Geological model generation method

Abstract

The invention provides a geologic model generation method, and relates to the technical field of geologic model generation, and the method comprises the steps: obtaining target geologic data of a target region; obtaining a plurality of low-dimensional noise vectors, and inputting each low-dimensional noise vector into the trained geologic model generation module to obtain a plurality of three-dimensional generation geologic models; screening the three-dimensional generation geologic models based on the target geologic data, and determining three-dimensional output geologic models in the three-dimensional generation geologic models; and performing uncertainty quantitative synthesis on each three-dimensional output geologic model to obtain an uncertainty geologic model corresponding to the target area. According to the method, uncertainty quantitative synthesis is carried out on the multiple three-dimensional output geologic models, the multiple three-dimensional output geologic models are integrated, the uncertainty geologic model is obtained, the situation that data in the obtained geologic model is single is avoided, and the problem that the real geologic situation of the target area cannot be reflected is avoided.

Description

Technical Field

[0001] This invention relates to the field of geological model generation technology, and more specifically, to a method for generating geological models. Background Technology

[0002] In the field of coal mine geological exploration and development, constructing an accurate three-dimensional geological model is the cornerstone for resource assessment, mining design, and disaster prevention. The existing technology mainly uses the Kriging interpolation method to interpolate geological data to obtain a complete three-dimensional geological model. This method calculates the best estimate of unknown points based on the optimal weight of each known data point, thereby obtaining the data of each node in the three-dimensional geological model and completing the data filling of the three-dimensional geological model.

[0003] However, the existing Kriging interpolation method is not a simple mathematical interpolation, but a "spatial optimal estimation" technique based on unbiasedness and optimality (minimum estimation variance). It is precisely because of the pursuit of "optimality" and "unbiasedness" in the Kriging interpolation method that the three-dimensional geological model generated by the Kriging interpolation method has a "smoothing effect". This results in the absence of extreme prediction values ​​in the three-dimensional geological model, which suppresses high-value anomalies (such as extra-thick coal seams) or low-value anomalies (such as fault fracture zones) that may actually exist in the geological model. As a result, the generated three-dimensional geological model is simplistic and cannot reflect the true geological situation. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a geological model generation method that can solve the technical problem that the geological models generated in the prior art are singular and cannot reflect the true nature of the situation.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows: In a first aspect, embodiments of the present invention provide a method for generating a geological model, comprising: Obtain target geological data for the target area; Multiple low-dimensional noise vectors are acquired and input into a trained geological model generation module to obtain multiple three-dimensional generated geological models. Based on the target geological data, each of the three-dimensional generated geological models is screened to determine the three-dimensional output geological model. The relative error between the output geological data corresponding to each three-dimensional output geological model and the target geological data is less than a preset error threshold. The trained geological model generation module is obtained through training with a real three-dimensional geological model and training noise vectors. The uncertainties of each of the three-dimensional output geological models are quantified and synthesized to obtain the uncertain geological model corresponding to the target area.

[0006] Furthermore, this embodiment of the invention provides a first possible implementation of the first aspect, wherein the step of performing uncertainty quantification synthesis on each of the three-dimensional output geological models to obtain an uncertain geological model corresponding to the target region includes: Uncertainty parameters and uncertainty indices are determined based on the output geological data corresponding to each of the three-dimensional output geological models; wherein, the uncertainty parameters include variance, mean, and confidence interval; and the uncertainty indices include the coefficient of variation. The uncertainty parameters and uncertainty indices are mapped to a preset geological grid corresponding to the target area to obtain the uncertainty geological model corresponding to the target area.

[0007] Furthermore, this embodiment of the invention provides a second possible implementation of the first aspect, wherein the step of determining uncertainty parameters and uncertainty indices based on the output geological data corresponding to each of the three-dimensional output geological models includes: The uncertainty parameters corresponding to the positions of each volume element are determined based on the output geological data corresponding to the volume elements at the same positions in each of the three-dimensional output geological models. The coefficient of variation for each voxel position is obtained by calculating the ratio of the variance to the mean.

[0008] Furthermore, this embodiment of the invention provides a third possible implementation of the first aspect, wherein the step of mapping the uncertainty parameter and the uncertainty index to a preset geological raster corresponding to the target area to obtain the uncertainty geological model corresponding to the target area includes: The uncertainty parameters and uncertainty indices corresponding to each of the aforementioned voxel positions are mapped to the voxels at the corresponding voxel positions of the preset geological raster to obtain the uncertainty geological model.

