A method for constructing an intelligent physical space world model and related equipment

Abstract

This invention provides a method and related equipment for constructing an intelligent physical space world model. By constructing a multi-source information latent feature spatial model, it achieves deep fusion of multi-source heterogeneous data such as geological outcrop profiles, stratigraphic geological models, and wavefield simulation distributions. At the same time, through multi-mode clustering and network topology optimization, it enhances the identification accuracy of stratigraphic structural boundaries and achieves joint and accurate identification of deep sedimentary facies. Based on the sedimentary facies identification results, it completes the segmentation of multiple sub-world regions. Through a multi-agent reinforcement learning algorithm, it adapts the prediction sub-model to each sub-world region, solving the spatial coupling problem of multiple sub-regions and improving the accuracy and reliability of the deep underground physical world model.

Description

Technical Field

[0001] This invention relates to the field of underground geological space modeling technology, and in particular to a method for constructing an intelligent physical space world model and related equipment. Background Technology

[0002] The core of underground physical space world modeling is the accurate three-dimensional digital representation of underground stratigraphic structure, attribute distribution, and physical characteristics, serving as the core foundation for resource exploration and development, and underground engineering planning. Currently, mainstream underground physical space modeling technologies are mainly divided into two categories: traditional geostatistical modeling methods and data-driven deep learning modeling methods. Meanwhile, the development of world model-related technologies has also provided new theoretical references for intelligent spatial modeling.

[0003] Traditional methods for modeling the underground physical world are primarily based on geostatistics, interpolation theory, and variational models. For example, geological modeling methods based on Kriging interpolation and sequential Gaussian models achieve spatial interpolation of stratigraphic attributes by introducing prior knowledge of spatial continuity. These methods are applicable to areas with simple geological conditions and abundant data, but their modeling quality heavily relies on the manual setting of empirical parameters such as variograms and regularization weights. For real-world scenarios with complex deep sedimentary environments, strong heterogeneity, and limited data, their adaptability and robustness are severely lacking, easily leading to problems such as multiple solutions and inconsistencies with actual geological characteristics. Furthermore, while traditional facies-controlled modeling methods introduce the constraint of sedimentary facies zones, the sedimentary identification and attribute modeling processes are disconnected, failing to establish an essential connection between sedimentary facies zone division and rock physical mechanisms, making it difficult to achieve differentiated and accurate modeling.

[0004] With the development of deep learning technology, data-driven subsurface space modeling methods can learn complex nonlinear mappings from observational data to subsurface stratigraphic properties from massive drilling and seismic data, surpassing traditional methods in multiple spatial attribute prediction tasks. Researchers have further introduced advanced architectures such as attention mechanisms, reinforcement learning, and tree search to improve the accuracy of spatial modeling and the rationality of reasoning. For example, Monte Carlo tree search can be used to efficiently explore complex reasoning spaces and balance exploration and utilization in the modeling process. However, most of these methods adopt a globally unified network architecture and modeling paradigm, which is not adaptable to the differentiated characteristics of different sedimentary facies zones and lacks the ability to specifically perceive and differentiate specific sedimentary areas. It is difficult to achieve the optimal balance between completely restoring local heterogeneity and perfectly maintaining global spatial continuity.

[0005] In recent years, world modeling technologies have made groundbreaking progress in intelligent decision-making and environmental modeling. Related research, by constructing internal mental representations of the environment, has achieved accurate simulations of future states after actions are performed, providing a new paradigm for modeling and planning complex spaces. These technologies have been successfully applied in scenarios such as game agents and embodied decision-making in robots. However, most existing world modeling technologies focus on discrete-state text reasoning and virtual game environment modeling, lacking the ability to characterize continuous three-dimensional underground physical spaces, multi-regional spatial coupling relationships, and geophysical mechanisms. They cannot adapt to the strong heterogeneity, diverse information heterogeneity, and complex spatial relationships of deep sedimentary environments, making them difficult to directly apply to high-precision construction of the underground physical world. Furthermore, existing methods cannot effectively transform the differences in deep sedimentary environments into prior constraints for physical world modeling, leading to a disconnect between sedimentary facies identification and physical world modeling processes, failing to fundamentally reduce the model's ambiguity. Summary of the Invention

[0006] This invention provides a method and related equipment for constructing an intelligent physical space world model, with the aim of improving the accuracy and reliability of deep underground physical world models.

[0007] To achieve the above objectives, the present invention provides a method for constructing an intelligent physical space world model, comprising: Step 1: Collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multivariate effective dataset; Step 2: Use geological outcrop profile data and drilling engineering data to classify sedimentary spatial models, and obtain the classification results of sedimentary spatial models and a prior knowledge base of sedimentary facies models; Step 3: Using the sedimentary spatial pattern classification results as constraints, the constructed deep neural network is optimized, and then the geophysical exploration data is input into the optimized identification network for identification to obtain the structural identification results. Step 4: Generate a three-dimensional wavefield simulation distribution data volume of the target work area based on the structural identification results and the prior knowledge base of sedimentary facies models. Calculate the wavefield characteristic deviation between the wavefield simulation distribution data volume and the geophysical exploration data. After correcting the wavefield characteristic deviation, fuse the wavefield simulation distribution data volume with the structural identification results to obtain a comprehensive geophysical feature volume. Step 5: Input the multivariate valid dataset into the constructed multivariate information latent feature space model for mapping to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. Step 6: The sedimentary facies classification network constructed based on the multivariate information latent feature spatial model is identified by integrating latent feature inputs to obtain sedimentary facies distribution data volume. The sedimentary facies distribution data volume is then used to segment the three-dimensional underground space of the target work area to obtain multiple sub-world regions. Step 7: Based on the prior knowledge base of sedimentary facies patterns, set up an agent for each sub-world region, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region; Step 8: Integrate the optimal prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

[0008] Furthermore, step 2 includes: Sedimentological features were extracted from geological outcrop profile data to obtain lithological assemblage features, sedimentary structural features, and vertical sequence stratigraphy features; The core feature combination was selected using a dual-criteria joint feature screening method. Hierarchical clustering is used to hierarchically divide the core feature combinations to obtain hierarchical clustering results; The density clustering method is used to further subdivide the hierarchical clustering results, resulting in multiple feature clusters; By combining drilling engineering data, each feature cluster is validated to obtain the classification results of sedimentary spatial patterns. A prior knowledge base for sedimentary facies models is constructed based on the classification results of sedimentary spatial models.

[0009] Furthermore, deep neural networks include: An encoder composed of multi-scale convolutional modules is used to extract stratigraphic reflection features at different scales from geophysical exploration data. The kernel size and number of convolutional layers of the multi-scale convolutional modules are determined based on the sedimentary spatial pattern classification results. The decoder employs an upsampling and skip connection structure. The decoder also includes a spatial edge attention module, which is used to fuse stratigraphic reflection features at different scales to obtain structural identification results. The multi-scale feature fusion method of the decoder is determined by the sedimentary spatial pattern classification results.

