Geological-geophysical modeling system and method thereof
Patent Information
- Application Number
- CN202610938525.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-26
- Publication Date
- 2026-10-09
AI Technical Summary
然而，传统的地球物理反演方法通常面临强烈的多解性问题
[0051]本发明通过构建包含数据驱动和物理约束的双分支物理信息多模态神经网络，并利用交叉梯度算子和符号距离函数构建了融合数据拟合误差、物理约束残差和地质构造约束的多重联合损失函数。使得最终输出的三维物理属性场在空间上能够严格满足观测数据拟合、客观物理定律以及隐式地质界面的多重约束，不仅获得极高的空间分辨率，还消除了常规模型在断层等构造边界处的平滑效应等多解性问题，为精确识别岩性异常体、圈定油气储层的空间展布范围以及为固体矿产提供高精度的靶区定位依据提供强有力的工业级模型支撑。
Smart Images
Figure CN122888151A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysical exploration technology, and in particular to a geological-geophysical modeling system and method. Background Technology
[0002] With the increasing difficulty of underground resource exploration, high-precision 3D geological-geophysical modeling has become a crucial prerequisite for oil and gas reservoir identification and solid mineral location. However, traditional geophysical inversion methods often face severe ambiguity issues. Existing modeling methods struggle to deeply and effectively integrate multi-source geophysical observation data with multimodal geological prior knowledge such as well logging and fault analysis. Furthermore, conventional joint inversion algorithms often inevitably produce smoothing effects when dealing with faults and lithological unconformities, leading to blurred boundary delineations. Simultaneously, traditional pure data-driven deep learning models cannot guarantee that the output strictly adheres to objective physical mechanisms, which significantly limits the accuracy and engineering practical value of high-precision modeling in complex geological structures. Summary of the Invention
[0003] This invention provides a geological-geophysical modeling system and method. By using a dual-branch physical information multimodal neural network and multiple joint loss functions, it achieves high-precision three-dimensional modeling that strictly follows objective physical laws and implicit geological interfaces, effectively eliminating multi-solution problems such as smoothing effects at boundaries such as faults.
[0004] This invention provides a geological-geophysical modeling method, comprising:
[0005] S1. Acquire multi-source geophysical observation data and geological prior interpretation data of the target exploration area, and map them to a pre-constructed three-dimensional spatial grid covering the target exploration area for alignment and standardization processing to obtain standardized geophysical data and standardized geological data respectively.
[0006] S2. Extract discrete geological structural interfaces from the standardized geological data, and use the symbolic distance function to convert the discrete geological structural interfaces into a continuous three-dimensional spatial scalar field.
[0007] S3. Construct a two-branch physical information multimodal neural network including a data-driven branch and a physical constraint branch. Input the standardized geophysical data into the data-driven branch for feature extraction to obtain an initial three-dimensional physical property field. The physical constraint branch is equipped with geophysical forward partial differential equations.
[0008] S4. Construct a joint loss function for network optimization that includes a data fitting error term, a physical constraint residual term, and a geological structure constraint error term; wherein, the data fitting error term is constructed based on the initial three-dimensional physical property field and standardized geophysical data, the physical constraint residual term is constructed based on the initial three-dimensional physical property field and geophysical forward partial differential equations, and the geological structure constraint error term is constructed based on the initial three-dimensional physical property field and a three-dimensional spatial scalar field;
[0009] S5. The network parameters of the dual-branch physical information multimodal neural network are iteratively backpropagated and updated using the joint loss function until the preset convergence condition is met, and the three-dimensional physical attribute field output by the converged dual-branch physical information multimodal neural network is used as the target three-dimensional physical attribute model of the target exploration area.
[0010] Furthermore, S1 specifically includes:
[0011] S101. Acquire multi-source geophysical observation data of the target exploration area and collect geological prior interpretation data of the target exploration area; wherein, the multi-source geophysical observation data includes at least three-dimensional seismic data volume, and the geological prior interpretation data includes well logging petrophysical curves, manually or semi-automatically interpreted discrete geological structural horizons, and fault polygon surfaces.
[0012] S102. Based on the spatial range of the target exploration area and the preset modeling accuracy, construct a unified three-dimensional spatial grid covering the target exploration area;
[0013] S103. The three-dimensional seismic data volume is mapped to the three-dimensional spatial grid by a resampling method, and the well logging rock physics curves, discrete geological structural layers and fault polygon surfaces are mapped and assigned to the nearest three-dimensional spatial grid nodes by a three-dimensional spatial interpolation algorithm.
[0014] S104. The maximum-minimum normalization method is used to linearly map each physical quantity on the grid node to the target interval, and the missing values in the three-dimensional spatial grid are cleaned by background value filling or masking mechanism, and the normalized geophysical data and normalized geological data in tensor format are output.
[0015] Furthermore, S2 specifically includes:
[0016] S201. Remove the one-dimensional well logging petrophysical curves from the standardized geological data, extract discrete geological structural layers and fault polygon surfaces as the discrete geological structural interfaces, and use surface reconstruction algorithms to reconstruct the discrete geological structural interfaces into closed or semi-closed triangular mesh geometric models.
