Three-dimensional geological model construction method based on potential field interpolation and deep learning coupling

Abstract

The invention provides a three-dimensional geologic model construction method based on potential field interpolation and deep learning coupling. The method comprises the following steps: S1, constructing a sample data set and carrying out normalization processing; s2, according to the characteristics of the sample data, determining a stratum contact relation in a region by using a variation coefficient between upper and lower stratums; s3, designing a loss function based on a scalar field construction rule and an adaptive stratigraphic sequence; s4, designing a three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture, wherein the three-dimensional geological prediction model comprises a concise full-connection module, a complex convolutional residual module and a hierarchical integration module; s5, fitting a pre-training plane according to sample data features, and capturing stratum features in the region to obtain a similar ideal stratum plane; and S6, predicting a three-dimensional geologic model, loading the pre-trained network parameters to each layer of convolutional residual implicit network for training, enabling the model to converge and output a predicted scalar field value, and then performing contour surface extraction to obtain the three-dimensional geologic model.

Description

Technical Field

[0001] The present invention belongs to the field of surveying and mapping science and technology, and in particular to a three-dimensional geological model construction method based on potential field interpolation and deep learning coupling. Background Art

[0002] Three-dimensional geological modeling is a method of representing underground geometry based on observable geological data. Common data sources include drilling data, outcrop observations, occurrence observations, stratigraphic unit sampling, stratigraphic contact interfaces, etc. As humans gradually deepen their development of the surface, the solution of underground problems has also attracted increasing attention. However, underground space is often invisible, and the underground structure is complex and difficult to detect. Therefore, three-dimensional geological modeling is carried out based on effective geological data. The resulting three-dimensional geological model is an important basis for human resources estimation, urban planning, disaster prediction and other applications in the earth's interior.

[0003] At present, geological modeling has a perfect work plan and method system in different geological structures and application scenarios. Three-dimensional geological modeling methods are mainly divided into explicit modeling, implicit modeling, and related deep learning modeling algorithms that have begun to attract attention in the field of geological modeling with the rapid development of machine learning methods. Traditional explicit modeling methods are based on relevant knowledge of geology, geophysics and other disciplines to collect, process, analyze and interpret geological data, and construct geological interfaces by interpolating relevant geological data. However, in the case of insufficient modeling experience and incomplete geological knowledge, the geometric quality of the established model is difficult to meet the standards, which takes a lot of time and is not easy to update and reproduce. The implicit modeling method uses isosurfaces in a three-dimensional scalar field to represent geological interfaces, and uses implicit functions to interpolate discrete geological information such as geological interface points and occurrence observations. Compared with explicit modeling methods, implicit modeling methods have the advantages of higher mesh quality, reproducibility, and dynamic updating. However, artifacts are prone to occur when modeling in areas with complex geological structures. To solve modeling artifacts, the spatial relationship between relevant geological history and geological characteristics can be constrained into the implicit modeling process, forming an implicit-explicit hybrid method, but reproducibility and modeling efficiency will be reduced.

[0004] The difficulty in introducing a large amount of geological knowledge and data has always been a pain point in the study of 3D geological modeling methods. With the development of deep learning methods, its diverse data input, output types and rich network structures have provided possibilities for the integration of geological knowledge and geological observation data. However, there is still a lack of detailed description of geological details and attribute information in complex geological structures. The sparse and uneven distribution of geological sampling data has led to the lack of constraints on geological evolution laws in the constructed 3D geological model, which is also an urgent problem to be solved in the study of 3D geological modeling methods. Summary of the invention

[0005] In response to existing problems, the present invention proposes a three-dimensional geological modeling method based on the coupling of potential field interpolation and deep learning, a flexible modeling framework that combines different types of geological constraints with the modeling process, and constructs a loss function based on geological prior knowledge and inputs hard constraint data, making it possible to obtain geological predictions that are consistent with geological knowledge, further solving the problem that current geological conditions and knowledge are difficult to constrain in the modeling process.

[0006] In order to achieve the above-mentioned and other related purposes, the present invention provides a method for constructing a three-dimensional geological model based on potential field interpolation and deep learning coupling, comprising the following steps:

[0007] S1 constructs a sample data set, discretizes various geological data into geological interface points and stratigraphic unit sampling point data, and obtains the sample data set after normalization processing;

[0008] S2 analyzes the contact relationship of the strata in the region. Based on the sample data set constructed in the modeling area, the upper and lower strata data are calculated to obtain their coefficient of variation. A threshold is defined to compare whether the upper and lower strata are consistent in geometry, thereby confirming the contact relationship of the strata. The information is collated, including the strata code, strata name, and contact relationship between strata (with the upper strata);

[0009] S3 designs a loss function that takes into account geological implicit constraints. It groups geological interfaces according to stratigraphic contact relationships, determines the number of scalar fields to be constructed, and the correspondence between scalar fields and geological interfaces. Then, based on the correspondence between scalar fields and geological interfaces, it constructs stratigraphic sequence relationships, and uses inequalities to express implicit constraints on geological conditions such as stratigraphic sequence relationships and occurrence. Finally, it constructs relevant loss functions based on geological implicit inequality constraints, and adds global regularized smoothing losses to constrain the training process of the network model, introducing geological prior knowledge into the modeling process.

[0010] S4 designs a three-dimensional geological prediction model based on a convolutional residual implicit hierarchical network architecture, wherein the network architecture includes a simple fully connected module, a complex convolutional residual module, and a hierarchical integration module, wherein the simple fully connected module is used to integrate and transform the extracted geological features, the complex convolutional residual module includes multiple convolutional layers and residual connections, and the hierarchical integration module integrates multiple basic convolutional residual implicit networks to achieve the prediction of multiple scalar field values;

[0011] S5 plane geometry pre-training, plane fitting is performed according to the characteristics of the formation interface point data, and a quasi-ideal formation plane containing the formation characteristics in the region is constructed from bottom to top. The HuberRegressor regression algorithm is used when constructing the quasi-ideal formation plane, and the formation dip angle and the formation up-down relationship are considered as constraints. Finally, the quasi-ideal formation plane is used for initialization training of the network model;

[0012] S6 uses the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture to predict the three-dimensional geological model, loads the pre-trained network parameters onto each layer of the convolutional residual implicit network for training, and stops training when the loss functions no longer decrease. The trained network model can then be used to predict the scalar field and extract the isosurface to obtain the corresponding three-dimensional geological model.