[0010] Furthermore, multimodal branch coding networks include: The geological outcrop sedimentary feature encoding branch is used to extract features from geological outcrop profile data in a multivariate valid dataset; The stratigraphic geological model semantic encoding branch is used to extract features from the semantics of geological models in multivariate valid datasets; The seismic wavefield simulation feature coding branch is used to extract features from geophysical exploration data in multivariate valid datasets; The well logging rock physics feature encoding branch is used to extract features from drilling engineering data in a multivariate valid dataset.

[0011] Furthermore, the sedimentary facies classification network, constructed based on a multivariate information latent feature space model and fused with latent feature inputs, is used for identification to obtain sedimentary facies distribution data, including: A sedimentary facies classification network was constructed based on a multivariate information latent feature space model. A training dataset is constructed to train the sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network. The latent features are input into the trained sedimentary facies classification network and processed to obtain the sedimentary facies probability volume. An adaptive thresholding method is used to determine the sedimentary facies type for each grid in the sedimentary facies probability volume, generating an initial sedimentary facies distribution data volume. Three-dimensional morphological operations were used to remove isolated noise points and anomalous patches in the initial sedimentary facies distribution data volume, and the sedimentary facies boundaries were smoothed and optimized to obtain the first sedimentary facies distribution data volume. Based on sequence stratigraphy and the spatial distribution pattern of sedimentary facies, the first sedimentary facies distribution data volume was optimized with three-dimensional spatial constraints to obtain sedimentary facies distribution data.

[0012] Furthermore, the training process for the sedimentary facies classification network includes: High-confidence labeled data are obtained from drilling engineering data and geological outcrop profile data; High-confidence pseudo-labels were generated for areas without well calibration using the results of sedimentary spatial pattern classification, resulting in semi-supervised pseudo-label data; The sedimentary facies classification network was pre-trained using high-confidence labeled data to obtain the pre-trained sedimentary facies classification network. By introducing consistency regularization constraints to apply feature perturbations to semi-supervised pseudo-label data and inputting it into the pre-trained sedimentary facies classification network, a preliminary optimization of the sedimentary facies classification network is obtained. A composite loss function combining cross-entropy loss and similarity loss is used to iteratively optimize the initially optimized sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network.

[0013] Furthermore, before driving each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm, the following steps are also included: Using the goodness-of-fit criterion as a hard constraint, a hierarchical reward and punishment system is constructed to optimize each agent.

[0014] Furthermore, goodness-of-fit criteria include: A goodness-of-fit criterion for drilling data is used to constrain the measured fit of each agent. A rock physics mechanism goodness-of-fit criterion is used to constrain the geological rationality of each agent; The spatial continuity fit criterion is used to constrain the spatial coupling coordination of each sub-world region.

[0015] This invention also provides a device for constructing an intelligent physical space world model, and applies a method for constructing an intelligent physical space world model. The device includes: The data collection module is used to collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multi-dimensional and effective dataset. The partitioning module is used to classify sedimentary spatial patterns using geological outcrop profile data and drilling engineering data, and to obtain the sedimentary spatial pattern classification results and a prior knowledge base of sedimentary facies patterns. The identification module is used to optimize the constructed deep neural network based on the classification results of sedimentary spatial patterns, and then input geophysical exploration data into the optimized identification network for identification to obtain the structural identification results. The fusion module is used to generate a three-dimensional wavefield simulation distribution data volume of the target work area based on the structure identification results and the prior knowledge base of sedimentary facies models. It also calculates the wavefield characteristic deviation between the wavefield simulation distribution data volume and the geophysical exploration data. After correcting the wavefield characteristic deviation, the wavefield simulation distribution data volume is fused with the structure identification results to obtain a comprehensive geophysical feature volume. The mapping module is used to map the multivariate information latent feature space model constructed from the input of multivariate valid datasets to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. The segmentation module is used to identify the sedimentary facies classification network constructed based on the multivariate information latent feature spatial model by integrating latent feature inputs, obtain sedimentary facies distribution data volume, and use the sedimentary facies distribution data volume to segment the three-dimensional underground space of the target work area to obtain multiple sub-world regions. The driving module is used to set an agent for each sub-world region based on the prior knowledge base of sedimentary facies patterns, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region. The integration module is used to integrate the best prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

[0016] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for constructing an intelligent physical space world model.

[0017] The above-described solution of the present invention has the following beneficial effects: This invention utilizes geological outcrop profile data and drilling engineering data to classify sedimentary spatial patterns, obtaining the classification results and a prior knowledge base of sedimentary facies patterns. Using the sedimentary spatial pattern classification results as constraints, a constructed deep neural network is optimized. Geophysical exploration data is then input into the optimized recognition network for identification, yielding structural identification results. Based on the structural identification results and the prior knowledge base of sedimentary facies patterns, a wavefield simulation distribution data volume of the target work area in three-dimensional space is generated. The wavefield feature deviation between the wavefield simulation distribution data volume and the geophysical exploration data is calculated. The wavefield simulation distribution data volume is corrected based on the wavefield feature deviation and then fused with the structural identification results to obtain a comprehensive geophysical feature volume. The multivariate effective dataset is input into a constructed multivariate information latent feature space model for mapping, obtaining fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. The fused latent features are input into a sedimentary facies classification network constructed based on the multivariate information latent feature space model for identification, yielding sedimentary facies. The invention utilizes a distributed data volume to segment the 3D subsurface space of the target work area, resulting in multiple sub-world regions. Based on a prior knowledge base of sedimentary facies patterns, an agent is assigned to each sub-world region, and a multi-agent reinforcement learning algorithm drives each agent to iteratively optimize, yielding the optimal prediction sub-model for each sub-world region. These optimal prediction sub-models are then integrated to obtain a 3D subsurface intelligent physical space world model, used for the entire exploration process of deep oil and gas reservoirs. Compared to existing technologies, this invention achieves deep fusion of multi-source heterogeneous data, including geological outcrop profiles, stratigraphic geological models, and wavefield simulation distributions, by constructing a multi-source information latent feature spatial model. Simultaneously, multi-mode clustering and network topology optimization enhance the accuracy of stratigraphic boundary identification, enabling joint and accurate identification of deep sedimentary facies. Based on the sedimentary facies identification results, the multi-sub-world region segmentation is completed, and a multi-agent reinforcement learning algorithm is used to adapt the prediction sub-model to each sub-world region, solving the spatial coupling problem of multiple sub-regions and improving the accuracy and reliability of the deep subsurface physical world model.