[0017] S202. Traverse the three-dimensional spatial grid, calculate the absolute shortest distance from each spatial node to the triangular mesh geometric model, and assign positive or negative signs to the absolute shortest distance according to the preset geological topological relationship rules to construct a signed distance function.
[0018] S203. Based on the symbolic distance function, calculate the symbolic distance value of all spatial nodes in the three-dimensional spatial grid, and generate the set of symbolic distance values as a three-dimensional tensor, which is then output as the continuous three-dimensional spatial scalar field.
[0019] Furthermore, S202 specifically includes:
[0020] For any spatial node P(x,y,z) in the three-dimensional spatial mesh, calculate its Euclidean distance to all points on the surface of the triangular mesh geometric model, and take the minimum value as the absolute shortest distance d(x,y,z). The calculation formula is as follows:
[0021]
[0022] Where Q is the discrete geological structural interface. Any point on the corresponding triangular mesh geometric model, Represents the L2 norm;
[0023] By setting the preset geological topological relationship rules, the absolute shortest distance of the spatial node P(x,y,z) is assigned a positive or negative sign, resulting in the signed distance function SDF(x,y,z), whose formula is:
[0024]
[0025] Specifically, for the discrete geological structural strata, the space where the overlying strata are located is defined as the positive region, and the space where the underlying strata are located is defined as the negative region; for the fault polygonal surface, the hanging wall is defined as the positive region, and the footwall as the negative region.
[0026] Furthermore, S3 specifically includes:
[0027] S301. Construct the deep learning network architecture for the data-driven branch; wherein the deep learning network architecture includes at least an encoder and a decoder;
[0028] S302. In the physical constraint branch, based on the physical modes of the standardized geophysical data, the corresponding geophysical forward partial differential equation is dynamically selected and embedded to construct a physical differential calculation graph that does not contain trainable network weight parameters; wherein, if the standardized geophysical data is a three-dimensional seismic data volume characterizing kinematic characteristics, then a three-dimensional acoustic wave equation or an elastic wave equation is embedded in the physical constraint branch; if the standardized geophysical data is gravity anomaly data or magnetic anomaly data characterizing potential field characteristics, then a Poisson equation is embedded in the physical constraint branch.
[0029] S303. The standardized geophysical data is transmitted as an input tensor to the data-driven branch. The input tensor undergoes spatial feature extraction and dimensionality reduction through multiple layers of three-dimensional convolution kernels in the encoder to obtain a deep high-dimensional latent tensor containing multi-scale local structural features. The deep high-dimensional latent tensor enters the decoder, and the spatial resolution is gradually restored through three-dimensional deconvolution operations or three-dimensional upsampling layers. It is then spliced with the shallow feature map of the corresponding scale in the encoder through skip connections. Channel mapping is performed through the output layer at the end of the data-driven branch, and the numerical range of the output tensor is mapped to the real physical dimension range of the target physical parameters using a linear scaling function to generate the initial three-dimensional physical property field.
[0030] Furthermore, S4 specifically includes:
[0031] S401. Using numerical forward modeling operators, perform forward modeling calculations on the initial three-dimensional physical property field to obtain predicted geophysical response data, and calculate the difference between the geophysical response data and the standardized geophysical data to obtain the data fitting error term. The calculation formula is:
[0032]
[0033] Where m is the initial three-dimensional physical property field. For the numerical forward modeling operator, For geophysical response data, For the standardized geophysical data, Represents the L2 norm;
[0034] S402. Substitute the initial three-dimensional physical property field into the geophysical forward partial differential equation in the physical constraint branch, calculate the physical residuals that do not conform to objective physical laws, and obtain the physical constraint residual term. ;
[0035] S403. Calculate the spatial attribute gradient of the initial three-dimensional physical attribute field and the spatial geometric gradient of the continuous three-dimensional spatial scalar field, respectively, and use the cross gradient operator to make the spatial attribute gradient and the spatial geometric gradient tend to be collinear, so as to construct the geological structure constraint error term. The calculation formula is:
[0036]
[0037] in, The spatial property gradient of the initial three-dimensional physical property field m. For the continuous three-dimensional spatial scalar field, Let × be the spatial geometric gradient of the continuous three-dimensional scalar field, and let × denote the vector cross product operator.
[0038] S404. By assigning weight coefficients, the data fitting error term, physical constraint residual term, and geological structure constraint error term are weighted and fused to construct the joint loss function that guides the parameter update of the dual-branch physical information multimodal neural network. The calculation formula is:
[0039]
[0040] in, These are the dynamic weight coefficients corresponding to each error term; the dynamic weight coefficients are adaptively and dynamically adjusted based on the number of iterations in network training or the backpropagation variance of the gradient of each error term.
