phase. The inlet three-way switching valve is opened, and the outlet back pressure is gradually reduced based on a feedforward-feedback composite control strategy. This strategy pre-calculates the current temperature and fluid composition conditions based on the Peng-Robinson equation of state combined with fluid composition data. At the bubble nucleation pressure node, when the outlet back pressure drops to 0.5 MPa from the bubble nucleation pressure node, the system actively switches to a small-step pressure reduction mode with a step size not exceeding 0.1 MPa. Simultaneously, the fast-response electromagnetic proportional valve performs closed-loop correction of the pressure reduction rate based on real-time feedback from the pressure sensor, maintaining pressure control accuracy within ±10 kPa. The outlet back pressure is gradually reduced from 10 MPa to 6 MPa, with the target back pressure based on measured data of the crude oil bubble point pressure. The miscibility pressure calculation results are determined iteratively. After the crude oil in the chip stabilizes, the inlet three-way switching valve is closed, and the distribution of remaining oil during the production stage is recorded by a high-speed camera.

[0083] The specific implementation of step S07 is as follows: Open the inlet three-way switching valve and the outlet three-way switching valve, gradually reduce the back pressure at the outlet to release the pressure to atmospheric pressure, clean the pipeline and the microfluidic chip of the fractured reservoir with anhydrous ethanol, and the experiment ends.

[0084] The specific implementation of step S08 is as follows: The remaining oil distribution during the injection stage, the well-closing stage, and the production stage, recorded by the high-speed camera, are loaded into image processing software. Two-phase segmentation is performed on the images of each stage, and the degree of crude oil recovery and the type of remaining oil are statistically analyzed. When establishing the equivalent mapping framework from the two-dimensional experimental results to the three-dimensional reservoir, the lattice Boltzmann method is used for benchmarking simulation on a unified digital core, using dimensionless capillary numbers... With dimensionless Bond number As a similarity criterion, each reference value is taken from the corresponding physical property parameter values ​​under the measured formation conditions of the target block. A conversion relationship between two-dimensional microfluidic observations and three-dimensional equivalent parameters is established to eliminate systematic deviations in geometric topological differences, missing gravity effects, and capillary force anisotropy. By changing the displacement pump displacement pressure differential, outlet back pressure, and well shut-in time, S01 to S07 are repeated to simulate the impact of different throughput parameters on crude oil mobilization in fractured reservoirs.

[0085] It should be noted that the key technologies of this invention include: The seepage topological invariant fracture network connectivity evaluation algorithm transforms the binary image of the fracture network into a continuous homology computational framework using an algebraic topological method. This directly extracts the topological information of effective seepage paths from the image without relying on a complete pore network model, providing a quantitative basis for selecting the injection pressure. The diffusion-convection hybrid encoding and decoding model embeds physical operators based on Fick's second law and convection equations into the latent space, embedding physical constraints as prior knowledge into the network structure. This allows the model to output physically reasonable phase interface prediction results even with limited samples, thus enabling real-time prediction of the diffusion field, which is difficult to observe directly during well shut-in. The feedforward-feedback composite control strategy predicts bubble nucleation pressure nodes through thermodynamic equations and switches the pressure reduction mode in advance. Combined with closed-loop correction using an electromagnetic proportional valve, it overcomes the overshoot defect of traditional proportional-integral-derivative control during fluid phase transitions. The three components are sequentially linked in the experimental process: the topology algorithm guides the selection of injection pressure, the physical constraint model realizes well stagnation monitoring, and the composite control strategy ensures the accuracy of pressure reduction. Together, they constitute a quantitative control and prediction system for the entire process from injection to production, making the experimental results more repeatable and interpretable.

[0086] It should be noted that this invention also solves the following technical problem: In current microfluidic experiments on fractured reservoirs, due to the systematic differences between the two-dimensional geometry of the microfluidic chip and the actual three-dimensional reservoir in terms of geometric topology, gravity effects, and capillary force anisotropy, the two-dimensional experimental observations are difficult to directly map to three-dimensional reservoir conditions, thus limiting the applicability of the experimental conclusions. This invention establishes an equivalent mapping framework from two-dimensional experimental results to three-dimensional reservoirs, uses the lattice Boltzmann method for benchmarking simulation on a unified digital core, and uses dimensionless capillary number and dimensionless Bond number as similarity criteria. It converts two-dimensional microfluidic observations into three-dimensional equivalent parameters through dimensional analysis, systematically eliminating the biases introduced by geometric topological differences, missing gravity effects, and capillary force anisotropy. This allows two-dimensional experimental conclusions to be quantitatively extended to three-dimensional reservoirs, thereby improving the applicability and accuracy of the experimental results to actual reservoirs.