[0018] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the intelligent physical space world model construction device in an embodiment of the present invention; Figure 3This is a schematic diagram of the structure of the terminal device in an embodiment of the present invention. Detailed Implementation

[0020] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a locking connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] This invention addresses existing problems by providing a method for constructing an intelligent physical space world model and related equipment.

[0025] like Figure 1 As shown, embodiments of the present invention provide a method for constructing an intelligent physical space world model, including: Step 1: Collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multivariate effective dataset; Step 2: Use geological outcrop profile data and drilling engineering data to classify sedimentary spatial models, and obtain the classification results of sedimentary spatial models and a prior knowledge base of sedimentary facies models; Step 3: Using the sedimentary spatial pattern classification results as constraints, the constructed deep neural network is optimized, and then the geophysical exploration data is input into the optimized identification network for identification to obtain the structural identification results. Step 4: Generate a three-dimensional wavefield simulation distribution data volume of the target work area based on the structural identification results and the prior knowledge base of sedimentary facies models. Calculate the wavefield characteristic deviation between the wavefield simulation distribution data volume and the geophysical exploration data. After correcting the wavefield characteristic deviation, fuse the wavefield simulation distribution data volume with the structural identification results to obtain a comprehensive geophysical feature volume. Step 5: Input the multivariate valid dataset into the constructed multivariate information latent feature space model for mapping to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. Step 6: The sedimentary facies classification network constructed based on the multivariate information latent feature spatial model is identified by integrating latent feature inputs to obtain sedimentary facies distribution data volume. The sedimentary facies distribution data volume is then used to segment the three-dimensional underground space of the target work area to obtain multiple sub-world regions. Step 7: Based on the prior knowledge base of sedimentary facies patterns, set up an agent for each sub-world region, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region; Step 8: Integrate the optimal prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

[0026] Specifically, step 1 includes: Collect all types of basic data for the target work area and divide the core data into four categories: The first category is initial geological outcrop profile data, including high-resolution images of outcrops around the work area, measured profile columnar diagrams, lithological identification reports, and sedimentary structure description data; The second category is initial drilling engineering data, including wellbore trajectory data of all wells, full logging curve data, core sampling and experimental analysis data, and logging stratification data; The third category is initial geophysical exploration data, including 3D seismic data volumes of the target work area, seismic stacking velocity data, seismic attribute volume data, and seismic inversion data volumes; First, the initial data was standardized and sorted out to unify the naming rules, units of measurement, and stratigraphic classification standards, thereby eliminating conflicts in basic information from different sources. Then, targeted preprocessing and noise suppression are carried out based on the characteristics of different types of data; Next, based on the geodetic coordinate system of the target work area and combined with the exploration and development coordinate system of the work area, a unified three-dimensional spatial coordinate benchmark covering the entire work area is established, clarifying the coordinate origin, projection method, and vertical depth benchmark surface. At the same time, the unified conversion of surface elevation, core elevation, and stratum depth is completed to ensure that all data are completely unified in terms of vertical depth and horizontal plane benchmarks, eliminating depth misalignment and plane offset problems between different data. Then, three-dimensional spatial registration of multi-source data is carried out based on a unified spatial coordinate benchmark; Finally, all data underwent quality verification to remove invalid data with excessively low signal-to-noise ratios, spatial registration deviations exceeding thresholds, or serious conflicts with regional geological patterns. Simultaneously, the valid data was gridded, mapping all data to a unified three-dimensional grid system according to the modeling grid size of the target work area. Each grid cell corresponds to a unique spatial coordinate and is associated with all valid data information corresponding to that location. Ultimately, geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area were generated, resulting in a multivariate valid dataset.

[0027] In this embodiment of the invention, targeted preprocessing and noise suppression are carried out based on the characteristics of different types of data, including: An abnormal jump point in the standardized logging curve data was removed by using a sliding window filtering method. The missing curve segments were reasonably supplemented by combining regional rock physics laws. At the same time, environmental correction and standardization were performed on logging data from different wells to eliminate systematic errors caused by well conditions and mud invasion. The random noise and coherent noise in the standardized and sorted 3D seismic data volume are suppressed by alternative filtering and random noise attenuation methods, thereby improving the signal-to-noise ratio of the seismic data. At the same time, phase correction is used to ensure the phase consistency of the seismic reflection axis. The standardized geological outcrop profile data are processed by image distortion removal, intelligent extraction of lithological boundaries, removal of weathering zones and invalid data caused by human disturbance, and preservation of true sedimentological characteristics.

[0028] This invention embodiment performs three-dimensional spatial registration of multi-source data based on a unified spatial coordinate reference, including: First, using wellbore trajectory data as control points, we complete the precise positioning of the depth domain for logging data, core sampling, and experimental analysis data, ensuring that each set of logging data, core sampling, and experimental analysis data can be accurately mapped to specific coordinate points in three-dimensional space. Then, using well logging layer data as calibration, the layer calibration and time-depth conversion of the seismic data volume are completed, converting the time-domain seismic data into depth-domain data, and achieving precise vertical alignment between the seismic data and the drilling data; Finally, by analyzing the profile direction and stratigraphic correlation, the spatial mapping and registration of geological outcrop profile data with underground strata are completed, accurately mapping the sedimentological characteristics of the outcrops to the corresponding strata in the underground three-dimensional space, thus achieving spatial fusion of surface outcrops and underground data.

[0029] Specifically, step 2 includes: Sedimentological features were extracted from geological outcrop profile data to obtain lithological assemblage features, sedimentary structural features, and vertical sequence stratigraphy features; The core feature combination was selected using a dual-criteria joint feature screening method. Hierarchical clustering is used to hierarchically divide the core feature combinations to determine the number of major categories and hierarchical relationships of sedimentary spatial patterns, thus obtaining hierarchical clustering results; Density clustering is used to further subdivide the hierarchical clustering results, identify the characteristic clusters corresponding to different sedimentary microstructures, and obtain multiple characteristic clusters to solve the problem that traditional single clustering methods cannot adapt to complex sedimentary feature distributions. By combining drilling engineering data, each feature cluster is validated to obtain the classification results of sedimentary spatial patterns. A prior knowledge base for sedimentary facies models is constructed based on the classification results of sedimentary spatial models.

[0030] It should be noted that the lithological features obtained by extracting sedimentary features from geological outcrop profile data using conventional feature extraction methods in this embodiment of the invention include lithological type, lithological sequence, lithological proportion, lithological percentage, grain sequence transformation law, mudstone interlayer distribution characteristics, sedimentary structural features including bedding type, cross-bedding scale, scour surface characteristics, degree of bioturbation, sand body stacking pattern, and vertical sequence features including single sand body thickness, cycle type, positive / negative rhythm characteristics, and stratigraphic contact relationship.