[0013] Preferably, in step S1, constructing the sample data set includes the following steps:

[0014] S11 discretizes various geological data, including: discretizing drilling data, geological profiles, geological map data, etc. into spatial scattered points, and classifying and coding them according to existing geological prior knowledge;

[0015] S12 constructs a geological interface point file, and selects corresponding discrete points according to the drilling layer, geological profile, etc. to construct a geological interface point file;

[0016] S13 constructs a stratigraphic unit sampling point file, and selects corresponding discrete points according to the drilling layer, geological profile, etc. to construct the stratigraphic unit sampling point file;

[0017] S14 coordinate normalization processing is used to normalize the geological interface points and stratigraphic unit sampling points to obtain the final data set for network model training.

[0018] Furthermore, in step S2, the analysis of the contact relationship of the strata in the region includes the following steps:

[0019] S21 Calculation of coefficient of variation between upper and lower strata: According to the coordinate values ​​of the upper and lower strata data, the corresponding strata spacing is obtained, and the coefficient of variation of the spacing is calculated to determine whether the geometric shapes between the upper and lower strata are similar;

[0020] S22 determines the contact relationship between the upper and lower strata according to the coefficient of variation. According to the characteristics of the modeling area, a threshold of the coefficient of variation is set. If the coefficient of variation is less than the set threshold, the contact relationship between the upper and lower strata is integrated, and if the coefficient of variation is greater than the set threshold, the contact relationship between the upper and lower strata is unconformable.

[0021] S23 outputs a stratum contact relationship information table, where the organized information includes stratum code, stratum name, and stratum contact relationship (with the upper stratum);

[0022] Furthermore, in step S3, the loss function considering the geological implicit constraints includes the following steps:

[0023] S31 determines the number of scalar fields that need to be constructed, and groups the geological interfaces that need to be modeled according to the stratigraphic contact relationship, so that the upper and lower strata that are integrated use the same scalar field for prediction, and the upper and lower strata that are unconformable use different scalar fields for prediction, thereby obtaining the number of scalar fields that need to be constructed in the modeling area and the corresponding relationship between the scalar field and the geological interface;

[0024] S32 constructs stratigraphic sequence relationships, sorting out stratigraphic sequence relationships based on the correspondence between scalar fields and geological interfaces, referred to as geological interface-geological interface sequence relationships and geological unit-geological interface sequence relationships;

[0025] S33 constructs geological implicit constraint inequalities. Geological implicit constraints include constraints related to geological interfaces and stratigraphic unit sampling points, as well as constraints on occurrence points and geological tangents. The above implicit constraints are organized into inequality expressions.

[0026] S34 constructs the loss function of the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture. The loss function includes two parts, one part is transformed from the above-mentioned geological implicit inequality, and the other part is the global loss function, which uses the Eikonal unit norm constraint, that is, the regularization loss.

[0027] Furthermore, in step S4, the construction of the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture includes the following steps:

[0028] S41 builds a concise fully connected module, which consists of a fully connected layer and a Softplus activation function;

[0029] S42 builds a complex convolutional residual module, which consists of a convolutional layer, a residual connection block, and a Softplus activation function;

[0030] S43 constructs a three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture according to the hierarchical integration module, combines the above-mentioned simple fully connected module with the complex convolutional residual module to construct a basic convolutional residual implicit network, and then integrates multiple basic convolutional residual implicit networks according to the number of scalar fields to be constructed to obtain the final prediction model.

[0031] Furthermore, in step S5, the plane geometry pre-training process includes the following steps:

[0032] S51 calculates the average dip of the entire formation, considers all the formation interface point data for plane fitting, uses the HuberRegressor regression algorithm, then evaluates the dip of all the formation fitting planes, and finally averages them to get the average dip of the entire formation;

[0033] S52 constructs a constrained fitting plane, and uses the average dip angle of the entire formation to perform HuberRegressor regression fitting on the constrained plane layer by layer from bottom to top. The constructed constrained fitting plane contains regional formation characteristics and avoids intersection to a certain extent;

[0034] S53 checks and corrects the construction of a quasi-ideal formation plane, and checks and corrects each layer from bottom to top. The constrained fitting plane after the inspection and correction is the quasi-ideal formation plane, which integrates the characteristics of the ideal formation such as relative flatness in the horizontal direction, layered distribution, and formation dip angle in the region;

[0035] S54 network model parameter geometry initialization, use the constructed ideal-like stratigraphic plane to pre-train the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture, retain the pre-trained network parameters as the starting point of the model parameters for formal training, and finally the output scalar field will have a reasonable starting geometry of the geological interface.

[0036] Furthermore, in step S6, the process of performing three-dimensional geological model prediction using the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture includes the following steps:

[0037] S61 trains a three-dimensional geological prediction model based on a convolutional residual implicit hierarchical network architecture, loads the network parameters pre-trained by geometric initialization into each layer of the convolutional residual implicit network for training, and stops training when each loss function no longer decreases, thus obtaining a trained three-dimensional geological prediction model;

[0038] S62 constructs a predicted geological grid, obtains its bounding box grid according to the stratigraphic unit interface point data, customizes the resolution in the x, y, and z directions, and obtains the predicted geological grid structure in the region and the grid nodes of the grid structure;

[0039] S63 predicts the scalar field in the region, using the trained three-dimensional geological prediction model to predict the constructed prediction geological grid, and predicts the scalar field value in the regional geological grid;

[0040] S64 constructs a three-dimensional geological model, extracts isosurfaces from the predicted scalar field grid, and interpolates the three-dimensional geological model based on the average scalar field value of each stratum obtained during the model training phase. The isosurface extraction process uses the isosurface extraction algorithm in the PyVista open source library.

[0041] The beneficial effects of the present invention are as follows:

[0042] (1) According to the stratigraphic characteristics of the modeling area, ideal stratigraphic features are constructed to participate in the pre-training process, which better considers factors such as the shape, size, and position of the geological interface, thereby improving the model accuracy;

[0043] (2) Combining advanced implicit geological modeling methods with deep learning 3D modeling methods improves modeling efficiency;

[0044] (3) In order to address the problem that geological prior knowledge is difficult to incorporate into the modeling process, geological data and conditions are converted into loss functions in deep learning and constrained in the modeling process to improve the accuracy and rationality of the model.