[0087] Specifically, the principle of this invention is as follows: The reason this invention can solve the aforementioned technical problems lies in the fact that the diffusion-convection hybrid encoding / decoding model embeds fixed-weight convolutional kernels constructed based on Fick's second law and the discretization of the convection equation in the latent space. These physical operators, as non-trainable constraint terms, directly act on the temporal evolution of the latent variables, ensuring that the model's feature update direction is always constrained by the physical laws of diffusion and convection. This design fundamentally differs from purely data-driven methods because the physical constraints do not depend on the number of samples but are embedded in the network structure as prior knowledge, thus maintaining the physical rationality of the prediction results even with limited samples. Simultaneously, the three-dimensional convolutional structure extracts local texture features of pores in the spatial dimension and extracts the inter-frame evolution trend of the phase interface in the temporal dimension. The combined effect of these two aspects enables the model to infer the diffusion field distribution during well stagnation from continuous frame images, which is difficult to observe directly, achieving real-time inference frame by frame. This provides a basis for the quantitative optimization of well stagnation time, and is logically sound.

[0088] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0089] The specific implementation of step S01 is as follows: A cast thin section is prepared from an outcrop core of the target block. The thin section image is binarized and smoothed. The main fracture and several microfractures are superimposed. The etching depth is determined to be 15 μm, and the etching area size is 15 mm × 10 mm. A microfluidic chip for the fractured reservoir is fabricated using a masking method. Macroscopic CT scanning is used to obtain the spatial topology of the fractured body, with a scanning spatial resolution of 1–10 μm. Representative voxels are selected from the fracture zone and matrix zone, and the three-dimensional structure of the matrix pores is reconstructed using focused ion beam scanning electron microscopy at a resolution of 50–100 nm. A scale-up algorithm is used to map the pore permeability tensor at the focused ion beam scanning electron microscopy scale to the grid nodes at the macroscopic CT scanning scale, forming a unified digital core covering a scale span from 50 nm to 500 μm. The scale-up algorithm performs a volume-weighted average of the permeability tensors of all focused ion beam scanning electron microscopy daughter voxels within each macroscopic CT scanning grid cell. The formula is as follows:

[0090] ;

[0091] In the formula, For the first The equivalent permeability tensor of a macroscopic CT scan grid cell, in units of Indexing of macroscopic CT scan grid cells For the first The total number of daughter voxels in a macroscopic CT scan grid cell using a focused ion beam scanning electron microscope, dimensionless. indexing daughter voxels For the first Within the macroscopic CT scan grid cell, the first The volume of a voxel, in units of This is the permeability tensor of the corresponding daughter voxel, in units of .

[0092] The specific implementation methods for steps S02 to S03 are as follows: Prepare simulated oil and stain it with Oil Red O dye, then use a gas booster pump to... Boost the pressure to simulate oil, Anhydrous ethanol was placed in three intermediate containers. The microfluidic chip of the fractured reservoir was installed in the microfluidic chip holder. The back pressure was set to 10 MPa, the confining pressure was 2 MPa greater than the injection pressure, and the reservoir temperature was 55 °C. Simulated oil was saturated into the microfluidic chip of the fractured reservoir at a displacement pump flow rate of 0.001 mL / min. After saturation, the chip was aged for 24 h.