[0031] Specifically, a dual-criteria joint feature screening method is used to select core feature combinations, including: The geological feature importance index is used to quantify the ability of lithological assemblage features, sedimentary structural features, and vertical sequence features to distinguish sedimentary facies. Its expression is: ; in, Indicators representing the importance of geological features Indicates information gain. Indicates geological contribution. This refers to a single sedimentological characteristic variable to be evaluated, specifically any one of the following characteristics extracted from geological outcrop profiles, well core data, lithological assemblage characteristics, sedimentary structural characteristics, or vertical sequence characteristics; This represents the set of sedimentary facies categories for the target work area, i.e., the complete set of all sedimentary facies / subfacies / microfacies types that need to be distinguished in this invention; Calculate feature pair mutual information The expression is: ; in, , Features , Marginal probability distribution, for and The joint probability distribution, Representation of features The range of values, Representation of features The range of values ​​for .

[0032] In this embodiment of the invention, weak distinguishing features with a geological feature importance index of less than 0.15 are removed, and highly redundant features with mutual information greater than 0.75 are removed, resulting in 12 core features, thus obtaining a core feature combination.

[0033] Specifically, hierarchical clustering is used to hierarchically divide the core feature combinations, including: First, the core feature vector of each emerging sample is initialized as an independent cluster; The Ward method was used to calculate the sum of squared errors between clusters as a distance metric. Iteratively merge the two closest independent clusters until a three-level hierarchical structure of sedimentary facies, subfacies, and microfacies is formed; Finally, the geological rationality of the clustering hierarchy was verified by combining regional sedimentary models.

[0034] Specifically, the process of further subdividing the hierarchical clustering results using density clustering is as follows: Using each major cluster obtained from hierarchical clustering as input, the neighborhood radius and minimum number of samples are set to identify core points, boundary points and noise points; Connecting core points with a distance less than the neighborhood radius to form sub-clusters, assigning boundary points to the corresponding sub-clusters, eliminating noise points with no geological significance, and finally completing the fine subdivision at the sedimentary microfacies level.

[0035] Specifically, by combining drilling engineering data to verify each feature cluster, the results of sedimentary spatial pattern classification are obtained, including: Based on prior knowledge of sedimentary geology experts, each characteristic cluster is geologically verified, focusing on whether the lithological assemblage, sedimentary structure, and vertical sequence of each characteristic cluster conform to the sedimentary laws of the corresponding sedimentary facies, and invalid clusters with unclear geological significance and overlapping characteristics are eliminated; Clusters of features with insufficient distinguishability are merged, and clusters of features with significant differences in geological features are further subdivided; By combining core sampling and experimental analysis data from drilling engineering data, the subdivided feature clusters are quantitatively verified. The parameters of the density clustering method are adjusted to ensure that the final sedimentary spatial pattern category matches the drilling measurement results to a preset threshold, and the sedimentary spatial pattern category classification result is finally determined.

[0036] The prior knowledge base of sedimentary facies models in this embodiment of the invention stores the corresponding core feature combinations, lithological combination rules, statistical distribution range of rock physical parameters, well logging response characteristics, seismic reflection characteristics, spatial distribution rules, and sedimentary genesis interpretations for each type of sedimentary spatial model. At the same time, it also records prior information such as sequence stratigraphy rules, sand body stacking patterns, and planar distribution range of the corresponding sedimentary facies.

[0037] Specifically, deep neural networks include: An encoder composed of multi-scale convolutional modules is used to extract stratigraphic reflection features at different scales from geophysical exploration data. The kernel size and number of convolutional layers of the multi-scale convolutional modules are determined based on the sedimentary spatial pattern classification results. The decoder employs an upsampling and skip connection structure. The decoder also includes a spatial edge attention module, which is used to fuse stratigraphic reflection features at different scales to obtain structural identification results. The multi-scale feature fusion method of the decoder is determined by the sedimentary spatial pattern classification results.

[0038] In this embodiment of the invention, the multi-scale convolution module is used to extract stratigraphic reflection features at different scales from geophysical exploration data, including large-scale stratigraphic interfaces, unconformities, and fault features, as well as small-scale sand body boundaries and stratigraphic pinch-out features.

[0039] Specifically, the kernel size and number of convolutional layers of the multi-scale convolutional module are determined based on the sedimentation spatial pattern classification results, including: For narrow channel sand bodies in braided river sedimentary facies, the convolution kernel size is reduced to improve the network's ability to identify small-scale lateral boundaries; For the large-area sheet sand deposited at the delta front, the receptive field of the convolution nucleus is expanded, and the network's ability to characterize large-scale continuous strata is enhanced.

[0040] In this embodiment of the invention, the spatial edge attention module is used to enhance the edge correlation between adjacent grid cells. By adaptively adjusting the attention weight, it improves the network's sensitivity to identifying stratigraphic boundaries, sand body pinch-outs, and fault locations, solving the problems of fuzzy boundary identification and poor continuity in traditional networks, and making the network output structural results highly consistent with the actual sedimentary facies boundaries.

[0041] Specifically, step 4 includes: Based on the structural identification results, a three-dimensional layered medium is established, and each stratigraphic unit is assigned corresponding sedimentary facies petrophysical parameters, including P-wave velocity, S-wave velocity, density, porosity, etc. The finite difference forward modeling method is then used to simulate the propagation process of seismic waves in the underground three-dimensional medium. The seismic wave field response characteristics of each spatial grid in the entire work area are calculated, including key parameters such as wave field amplitude, frequency, phase, and propagation velocity. Finally, the wave field simulation distribution data volume of the target work area in three-dimensional space is generated. The wavefield simulation distribution data volume is compared with the geophysical exploration data, and the wavefield characteristic deviation between the two is calculated. The wavefield simulated distribution data volume is corrected based on wavefield characteristic deviation until the matching degree between the wavefield simulated distribution data volume and the measured seismic data reaches a preset threshold, thus obtaining the corrected wavefield simulated distribution data volume. The corrected wavefield simulation distribution data volume is fused with the structural identification results to obtain a comprehensive geophysical feature volume that simultaneously contains stratigraphic boundary information and wavefield propagation response characteristics. This feature volume is used to characterize the propagation and response patterns of seismic waves in different sedimentary facies strata.

[0042] To address the problem of poor compatibility and difficulty in deep integration of diverse information such as geological outcrops, stratigraphic models, and wavefield simulations, this invention constructs a multi-source latent feature spatial model through multimodal coding and cross-modal attention fusion. This achieves complementarity and fusion of information from different sources, providing a complete feature foundation for accurate identification of sedimentary facies.

[0043] Specifically, multimodal branch coding networks include: The geological outcrop sedimentary feature encoding branch is used to extract features from geological outcrop profile data in a multivariate valid dataset; The stratigraphic geological model semantic encoding branch is used to extract features from the semantics of geological models in multivariate valid datasets; The seismic wavefield simulation feature coding branch is used to extract features from geophysical exploration data in multivariate valid datasets; The well logging rock physics feature encoding branch is used to extract features from drilling engineering data in a multivariate valid dataset.