[0041] Furthermore, S5 specifically includes:
[0042] S501. Set a gradient descent optimizer for the dual-branch physical information multimodal neural network, calculate the gradient of the joint loss function using the backpropagation algorithm, and update the trainable network parameters layer by layer. The formula is as follows:
[0043]
[0044] in, and Let represent the trainable network parameters at the t-th and t+1-th iterations, respectively. The preset learning rate, Let the joint loss function be... Let be the gradient vector of the joint loss function with respect to the trainable network parameters;
[0045] S502. Perform multiple rounds of iterative training, monitor the changing trend of the joint loss function in real time, and stop training and solidify the network parameters when a preset convergence condition is met; wherein, the preset convergence condition includes at least one of the following: the number of iterations reaches a preset maximum iteration threshold; the current value of the joint loss function is lower than a preset tolerance lower limit; in a preset number of consecutive iterations, the absolute value of the rate of change of the joint loss function is less than a preset early stopping threshold;
[0046] S503. Using the standardized geophysical data as input, forward inference is performed through the converged dual-branch physical information multimodal neural network to extract the three-dimensional physical attribute field output by the data-driven branch, and de-standardization is performed on it to obtain the target three-dimensional physical attribute model.
[0047] S504. The target three-dimensional physical property model is stored and three-dimensionally visualized according to the standard geological data format for application in the geological interpretation and underground resource exploration of the target exploration area.
[0048] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.
[0049] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.
[0050] The beneficial effects of this invention are as follows:
[0051] This invention constructs a dual-branch physical information multimodal neural network incorporating data-driven and physical constraints, and utilizes a cross-gradient operator and a signed distance function to build a multi-joint loss function that integrates data fitting errors, physical constraint residuals, and geological structural constraints. This ensures that the final output three-dimensional physical property field strictly satisfies multiple constraints in space, including observational data fitting, objective physical laws, and implicit geological interfaces. This not only achieves extremely high spatial resolution but also eliminates the multi-solution problem of conventional models, such as the smoothing effect at structural boundaries like faults. It provides strong industrial-grade model support for accurately identifying lithological anomalies, delineating the spatial distribution of oil and gas reservoirs, and providing high-precision target area location data for solid minerals. Attached Figure Description
[0052] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.
[0053] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.
[0054] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.
[0055] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0056] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0057] like Figure 1 As shown, this invention provides a geological-geophysical modeling method, comprising:
[0058] S1. Acquire multi-source geophysical observation data and geological prior interpretation data of the target exploration area, and map them to a pre-constructed three-dimensional spatial grid covering the target exploration area for alignment and standardization processing to obtain standardized geophysical data and standardized geological data respectively.
[0059] In a specific embodiment of the present invention, step S1 specifically includes the following sub-steps:
[0060] S101. Obtain multi-source heterogeneous raw data of the target exploration area.
[0061] Specifically, multi-source geophysical observation data of the target exploration area are collected. This multi-source geophysical observation data includes at least a three-dimensional seismic data volume, and preferably, one or more of gravity anomaly data, magnetic anomaly data, or electromagnetic exploration data. Simultaneously, geological prior interpretation data of the target exploration area are collected. This geological prior interpretation data includes well logging petrophysical curves (such as sonic transit time and density logging data), manually or semi-automatically interpreted discrete geological structural horizons, and fault polygon surfaces.
[0062] S102. Construct a unified three-dimensional spatial grid covering the target exploration area.
[0063] Because geophysical observation data and geological prior data typically have different spatial resolutions and coordinate systems, a unified discretized grid space needs to be established. Based on the spatial extent of the target exploration area and the required modeling accuracy, the boundary coordinates of the 3D grid are set. And the spatial step size in three mutually perpendicular directions. The entire target exploration area is divided into sections consisting of... A three-dimensional spatial grid model composed of grid nodes.
[0064] S103. Perform spatial mapping and alignment processing on multi-source data.
[0065] The heterogeneous data obtained in step S101 is interpolated or resampled onto each node of the three-dimensional spatial mesh constructed in step S102, thereby achieving spatial physical location alignment of multi-source data. Specifically,
[0066] For geophysical observation data, the 3D seismic data volume is mapped to a 3D spatial grid using resampling technology; if it contains 2D gravity and magnetic surface data, the dimensions are expanded along the depth direction or mapped as boundary constraints of the top layer of the 3D grid.
[0067] For geological prior interpretation data, three-dimensional spatial interpolation algorithms (such as Kriging interpolation or inverse distance weighted interpolation) are used to project and assign one-dimensional well logging petrophysical curves, scattered discrete geological structural horizons, and fault polygons to the nearest three-dimensional spatial grid nodes.
[0068] S104. Clean and standardize the mapped data.
[0069] Because data from different modalities have different physical dimensions and numerical magnitudes (for example, the magnitude of earthquake amplitude differs greatly from that of gravity anomalies), directly inputting them into a neural network can lead to gradient explosion or network convergence difficulties. Therefore, after grid alignment, the data at each grid node needs to be dimensionless.
[0070] In a preferred embodiment, the Min-Max Normalization method is used to linearly map each physical quantity to the interval [-1,1] or [0,1], and the transformation formula is as follows:
[0071]
[0072] Where x is the original data value on the grid node, and These are the minimum and maximum values of this type of data within the entire three-dimensional space grid, respectively. These are the standardized data values. Furthermore, for missing values in the grid due to a lack of observations, background value filling or masking mechanisms are used for cleaning. The final output is standardized geophysical and geological data in tensor format suitable for direct reading by neural networks.