[0093] The specific implementation of step S04 is as follows: injecting a low pressure differential of 0.2 MPa into the microfluidic chip of the fractured reservoir. After the simulated oil is largely displaced and its flow ceases, a high-speed camera records the distribution of the remaining oil during the injection phase. Based on a seepage topological invariant fracture network connectivity evaluation algorithm, the topological connectivity index is calculated to determine the number of effective seepage paths and... Breakthrough in time. The seepage topological invariant fracture network connectivity evaluation algorithm takes a binarized image of the fracture network as input, constructs a simple complex sequence at different scales using the Vitoris-Lipps complex, and uses the fracture skeleton nodes as the point set. When the distance between two nodes does not exceed the current scale parameter... Establish edges as needed By gradually increasing the size of the simplex to form edges, triangular faces, and tetrahedrons, the topological evolution of the crack network can be captured at different resolutions, with a computational complexity of O(n log n). ,in Let be the total number of nodes in the crack network, dimensionless. Calculate the Betty number at each scale. (Number of independent connected components) and (Number of independent loops) Generate a persistent barcode map to quantify multi-scale connectivity features. The horizontal axis of the persistent barcode map represents the scale parameter, and the vertical axis represents the birth and death events of topological features. The barcode length reflects the stability of the corresponding topological feature. Defined as a scale in a persistent barcode image The ratio of the persistence length of a corresponding topological feature to the total persistence length is expressed by the following formula:

[0094] ;

[0095] In the formula, For scale The duration of a topological feature, in units of For integration dummy variables, the unit is The molecular unit is The denominator unit is , The weights are dimensionless. Topological connectivity index. The calculation formula is expressed as follows:

[0096] ;

[0097] In the formula, This is a scale parameter, in units of Dimensionless weighting function In order to scale The number of independent loops under the given condition is dimensionless. The topological connectivity index is dimensionless. and These are the lower and upper limits of the scale integral, respectively, in units of The crack aperture was determined based on the statistical distribution of crack aperture in macroscopic CT scan images, using the 5th and 95th quantiles; the numerator integral result was dimensionless, and the denominator integral result was dimensionless. The whole is dimensionless. and Breaking through time to establish regression relationships, achieving Topological prediction of throughput leading edge arrival time.

[0098] The specific implementation of step S05 is as follows: Close the outlet three-way switching valve and the displacement pump, and begin well shut-in for 1 hour. During well shut-in, a diffusion-convection hybrid encoding / decoding model is used to... The model predicts transport within cracks and the matrix in real time, adjusting the learning rate parameters based on a dynamic learning rate adjustment function. The overall architecture of the diffusion-convection hybrid encoder-decoder model is a spatiotemporal convolutional autoencoder. Taking microfluidic image sequences as input, the encoder consists of four cascaded improved 3D convolutional modules. Each module contains a 3D convolutional layer with a kernel size of 3×3×3, a batch normalization layer, and a linear rectified activation function. Each module is followed by a max-pooling layer with a stride of 2, reducing the feature map resolution sequentially to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original input. The bottleneck layer embeds a physical prior module, constructing a fixed-weight, non-trainable, differentiable operator by discretizing the diffusion operator (Fick's second law) and the convection operator (convection equation), with a latent variable dimension of 512. The decoder consists of four deconvolutional modules, performing three iterations of self-calibration. In each iteration, the previous prediction result is added to the current decoded feature map via a residual connection, resulting in a final output channel of 2, corresponding to... Phase probability distribution maps and oil phase probability distribution maps; skip connections are set between corresponding layers of the encoder and decoder. The training dataset uses the lattice Boltzmann method on a unified digital core under different injection pressures and well-drainage times. Numerical simulations of oil two-phase flow were performed to generate a synthetic image sequence, which was then blended with a pre-experimental microfluidic image sequence at an 8:2 ratio. Each sample consisted of 10 consecutive frames, labeled as the 11th frame's two-phase distribution map. Total loss function... The weighted sum of pixel-level cross-entropy loss and physical residual loss is expressed by the following formula:

[0099] ;

[0100] In the formula, The total loss function value is dimensionless. and The weighting coefficients are dimensionless and determined through a pre-experimental grid search; the empirical values ​​satisfy... Pixel-level cross-entropy loss, dimensionless The physical residual loss is dimensionless, and represents the mean square value of the residual obtained by substituting the latent variable time evolution results into the discretized equation based on Fick's second law. Backpropagation is performed using the Adam optimizer with a batch size of 8 and 200 training epochs. Prediction accuracy is evaluated on the validation set every 20 epochs, and the optimal weights are saved. The dynamic learning rate adjustment function defines the comprehensive evaluation index. The formula is expressed as follows:

[0101] ;