[0044] In this embodiment of the invention, each coding branch adopts a network structure adapted to the corresponding data type. For example, a one-dimensional convolutional coding structure is used for numerical geophysical exploration data, a two-dimensional convolutional coding structure is used for image-type geological outcrop profile data, and a Transformer coding structure is used for semantic feature extraction of text-type geological patterns. Each coding branch independently completes the feature extraction of the corresponding data, maps the original data to a high-dimensional feature space, and extracts the deep latent features of various types of data.

[0045] In this embodiment of the invention, in order to eliminate the semantic gap and distribution differences between different modal features, a linear projection layer is used as a cross-modal feature alignment module to map the latent features output by different branches to the same feature dimension space, ensuring that all modal features have the same dimension. Then, a contrastive learning constraint is introduced into the cross-modal feature alignment module to make the distance between different modal features of the same spatial location and the same sedimentary facies type as close as possible in the latent space, and to make the distance between features of different sedimentary facies types as far apart as possible in the latent space, thereby achieving semantic alignment of different modal features. Finally, a multivariate information compatibility constraint is added to the cross-modal feature alignment module to penalize semantic conflicts between different modal features of the same spatial location, forcing the network to learn the intrinsic correlation between different modal features and eliminating compatibility contradictions between multivariate information.

[0046] In this embodiment of the invention, in order to achieve deep fusion of its multimodal latent features, a cross-modal attention fusion module is set up. This cross-modal attention fusion module uses the features of one modality as the query vector and the features of other modalities as the key vector and value vector. Through adaptive calculation of attention weights, it automatically mines the complementary information between features of different modalities, assigns higher weights to features that contribute more to sedimentary facies identification, weakens the interference of redundant and low-confidence features, and thus obtains fused latent features that simultaneously include geological sedimentary features, petrophysical features, geophysical features, and stratigraphic model priors.

[0047] Specifically, the sedimentary facies classification network, constructed based on a multivariate information latent feature spatial model, is used to identify sedimentary facies distribution data, including: A sedimentary facies classification network was constructed based on a multivariate information latent feature space model. A training dataset is constructed to train the sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network. The latent features are input into the trained sedimentary facies classification network and processed to obtain the sedimentary facies probability volume. An adaptive thresholding method is used to determine the sedimentary facies type for each grid in the sedimentary facies probability volume, generating an initial sedimentary facies distribution data volume. Three-dimensional morphological operations were used to remove isolated noise points and anomalous patches in the initial sedimentary facies distribution data volume, and the sedimentary facies boundaries were smoothed and optimized to obtain the first sedimentary facies distribution data volume. Based on sequence stratigraphy and the spatial distribution pattern of sedimentary facies, the first sedimentary facies distribution data volume was optimized with three-dimensional spatial constraints to obtain sedimentary facies distribution data.

[0048] In this embodiment of the invention, the sedimentary facies classification network is obtained by combining a fully connected neural network and a lightweight convolutional network. Its input is the fused latent features, and its output is the probability of different sedimentary facies types corresponding to each spatial grid. Finally, a sedimentary facies probability volume of the three-dimensional space of the entire target work area is generated to characterize the probability distribution of each location in the three-dimensional space belonging to different sedimentary facies types.

[0049] Specifically, the training process for the sedimentary facies classification network includes: High-confidence labeled data are obtained from drilling engineering data and geological outcrop profile data; High-confidence pseudo-labels were generated for areas without well calibration using the results of sedimentary spatial pattern classification, resulting in semi-supervised pseudo-label data; The sedimentary facies classification network was pre-trained using high-confidence labeled data to obtain the pre-trained sedimentary facies classification network. By introducing consistency regularization constraints to apply feature perturbations to semi-supervised pseudo-label data and inputting it into the pre-trained sedimentary facies classification network, a preliminary optimization of the sedimentary facies classification network is obtained. A composite loss function combining cross-entropy loss and similarity loss is used to iteratively optimize the initially optimized sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network.

[0050] Specifically, the three-dimensional underground space of the target work area is segmented using the sedimentary facies distribution data volume to obtain multiple sub-world regions. The specific process is as follows: First, based on sedimentary facies type, continuous spatial grids of the same sedimentary facies type are divided into the same sub-world region, and grids of different sedimentary facies types are divided into different sub-world regions. Finally, the entire three-dimensional subsurface space is divided into multiple independent sub-world regions, each sub-world region corresponding to a specific sedimentary facies zone, possessing unified geological characteristics and rock physical laws. Then, the boundaries of each sub-world region are refined to ensure that the boundaries of the sub-world regions are strictly aligned with the stratigraphic boundaries and sedimentary facies transition zones, and that the boundaries of adjacent sub-world regions do not overlap, misalign, or have spatial gaps. Finally, multiple sub-world regions of the target work area are output, along with the spatial range, boundary coordinates, and corresponding sedimentary facies type of each sub-world region.

[0051] Specifically, before driving each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm, the following steps are also included: Based on the sedimentary facies type corresponding to each sub-world region, the petrophysical laws, attribute distribution characteristics, and well logging response relationships of the corresponding sedimentary facies are retrieved from the prior knowledge base of sedimentary facies models. Adaptive model architectures are selected for different sub-world regions. For example, for sub-world regions with strong heterogeneity of channel sand bodies, a multi-scale convolutional neural network is used as the basic model to enhance the ability to characterize local heterogeneity; for sub-world regions with shallow lacustrine mudstone deposits, a generalized regression neural network is used as the basic model to ensure the spatial continuity of attribute distribution. The basic models of all sub-world regions take seismic attributes and spatial coordinates as inputs and key attributes such as formation porosity, permeability, and clay content as outputs.

[0052] It should be noted that the training set for all basic models is based on the actual drilling data within the sub-world region, including formation attribute data obtained from well logging interpretation, core sampling and experimental analysis data, and also incorporates the corresponding seismic attribute data and sedimentary facies prior constraint data for that sub-world region. The training set is standardized to unify the dimensions and distribution range of the input features. For sub-world regions with limited drilling data, data augmentation methods are used to generate virtual samples that conform to geological laws by combining the petrophysical laws of the corresponding sedimentary facies, thereby expanding the scale of the training dataset and avoiding model overfitting. Finally, a dedicated training set for each basic model is generated.

[0053] Specifically, before driving each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm, the following steps are also included: Using the goodness-of-fit criterion as a hard constraint, a hierarchical reward and punishment system is constructed to optimize each agent.

[0054] Specifically, the goodness-of-fit criteria include: A goodness-of-fit criterion for drilling data is used to constrain the measured fit of each agent. A rock physics mechanism goodness-of-fit criterion is used to constrain the geological rationality of each agent; The spatial continuity fit criterion is used to constrain the spatial coupling coordination of each sub-world region.

