A Controllable 3D Geological Body Generation Method Based on Sparse Geological Constraints
By obtaining sparse geological constraints in 3D geological modeling and encoding them as unified control voxel conditions, and using a geological residual control network to generate multi-scale control residuals, the problem of unified encoding of multiple types of sparse geological constraints is solved, and the controllable generation of 3D geological bodies is realized.
Patent Information
- Application Number
- CN202611436998.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
Existing 3D geological modeling methods struggle to uniformly encode multiple types of sparse geological constraints, making it difficult to achieve multi-scale hierarchical constraint control for the generation of 3D geological bodies.
By acquiring sparse geological constraints within the target area, mapping them to the target 3D voxel grid and encoding them as unified control voxel conditions, and using a geological residual control network for step-by-step encoding, multiple control features are output and converted into multi-scale control residuals injected into the latent space to generate the backbone, thus generating the semantic voxel results of the target 3D geological body.
It achieves unified compatibility with multiple sparse geological constraints, generates multi-scale control residuals that match the various levels of the latent space generation backbone, and thus generates controllable three-dimensional geological body semantic voxel results.
Smart Images

Figure CN123071571A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional geological modeling technology, and specifically discloses a method for generating controllable three-dimensional geological bodies based on sparse geological constraints. Background Technology
[0002] Existing 3D geological modeling requires handling data with varying spatial distributions and meanings, such as boreholes, profiles, outcrops, and interpreted structural lines. Stratigraphic category observations and fault or stratigraphic location hints have different data types. Directly stitching data or only correcting at the output makes it difficult to distinguish the effective range of various information types and to continuously utilize local sparsity conditions during the generation process. Therefore, it is necessary to unify multiple types of constraints into a voxel mesh and establish a correspondence between them and different levels of the latent space generation backbone. Summary of the Invention
[0003] This invention aims to solve the technical problems of existing 3D geological modeling, which cannot uniformly encode multiple types of sparse geological constraints and is difficult to achieve multi-scale hierarchical constraint control for the generation of 3D geological bodies.
[0004] This invention provides a method for generating controllable three-dimensional geological bodies based on sparse geological constraints, comprising:
[0005] Obtain at least one sparse geological constraint within the target area, wherein the sparse geological constraint includes at least one of borehole constraint, profile constraint, outcrop constraint, fault line constraint, and stratigraphic line constraint;
[0006] The sparse geological constraints are mapped to the target 3D voxel grid and encoded as unified control voxel conditions. The unified control voxel conditions include semantic cue channels that store known stratigraphic categories and semantic effective region mask channels that mark the known locations of the categories. When the constraints contain fault lines or stratigraphic information, they also include corresponding structural cue channels and structural effective region mask channels. Each cue channel stores data of the corresponding type, and each effective region mask channel marks the effective location of the data of that type.
[0007] The unified control voxel conditions are input into the geological residual control network. The geological residual control network encodes the unified control voxel conditions step by step and outputs multiple control features. The multiple control features correspond to the downsampling layer, intermediate layer and upsampling layer of the latent space generation backbone in terms of hierarchical position.
[0008] The multiple control features are converted into multi-scale control residuals, the number of channels and spatial resolution of the multi-scale control residuals are matched with the features of the corresponding level of the latent space generation backbone, and the multi-scale control residuals are injected into the latent space generation backbone;
[0009] The target latent variables are generated by the latent space generation backbone under the multi-scale control residual. The target latent variables are then decoded to obtain the semantic voxel results of the target three-dimensional geological body.
[0010] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the method when it invokes the computer program.
[0011] The beneficial technical effects of this invention are: it can uniformly and compatiblely support multiple sparse geological constraints and standardize their coding, generate multi-scale control residuals that match the various levels of the latent space generation backbone, impose constraints on the latent variable generation process from multiple scale levels, and realize the controllable generation of three-dimensional geological body semantic voxel results based on sparse geological constraints. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of the overall process of the controllable three-dimensional geological body generation method based on sparse geological constraints of the present invention.
[0013] Figure 2 This is a schematic diagram illustrating the construction of unified control voxel conditions in this invention.
[0014] Figure 3 This is a schematic diagram of the geological residual control network structure of the present invention.
[0015] Figure 4 This is a schematic diagram of the multi-scale controlled residual injection latent space generation backbone of the present invention.