[0102] In the formula, Dimensionless comprehensive evaluation index The initial reference value for the physical residual loss is dimensionless and is taken from the value at the end of the first training round. value To validate the pixel-level cross-entropy loss, an initial reference value is used, which is dimensionless and taken from the value at the end of the first training round. value The current consecutive round loss reduction rate is dimensionless and is defined as the ratio of the difference between the total losses of two adjacent rounds to the total loss of the previous round. This is an initial reference value for the rate of decrease in loss over consecutive rounds, dimensionless, taken from the first to the second round of training. value , , The weighting coefficients are dimensionless and satisfy the following conditions: The parameter was determined within a 0.1 interval parameter space using a grid search method. Based on... Different learning rate adjustment strategies are used for the range of values: when When the current learning rate is multiplied by 0.5, when... When, maintain the current learning rate; when When the current learning rate is multiplied by 1.2, then... At that time, multiply the current learning rate by 2.0. The above threshold is based on the comprehensive evaluation index in the pre-experiment data. The 25th, 60th, and 90th percentile values ​​of the distribution were determined.

[0103] The specific implementation of step S06 is as follows: After the well is shut-in is completed, the distribution of remaining oil during the shut-in stage is recorded using a high-speed camera. The inlet three-way switching valve is opened, and a feedforward-feedback composite control strategy is used to gradually reduce the outlet back pressure from 10 MPa to 6 MPa. After the crude oil in the microfluidic chip of the fractured reservoir stabilizes, the inlet three-way switching valve is closed, and the distribution of remaining oil during the production stage is recorded using a high-speed camera. The feedforward-feedback composite control strategy pre-calculates the current reservoir temperature and fluid composition conditions based on the Penn-Robinson equation of state combined with fluid composition data. Bubble nucleation pressure node, i.e. The critical pressure value corresponding to the transition from dissolved state to bubble state, and the back pressure at the outlet drops to a distance from the point where the solution changes. When the bubble nucleation pressure node is 0.5MPa, it actively switches to a small step pressure reduction mode with a step size not exceeding 0.1MPa. At the same time, the fast-response electromagnetic proportional valve performs closed-loop correction of the pressure reduction rate based on real-time feedback from the pressure sensor, maintaining the pressure control accuracy within ±10kPa.

[0104] The specific implementation of steps S07 to S08 is as follows: The outlet back pressure is gradually reduced to atmospheric pressure, and the pipeline and chip are cleaned with anhydrous ethanol, ending the experiment. The remaining oil distribution at each stage recorded by the high-speed camera is loaded into image processing software to statistically analyze the crude oil recovery rate and remaining oil type, establishing an equivalent mapping framework from the two-dimensional experimental results to the three-dimensional reservoir. The equivalent mapping framework is simulated using the lattice Boltzmann method on a unified digital core, using dimensionless capillary numbers... With dimensionless Bond number As a similarity criterion, a conversion relationship between two-dimensional microfluidic observables and three-dimensional equivalent parameters is established to eliminate systematic biases related to geometric topological differences, missing gravitational effects, and capillary force anisotropy. The dimensionless capillary number formula is expressed as follows:

[0105] ;

[0106] The formula for dimensionless Bond numbers is as follows:

[0107] ;

[0108] In the formula, The fluid seepage velocity is expressed in units of... For reference seepage velocity, the unit is... Fluid viscosity, in units of Reference fluid viscosity, in units of Interfacial tension, unit: For reference interface tension, the unit is... The density difference between the two phases, in units of For reference, the density difference between the two phases is expressed in units of... This is the acceleration due to gravity, in units of 1. For reference to gravitational acceleration, the unit is . The characteristic length is expressed in units of 10 ... For reference feature length, the unit is... Each reference value is taken from the corresponding physical property parameter values ​​under the measured stratigraphic conditions of the target block; each item is a ratio of the same type of quantity. and The overall parameters are dimensionless. By varying the displacement pump pressure differential, outlet back pressure, and well shut-in time, and repeating steps S01 to S07, the effects of different throughput parameters on crude oil mobilization in fractured reservoirs were simulated.

[0109] To better understand and implement this invention, the following is a specific application scenario of embodiment 2: In order to solve practical problems using the technical solution of this invention, technicians selected a block of the Chang 8 reservoir in the Triassic Yanchang Formation on the edge of the Ordos Basin as the target block for experiments. The permeability of this block is approximately The crude oil has a surface area of ​​μm², a porosity of 8%–10%, a density of approximately 0.85 g / cm³, a viscosity of approximately 10 mPa·s, a formation pressure of 10 MPa, and a geothermal gradient corresponding to a reservoir temperature of 55℃. The experimental setup consists of six modules, including a microfluidic chip model, an injection system, a throughput control system, a confining pressure and temperature control system, a visualization monitoring system, and a data acquisition system. The overall structure of the setup is as follows: Figure 2 As shown.