[0055] In this embodiment of the invention, the quantitative indicators of the goodness-of-fit criterion for drilling measured data include: the mean absolute error, root mean square error, and correlation coefficient of the formation attribute prediction at the well location. It requires that the mean absolute error of the prediction at all drilling points be lower than a preset threshold, and that the correlation coefficient between the predicted and measured values ​​be higher than a preset threshold, ensuring a high degree of match between the model prediction results and the actual underground conditions. The quantitative indicators of the goodness-of-fit criterion for rock physics mechanisms include: whether the distribution range of formation attribute parameters conforms to the statistical laws of the prior knowledge base of the corresponding sedimentary facies; whether the correlation between attributes conforms to rock physics theory; and whether the vertical attribute changes conform to the sedimentary cycle law. It requires that more than 95% of the grid of attribute parameters predicted by the model conforms to the constraints of rock physics mechanisms, preventing abnormal prediction results that violate geological laws. The quantitative indicators of the goodness-of-fit criterion for spatial continuity include: the attribute gradient change value at the boundary of adjacent sub-regions; the lateral variation coefficient of attributes within the same stratum; and the attribute misalignment amplitude on both sides of a fault. It requires that attributes at the boundary of adjacent sub-regions achieve a smooth transition without abrupt abrupt changes, and that the attribute distribution of the same continuous sedimentary facies zone conforms to the spatial continuity law, maintaining the spatial structural integrity of the global strata.

[0056] Specifically, using the goodness-of-fit criterion as a hard constraint, a hierarchical reward and penalty system is constructed, including: Based on three goodness-of-fit criteria, an immediate reward and penalty function is designed for each agent. After each action is performed by the agent, the environment calculates the corresponding immediate reward or penalty based on the model state after the action. The immediate reward includes three levels: The first level is the accuracy improvement reward. When the agent performs an action, the prediction error and forward modeling error of the basic model decrease compared to before. A corresponding positive reward is given according to the magnitude of the error decrease. The greater the magnitude of the error decrease, the higher the reward value. The second level is constraint satisfaction reward. When a certain goodness-of-fit criterion of the base model changes from not being satisfied to being satisfied, a large positive reward is given to incentivize the agent to satisfy all hard constraints. The third level is the boundary collaboration reward. When the smoothness of the boundary attributes of adjacent sub-world regions is improved and the spatial coupling coordination is optimized, the corresponding two agents are given synchronous positive rewards to guide the agents to collaboratively optimize the boundary relationship of adjacent sub-world regions. The penalty mechanism is also divided into three levels. The first level is the error increase penalty. When the agent performs an action, the prediction error and forward modeling error of the basic model increase. The penalty is given according to the magnitude of the error increase. The second level is constraint violation penalty. When a sub-model violates the three types of goodness-of-fit criteria, a large penalty is imposed according to the severity of the violation. If there is an abnormal prediction that seriously violates the rock physics mechanism, an extreme penalty is imposed to prevent the agent from generating a model that is geologically invalid. The third level is boundary conflict penalty. When there are serious attribute mutations or spatial misalignments at the boundaries of adjacent sub-regions, a penalty is imposed on the corresponding two agents to synchronize, so as to prevent the local optimization of the agents from destroying the global spatial continuity.

[0057] It should be noted that, in order to achieve simultaneous optimization of the prediction accuracy, geological rationality, and spatial coordination of the basic model, this embodiment of the invention designs a cumulative reward function for each agent in addition to the immediate reward. The cumulative reward is the discounted cumulative value of the immediate rewards obtained by the agent throughout the entire iteration process. A discount factor is introduced to make the agent focus more on the long-term model optimization effect rather than short-term local gains. At the same time, the long-term optimization goal of each agent is clearly defined, namely, maximizing its own cumulative reward, while ensuring that the model of the corresponding sub-region satisfies the three types of goodness-of-fit hard constraints.

[0058] This invention addresses the spatial coupling problem in multi-sub-world regions by designing a global collaborative reward mechanism. After each iteration, if all sub-region models satisfy the three types of goodness-of-fit constraints and the average prediction error of the global model decreases compared to the previous iteration, a global collaborative reward is awarded to all agents. If the overall performance of the global model reaches a preset optimal target, a final large reward is awarded to all agents. This global collaborative reward mechanism guides all agents to optimize the model from a globally optimal perspective, avoiding the problem of individual agents achieving local optima while overall global performance declines.

[0059] In this embodiment of the invention, the multi-agent reinforcement learning algorithm is a multi-agent reinforcement learning algorithm based on deep Q-networks. This algorithm is used to fit the state-action value function of each agent and predict the future cumulative reward corresponding to different actions. At the same time, a global centralized critique network is built, which takes the global state and all actions of all agents as input and outputs the global state value evaluation, providing global guidance for the policy optimization of each agent and solving the non-stationarity problem of the multi-agent environment. The algorithm adopts an experience replay pool mechanism to store the interaction samples of each agent's state, action, reward, and next state. The network is trained by random sampling and replay, which improves the stability of training.

[0060] In this embodiment of the invention, the multi-intelligence reinforcement learning algorithm based on deep Q-networks completes pre-training using the initial base model and training dataset for each sub-world region, specifically as follows: Using the parameter adjustments and error reduction of the initial model as samples, the agent learns basic model optimization strategies and masters the basic action logic of reducing prediction errors and satisfying constraints by adjusting model parameters. At the same time, the network parameters of all agents are uniformly initialized, and the same hyperparameters such as exploration rate, learning rate, and discount factor are set to ensure that all agents start iterative optimization under the same initial conditions.

[0061] Specifically, a multi-agent reinforcement learning algorithm drives each agent to perform alternating iterative optimization, which includes four stages, as follows: During the state observation phase, each agent independently observes the state of the current environment, including the model error, constraint satisfaction, and uncertainty results of its corresponding sub-world region, as well as the model state and boundary attribute distribution of neighboring agents, forming the observation state of the current round. During the action execution phase, each agent selects an action to execute based on the observed state using an ε-greedy strategy. It randomly explores new actions with a preset probability and selects the action with the highest current cumulative benefit with the remaining probability, balancing the relationship between exploration and utilization. All agents execute actions sequentially according to the stratigraphic deposition order of the sub-world region to avoid environmental state conflicts caused by simultaneous actions. After executing an action, the agent completes the parameter update of the basic model of the corresponding sub-world region. During the reward calculation phase, after each agent performs an action, the environment calculates the immediate reward obtained by the agent based on the updated state of the model, and at the same time determines whether the conditions for issuing global collaborative rewards are met, and issues collaborative rewards to the corresponding agents. In the experience storage and network update phase, the state, action, reward, and next state sample of each agent are stored in the experience replay pool. A batch of samples are randomly sampled from the replay pool to update the parameters of the deep Q network and global criticism network of each agent, thereby optimizing the network's policy and value evaluation capabilities.