[0016] Figure 5 This is a schematic diagram illustrating the generation of mixed constraints for boreholes, profiles, outcrops, and construction lines according to the present invention. Detailed Implementation
[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0018] like Figure 1 As shown, this embodiment first obtains at least one sparse geological constraint within the target area. Borehole constraints may consist of one or more boreholes, each providing discrete known stratigraphic categories or structural identifiers along the depth direction; profile constraints may consist of one or more two-dimensional profiles, used to provide continuous or discrete known stratigraphic categories within the profile plane; outcrop constraints may originate from surface survey data or shallow observation results; fault line and horizon line constraints may originate from seismic interpretation data, engineering interpretation data, or manual annotation results.
[0019] like Figure 2As shown, the unified control voxel conditions consist of cue channels and corresponding effective region masks. In an embodiment with eight stratigraphic categories, thirteen actual tensor channels are used: channels one through eight are uniquely thermal semantic cues for stratigraphic categories, channel nine is the semantic effective region mask, channels ten and eleven are fault cue and fault effective region mask, respectively, and channels twelfth and thirteenth are stratigraphic cue and stratigraphic effective region mask, respectively. When the category is known, the corresponding category channel is set to one, other category channels are set to zero, and the semantic effective mask is set to one; when the category is unknown, all semantic cue and semantic effective mask are set to zero. Structural cue uses binary labeling: confirmed structural locations are set to one, confirmed non-structural locations are set to zero, and their corresponding effective masks are both set to one; cue and effective mask for unobserved locations are both set to zero. When the number of categories changes, the semantic channels are adjusted accordingly; in this example, the remaining five channels remain unchanged.
[0020] When combining borehole and profile constraints, known categories are written into the semantic cue channel, and the known locations of these categories are set to 1 in the semantically valid region mask. Fault and stratigraphic information from the data are written into fault cue, stratigraphic cue, and their respective valid masks, respectively. Figure 2 The terms “semantic effective mask,” “fault effective mask,” and “stratum effective mask” correspond to the three types of effective region mask channels mentioned above, and each only indicates whether the corresponding data is valid.
[0021] When outcrops, fault lines, and bedding planes are used as joint constraints, the outcrop stratigraphic category is written into the semantic cue channel, and the semantic effective mask for the known location of the corresponding category is set to 1. Fault lines and bedding planes are rasterized to their respective structural cue channels, and their respective effective masks are set. When there are only fault lines or bedding planes without a known category, both the semantic cue and the semantic effective mask are set to zero, and the control is provided by the structural cue and the structural effective mask. If there are no non-structural observations at the location outside the line, its effective mask remains zero.
[0022] like Figure 3 As shown, the geological residual control network performs 3D convolution and stepwise downsampling on uniform control voxel conditions to obtain multi-scale control features. Flow time, physical scale, and normalized depth are embedded and then used for feature modulation. Features at each scale are transformed by independent residual mapping layers and matched with the corresponding downsampling, intermediate, and upsampling layers of the backbone through channel mapping and spatial resampling to output control residuals at each level. The same observation conditions can produce different residuals at different flow times to participate in the velocity prediction of the current latent variable.
[0023] like Figure 4As shown, the backbone features at each level are element-wise added to the matched control residuals, and then propagated forward in the backbone order. The injection positions are expanded in the order of downsampling, intermediate layers, and upsampling, with the plus sign indicating the addition operation of the corresponding features. One implementation uses a 3D convolution with a kernel size of one voxel for channel mapping. At the start of training, the mapping layer weights and biases are set to zero, making the initial control residuals zero, and then updated through training. The size of the backbone features and the tensor specifications of the input latent variables and output velocity field remain unchanged at each injection point.
[0024] During training, at least two scenarios are selected from those such as boreholes, profiles, outcrops, fault lines, stratigraphic lines, random sparse points, random sparse lines, random sparse blocks, and half-block extensions. Scenarios and observation locations are randomly sampled to construct training batches; different batches can use different combinations of scenarios. The half-block extension scenario retains a portion of the sample as a known region and sets the adjacent portion as the region to be generated, used to learn the generation conditions near the boundaries of existing models.