[0110] The microfluidic chip model consists of two square, temperature- and pressure-resistant glass plates, each 76 mm on each side and 2 mm thick. Each plate is divided into an upper capping layer and a lower etched layer, which are bonded together via plasma reaction cleaning and vacuum bonding. Through-holes are made at both ends of the capping layer to connect with the etched layer. Wet etching is used to form an etched region with a depth of 15 μm and a size of 15 mm × 10 mm. The fracture zone is located on the far left of the etched region, measuring 3 mm × 10 mm; the crack zone is located in the middle of the etched region, measuring 3 mm × 10 mm; and the matrix zone is located on the far right of the etched region, measuring 9 mm × 10 mm. The fracture zone contains two main cracks and several microcracks in addition to the matrix, with crack apertures ranging from 50 to 500 μm. The crack zone contains several microcracks in addition to the matrix, with the microcrack density increasing closer to the fracture zone and decreasing closer to the matrix zone. The matrix pore throat radius ranges from 5 to 50 μm. The injection system includes a constant speed and pressure pump, a gas booster pump, an intermediate container, and... Gas cylinders. The throughput control system includes a three-way switching valve. Opening the outlet three-way switching valve allows fluid to flow from the inlet to the outlet. Closing the outlet three-way switching valve and opening the inlet three-way switching valve connects the inlet directly to the outlet, completing the throughput process. The confining pressure and temperature control system includes a high-temperature and high-pressure chamber, a microfluidic chip holder, a confining pressure tracking pump, pressure sensors, a temperature control system, and temperature sensors. The visual monitoring system includes a high-speed camera and an image acquisition system. The acquisition system includes a collection device and a backpressure tracking pump.

[0111] Select an 8-long outcrop core from this block to prepare a cast thin section, such as Figure 3 As shown, the thin-film casting clearly reveals the matrix pore structure and crack distribution characteristics. For example... Figure 4 The thin section image of the cast body is binarized and smoothed to obtain the base image. The base image is then overlaid and details are processed to obtain the matrix image of the fractured reservoir. Based on zoning, the main fracture and several microfractures are overlaid to obtain the microfluidic chip effect image of the fractured reservoir, as shown. Figure 5 As shown, the partitioned structure of the fracture zone, crack zone, and matrix zone is clearly visible in the design drawing. A physical microfluidic chip for the fractured reservoir was fabricated using the masking method, as shown below. Figure 6 As shown.

[0112] When using multi-scale fusion reconstruction technology to establish a unified digital core, the spatial resolution of macroscopic CT scanning was set to 5 μm, and the resolution of focused ion beam scanning electron microscopy was set to 80 nm. The permeability tensor of the focused ion beam scanning electron microscopy scale was mapped to the macroscopic CT scanning grid nodes through the scale upgrade algorithm to form a unified digital core covering 80 nm to 500 μm. The relevant parameters are shown in Table 1.

[0113] Table 1 Key Parameters for Multi-Scale Fusion Reconstruction

[0114]

[0115] When preparing the simulated oil, a mixture of white oil and formation crude oil was used to prepare a simulated oil with a viscosity of 10 mPa·s and a density of 0.85 g / cm³. Oil Red O dye was added to color the simulated oil red, and the residue was filtered out before use. A gas booster pump was then used to... Pressurized to 10MPa, the simulated oil, Anhydrous ethanol was placed into three intermediate containers. The pipeline was cleaned by a displacement pump through the intermediate containers containing anhydrous ethanol. After the liquid flowing out of the outlet was pure anhydrous ethanol, the displacement pump was turned off. After the anhydrous ethanol had fully evaporated, the microfluidic chip was installed into the holder, and the back pressure was set to 10 MPa, the confining pressure to 12 MPa, and the reservoir temperature to 55°C.