[0062] It should be noted that the termination conditions for alternating iterative optimization in this embodiment of the invention include: The number of iterations has reached the preset maximum limit; All models in all sub-regions satisfy three types of goodness-of-fit hard constraints, and the cumulative gains of all agents do not fluctuate significantly over 10 consecutive iterations. The average prediction error of the global model tends to stabilize. When any termination condition is met, the iterative optimization stops, and the optimal prediction sub-model after convergence for each sub-world region is output.

[0063] Specifically, the optimal prediction sub-models for all sub-world regions are integrated to obtain a three-dimensional underground intelligent physical space world model, including: The stratigraphic attribute data output by the optimal prediction sub-model for each sub-world region is accurately mapped to the corresponding 3D spatial grid, completing the spatial stitching of all optimal prediction sub-models. During the stitching process, the mapping is strictly performed according to the boundary coordinates of the sub-world region to ensure that the attribute data of each spatial grid is unique, without overlap, missing data, or spatial misalignment. At the same time, metadata such as sedimentary facies type, model parameters, and accuracy indicators corresponding to each sub-world region are preserved.

[0064] This invention utilizes geological outcrop profile data and drilling engineering data to classify sedimentary spatial patterns, obtaining sedimentary spatial pattern classification results and a prior knowledge base of sedimentary facies patterns. Using the sedimentary spatial pattern classification results as constraints, a constructed deep neural network is optimized, and geophysical exploration data is input into the optimized recognition network for identification, yielding structural identification results. Based on the structural identification results and the prior knowledge base of sedimentary facies patterns, a wavefield simulation distribution data volume of the target work area in three-dimensional space is generated. The wavefield feature deviation between the wavefield simulation distribution data volume and the geophysical exploration data is calculated. The wavefield simulation distribution data volume is corrected based on the wavefield feature deviation and then fused with the structural identification results to obtain a comprehensive geophysical feature volume. The multivariate effective dataset is input into a constructed multivariate information latent feature space model for mapping, obtaining fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. The fused latent features are input into a sedimentary facies classification network constructed based on the multivariate information latent feature space model for identification, yielding sedimentary facies. The invention distributes data volumes and uses these sedimentary facies distribution data volumes to segment the three-dimensional subsurface space of the target work area, resulting in multiple sub-world regions. Based on a prior knowledge base of sedimentary facies patterns, an agent is assigned to each sub-world region, and a multi-agent reinforcement learning algorithm drives each agent to iteratively optimize, yielding the optimal prediction sub-model for each sub-world region. The optimal prediction sub-models of all sub-world regions are integrated to obtain a three-dimensional subsurface intelligent physical space world model, used for the entire exploration process of deep oil and gas reservoirs. Compared with existing technologies, this embodiment of the invention achieves deep fusion of multi-source heterogeneous data such as geological outcrop profiles, stratigraphic geological models, and wavefield simulation distributions by constructing a multi-source information latent feature spatial model. Simultaneously, through multi-mode clustering and network topology optimization, the accuracy of stratigraphic structural boundary identification is enhanced, achieving joint and accurate identification of deep sedimentary facies. Based on the sedimentary facies identification results, the multi-sub-world region segmentation is completed, and a prediction sub-model is adapted to each sub-world region using a multi-agent reinforcement learning algorithm, solving the spatial coupling problem of multiple sub-regions and improving the accuracy and reliability of the deep subsurface physical world model.

[0065] The present invention also provides a device 100 for constructing an intelligent physical space world model, such as... Figure 2As shown, the method for constructing an intelligent physical space world model includes a construction device 100 comprising: The collection module 101 is used to collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multi-dimensional effective dataset. The partitioning module 102 is used to classify sedimentary spatial patterns using geological outcrop profile data and drilling engineering data, and to obtain the sedimentary spatial pattern classification results and a prior knowledge base of sedimentary facies patterns. The identification module 103 is used to optimize the constructed deep neural network based on the classification results of sedimentary spatial patterns, and then input geophysical exploration data into the optimized identification network for identification to obtain the structural identification results. The fusion module 104 is used to generate a wavefield simulation distribution data volume of the target work area in three-dimensional space based on the structure identification results and the prior knowledge base of sedimentary facies models, and to calculate the wavefield characteristic deviation between the wavefield simulation distribution data volume and the geophysical exploration data. After correcting the wavefield simulation distribution data volume based on the wavefield characteristic deviation, it is fused with the structure identification results to obtain a comprehensive geophysical feature volume. The mapping module 105 is used to map the multivariate information latent feature space model constructed by inputting the multivariate valid dataset to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. The segmentation module 106 is used to identify the sedimentary facies classification network constructed based on the multivariate information latent feature spatial model by integrating the latent feature input, to obtain the sedimentary facies distribution data volume, and to segment the three-dimensional underground space of the target work area using the sedimentary facies distribution data volume to obtain multiple sub-world regions. The driving module 107 is used to set up an agent for each sub-world region based on the prior knowledge base of sedimentary facies patterns, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region. The integration module 108 is used to integrate the best prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

[0066] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0067] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0068] This invention also provides a terminal device, such as... Figure 3 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 3 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, it implements the above-described method for constructing the intelligent physical space world model.

[0069] The terminal device D10 can be a desktop computer, laptop, handheld computer, server, server cluster, or cloud server, etc. This terminal device may include, but is not limited to, a processor D100 and a memory D101. Those skilled in the art will understand that... Figure 3 This is merely an example of terminal device D10 and does not constitute a limitation on terminal device D10. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0070] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0071] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0072] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0074] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing an intelligent physical space world model, characterized in that, include: Step 1: Collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multivariate effective dataset; Step 2: Use the geological outcrop profile data and the drilling engineering data to classify sedimentary spatial patterns, and obtain the sedimentary spatial pattern classification results and the prior knowledge base of sedimentary facies patterns; Step 3: Using the sedimentary spatial pattern classification results as constraints, the constructed deep neural network is optimized, and the geophysical exploration data is input into the optimized recognition network for recognition to obtain the structural recognition results. Step 4: Generate a wavefield simulation distribution data volume of the target work area in three-dimensional space based on the structure identification result and the sedimentary facies model prior knowledge base, calculate the wavefield feature deviation between the wavefield simulation distribution data volume and the geophysical exploration data, correct the wavefield simulation distribution data volume based on the wavefield feature deviation, and then fuse it with the structure identification result to obtain a comprehensive geophysical feature volume. Step 5: The multivariate valid dataset is input into the constructed multivariate information latent feature space model for mapping to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. Step 6: Input the fused latent features into the sedimentary facies classification network constructed based on the multivariate information latent feature spatial model for identification, obtain sedimentary facies distribution data volume, and use the sedimentary facies distribution data volume to segment the three-dimensional underground space of the target work area to obtain multiple sub-world regions; Step 7: Based on the prior knowledge base of sedimentary facies patterns, set up an agent for each sub-world region, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region; Step 8: Integrate the optimal prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

2. The method for constructing an intelligent physical space world model according to claim 1, characterized in that, Step 2 includes: Sedimentological features were extracted from the geological outcrop profile data to obtain lithological assemblage features, sedimentary structural features, and vertical sequence stratigraphy features; The core feature combination was selected using a dual-criteria joint feature screening method. The core feature combination is hierarchically divided using a hierarchical clustering method to obtain the hierarchical clustering results; The hierarchical clustering results are further subdivided using density clustering to obtain multiple feature clusters; By combining the drilling engineering data, each feature cluster is verified to obtain the sedimentation spatial pattern classification results; A prior knowledge base of sedimentary facies models is constructed based on the classification results of the sedimentary spatial models.