[0073] S2. Extract discrete geological structural interfaces from the standardized geological data, and use the symbolic distance function to convert the discrete geological structural interfaces into a continuous three-dimensional spatial scalar field.
[0074] In a specific embodiment of the present invention, step S2 specifically includes the following sub-steps:
[0075] S201. Extract discrete geological structural interface data and construct a triangular mesh geometric model of the geological structural interface.
[0076] Since the mathematical properties of the Signed Distance Function (SDF) strictly depend on the representation of spatial geometric boundaries, targeted extraction is performed from the standardized geological data output in step S1. Specifically, point or line data such as one-dimensional well logging petrophysical curves are removed, and discrete geological structural horizons and fault polygons with two-dimensional manifold characteristics and geometric topological relationships are extracted, collectively referred to as the discrete geological structural interfaces.
[0077] Since the extracted discrete geological structure interfaces are usually represented as irregular scattered point clouds or sparse polygon sets, in order to facilitate subsequent accurate distance calculations, surface reconstruction algorithms such as Delaunay triangulation are used to reconstruct the discrete geological structure interfaces into closed or semi-closed triangular mesh geometric models.
[0078] S202. Calculate the absolute shortest distance between nodes in the spatial grid and construct the signed distance function.
[0079] Traverse the three-dimensional spatial mesh covering the target exploration area constructed in step S1. For any spatial node P(x,y,z) in the mesh, calculate its Euclidean distance to all points on the surface of the triangular mesh geometric model constructed in step S202, and take the minimum value as the absolute shortest distance d(x,y,z). The formula is:
[0080]
[0081] Where Q is the discrete geological structural interface. any point on, This represents the L2 norm (i.e., Euclidean distance).
[0082] Since the absolute shortest distance d(x,y,z) only contains distance information and cannot express the spatial relative positional relationship of geological bodies, a sign function is needed to represent the inside-outside or above-below relationship of spatial nodes relative to geological interfaces. Specifically, geological topological determination rules are set; for example, for stratigraphic interfaces, the space containing the overlying strata is defined as the positive zone, and the space containing the underlying strata is defined as the negative zone; for fault planes, the hanging wall is defined as the positive zone, and the footwall as the negative zone. Based on these rules, each spatial node P(x,y,z) is assigned a positive or negative sign, thus obtaining the sign distance function SDF(x,y,z):
[0083]
[0084] S203. The SDF values of all the three-dimensional spatial grid nodes calculated above are aggregated into a three-dimensional matrix or tensor. This tensor is continuous and almost everywhere differentiable throughout the entire target exploration area, where the zero-value isosurface precisely corresponds to the physical location of the original discrete geological structural interface. This three-dimensional tensor is output as the continuous three-dimensional spatial scalar field, serving as an implicit geological representation and providing prior constraints on geological structures with geometric gradients for subsequent training of the neural network.
[0085] S3. Construct a two-branch physical information multimodal neural network including a data-driven branch and a physical constraint branch. Input the standardized geophysical data into the data-driven branch for feature extraction to obtain an initial three-dimensional physical property field. The physical constraint branch is equipped with geophysical forward partial differential equations.
[0086] In a specific embodiment of the present invention, step S3 specifically includes the following sub-steps:
[0087] S301. Construct a deep learning network architecture with a data-driven branch.
[0088] The data-driven branch establishes a nonlinear end-to-end mapping from multi-source observation data to underground rock physical properties, employing a deep neural network model suitable for 3D spatial feature extraction as its main architecture, preferably a 3D U-Net or a 3D Vision Transformer. The data-driven branch includes an encoder and a decoder; the encoder contains multiple layers of 3D convolutional kernels and downsampling layers, used to extract multi-scale spatial features from the standardized geophysical data under different receptive fields; the decoder contains upsampling layers and skip connections, used to restore spatial resolution and map the extracted high-dimensional features to an attribute space completely consistent with the dimensions of the 3D spatial grid.
[0089] S302. Embed the corresponding geophysical forward partial differential equations in the physical constraint branch according to the data modes.
[0090] A physical constraint branch is constructed parallel to the data-driven branch. This branch does not contain trainable network weight parameters, but instead constructs a physical differential computation graph based on automatic differentiation techniques. According to the specific physical modes of the standardized geophysical data obtained in step S1, the corresponding geophysical forward partial differential equations are dynamically selected and embedded. Specifically,
[0091] If the input is a 3D seismic data volume characterizing kinematic features, a 3D acoustic wave equation or elastic wave equation is embedded to describe the propagation law of the wave field in a non-homogeneous medium. If the input is gravity or magnetic anomaly data characterizing potential field features, a Poisson equation is embedded to describe the relationship between density or magnetic susceptibility and potential field distribution. This physical constraint branch receives the attribute field output by the data-driven branch and calculates physical residuals that do not conform to objective physical laws in subsequent training.
[0092] S303. Perform forward propagation through hierarchical tensor operations to obtain the initial three-dimensional physical property field.