[0045] (4) The present invention aims to solve the problem that geological knowledge is difficult to constrain in the modeling process, resulting in a certain discrepancy between the constructed geological model and people's geological understanding. The loss function of deep learning is optimized, and geological constraints such as the stratigraphic sequence relationship and occurrence of unconformable geological structures are introduced into the loss function, so that the generated three-dimensional geological model is more in line with the actual geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A flowchart of a three-dimensional geological modeling method based on potential field interpolation and deep learning coupling of the present invention;

[0048] Figure 2 This is an example diagram of constructing the scalar field number rule in step S31 of the present invention;

[0049] Figure 3 An example diagram of constructing stratigraphic sequence relationship rules in step S32 of the present invention;

[0050] Figure 4 This is a schematic diagram of the production state constraint in step S33 of the present invention;

[0051] Figure 5 A schematic diagram of a complete network architecture of a three-dimensional geological prediction model based on a convolutional residual implicit hierarchical network architecture constructed in step S4 of the present invention;

[0052] Figure 6 This is an example diagram of constructing a quasi-ideal stratigraphic plane and a predicted three-dimensional geological model in steps S5 and S6 of the present invention. DETAILED DESCRIPTION

[0053] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The various details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0054] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0055] See also Figure 1 As shown, the present invention discloses a three-dimensional geological modeling method based on potential field interpolation and deep learning coupling, comprising the following steps:

[0056] In step S1, the input data is geological data such as borehole data and geological profile data, and the output is a three-dimensional point file of geological interface sampling points and stratigraphic unit sampling points in vtp format directly passed into the model prediction part. The attribute parameters of each point in the three-dimensional point file include its three-dimensional coordinates x, y, z and stratigraphic code level. The construction of the three-dimensional point file is as follows: for the borehole data, firstly, according to the size of the modeling area, the borehole data is reasonably screened and thinned, and the top plate depth, coordinates and geological code of each stratum of each borehole are taken to construct the geological interface sampling point vtp format file, and then each stratum of each borehole is uniformly sampled in the bottom plate and top plate depth to obtain multiple depths, which are respectively formed into the stratigraphic unit sampling point vtp format file together with the coordinates and geological codes. Similarly, the geological boundaries and stratigraphic units can also be sampled for the profile. After that, the geological interface points and stratigraphic unit sampling points are normalized and the coordinates are mapped to the interval [-1,1].

[0057] In step S2, the input data is the sample data set constructed in S1, and the output data is a stratigraphic contact relationship file in csv format that can be directly passed into the prediction part of the model. The stratigraphic contact relationship file is constructed by performing stratigraphic sorting on the three-dimensional point data in the sample data set, determining the number of stratigraphic layers to be constructed, and calculating the coefficient of variation (CV) based on the distance between the upper and lower stratigraphic layers in the z direction on the same x and y coordinates (similar structural data can be obtained after sorting geological data such as boreholes and geological profiles). The coefficient of variation is a statistic used to measure the variability or discreteness of a data set. A higher coefficient of variation indicates that the data points of the data set are relatively dispersed, while a lower coefficient of variation indicates that the data points are relatively concentrated.

[0058] Then specify the threshold of the coefficient of variation. For example, if the threshold is 20%, the distance difference between the upper and lower strata with a coefficient of variation greater than 20% is more dispersed, and the geometric shape will be more obviously different, so the stratigraphic contact relationship is unconformable. The distance difference between the upper and lower strata with a coefficient of variation less than 20% is more concentrated, and the geometric shape is similar, so the stratigraphic contact relationship is integrated. Determine the contact relationship between the corresponding strata of each layer and the strata above, extract the stratigraphic layer to be constructed, name it, obtain the stratigraphic code, and judge the contact relationship, together to form a stratigraphic contact relationship csv format file.

[0059] In step S3, the input data is the stratigraphic contact relationship file constructed in S2, and the output is the loss function of the stratigraphic condition implicit constraint and the regularization loss function of the network model and other available different types of loss functions, which can be flexibly adjusted according to the geological structure of the modeling area and the modeling task. The specific implementation content of step S3 is relatively complex, and the specific contents are as follows:

[0060] In step S31, the number of scalar fields to be constructed is determined, and the implementation rule for grouping the geological interfaces to be modeled according to the stratum contact relationship is:

[0061] Stratigraphic contact relationships include conformity and unconformity. The upper and lower strata of the conformity relationship have experienced the same geological history and show quasi-parallel geometric shapes in the nearby geological interfaces and strata. The upper and lower strata of the unconformity relationship often have erosion surfaces produced by the cessation of erosion or sedimentation, indicating the transition of strata of different ages and discontinuities. A scalar field is constructed for each different geological structure. The upper and lower strata of the conformity relationship can be regarded as associated with the same scalar field and data constraints due to their similar stratigraphic occurrences. The upper and lower strata of the unconformity relationship cannot be represented by the same scalar field due to their different stratigraphic occurrences, and different scalar fields need to be used for construction. According to different stratigraphic contact relationships, different scalar fields can also be regarded as having corresponding geological rules, such as erosion or overlay relationships. The erosion relationship indicates that the scalar field corresponding to a certain unconformity interface contacts the scalar field below. The overlay relationship exists in the scalar fields corresponding to multiple continuous conformity interfaces overlapping the scalar field below. Two scalar fields with an overlying relationship can be considered to be under the same geological domain, because in the process of geological evolution, multiple consecutive conformable strata are normally deposited above the unconformable contact and belong to the same geological age and geological characteristics.