3. The method for constructing an intelligent physical space world model according to claim 2, characterized in that, The deep neural network includes: An encoder composed of multi-scale convolutional modules is used to extract stratigraphic reflection features at different scales from the geophysical exploration data. The kernel size and number of convolutional layers of the multi-scale convolutional modules are determined by the sedimentary spatial pattern classification results. The decoder employs an upsampling and skip connection structure, and further includes a spatial edge attention module for fusing stratigraphic reflection features at different scales to obtain structural identification results. The multi-scale feature fusion method of the decoder is determined by the sedimentary spatial pattern classification results.

4. The method for constructing an intelligent physical space world model according to claim 3, characterized in that, The multimodal branch coding network includes: The geological outcrop sedimentary feature encoding branch is used to extract features from the geological outcrop profile data in the multivariate valid dataset; The stratigraphic geological model semantic encoding branch is used to extract features of the geological model semantics in the multivariate valid dataset; The seismic wavefield simulation feature encoding branch is used to extract features from the geophysical exploration data in the multivariate effective dataset; The well logging rock physics feature encoding branch is used to extract features from the drilling engineering data in the multivariate valid dataset.

5. The method for constructing an intelligent physical space world model according to claim 1, characterized in that, The fused latent features are input into a sedimentary facies classification network constructed based on the multivariate information latent feature spatial model for identification, resulting in sedimentary facies distribution data, including: A sedimentary facies classification network is constructed based on the aforementioned multivariate information latent feature space model; A training dataset is constructed to train the sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network. The fused latent features are input into the trained sedimentary facies classification network for processing to obtain the sedimentary facies probability volume. An adaptive thresholding method is used to determine the sedimentary facies type for each grid in the sedimentary facies probability volume, generating an initial sedimentary facies distribution data volume. Isolated noise points and anomalous patches in the initial sedimentary facies distribution data volume are removed using three-dimensional morphological operations, and the sedimentary facies boundaries are smoothed and optimized to obtain the first sedimentary facies distribution data volume. Based on sequence stratigraphy and the spatial distribution pattern of sedimentary facies, the first sedimentary facies distribution data volume is optimized by three-dimensional spatial constraints to obtain sedimentary facies distribution data.

6. The method for constructing an intelligent physical space world model according to claim 5, characterized in that, The training process of the sedimentary facies classification network includes: High-confidence labeled data are obtained from the drilling engineering data and the geological outcrop profile data; The sedimentary spatial pattern classification results are used to generate high-confidence pseudo-labels for areas without well calibration, resulting in semi-supervised pseudo-label data; The sedimentary facies classification network is pre-trained using the high-confidence labeled data to obtain the pre-trained sedimentary facies classification network. After introducing consistency regularization constraints to apply feature perturbation to the semi-supervised pseudo-label data, the data is input into the pre-trained sedimentary facies classification network for preliminary optimization, resulting in a pre-optimized sedimentary facies classification network. A composite loss function combining cross-entropy loss and similarity loss is used to iteratively optimize the initially optimized sedimentary facies classification network, resulting in a well-trained sedimentary facies classification network.

7. The method for constructing an intelligent physical space world model according to claim 1, characterized in that, Before driving each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm, the following is also included: Using the goodness-of-fit criterion as a hard constraint, a hierarchical reward and punishment system is constructed to optimize each agent.

8. The method for constructing an intelligent physical space world model according to claim 7, characterized in that, The goodness-of-fit criteria include: A goodness-of-fit criterion for drilling data is used to constrain the measured fit of each agent. A rock physics mechanism goodness-of-fit criterion is used to constrain the geological rationality of each agent; The spatial continuity fit criterion is used to constrain the spatial coupling coordination of each sub-world region.

9. A device for constructing an intelligent physical space world model, characterized in that, The method for constructing an intelligent physical space world model as described in any one of claims 1-8, wherein the construction apparatus comprises: The data collection module is used to collect geological outcrop profile data, drilling engineering data, and geophysical exploration data of the target work area to obtain a multi-dimensional and effective dataset. The classification module is used to classify sedimentary spatial patterns using the geological outcrop profile data and the drilling engineering data, and to obtain the sedimentary spatial pattern classification results and a prior knowledge base of sedimentary facies patterns. The identification module is used to optimize the constructed deep neural network based on the classification results of the sedimentary spatial patterns, and then input the geophysical exploration data into the optimized identification network for identification to obtain the structural identification results. The fusion module is used to generate a three-dimensional wavefield simulation distribution data volume of the target work area based on the structure identification result and the sedimentary facies model prior knowledge base, calculate the wavefield feature deviation between the wavefield simulation distribution data volume and the geophysical exploration data, correct the wavefield simulation distribution data volume based on the wavefield feature deviation, and then fuse it with the structure identification result to obtain a comprehensive geophysical feature volume. The mapping module is used to map the multivariate effective dataset input into the constructed multivariate information latent feature space model to obtain fused latent features. The multivariate information latent feature space model includes a multimodal branch coding network, a cross-modal feature alignment module, and a cross-modal attention fusion module. The segmentation module is used to identify the sedimentary facies classification network constructed based on the multivariate information latent feature spatial model by inputting the fused latent features, to obtain sedimentary facies distribution data volume, and to segment the three-dimensional underground space of the target work area using the sedimentary facies distribution data volume to obtain multiple sub-world regions; The driving module is used to set up an agent for each sub-world region based on the prior knowledge base of the sedimentary facies pattern, and drive each agent to perform alternating iterative optimization through a multi-agent reinforcement learning algorithm to obtain the optimal prediction sub-model for each sub-world region. The integration module is used to integrate the best prediction sub-models of all sub-world regions to obtain a three-dimensional underground intelligent physical space world model, which is used for the entire process of deep oil and gas reservoir exploration.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for constructing the intelligent physical space world model as described in any one of claims 1 to 8.

Images