[0009] Furthermore, this embodiment of the invention provides a fourth possible implementation of the first aspect, wherein the step of inputting each of the low-dimensional noise vectors into the trained geological model generation module to obtain multiple three-dimensional generated geological models includes: Based on the trained geological model generation module, each of the low-dimensional noise vectors is mapped to a corresponding feature tensor. The three-dimensional generated geological model is obtained by upsampling each of the feature tensors and obtaining each of the low-dimensional noise vectors.

[0010] Furthermore, this embodiment of the invention provides a fifth possible implementation of the first aspect, wherein the training step of the geological model generation module includes: Obtain a realistic 3D geological model and multiple training noise vectors; A training dataset is established based on the real three-dimensional geological model and each of the training noise vectors. The training dataset is input into a 3D generative adversarial network for training to obtain the trained geological model generation module.

[0011] Furthermore, this embodiment of the invention provides a sixth possible implementation of the first aspect, wherein the step of inputting the training dataset into a three-dimensional generative adversarial network for training to obtain the trained geological model generation module includes: The generator in the three-dimensional generative adversarial network is trained based on each of the training noise vectors to obtain the training three-dimensional geological model and generator loss function corresponding to each of the training noise vectors. The discriminator in the three-dimensional generative adversarial network is trained based on the trained three-dimensional geological models and the real three-dimensional geological models to obtain the discriminator loss function; Based on the generator loss function and the discriminator loss function, it is determined whether the trained generator and the trained discriminator have reached Nash equilibrium until they do. If the Nash equilibrium is not reached, the generator and the discriminator are trained again. If the Nash equilibrium is reached, the training of the three-dimensional generative adversarial network is completed, and the trained geological model generation module is obtained.

[0012] Furthermore, this embodiment of the invention provides a seventh possible implementation of the first aspect, wherein the step of training the generator in the three-dimensional generative adversarial network based on each of the training noise vectors to obtain the training three-dimensional geological model and the generator loss function corresponding to each of the training noise vectors includes: Each of the training noise vectors is input into the generator in the three-dimensional generative adversarial network for training, thereby obtaining each of the training three-dimensional geological models. The first average relative error between the training geological data corresponding to each of the training three-dimensional geological models and the real geological data corresponding to the real three-dimensional geological models is calculated and used as the generator loss function.

[0013] Furthermore, this invention provides an eighth possible implementation of the first aspect, wherein the step of training the discriminator in the three-dimensional generative adversarial network based on each of the trained three-dimensional geological models and the real three-dimensional geological models to obtain the discriminator loss function includes: Determine the true variation function corresponding to the true three-dimensional geological model based on the real geological data; Each of the trained 3D geological models and the real 3D geological models are input into the discriminator in the 3D generative adversarial network to obtain the true probability and training variation function corresponding to each of the trained 3D geological models; wherein, the true probability is the probability value corresponding to the discriminator judging the trained 3D geological model as a real 3D geological model; Image quality indices are obtained between each of the trained 3D geological models and the real 3D geological models, and an average image quality index is obtained based on each of the image quality indices; wherein, the image quality index includes the Fraser initial distance score; The average true probability is determined based on each of the stated true probabilities; The function similarity loss between each training variation function and the true variation function is calculated; wherein, the function similarity loss includes the average relative error, mean square error, or L2 norm error between each training variation function and the true variation function; The average image quality index, the average true probability, and the function similarity loss are used as the discriminator loss function.