[0093] The standardized geophysical data in tensor format obtained in step S1 is used as the input tensor X and transmitted to the data-driven branch for forward propagation computation. The specific computation process is as follows:
[0094] First, the input tensor X is processed by multiple layers of 3D convolutional kernels in the encoder for spatial feature extraction. Dimensionality reduction is then achieved using nonlinear activation functions such as modified linear units and max-pooling layers, extracting deep, high-dimensional latent tensors containing local structural features at different spatial scales. Subsequently, these deep, high-dimensional latent tensors are fed into the decoder, where spatial resolution is gradually restored through 3D deconvolution or 3D upsampling layers. They are then concatenated with shallow feature maps of corresponding scales in the encoder via skip connections to preserve high-frequency edge details. Finally, channel mapping is performed through fully connected layers or 1×1×1 3D convolutional layers at the network's end, and a linear scaling function maps the numerical range of the output tensor to the true physical dimension range of the target physical parameters (e.g., mapping the values to the rock P-wave velocity range of 1500 m / s to 6500 m / s). The resulting 3D data tensor is the initial 3D physical property field, and the spatial dimension of this property field is... It is completely consistent with the three-dimensional space mesh in step S1.
[0095] S4. Construct a joint loss function for network optimization that includes a data fitting error term, a physical constraint residual term, and a geological structure constraint error term; wherein the data fitting error term is constructed based on the initial three-dimensional physical property field and standardized geophysical data, the physical constraint residual term is constructed based on the initial three-dimensional physical property field and geophysical forward partial differential equations, and the geological structure constraint error term is constructed based on the initial three-dimensional physical property field and a three-dimensional spatial scalar field.
[0096] In a specific embodiment of the present invention, step S4 specifically includes the following sub-steps:
[0097] S401. Construct a data fitting error term based on the observation space. The data fitting error term... This is used to measure the difference between the forward response of the physical properties predicted by the neural network and the actual observed data. The initial three-dimensional physical property field m, output from step S3, is extracted and processed using built-in conventional numerical forward modeling operators. The predicted geophysical response data is then calculated and compared with the standardized geophysical data output in step S1. For comparison, the mean squared error (MSE) or L2 norm is used to construct the data fitting error term, and its formula is as follows:
[0098]
[0099] in, Forward modeling data based on the initial three-dimensional physical property field, This represents the L2 norm.
[0100] S402. Construct physical constraint residuals based on partial differential equations.
[0101] The physical constraint residual term This is used to force the output of the neural network to strictly adhere to objective physical laws even without labeled data. The initial three-dimensional physical property field *m* is substituted into the physical constraint branch constructed in step S302, and its physical residual under the control of the geophysical forward partial differential equation is calculated. Using the Poisson equation... For example (where) (where u represents the potential field and f is the source term, and u represents the conductivity / density properties). The physical constraint residual term is defined as the L2 norm of the difference between the left and right sides of the equation.
[0102]
[0103] If it is a wave equation, the corresponding residual operator contains time and space second derivatives. The introduction of this term constitutes the core mechanism of the Physical Information Neural Network (PINN).
[0104] S403. Construct geological structural constraint error terms based on implicit scalar fields.
[0105] The geological structure constraint error term This is used to ensure a high degree of agreement between the physical property boundaries obtained from the inversion and the prior geological structural boundaries (such as faults and horizons). The continuous three-dimensional spatial scalar field output in step S2 is extracted. (i.e., SDF tensor), calculate its spatial geometric gradient. Simultaneously, the spatial property gradient of the initial three-dimensional physical property field m is calculated. To ensure that the boundaries of physical properties are strictly parallel to the geological interface, a cross-gradient operator is introduced to construct the geological structural constraint error term:
[0106]
[0107] Where × represents the cross product of vectors; when the cross gradient approaches 0, it means that the direction of change of the physical property field in space (gradient) is completely collinear with the normal of the geological interface (gradient of SDF), thereby realizing the strong constraint of the geological a priori structure on geophysical inversion and effectively eliminating the ambiguity of the smoothing effect at faults.
[0108] S404, construct the joint loss function by weighted fusion.
[0109] The data fitting error term, physical constraint residual term, and geological structure constraint error term obtained above are linearly weighted and fused to construct the joint loss function used to guide the parameter update of the entire dual-branch network. :
[0110]
[0111] in, These are the dynamic weighting coefficients for the corresponding error terms.
[0112] In a preferred embodiment, the dynamic weight coefficients can be adaptively and dynamically adjusted according to the number of iterations of network training or the backpropagation variance of the gradients of each error term, so as to avoid the phenomenon that a certain gradient dominates during the multi-objective optimization process, causing the network to get stuck in a local optimum.
[0113] S5. The network parameters of the dual-branch physical information multimodal neural network are iteratively backpropagated and updated using the joint loss function until the preset convergence condition is met, and the three-dimensional physical attribute field output by the converged dual-branch physical information multimodal neural network is used as the target three-dimensional physical attribute model of the target exploration area.
[0114] In a specific embodiment of the present invention, step S5 specifically includes the following sub-steps:
[0115] S501. Extract the joint loss function constructed in step S4. A gradient descent optimizer is set for the dual-branch physical information multimodal neural network.