[0062] For specific examples, see Figure 2, the upper figure shows the observation data input for 3D geological modeling, which has an intuitive display of geological interfaces and strata. The lower figure shows the stratigraphic column of the geological situation in the upper figure and the scalar field construction based on the stratigraphic contact relationship. As shown in the figure, there are six geological interfaces, including two unconformity interfaces (I2 and I5) and four conformity interfaces (I0, I1, I3 and I4). According to the stratigraphic contact relationship, four scalar fields are constructed to model them. The corresponding relationships between scalar fields, geological interfaces and stratigraphic units are: scalar field Corresponding to the construction of geological interfaces I0, I1 and stratigraphic units A, B, C; scalar field Corresponding to the construction of geological interface I2; scalar field Corresponding to the construction of geological interfaces I3, I4 and stratigraphic units D, E, F; scalar field The corresponding geological interface I5 is constructed. The geological regular relationships between scalar fields are: and There is an erosion relationship between them; scalar field and There is an overlying relationship between them; scalar field and There is an erosion relationship between them.

[0063] The implementation rules of step S32 for sorting out the stratigraphic sequence relationship based on the correspondence between the scalar field and the geological interface are as follows:

[0064] Since the scalar field is constructed based on the stratigraphic contact relationship in the unconformable strata, the stratigraphic sequence relationship takes into account the geological interface, which is divided into two types: the geological interface-geological interface sequence relationship and the stratigraphic unit-geological interface sequence relationship. Both of the sequence relationships consider the relationship above and below, while the geological interface-geological interface stratigraphic sequence relationship will judge the overlapping relationship, while the stratigraphic unit-geological interface sequence relationship will not. For specific examples, please refer to Figure 3 , the rules for constructing stratigraphic sequence relationships are as follows:

[0065] (1) For the above relationship: the existence of the above relationship is only considered within the same geological domain, because for the characteristic points of the newer geological interface that are not in the same geological domain, they can be regarded as measured at an older geological interface below its scalar field, and the scalar field and isosurface related to the older geological interface are used to determine its geometric shape. Figure 3 The unconformable geological interface I2 shown in the figure erodes the geological interface I1. Therefore, the scalar field φ that simulates I1 can be used in the eroded part. 0 The presence of a measured unconformity beneath I1 indicates an erosional part of I2.

[0066] (2) For the “below” relationship: it is considered in the entire domain and is not restricted by the conditions of the same geological domain, because the sampling points of the geological interface or stratigraphic unit must all be below a certain geological interface to be considered below the geological interface.

[0067] (3) For overlap (on) relationship: it is only considered in the geological interface-geological interface sequence relationship.

[0068] Step S33 constructs the implementation rule of the geological implicit constraint inequality as follows:

[0069] The specific contents of geological implicit constraints include constraints related to geological interfaces and stratigraphic unit sampling points, as well as constraints on occurrence points and geological tangents. The constraints related to geological interfaces include scalar field value constraints and geological interface-geological interface sequence relationship constraints. The constraints related to stratigraphic unit sampling points include stratigraphic unit-geological interface sequence relationship constraints. The specific construction rules and inequalities are described as follows:

[0070] a) Geological interface related constraints

[0071] The constraints related to geological interfaces are divided into two aspects: potential field and stratigraphic sequence relationship, namely, scalar field value constraints and geological interface-geological interface sequence relationship constraints. The specific constraint information is as follows:

[0072] (1) Scalar field value constraint

[0073] Geological Interface I K The sampling point set K, the corresponding scalar field value The variance of is approximately 0, which can ensure that the points in the point set are all on the same scalar scene, and the fitted scalar scene is closer to the actual situation of the geological interface. The relevant formula is expressed as follows:

[0074] Scalar field-valued constraint inequalities:

[0075]

[0076] in, Indicates geological interface I K The sampling point set K is on the geological interface I K The collection of scalar field values ​​within the associated scalar field.

[0077] (2) Constraints on geological interface-geological interface sequence relationships

[0078] For two geological interfaces I X and I K The stratigraphic relationship can be above, below and overlap. X The sampling point set X, any point x belongs to the point set X, the scalar field value corresponding to point x and geological interface IK The average scalar field value of The difference between The positive or negative value can determine the point x and the geological interface I K The relevant formula is as follows:

[0079] Geological Interface I K The average scalar field value of formula:

[0080]

[0081] Where |K| represents the geological interface I K The number of points in the sampling point set K, k belongs to the sampling point of the point set K, Indicates geological interface I K The average scalar field value of .

[0082] Geological interface-geological interface scalar field difference formula:

[0083]

[0084] in, Indicates that point x is on the geological interface I X The scalar field value in the corresponding scalar field.

[0085] Constrained inequality of geological interface-geological interface sequence relationship:

[0086]

[0087] in, When it is greater than 0, it indicates that point x is on the geological interface I K above; when When it is less than 0, it indicates that point x is on the geological interface I K below; when When it is equal to 0, it indicates that point x is at the geological interface I K Above, the two overlap.

[0088] b) Constraints on stratigraphic unit sampling points

[0089] The constraints of stratigraphic unit sampling points are mainly constraints on stratigraphic sequence relationships, that is, stratigraphic unit-geological interface sequence relationships. The specific constraint information is as follows:

[0090] For a stratigraphic unit U X and geological interface I K , its stratigraphic relationship can be either above or below. X The sampling point set X, any point x belongs to the point set X, the scalar field value corresponding to point x and the geological interface IK The average scalar field value of The positive or negative value of the difference between point x and geological interface I K The relevant formula is as follows:

[0091] Scalar field difference between stratigraphic unit and geological interface formula:

[0092]

[0093] in, Indicates that point x is in stratigraphic unit U X The scalar field value in the corresponding scalar field.

[0094] Stratigraphic unit-geological interface sequence relationship constraint inequality:

[0095]

[0096] in, When it is greater than 0, it indicates that point x is on the geological interface I K above; when When it is less than 0, it indicates that point x is on the geological interface I K Below.

[0097] c) Occurrence point related constraints

[0098] The occurrence describes the geometric characteristics of the stratum in space, such as the structural orientation, posture and shape, including information such as dip and inclination. For the scalar field associated with the geological interface IK There is a point set O of occurrence points on it. Any point o belongs to the point set O. The angle between the true occurrence direction vector and the scalar field gradient at point o is like Figure 4 As shown, the angle is the difference between the true occurrence direction vector and the scalar field gradient, that is, the angle constraint of the occurrence point. When the occurrence direction vector and the scalar field gradient at point o are closer, that is, When the angle approaches 0°, it means that the occurrence constraint of the constructed scalar field at point o is good. The relevant formula is as follows:

[0099] The point constraint inequality:

[0100]

[0101] in, Represents the scalar field associated with the direction vector at point o and the geological interface IK The angle between the scalar field gradients, point o belongs to the sampling point set O of the geological interface IK.