[0025] During inference, a unified control voxel condition with the same channel order, coordinate system, and numerical definition as the training voxel is constructed based on actual data. If a certain type of observation is missing, that type of cue and its effective mask are set to zero. The physical scale can be the ratio of the three-axis length of the target block to the corresponding length of the training reference block, and the depth can be the ratio of the depth of the target block's center from the top of the modeling region to the total depth of the region; the reference scale, positive depth direction, and embedding method remain consistent between training and inference.
[0026] In one inference implementation, a pre-trained 3D autoencoder establishes the correspondence between geological voxels and latent variables. When there are no known latent grid locations, standard normal noise is used as the initial latent variable; when there are known region conditions, the known locations use encoded conditional latent variables, and other locations use noise. The flow time is incremented from zero to one. At each step, the velocity is predicted based on the current latent variable, flow time, condition information, and corresponding control residual. The velocity is multiplied by a positive time increment and added to the current latent variable until the target latent variable is obtained. The sum of all time increments is one. The target latent variable is processed by a pre-trained decoder to obtain semantic voxel results, and structural edge results can also be obtained simultaneously. Figure 5 This illustrates the relationship between multiple source constraints in generation and result verification.
[0027] When constructing the conditional mesh, the origin, axis, length unit, and target voxel spacing of the data coordinates are first standardized. Boreholes are located based on trajectory sampling points, profiles are located based on the overlap of known pixels with the mesh, and outcrops are located based on their surface or shallow location. The location index is obtained by subtracting the mesh origin from the actual coordinates, dividing by the corresponding voxel spacing, and rounding down. Fault lines and stratigraphic lines are rasterized according to the voxels they pass through. Semantic categories, structural markers, and their respective effective masks are written separately. Conditional latent variables are obtained by encoding the original voxel mesh conditions, and the known location masks are separately mapped to the latent space mesh; both participate in the generation at the same origin and resolution.
[0028] When a voxel has a known category and fault or stratigraphic indication, each category is retained separately; observations with consistent values within the same channel are merged into a single valid value. In case of discrepancies, observations are selected based on the verification results and reliability recorded during the data processing phase; if the reliability is the same and cannot be determined, the valid mask for that channel at that location is set to zero, and the observation is written in only after confirmation. Coarse-scale priors only supplement locations lacking observations and retain source markers separately; the final original category retention step uses only truly known categories.
[0029] In one training embodiment, the parameters of the pre-trained 3D autoencoder and the latent space generation backbone are fixed. The mean value obtained from encoding real voxel samples is used as the sample latent variable; the initial latent variable uses the sample value at known latent grid positions and standard normal noise at unknown positions. A flow time between zero and one is selected, and the initial latent variable and the sample latent variable are weighted by a decrease in flow time and the flow time, respectively, to obtain intermediate latent variables. The control network outputs residuals according to unified control voxel conditions, flow time, scale, and depth. The backbone uses the intermediate latent variables and residuals to predict the velocity, using the difference between the sample latent variable and the initial latent variable as the target velocity, calculating the element-wise squared error, and averaging or weighted averaging at effective locations. The error is backpropagated to the control branch through the backbone to update the control coding layer and the residual mapping layer; the parameters of the backbone and autoencoder remain fixed. The pre-trained embedding parameters for time, scale, and depth can be fixed, while the remaining trainable parameters of the control branch are updated with loss.
[0030] During output verification, at the location marked by the mask in the semantically valid region, the semantic voxel results are restored to the original known categories. Stratigraphic hints are compared with category boundaries or corresponding strata edge channels; fault hints are compared with fault edge channels or known fault data. Structural verification records deviations by location, and semantic preservation and structural verification are processed separately. This preserves known categories while checking the correspondence between the generated structure guided by multi-scale residuals and the input structural hints.
[0031] This invention can be executed by an electronic device. The electronic device may include a processor and a memory, the memory storing a computer program for performing the above steps. When the processor invokes the computer program, it can complete sparse geological constraint encoding, control residual generation of the geological residual control network, multi-scale residual injection, and controlled three-dimensional geological body generation.