[0116] In a preferred embodiment, an Adaptive Moment Estimation (Adam) optimizer or a Finite-Memory Quasi-Newton (L-BFGS) optimizer is employed. The joint loss function is calculated using the backpropagation algorithm. For all trainable network parameters (i.e., weights and biases, denoted as) in the data-driven branch The partial derivatives (gradients) of ) are calculated according to the set learning rate. The network parameters are updated layer by layer. The basic formula for parameter update is:
[0117]
[0118] in, and These represent the network parameters at the t-th and t+1-th iterations, respectively. This is the gradient vector of the joint loss function with respect to the network parameters.
[0119] S502. Under a unified computing framework (such as a GPU-accelerated deep learning platform), perform multiple rounds of global iterative training. During each iteration, monitor the joint loss function in real time. The decreasing trend of each of its sub-items (data fitting error term, physical constraint residual term, and geological structure constraint error term) is analyzed, and it is determined whether a preset convergence condition is met. The preset convergence condition includes at least one of the following:
[0120] (1) The number of iterations for network training has reached the preset maximum iteration threshold. ;
[0121] (2) The joint loss function The current value is lower than the preset tolerance lower limit. ;
[0122] (3) In K consecutive iterations, the absolute value of the rate of change of the joint loss function is less than the preset early stopping threshold. .
[0123] When any of the above convergence conditions are met, iterative training will automatically stop and the current network parameters will be fixed.
[0124] S503. After the network stops training and the parameters are fixed, the dual-branch physical information multimodal neural network is determined to be in a convergent state. The standardized geophysical data obtained in step S1 is used as input, and the converged network performs a final forward inference to directly extract the final three-dimensional physical property field generated by the data-driven branch output layer. This property field spatially strictly satisfies multiple constraints including observation data fitting, objective physical laws, and implicit geological interfaces. It is then de-standardized to restore its actual physical dimensions, serving as the target three-dimensional physical property model for the target exploration area.
[0125] S504. The target three-dimensional physical property model is stored and output as a three-dimensional visualization according to a standard geological data format (e.g., SEGY format or standard three-dimensional mesh attribute volume format). This model has extremely high spatial resolution and eliminates the smoothing effect of fault boundaries. It can be directly applied to the subsequent geological interpretation of the target exploration area, including but not limited to identifying lithological anomalies, delineating the spatial distribution range of oil and gas reservoirs, or providing high-precision target area positioning basis for solid mineral exploration, thereby fully realizing the industrial application value of this invention.
[0126] like Figure 2 As shown, the present invention also provides a geological-geophysical modeling system, based on the geological-geophysical modeling method described above, the system comprising:
[0127] The acquisition module 1 is used to acquire multi-source geophysical observation data and geological prior interpretation data of the target exploration area, and map them to a pre-constructed three-dimensional spatial grid covering the target exploration area for alignment and standardization processing, so as to obtain standardized geophysical data and standardized geological data respectively.
[0128] The conversion module 2 is used to extract discrete geological structural interfaces from the standardized geological data, and to convert the discrete geological structural interfaces into a continuous three-dimensional spatial scalar field using a symbolic distance function.
[0129] The first construction module 3 is used to construct a two-branch physical information multimodal neural network including a data-driven branch and a physical constraint branch. The standardized geophysical data is input into the data-driven branch for feature extraction to obtain an initial three-dimensional physical property field. The physical constraint branch is equipped with geophysical forward partial differential equations.
[0130] The second construction module 4 is used to construct a joint loss function for network optimization, which includes a data fitting error term, a physical constraint residual term, and a geological structure constraint error term; wherein, the data fitting error term is constructed based on the initial three-dimensional physical property field and standardized geophysical data, the physical constraint residual term is constructed based on the initial three-dimensional physical property field and geophysical forward partial differential equations, and the geological structure constraint error term is constructed based on the initial three-dimensional physical property field and a three-dimensional spatial scalar field;
[0131] Output module 5 is used to iteratively backpropagate and update the network parameters of the dual-branch physical information multimodal neural network using the joint loss function until the preset convergence condition is met, and to use the three-dimensional physical attribute field output by the converged dual-branch physical information multimodal neural network as the target three-dimensional physical attribute model of the target exploration area.
[0132] Each of the above modules is used to perform the corresponding steps in the above geological-geophysical modeling method. The specific implementation method is as described in the above method embodiment, and will not be repeated here.
[0133] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all data required for the geological-geophysical modeling process. The network interface allows communication with external terminals via a network connection. The computer program is executed by the processor to implement the geological-geophysical modeling method.
[0134] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.
[0135] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described geological-geophysical modeling methods.
[0136] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).