[0102] d) Constraints related to geological boundary tangent points

[0103] The geological boundary tangent is mainly visible in specific geological structures, such as faults, folds, etc. For a set of sampling points C with the profile dip of a fault and the fold axis of a fold, the direction vector at any point c is K The scalar gradient of the associated scalar field at point c should be vertical, that is, the angle is 90°. The relevant formula is as follows:

[0104] Geological boundary tangent inequality:

[0105]

[0106] in, Represents the direction vector at point c and the geological interface I K Associated scalar field The angle between the scalar field gradients, point c belongs to the set of sampling points C that are tangent to the geological boundary.

[0107] The implementation rule for constructing the three-dimensional geological prediction model loss function in step S34 is:

[0108] The loss function of this prediction model consists of two parts: the geological implicit loss function and the global loss function. The former is transformed from the geological implicit constraint inequality constructed by S33, and the latter is the regularization loss, which uses the Eikonal unit norm constraint. The specific loss function is constructed as follows:

[0109] 1) Geological implicit loss function

[0110] a) Loss functions related to geological interfaces

[0111] The loss functions related to geological interfaces are similar to the constraints related to geological interfaces, including the loss functions related to potential fields and the loss functions related to stratigraphic sequences, namely the scalar field value loss function and the geological interface-geological interface sequence relationship loss function. The specific loss function is constructed as follows:

[0112] (1) Scalar field value loss function

[0113] The relevant constraint is defined as the geological interface I k The sampling point set K, the corresponding scalar field value The variance of is approximately 0, so it is necessary to ensure that the variance of the relevant scalar field values ​​of each point set on the geological interface is 0. The relevant formula is expressed as follows:

[0114] Scalar field value loss function formula:

[0115]

[0116] Where |I| represents the total number of geological interfaces, I KRepresents any scalar field. The formula means that the sum of the variances of the scalar field values ​​associated with all geological interfaces is the loss associated with the scalar field values ​​of the geological interfaces.

[0117] (2) Geological interface-geological interface sequence relationship loss function

[0118] The relevant constraint is defined as X and I K :

[0119] If the geological interface I X At the geological interface I K Above, the geological interface I X The scalar field value associated with any point x in the sampling point set X Greater than geological interface I K The average scalar field value of

[0120] If the geological interface I X At the geological interface I K Below, the geological interface I X The scalar field value associated with any point x in the sampling point set X Less than geological interface I K The average scalar field value of

[0121] If the geological interface I X At the geological interface I K On the geological interface I X The scalar field value associated with any point x in the sampling point set X Equal to geological interface I K The average scalar field value of

[0122] So if the scalar field value corresponding to point x is and geological interface I K The average scalar field value of The difference between The positive and negative of does not satisfy the inequality constraint of geological interface-geological interface sequence relationship, and the difference is the approximate signed distance in the scalar field is the geological interface I at point x K The average scalar field value of The relevant formula is as follows:

[0123] Approximate signed distance formula:

[0124]

[0125] in, Indicates that point x is on the geological interface IX The scalar field value in the corresponding scalar field, Indicates geological interface I K The average scalar field value of Indicates that point x is on the geological interface I X The scalar gradient in the corresponding scalar field can be used to represent the change in unit thickness between the interfaces in the scalar field: when the scalar gradient is small, the unit between the interfaces becomes thicker; when the scalar gradient is large, the unit between the interfaces becomes thinner. Scalar difference It can more accurately measure the above, below, and overlapping relationship of a point with a reference interface.

[0126] The loss function formula of geological interface-geological interface sequence relationship is:

[0127]

[0128]

[0129] The loss function formula of the above and below relationship means that when the above or below losses of all geological interfaces are required, it is necessary to calculate the above or below losses of a certain geological interface I X The sum of the approximate signed distances of all points x in the sampling point set X to the mean value of the scalar field of any geological interface is calculated, and then each geological interface I X The loss of the geological interface-geological interface sequence relationship above and below is the loss of the relationship; the loss function formula of the overlapping relationship means that when the geological interface I is found X Geological Interface I K When the overlap loss is X Find all points x in the sampling point set X and their correspondence to the geological interface I K The sum of the approximate signed distances of the scalar field mean is the overlap relationship loss of the geological interface-geological interface sequence relationship. |I| represents the number of geological interfaces, |X represents the number of points in the sampling point set X, and |K| represents the number of geological interfaces - 1, except for the current geological interface.

[0130] In summary, the loss function related to the geological interface is as follows:

[0131]

[0132] b) Loss function of stratigraphic unit sampling points

[0133] The loss function related to the stratigraphic unit sampling point is similar to the constraint related to the stratigraphic unit sampling point, that is, the loss function related to the stratigraphic sequence, that is, the loss function of the stratigraphic unit-geological interface sequence relationship.

[0134] The stratigraphic unit-geological interface sequence relationship constraint is defined asX and geological interface I K :

[0135] If the stratigraphic unit U X At the geological interface I K Above, the stratigraphic unit U X The scalar field value associated with any point x in the sampling point set X Greater than geological interface I K The average scalar field value of

[0136] If the stratigraphic unit U X At the geological interface I K Below, the stratigraphic unit U X The scalar field value associated with any point x in the sampling point set X Less than geological interface I K The average scalar field value of

[0137] So if the scalar field value corresponding to point x is and geological interface I K The average scalar field value of The difference between The positive and negative of does not satisfy the inequality constraint of the stratigraphic unit-geological interface sequence relationship, and the difference is the approximate signed distance in the scalar field is the geological interface I at point x K The average scalar field value of The relevant formula is as follows:

[0138] The loss function formula of stratigraphic unit-geological interface sequence relationship is:

[0139]

[0140]

[0141] The loss function formula of the above and below relationship means that when calculating the above or below loss of all stratigraphic units, it is necessary to calculate the above or below loss of a stratigraphic unit U X The sum of the approximate signed distances of all points x in the sampling point set X to the mean value of the scalar field of any geological interface is calculated, and then each stratigraphic unit U X The sum of the losses is the relationship loss above and below the stratigraphic unit-geological interface sequence relationship. |U represents the number of stratigraphic units, |X| represents the number of points in the sampling point set X, and |K represents the number of geological interfaces.