[0032] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for generating controllable three-dimensional geological bodies based on sparse geological constraints, characterized in that, include: Obtain at least one sparse geological constraint within the target area, wherein the sparse geological constraint includes at least one of borehole constraint, profile constraint, outcrop constraint, fault line constraint, and stratigraphic line constraint; The sparse geological constraints are mapped to the target three-dimensional voxel grid and encoded as unified control voxel conditions. The unified control voxel conditions include semantic cue channels that store known stratigraphic categories and semantic effective region mask channels that mark known locations of categories. When the constraints contain fault lines or stratigraphic line information, they also include corresponding structural cue channels and structural effective region mask channels. Each prompt channel stores data of the corresponding type, and each effective area mask channel marks the effective location of data of that type. The unified control voxel conditions are input into the geological residual control network. The geological residual control network encodes the unified control voxel conditions step by step and outputs multiple control features. The multiple control features correspond to the downsampling layer, intermediate layer and upsampling layer of the latent space generation backbone in terms of hierarchical position. The multiple control features are converted into multi-scale control residuals, the number of channels and spatial resolution of the multi-scale control residuals are matched with the features of the corresponding level of the latent space generation backbone, and the multi-scale control residuals are injected into the latent space generation backbone; The target latent variables are generated by the latent space generation backbone under the multi-scale control residual. The target latent variables are then decoded to obtain the semantic voxel results of the target three-dimensional geological body.
2. The method according to claim 1, characterized in that, The structural indication channel includes at least one of a fault indication channel and a stratigraphic indication channel; the fault indication channel is equipped with a fault effective region mask channel, and the stratigraphic indication channel is equipped with a stratigraphic effective region mask channel, wherein the effective position of each effective region mask is set to one, and the unobserved position is set to zero.
3. The method according to claim 1, characterized in that, The acquisition of at least one sparse geological constraint within the target area includes the acquisition of borehole constraints and profile constraints; The step of encoding the sparse geological constraints into unified control voxel conditions includes: encoding the known semantic categories at the corresponding positions of the borehole constraints and the profile constraints into the semantic cue channels, and encoding the known positions corresponding to the borehole constraints and the profile constraints into the semantic effective region mask channels.
4. The method according to claim 2, characterized in that, The acquisition of at least one sparse geological constraint within the target area includes the acquisition of outcrop constraints, fault line constraints, and stratigraphic line constraints. The step of encoding the sparse geological constraints into unified control voxel conditions includes: writing the known stratigraphic category of the outcrop location into the semantic hint channel and setting the corresponding semantic effective area mask to 1; writing the fault line location marker into the fault hint channel and marking the effective location in the corresponding fault effective area mask; writing the stratigraphic line location marker into the stratigraphic hint channel and marking the effective location in the corresponding stratigraphic effective area mask.
5. The method according to claim 1, characterized in that, The step of injecting the multi-scale control residuals into the latent space to generate the backbone includes: adding the features of each level to the matching control residuals element by element to obtain the updated features of that level, and keeping the tensor specifications of the input latent variables and the output velocity field of the latent space generation backbone unchanged after injection.
6. The method according to claim 5, characterized in that, The step of converting the multiple control features into multi-scale control residuals includes: converting the multiple control features into the multi-scale control residuals respectively through a convolutional mapping layer, wherein the weights and biases of the convolutional mapping layer are initialized to zero at the start of training and updated during training.
7. The method according to claim 1, characterized in that, When multiple sparse geological constraints are mapped to the same voxel, different types of valid data are retained in the corresponding prompt channels. For contradictory values within the same channel, verified observations are used first, and coarse-scale priors are used to supplement the locations where observations are missing. When there are contradictions in observations of equal confidence and the values have not yet been determined, the location is set to zero in the corresponding valid region mask.
8. The method according to claim 1, characterized in that, The geological residual control network is trained through a multi-scenario task sampling method. During the training process, at least two scenarios are sampled from boreholes, profiles, outcrops, fault lines, stratigraphic lines, random sparse points, random sparse lines, random sparse blocks, and half-block extended scenarios to construct training batches. The half-block extended scenario takes a part of the three-dimensional geological body sample as a known region and the other part adjacent to it as a region to be generated.
9. The method according to claim 1, characterized in that, Also includes: For the semantic voxel results, a position correspondence check is performed at the known locations corresponding to the sparse geological constraints. For voxels with known stratigraphic categories, the original constraint values corresponding to the sparse geological constraints are retained according to the principle of prioritizing known values.
10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 1 to 9 when it invokes the computer program.