[0137] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0138] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A geological-geophysical modeling method, characterized in that, include: S1. Acquire multi-source geophysical observation data and geological prior interpretation data of the target exploration area, and map them to a pre-constructed three-dimensional spatial grid covering the target exploration area for alignment and standardization processing to obtain standardized geophysical data and standardized geological data respectively. S2. Extract discrete geological structural interfaces from the standardized geological data, and use the symbolic distance function to convert the discrete geological structural interfaces into a continuous three-dimensional spatial scalar field. S3. Construct a two-branch physical information multimodal neural network including a data-driven branch and a physical constraint branch. Input the standardized geophysical data into the data-driven branch for feature extraction to obtain an initial three-dimensional physical property field. The physical constraint branch is equipped with geophysical forward partial differential equations. S4. Construct a joint loss function for network optimization that includes a data fitting error term, a physical constraint residual term, and a geological structure constraint error term; wherein, the data fitting error term is constructed based on the initial three-dimensional physical property field and standardized geophysical data, the physical constraint residual term is constructed based on the initial three-dimensional physical property field and geophysical forward partial differential equations, and the geological structure constraint error term is constructed based on the initial three-dimensional physical property field and a three-dimensional spatial scalar field; S5. The network parameters of the dual-branch physical information multimodal neural network are iteratively backpropagated and updated using the joint loss function until the preset convergence condition is met, and the three-dimensional physical attribute field output by the converged dual-branch physical information multimodal neural network is used as the target three-dimensional physical attribute model of the target exploration area.
2. The geological-geophysical modeling method according to claim 1, characterized in that, S1 specifically includes: S101. Acquire multi-source geophysical observation data of the target exploration area and collect geological prior interpretation data of the target exploration area; wherein, the multi-source geophysical observation data includes at least three-dimensional seismic data volume, and the geological prior interpretation data includes well logging petrophysical curves, manually or semi-automatically interpreted discrete geological structural horizons, and fault polygon surfaces. S102. Based on the spatial range of the target exploration area and the preset modeling accuracy, construct a unified three-dimensional spatial grid covering the target exploration area; S103. The three-dimensional seismic data volume is mapped to the three-dimensional spatial grid by a resampling method, and the well logging rock physics curves, discrete geological structural layers and fault polygon surfaces are mapped and assigned to the nearest three-dimensional spatial grid nodes by a three-dimensional spatial interpolation algorithm. S104. The maximum-minimum normalization method is used to linearly map each physical quantity on the grid node to the target interval, and the missing values in the three-dimensional spatial grid are cleaned by background value filling or masking mechanism, and the normalized geophysical data and normalized geological data in tensor format are output.
3. The geological-geophysical modeling method according to claim 2, characterized in that, S2 specifically includes: S201. Remove the one-dimensional well logging petrophysical curves from the standardized geological data, extract discrete geological structural layers and fault polygon surfaces as the discrete geological structural interfaces, and use surface reconstruction algorithms to reconstruct the discrete geological structural interfaces into closed or semi-closed triangular mesh geometric models. S202. Traverse the three-dimensional spatial grid, calculate the absolute shortest distance from each spatial node to the triangular mesh geometric model, and assign positive or negative signs to the absolute shortest distance according to the preset geological topological relationship rules to construct a signed distance function. S203. Based on the symbolic distance function, calculate the symbolic distance value of all spatial nodes in the three-dimensional spatial grid, and generate the set of symbolic distance values as a three-dimensional tensor, which is then output as the continuous three-dimensional spatial scalar field.
4. The geological-geophysical modeling method according to claim 3, characterized in that, S202 specifically includes: For any spatial node P(x,y,z) in the three-dimensional spatial mesh, calculate its Euclidean distance to all points on the surface of the triangular mesh geometric model, and take the minimum value as the absolute shortest distance d(x,y,z). The calculation formula is as follows: Where Q is the discrete geological structural interface. Any point on the corresponding triangular mesh geometric model, Represents the L2 norm; By setting the preset geological topological relationship rules, the absolute shortest distance of the spatial node P(x,y,z) is assigned a positive or negative sign, resulting in the signed distance function SDF(x,y,z), whose formula is: Specifically, for the discrete geological structural strata, the space where the overlying strata are located is defined as the positive region, and the space where the underlying strata are located is defined as the negative region; for the fault polygonal surface, the hanging wall is defined as the positive region, and the footwall as the negative region.
5. The geological-geophysical modeling method according to claim 1, characterized in that, S3 specifically includes: S301. Construct the deep learning network architecture for the data-driven branch; wherein the deep learning network architecture includes at least an encoder and a decoder; S302. In the physical constraint branch, based on the physical modes of the standardized geophysical data, the corresponding geophysical forward partial differential equation is dynamically selected and embedded to construct a physical differential calculation graph that does not contain trainable network weight parameters; wherein, if the standardized geophysical data is a three-dimensional seismic data volume characterizing kinematic characteristics, then a three-dimensional acoustic wave equation or an elastic wave equation is embedded in the physical constraint branch; if the standardized geophysical data is gravity anomaly data or magnetic anomaly data characterizing potential field characteristics, then a Poisson equation is embedded in the physical constraint branch. S303. The standardized geophysical data is transmitted as an input tensor to the data-driven branch. The input tensor undergoes spatial feature extraction and dimensionality reduction through multiple layers of three-dimensional convolution kernels in the encoder to obtain a deep high-dimensional latent tensor containing multi-scale local structural features. The deep high-dimensional latent tensor enters the decoder, and the spatial resolution is gradually restored through three-dimensional deconvolution operations or three-dimensional upsampling layers. It is then spliced with the shallow feature map of the corresponding scale in the encoder through skip connections. Channel mapping is performed through the output layer at the end of the data-driven branch, and the numerical range of the output tensor is mapped to the real physical dimension range of the target physical parameters using a linear scaling function to generate the initial three-dimensional physical property field.