[0142] In summary, the relevant loss function of the stratigraphic unit is as follows:

[0143]

[0144] c) Direction-related loss function

[0145] The direction-related loss function is obtained from the direction-related constraints. The direction-related constraints are divided into two parts, which have been explained in S33, including the occurrence point related constraints and the geological boundary tangent point related constraints. The geological boundary tangent point can be regarded as a special occurrence point. In the unconformity geological structure mainly considered in this paper, it can be ignored. The occurrence point constraint is defined as K Associated scalar field For any point o in the set of direction points O, the angle difference between the direction vector at point o and the scalar field gradient is If the angle is different The closer it is to 0, the better the constraint is and the more it conforms to the actual geological structure. The relevant formula is as follows:

[0146] Scalar field angle formula:

[0147]

[0148] in, Indicates geological interface I K The cosine of the angle between the scalar field gradient and the scalar field course at point o in the associated scalar field.

[0149] The loss function formula of the production point is:

[0150]

[0151] Among them, the occurrence point related loss needs to calculate the difference between the angle cosine value between the true occurrence direction vector and the scalar field direction and the angle cosine value between the scalar field gradient and the scalar field direction for all F scalar fields containing the occurrence direction point o. After the average value of the direction point set O of each scalar field is calculated, the average value of all scalar fields is added to obtain the occurrence point related loss. F represents the number of scalar fields, and |O| represents the number of points in the direction point set O of the scalar field.

[0152] 2) Global loss function

[0153] The energy minimization principle is used to globally constrain the modeling area. A unit norm constraint, the Eikonal constraint, is adopted. This constraint is defined as X A set of sampled points X, where the scalar field gradient of any point x is close to the unit norm, the smaller the difference between the two, the better the quasi-parallel characteristics of the nearby strata, making the values ​​in the scalar field more smooth and continuous to a certain extent. The relevant formula is as follows:

[0154] The unit norm of the scalar field in the modeling area is 1:

[0155]

[0156] in, Represents the modeling area Ω X The scalar field of represents the scalar field value of the modeling area at any point x in the sampling point set X, represents the scalar field gradient of the modeling area at any point x in the sampling point set X, Represents the unit norm of the scalar field gradient of the modeling region at any point x in the set of sampling points X.

[0157] Global smooth loss function formula:

[0158]

[0159] Among them, the global smooth loss function needs to calculate the difference between the scalar field gradient and the unit norm, i.e. 1, of any point x in the sampling point set X in the scalar field of all F modeling regions, and then calculate the average value of the sampling point set X in each modeling region scalar field, and then add the average values ​​of all modeling region scalar fields to obtain the global loss. F represents the number of modeling region scalar fields, |Ω X | represents the modeling area Ω X The number of points in the sampling point set X in the scalar field.

[0160] Therefore, the total loss function constructed in step S34 is the synthetic loss of all the above geological implicit losses plus the regularized global loss. The relevant formula is as follows:

[0161] The total loss function formula is:

[0162] L=L I +L U +L O +λL ξ (twenty one)

[0163] From left to right, they are the geological interface related loss function, stratigraphic unit related loss function, direction related loss function and global constraint loss function. The global constraint loss function is weighted by λ. The larger the value of λ, the more smoothed scalar fields are.

[0164] In step S4, the complete network architecture of the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture is constructed as follows: Figure 5 As shown in the figure, the specific construction rules of each module are as follows:

[0165] The concise fully connected module constructed in step S41 is mainly used to introduce nonlinear mapping into the neural network. It includes a fully connected layer that maps the input data to the intermediate feature representation and introduces nonlinearity using the Softplus activation function. The parameters include the number of input features, the number of output features, whether to use a bias term, whether it is the last layer, and whether to perform a concatenation operation. The activation function uses Softplus, where beta is set to 100 and the weights are initialized.

[0166] The complex convolution residual module constructed in step S42 is mainly used to extract high-level features of the input data, including convolutional layers and residual connection blocks, and the Softplus activation function introduces nonlinearity. The module includes two parts: a convolutional feature extraction part and a feature mapping and prediction part. The convolutional feature extraction part includes a series of convolution and full connection operations, in which the residual connection block provides a jump connection. The feature mapping and prediction part further processes the output of the convolutional feature extraction part and finally maps it to the required number of output features. The parameters include the number of input features, the number of hidden features, the number of hidden layers, and the number of output features.

[0167] The hierarchical integration module constructed in step S43 helps to build a multi-level neural network for tasks such as 3D geological prediction. It includes multiple instances of complex convolution residual modules that share the same parameters and are initialized by loading weight files. The input data is processed by multiple complex convolution residual modules, and the outputs are spliced ​​together in dimensions to finally generate an output, where each output represents the prediction of a scalar field. The parameters include the number of input features, the number of hidden features, the number of hidden layers, the number of output features, and the path to the weight file. This module enables simultaneous prediction of multiple scalar field values.

[0168] In step S5, the input data is the stratigraphic interface data point, and the output data is the quasi-ideal stratigraphic plane. The constructed quasi-ideal stratigraphic plane is then used for pre-training of plane geometry. First, all stratigraphic interface data points of all stratigraphic layers are considered for HuberRegressor regression interface fitting, and the stratigraphic dip of each interface is calculated. The stratigraphic dip at this time takes into account the influence of all extreme values, and then the stratigraphic dip is averaged. The average stratigraphic dip obtained can be used as a constraint to control the generation of the quasi-ideal plane, so that the quasi-ideal plane constructed from bottom to top reflects the basic characteristics of the stratigraphic geometry in the region. The specific rules for constructing quasi-ideal stratigraphic planes are as follows:

[0169] 1) Get the coordinate data of the current formation.

[0170] 2) Extract the X, Y, and Z coordinates from the coordinate data.

[0171] 3) Calculate the percentile values ​​in the Z direction, with initial values ​​of p1=0 and p3=100, to determine the vertical range.

[0172] 4) Enter the loop until a prediction plane that meets the conditions is found: the inclination of the prediction plane is not greater than the average inclination.

[0173] a) In each iteration, the quantile values ​​p1 and p3 are calculated, and the data points whose Z coordinates are within this range are selected.