[0116] Simulated oil was saturated into the microfluidic chip at a displacement pump flow rate of 0.001 mL / min, and aged for 24 hours after saturation. The outlet three-way switching valve was opened, and the inlet three-way switching valve was closed, injecting the simulated oil at a low pressure differential of 0.2 MPa. The simulated oil was displaced, and after it had essentially stopped flowing, the distribution of the remaining oil during the injection phase was recorded using a high-speed camera. A seepage topological invariant fracture network connectivity evaluation algorithm was used to calculate the topological connectivity index, where... Taking the 5th percentile value of the crack aperture distribution, which is approximately 55 μm, Taking the 95th percentile value, which is approximately 480 μm, the calculated topological connectivity index is 0.73, indicating that the number of effective seepage paths is sufficient. The breakthrough time is approximately 18 minutes after the injection begins.

[0117] During the well-sealing process, the diffusion-convection hybrid encoding / decoding model performs frame-by-frame inference on the real-time microfluidic image sequence acquired by the high-speed camera, such as... Figure 3 As can be observed The model output describes the diffusion-convective mixed transport process in a fractured network. Phase probability distribution diagram and oil phase probability distribution diagram, dynamic learning rate adjustment function, comprehensive evaluation index at the 80th training round. The value is 1.35, indicating entry is possible. The learning rate was adjusted from the initial value of 0.001 multiplied by 0.5 to 0.0005 within the interval. By the 200th training round, the pixel-level cross-entropy loss on the validation set converged to 0.042. The well-clogging time was set to 1 hour, and after the well-clogging was completed, the distribution of remaining oil during the clogging stage was recorded using a high-speed camera.

[0118] By opening the inlet three-way switching valve, the outlet back pressure is gradually reduced from 10MPa to 6MPa using a feedforward-feedback composite control strategy. During the pressure reduction process, the pressure under the current conditions is calculated based on the Peng-Robinson state equation. The bubble nucleation pressure node is approximately 7.2 MPa. When the back pressure drops to 7.7 MPa, it actively switches to a small-step pressure reduction mode with a step size of 0.08 MPa. The electromagnetic proportional valve performs closed-loop correction based on real-time feedback from the pressure sensor, maintaining pressure control accuracy within ±8 kPa. After the crude oil in the chip stabilizes, the inlet three-way switching valve is closed, and a high-speed camera records the distribution of remaining oil during the production stage.

[0119] High-speed camera images from the injection, well-clogging, and production stages were loaded into image processing software for two-phase segmentation. The degree of crude oil recovery and the type of remaining oil were then statistically analyzed. The results are as follows: Figure 7 As shown in the figure, the distribution characteristics of residual oil in the three stages are clearly displayed. The saturation distribution of different types of residual oil is shown in Table 2.

[0120] Table 2. Statistical table of residual oil saturation of different types after injection stage

[0121]

[0122] Crude oil recovery levels at different production stages, such as Figure 8 As shown, the crude oil recovery rate was 38.2% during the injection phase and reached 61.5% during the production phase after well shut-in, indicating that the well shut-in process... Diffusion into the matrix carried away a significant amount of residual oil that failed to be displaced during the injection phase. The bar chart showing the saturation of different types of residual oil after the injection phase is shown below. Figure 9 As shown, clustered residual oil accounts for the highest proportion and is mainly distributed in the low-permeability areas of the matrix zone.

[0123] When establishing an equivalent mapping framework from two-dimensional experimental results to three-dimensional reservoirs, the dimensionless capillary number is used. With dimensionless Bond number For similarity criteria, refer to seepage velocity. Take the measured seepage velocity under actual formation conditions in the target block, and refer to the interfacial tension. Take actual measurements - Crude oil interfacial tension values, with reference values ​​shown in Table 3, were simulated on a unified digital core using the lattice Boltzmann method. This established the conversion relationship between two-dimensional microfluidic observations and three-dimensional equivalent parameters, eliminating systematic biases in two-dimensional chip experiments regarding geometric topological differences, lack of gravity effects, and capillary force anisotropy.

[0124] Table 3. Reference Material Properties of the Two-Dimensional to Three-Dimensional Equivalent Mapping Framework

[0125]

[0126] By varying the displacement pump differential pressure, outlet back pressure, and well shut-in time, the above steps were repeated to simulate the impact of different huff and puff parameters on crude oil recovery in fractured reservoirs. Table 4 shows the comparison of crude oil recovery rates under different huff and puff parameters.

[0127] Table 4 Comparison of crude oil recovery rates under different throughput parameters

[0128]

[0129] During the cleaning phase of the experiment, the inlet three-way switching valve and the outlet three-way switching valve were opened, and the back pressure at the outlet end was gradually reduced to atmospheric pressure. The pipeline and microfluidic chip were cleaned with anhydrous ethanol, and the experiment was completed.