6. The geological-geophysical modeling method according to claim 1, characterized in that, S4 specifically includes: S401. Using numerical forward modeling operators, perform forward modeling calculations on the initial three-dimensional physical property field to obtain predicted geophysical response data, and calculate the difference between the geophysical response data and the standardized geophysical data to obtain the data fitting error term. The calculation formula is: Where m is the initial three-dimensional physical property field. For the numerical forward modeling operator, For geophysical response data, For the standardized geophysical data, Represents the L2 norm; S402. Substitute the initial three-dimensional physical property field into the geophysical forward partial differential equation in the physical constraint branch, calculate the physical residuals that do not conform to objective physical laws, and obtain the physical constraint residual term. ; S403. Calculate the spatial attribute gradient of the initial three-dimensional physical attribute field and the spatial geometric gradient of the continuous three-dimensional spatial scalar field, respectively, and use the cross gradient operator to make the spatial attribute gradient and the spatial geometric gradient tend to be collinear, so as to construct the geological structure constraint error term. The calculation formula is: in, The spatial property gradient of the initial three-dimensional physical property field m. For the continuous three-dimensional spatial scalar field, Let × be the spatial geometric gradient of the continuous three-dimensional scalar field, and let × denote the vector cross product operator. S404. By assigning weight coefficients, the data fitting error term, physical constraint residual term, and geological structure constraint error term are weighted and fused to construct the joint loss function that guides the parameter update of the dual-branch physical information multimodal neural network. The calculation formula is: in, These are the dynamic weight coefficients corresponding to each error term; the dynamic weight coefficients are adaptively and dynamically adjusted based on the number of iterations in network training or the backpropagation variance of the gradient of each error term.
7. The geological-geophysical modeling method according to claim 1, characterized in that, S5 specifically includes: S501. Set a gradient descent optimizer for the dual-branch physical information multimodal neural network, calculate the gradient of the joint loss function using the backpropagation algorithm, and update the trainable network parameters layer by layer. The formula is as follows: in, and Let represent the trainable network parameters at the t-th and t+1-th iterations, respectively. The preset learning rate, Let the joint loss function be... Let be the gradient vector of the joint loss function with respect to the trainable network parameters; S502. Perform multiple rounds of iterative training, monitor the changing trend of the joint loss function in real time, and stop training and solidify the network parameters when a preset convergence condition is met; wherein, the preset convergence condition includes at least one of the following: the number of iterations reaches a preset maximum iteration threshold; the current value of the joint loss function is lower than a preset tolerance lower limit; in a preset number of consecutive iterations, the absolute value of the rate of change of the joint loss function is less than a preset early stopping threshold; S503. Using the standardized geophysical data as input, forward inference is performed through the converged dual-branch physical information multimodal neural network to extract the three-dimensional physical attribute field output by the data-driven branch, and de-standardization is performed on it to obtain the target three-dimensional physical attribute model. S504. The target three-dimensional physical property model is stored and three-dimensionally visualized according to the standard geological data format for application in the geological interpretation and underground resource exploration of the target exploration area.
8. A geological-geophysical modeling system, based on the geological-geophysical modeling method according to any one of claims 1 to 7, characterized in that, The system includes: The acquisition module is used to acquire multi-source geophysical observation data and geological prior interpretation data of the target exploration area, and map them to a pre-constructed three-dimensional spatial grid covering the target exploration area for alignment and standardization processing, so as to obtain standardized geophysical data and standardized geological data respectively. The conversion module is used to extract discrete geological structural interfaces from the standardized geological data, and to convert the discrete geological structural interfaces into a continuous three-dimensional spatial scalar field using a signed distance function. The first construction module is used to construct a two-branch physical information multimodal neural network including a data-driven branch and a physical constraint branch. The standardized geophysical data is input into the data-driven branch for feature extraction to obtain an initial three-dimensional physical property field. The physical constraint branch is equipped with geophysical forward partial differential equations. The second construction module is used to construct a joint loss function for network optimization, which includes a data fitting error term, a physical constraint residual term, and a geological structure constraint error term. The data fitting error term is constructed based on the initial three-dimensional physical property field and standardized geophysical data, the physical constraint residual term is constructed based on the initial three-dimensional physical property field and geophysical forward partial differential equations, and the geological structure constraint error term is constructed based on the initial three-dimensional physical property field and a three-dimensional spatial scalar field. The output module is used to iteratively backpropagate and update the network parameters of the dual-branch physical information multimodal neural network using the joint loss function until the preset convergence condition is met, and to use the three-dimensional physical attribute field output by the converged dual-branch physical information multimodal neural network as the target three-dimensional physical attribute model of the target exploration area.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.