[0174] b) Use the HuberRegressor regression algorithm to fit the plane to the screened data points and obtain the coefficients of the plane equation.

[0175] c) Calculate the plane normal vector obtained by fitting, and calculate the angle between it and the horizontal normal vector.

[0176] d) If the angle is greater than the average inclination, the percentile range is gradually reduced to find a prediction plane that meets the conditions until the angle is less than or equal to the average inclination. The reduction rule is p1=p1+5, p3=p3-5, and the extreme values ​​are gradually reduced.

[0177] 5) After finding a prediction plane that meets the conditions, create a uniformly distributed point grid to represent the ideal stratum plane, and then add the point cloud to the ideal stratum plane.

[0178] 6) Save the final point cloud object as a VTP file.

[0179] The HuberRegressor regression algorithm used is designed to reduce the sensitivity of ordinary linear regression to outliers. Huber regression is a robust regression method that performs better when there are outliers in the data.

[0180] After that, the quasi-ideal stratigraphic plane was checked and corrected, and the stratigraphic plane was checked and corrected layer by layer from bottom to top to create a three-dimensional geological model with high geological accuracy. First, the quasi-ideal stratigraphic plane was constructed layer by layer using the S53 inspection and correction method. According to the rule of self-defined distance, the coordinates of a certain point in the upper constraint fitting plane were corrected to the top of the lower constraint fitting plane, so that the upper and lower constraint fitting planes could learn the upper and lower relationship of the stratigraphic formation more accurately. The constrained fitting plane finally formed included the fusion of the characteristics of the ideal stratum being relatively flat in the horizontal direction, layered distribution, and stratigraphic dip in the region. Then, in S54, we used the constructed quasi-ideal stratigraphic plane to perform geometric initialization of the network model parameters, pre-trained a three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture, and retained the pre-trained network parameters as the starting point of the model parameters for formal training. This ensures that the final output scalar field has a reasonable starting geometry of the geological interface and ensures the reproducibility of the modeling results.

[0181] Step S6 input data is the sample data point VTP file, the stratigraphic contact relationship CSV file, the plane geometry pre-trained network parameters, and the constructed 3D geological prediction model based on the convolutional residual implicit hierarchical network architecture. The output data is the predicted 3D geological model. The conversion from input to output goes through the model training process, the prediction grid construction process, the model prediction process, and the scalar field isosurface extraction process. There is a lot of content. For specific prediction model example diagrams, please refer to Figure 6 , the detailed implementation rules are as follows:

[0182] Step S61: During the training process of the model, the input data is a set of scattered data points, such as a three-dimensional point file of geological interface sampling points and stratigraphic unit sampling points in vtp format, and a set of stratigraphic rules of conformity or unconformity relationship (i.e., geological interface) between stratigraphic units, such as a stratigraphic contact relationship file in csv format. The output is an implicit model. is a set of scattered data points, is the implicit scalar field set under the F index, and ξ is the global interpolation constraint, which can ensure the global rationality of the geological structure model. When the model starts training, the network parameters pre-trained by plane geometry initialization are loaded onto each layer of convolution residual implicit network. When a good convergence effect is achieved or the number of iterations is reached, the training is stopped, and a trained three-dimensional geological prediction model can be obtained.

[0183] In the process of constructing the prediction grid in step S62, the modeling resolution needs to be customized. The bounding box of the sample data point is encrypted according to the customized resolution to obtain the corresponding grid nodes, which are then normalized to obtain the prediction nodes of the three-dimensional geological model.

[0184] In the prediction process of the model in step S63, the input data is the grid nodes constructed in step S62, and the output is the scalar field value corresponding to each grid node, from which the overall scalar field value in the prediction area can be obtained.

[0185] In step S64, the isosurface extraction process of the scalar field can interpolate the scalar field grid predicted in step S63 by the average scalar field value of each stratum obtained in the training process of the model in step S61, extract the corresponding stratum interface, and further obtain a three-dimensional geological model. Among them, the isosurface extraction uses an algorithm developed by the open source library PyVista, which can generate all isosurfaces cut by the unconformity surface.

[0186] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for constructing a three-dimensional geological model based on coupling potential field interpolation and deep learning, characterized in that: The following steps are involved: S1 constructs a sample data set, discretizes various geological data into geological interface points and stratigraphic unit sampling point data, and obtains the sample data set after normalization processing; S2 analyzes the stratigraphic contact relationship in the region. Based on the sample data set constructed in the modeling area, the upper and lower stratigraphic data are calculated to obtain their coefficient of variation. A threshold is defined to compare whether the upper and lower stratigraphic layers are consistent in geometry, thereby determining the stratigraphic contact relationship. The information is organized to include stratigraphic code, stratigraphic name, and contact relationship between stratigraphic layers. S3 designs a loss function that takes into account geological implicit constraints. It groups geological interfaces according to stratigraphic contact relationships, determines the number of scalar fields to be constructed, and the correspondence between scalar fields and geological interfaces. Then, based on the correspondence between scalar fields and geological interfaces, it constructs stratigraphic sequence relationships, and uses inequalities to express implicit constraints on geological conditions such as stratigraphic sequence relationships and occurrence. Finally, it constructs relevant loss functions based on geological implicit inequality constraints, and adds global regularized smoothing losses to constrain the training process of the network model, introducing geological prior knowledge into the modeling process. S4 designs a three-dimensional geological prediction model based on a convolutional residual implicit hierarchical network architecture, wherein the network architecture includes a simple fully connected module, a complex convolutional residual module, and a hierarchical integration module, wherein the simple fully connected module is used to integrate and transform the extracted geological features, the complex convolutional residual module includes multiple convolutional layers and residual connections, and the hierarchical integration module integrates multiple basic convolutional residual implicit networks to achieve the prediction of multiple scalar field values; S5 plane geometry pre-training, plane fitting is performed according to the characteristics of the formation interface point data, and a quasi-ideal formation plane containing the formation characteristics in the region is constructed from bottom to top. The HuberRegressor regression algorithm is used when constructing the quasi-ideal formation plane, and the formation dip angle and the formation up-down relationship are considered as constraints. Finally, the quasi-ideal formation plane is used for initialization training of the network model; S6 uses the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture to predict the three-dimensional geological model, loads the pre-trained network parameters onto each layer of the convolutional residual implicit network for training, and stops training when the loss functions no longer decrease. The trained network model can then be used to predict the scalar field and extract the isosurface to obtain the corresponding three-dimensional geological model.