[0130] Compared to traditional methods, this invention achieves the following technological advancements: Traditional microfluidic experiments cannot obtain data during well shut-in. The dynamic information of migration can only be inferred from static images before and after well sumption, thus revealing the diffusion effect. However, this invention embeds discretized operators of Fick's second law and the convection equation into the latent space of a diffusion-convection hybrid encoding / decoding model. This constrains the model's characteristic evolution process with physical laws, enabling the inference of the diffusion field distribution during well sumption from a limited number of consecutive frames. This allows for real-time frame-by-frame prediction, filling the information gap in well sumption observations. Traditional fracture network connectivity evaluation relies on the construction of a complete pore network model, which is computationally intensive and depends on complete three-dimensional structural data. In contrast, the seepage topological invariant fracture network connectivity evaluation algorithm used in this invention directly extracts the topological connectivity index from the binarized image using an algebraic topological method, achieving a computational complexity of only [missing information]. This invention allows for the quantitative assessment of the seepage capacity of fracture networks without relying on a complete pore network model, providing a basis for selecting injection pressure. Traditional pressure reduction control uses proportional-integral-derivative control, which is prone to severe overshoot during fluid phase transitions. In contrast, the feedforward-feedback composite control strategy of this invention predicts the bubble nucleation pressure node through thermodynamic equations and switches the pressure reduction mode in advance. Combined with closed-loop correction by an electromagnetic proportional valve, the pressure control accuracy is maintained within ±10 kPa, thereby ensuring high-precision control of the pressure transients during the pressure reduction production stage in multiple throughput experiments.

[0131] It should be noted that the variables involved in this invention are explained in detail in Table 5.

[0132] Table 5. Variable Explanation Table

[0133]

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for determining the carbon dioxide huff and puff mechanism in fractured reservoirs, characterized in that, Includes the following steps: Select outcrop cores from the target block to produce cast thin sections, use multi-scale fusion reconstruction technology to establish a unified digital core, and fabricate a fractured reservoir microfluidic chip based on the unified digital core. Simulated oil was prepared and dyed. Carbon dioxide was pressurized using a gas booster pump. Simulated oil, carbon dioxide, and anhydrous ethanol were placed in three intermediate containers respectively. After cleaning the pipeline, the microfluidic chip of the fractured reservoir was installed into the microfluidic chip holder to establish back pressure, confining pressure, and reservoir temperature. The displacement pump saturates the simulated oil into the microfluidic chip of the fractured reservoir through an intermediate container containing simulated oil. After saturation, the simulated oil is aged in the microfluidic chip of the fractured reservoir. Open the outlet three-way switching valve and close the inlet three-way switching valve to inject carbon dioxide into the fractured reservoir microfluidic chip through the intermediate container containing carbon dioxide. Use the displacement pump to displace the simulated oil in the fractured reservoir microfluidic chip with low pressure difference. After the simulated oil stops flowing, use a high-speed camera to record the distribution of the remaining oil during the injection stage. Calculate the topological connectivity index based on the fracture network connectivity evaluation algorithm based on the seepage topological invariant. Determine the number of effective seepage paths and the carbon dioxide breakthrough time based on the topological connectivity index. Close the outlet three-way switching valve and displacement pump, and start well shut-in. During well shut-in, a diffusion-convection hybrid encoding and decoding model is used to predict the transport of carbon dioxide in fractures and matrix in real time. The learning rate parameter of the diffusion-convection hybrid encoding and decoding model is adjusted according to the dynamic learning rate adjustment function. After the well is shut down, the distribution of remaining oil during the shut-down phase is recorded by a high-speed camera. The inlet three-way switching valve is opened, and the outlet back pressure is gradually reduced to the specified pressure based on the feedforward-feedback composite control strategy. After the crude oil in the microfluidic chip of the fractured reservoir stabilizes, the inlet three-way switching valve is closed, and the distribution of remaining oil during the production phase is recorded by a high-speed camera. Open the inlet three-way switching valve and the outlet three-way switching valve, gradually reduce the back pressure at the outlet to release the pressure to atmospheric pressure, and clean the pipeline and the microfluidic chip of the fractured reservoir with anhydrous ethanol. The experiment ends. The remaining oil distribution during the injection stage, the simmering stage, and the production stage, as recorded by high-speed cameras, were loaded into image processing software. The degree of crude oil recovery and the type of remaining oil were statistically analyzed. An equivalent mapping framework from two-dimensional experimental results to three-dimensional reservoirs was established. The carbon dioxide huff and puff mechanism and the characteristics of remaining oil distribution were analyzed. The displacement pump displacement pressure differential, outlet back pressure, and simmering time were changed. The aforementioned steps were repeated to simulate the impact of different huff and puff parameters on crude oil utilization in fractured reservoirs.

2. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 1, characterized in that, The multi-scale fusion reconstruction technology specifically uses macroscopic CT scanning to obtain the spatial topological structure of fractures in the fractured body, and uses focused ion beam scanning electron microscopy to reconstruct the three-dimensional structure of matrix pores in the fracture zone and matrix zone respectively. Through a scale upgrade algorithm, the pore permeability tensor at the focused ion beam scanning electron microscopy scale is mapped to the grid nodes at the macroscopic CT scanning scale to form a unified digital core across scales.

3. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 2, characterized in that, The scale-up algorithm specifically involves performing a volume-weighted average of the permeability tensors of all focused ion beam scanning electron microscope daughter voxels within each macroscopic CT scan grid cell to obtain an equivalent permeability tensor representing that macroscopic CT scan grid cell.

4. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 3, characterized in that, The seepage topological invariant fracture network connectivity evaluation algorithm specifically uses a binarized image of the fracture network as input, constructs a simple complex sequence at different scales using Vietoris-Rips complexes, calculates the Betti number at each scale, generates a persistent barcode image, and calculates the topological connectivity index.

5. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 4, characterized in that, The overall architecture of the diffusion-convection hybrid encoding / decoding model is a spatiotemporal convolutional autoencoder. The encoder consists of multiple improved 3D convolutional modules connected in series. A physical prior module is embedded in the bottleneck layer. The diffusion operator discretized from Fick's second law and the convection operator discretized from the convection equation are respectively constructed as convolutional kernels with fixed weights. The decoder consists of multiple deconvolutional modules, and the final output... Phase probability distribution diagram and oil phase probability distribution diagram.

6. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 5, characterized in that, The training dataset for the diffusion-convection hybrid encoding and decoding model is specifically generated by numerical simulation on a unified digital core using the lattice Boltzmann method, and then mixed with the microfluidic image sequence collected in the pre-experiment according to the training ratio. Each sample consists of multiple consecutive frames of images, labeled as the two-phase distribution map of the next frame.

7. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 6, characterized in that, The dynamic learning rate adjustment function is specifically defined as the comprehensive evaluation index. The loss is calculated by weighting three factors: physical residual loss, validation set pixel-level cross-entropy loss, and loss reduction rate across consecutive rounds, based on a comprehensive evaluation index. The range of values ​​is adjusted by using different learning rate multiplication factors to adjust the current learning rate.

8. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 7, characterized in that, The feedforward-feedback composite control strategy is specifically based on pre-calculation of thermodynamic equations. When the back pressure at the outlet drops to a threshold distance ahead of the bubble nucleation pressure node, the system switches to a small-step pressure reduction mode. At the same time, the fast-response electromagnetic proportional valve performs closed-loop correction of the pressure reduction rate based on real-time feedback from the pressure sensor.

9. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 8, characterized in that, The equivalent mapping framework from the two-dimensional experimental results to the three-dimensional reservoir is specifically established by using the lattice Boltzmann method to perform benchmark simulation on a unified digital core, and using the dimensionless capillary number and dimensionless Bond number as similarity criteria to establish the conversion relationship between two-dimensional microfluidic observations and three-dimensional equivalent parameters.

10. The method for determining the carbon dioxide huff and puff mechanism of fractured reservoirs according to claim 9, characterized in that, The fabrication of the cast thin section and the microfluidic chip for the fractured reservoir are specifically carried out by selecting a long 8 outcrop core that meets the porosity and permeability standard to fabricate a cast thin section, performing binarization and smoothing on the result image to obtain a base map, superimposing the base map to process details to obtain a matrix map of the fractured reservoir, superimposing the main fracture and several micro fractures according to the zoning, and using the masking method to fabricate the microfluidic chip for the fractured reservoir.

Images