2. The method for constructing a three-dimensional geological model based on coupling of potential field interpolation and deep learning according to claim 1, characterized in that: In step S1, constructing the sample data set includes the following steps: S11 discretizes various geological data, including: discretizing drilling data, geological profiles, geological map data, etc. into spatial scattered points, and classifying and coding them according to existing geological prior knowledge; S12 constructs a geological interface point file, and selects corresponding discrete points according to the drilling layer, geological profile, etc. to construct a geological interface point file; S13 constructs a stratigraphic unit sampling point file, and selects corresponding discrete points according to the drilling layer, geological profile, etc. to construct the stratigraphic unit sampling point file; S14 coordinate normalization processing is used to normalize the geological interface points and stratigraphic unit sampling points to obtain the final data set for network model training.

3. The method for constructing a three-dimensional geological model based on coupling of potential field interpolation and deep learning according to claim 2, characterized in that: In step S2, the analysis of the contact relationship of the strata in the region includes the following steps: S21 Calculation of coefficient of variation between upper and lower strata: According to the coordinate values ​​of the upper and lower strata data, the corresponding strata spacing is obtained, and the coefficient of variation of the spacing is calculated to determine whether the geometric shapes between the upper and lower strata are similar; S22 determines the contact relationship between the upper and lower strata according to the coefficient of variation. According to the characteristics of the modeling area, a threshold of the coefficient of variation is set. If the coefficient of variation is less than the set threshold, the contact relationship between the upper and lower strata is integrated, and if the coefficient of variation is greater than the set threshold, the contact relationship between the upper and lower strata is unconformable. S23 outputs a stratum contact relationship information table, and the organized information includes stratum code, stratum name, and contact relationship between strata.

4. The method for constructing a three-dimensional geological model based on coupling of potential field interpolation and deep learning according to claim 3 is characterized in that: In step S3, the loss function considering the geological implicit constraints includes the following steps: S31 determines the number of scalar fields that need to be constructed, and groups the geological interfaces that need to be modeled according to the stratum contact relationship, thereby obtaining the number of scalar fields that need to be constructed in the modeling area and the corresponding relationship between the scalar fields and the geological interfaces; S32 constructs stratigraphic sequence relationships, sorting out stratigraphic sequence relationships based on the correspondence between scalar fields and geological interfaces, referred to as geological interface-geological interface sequence relationships and geological unit-geological interface sequence relationships; S33 constructs geological implicit constraint inequalities. Geological implicit constraints include constraints related to geological interfaces and stratigraphic unit sampling points, as well as constraints on occurrence points and geological tangents. The above implicit constraints are organized into inequality expressions. S34 constructs the loss function of the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture. The loss function includes two parts, one part is transformed from the above-mentioned geological implicit inequality, and the other part is the global loss function, which uses the Eikonal unit norm constraint, that is, the regularization loss.

5. The method for constructing a three-dimensional geological model based on potential field interpolation and deep learning coupling according to claim 4, characterized in that: The construction of a 3D geological prediction model based on the convolutional residual implicit hierarchical network architecture includes the following steps: S41 builds a concise fully connected module, which consists of a fully connected layer and a Softplus activation function; S42 builds a complex convolutional residual module, which consists of a convolutional layer, a residual connection block, and a Softplus activation function; S43 constructs a three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture according to the hierarchical integration module, combines the above-mentioned simple fully connected module with the complex convolutional residual module to construct a basic convolutional residual implicit network, and then integrates multiple basic convolutional residual implicit networks according to the number of scalar fields to be constructed to obtain the final prediction model.

6. The method for constructing a three-dimensional geological model based on coupling of potential field interpolation and deep learning according to claim 5, characterized in that: In step S5, the plane geometry pre-training process includes the following steps: S51 calculates the average dip of the entire formation, considers all the formation interface point data for plane fitting, uses the HuberRegressor regression algorithm, then evaluates the dip of all the formation fitting planes, and finally averages them to get the average dip of the entire formation; S52 constructs a constrained fitting plane, and uses the average dip angle of the entire formation to perform HuberRegressor regression fitting on the constrained plane layer by layer from bottom to top. The constructed constrained fitting plane contains regional formation characteristics and avoids intersection to a certain extent; S53 checks and corrects the construction of a quasi-ideal formation plane, and checks and corrects each layer from bottom to top. The constrained fitting plane after the inspection and correction is the quasi-ideal formation plane, which integrates the characteristics of the ideal formation such as relative flatness in the horizontal direction, layered distribution, and formation dip angle in the region; S54 network model parameter geometry initialization, use the constructed ideal-like stratigraphic plane to pre-train the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture, retain the pre-trained network parameters as the starting point of the model parameters for formal training, and finally the output scalar field will have a reasonable starting geometry of the geological interface.

7. The method for constructing a three-dimensional geological model based on coupling of potential field interpolation and deep learning according to claim 6, characterized in that: In step S6, the process of performing three-dimensional geological model prediction using the three-dimensional geological prediction model based on the convolutional residual implicit hierarchical network architecture includes the following steps: S61 trains a three-dimensional geological prediction model based on a convolutional residual implicit hierarchical network architecture, loads the network parameters pre-trained by geometric initialization into each layer of the convolutional residual implicit network for training, and stops training when each loss function no longer decreases, thus obtaining a trained three-dimensional geological prediction model; S62 constructs a predicted geological grid, obtains its bounding box grid according to the stratigraphic unit interface point data, customizes the resolution in the x, y, and z directions, and obtains the predicted geological grid structure in the region and the grid nodes of the grid structure; S63 predicts the scalar field in the region, using the trained three-dimensional geological prediction model to predict the constructed prediction geological grid, and predicts the scalar field value in the regional geological grid; S64 constructs a three-dimensional geological model, extracts isosurfaces from the predicted scalar field grid, and interpolates the three-dimensional geological model based on the average scalar field value of each stratum obtained during the model training phase. The isosurface extraction process uses the isosurface extraction algorithm in the PyVista open source library.

Images