[0014] Furthermore, this embodiment of the invention provides a ninth possible implementation of the first aspect, wherein the step of determining whether the trained generator and the trained discriminator have reached Nash equilibrium based on the generator loss function and the discriminator loss function, and if the Nash equilibrium has not been reached, then repeatedly training the generator and the discriminator until the trained generator and the trained discriminator reach Nash equilibrium; and if the Nash equilibrium has been reached, then completing the training of the three-dimensional generative adversarial network to obtain the trained geological model generation module, includes: Training completion requirements are determined based on the generator loss function and the discriminator loss function; wherein, the training completion requirements include a first completion requirement, a second completion requirement, a third completion requirement, a fourth completion requirement, and a fifth completion requirement; the completion condition for the first completion requirement is that the fluctuation amplitude of the first average relative error in a preset training batch is less than a first preset fluctuation amplitude threshold; the completion condition for the second completion requirement is that the fluctuation amplitude of the function similarity loss in a preset training batch is less than a second preset fluctuation amplitude threshold; the completion condition for the third completion requirement is that the fluctuation amplitude of the average image quality index in a preset training batch is less than a third preset fluctuation amplitude threshold; the completion condition for the fourth completion requirement is that the value of the average true probability is within a preset average true probability range; the completion condition for the fifth completion requirement is that the function similarity loss is less than a preset similarity loss threshold; Based on the training completion requirements, it is determined whether the trained generator and the trained discriminator have reached Nash equilibrium. If any one or more of the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement are not met, it is determined that the trained generator and the trained discriminator have not reached the Nash equilibrium. The generator and the discriminator are trained repeatedly until the trained generator and the trained discriminator reach the Nash equilibrium. If the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement all meet the completion conditions, then it is determined that the trained generator and the trained discriminator have reached the Nash equilibrium, completing the training of the three-dimensional generative adversarial network and obtaining the trained geological model generation module.

[0015] This invention provides a geological model generation method, which includes: acquiring target geological data of a target area; acquiring multiple low-dimensional noise vectors and inputting each low-dimensional noise vector into a trained geological model generation module to obtain multiple three-dimensional generated geological models; filtering each three-dimensional generated geological model based on the target geological data to determine the three-dimensional output geological model among the three-dimensional generated geological models; wherein the relative error between the output geological data corresponding to each three-dimensional output geological model and the target geological data is less than a preset error threshold; the trained geological model generation module is trained using a real three-dimensional geological model and training noise vectors; and performing uncertainty quantification synthesis on each three-dimensional output geological model to obtain an uncertain geological model corresponding to the target area. This invention acquires target geological data for a target area. Since the trained geological model generation module has been trained using a real 3D geological model and training noise vectors, inputting multiple low-dimensional noise vectors into the module directly yields multiple 3D generated geological models corresponding to the target area. The geological data corresponding to each 3D generated geological model is compared with the target geological data of the target area. The 3D generated geological model whose relative error with the target geological data is less than a preset error threshold is used as the 3D output geological model. Uncertainty quantification and synthesis are performed on each 3D output geological model to obtain an uncertain geological model corresponding to the target area. This ensures that the geological data in the uncertain geological model combines the output geological data corresponding to multiple 3D output geological models, avoiding the uniformity of geological model data. Furthermore, the uncertain geological model includes all possible real geological conditions in the target area, and the geological data corresponding to the highest probability is taken as the most likely real geological condition in the target area.

[0016] Other features and advantages of the embodiments of the present invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above in the embodiments of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of a geological model generation method provided by an embodiment of the present invention is shown; Figure 2 This illustration shows a flowchart of a geological model generation method for generating a three-dimensional model of a target area, as provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] This embodiment provides a method for generating geological models, which can be applied to electronic devices such as computers. See [link to documentation]. Figure 1 The diagram shows a flowchart of a geological model generation method, which mainly includes the following steps: Step S100: Obtain target geological data for the target area; The above steps involve collecting all available geological data within the target area (typically discrete, limited-quantity borehole data obtained during geological exploration of the mine, including coordinates, depth, lithology, coal thickness, grade, etc., and geological profiles corresponding to the target area), standardizing the target geological data (including coordinate system 1, data cleaning (outlier removal), and data normalization), and mapping the processed target geological data to a target 3D voxel grid using spatial interpolation (e.g., nearest neighbor interpolation to fill blank areas) or rasterization algorithms. The resolution of this 3D voxel grid has been preset (for example, the voxel size in all three directions can be set to 10m; or the voxel size in each of the three directions can be set to 1km×1km×0.5km respectively). Each voxel in the mapped target 3D voxel grid contains its position information. If the voxel in the target 3D voxel grid has corresponding target geological data, the voxel attribute value of the voxel is the measured value; if not, the voxel attribute value of the corresponding voxel is marked as a missing value (NaN).

[0022] Step S200: Obtain multiple low-dimensional noise vectors and input each low-dimensional noise vector into the trained geological model generation module to obtain multiple three-dimensional generated geological models; screen each three-dimensional generated geological model based on the target geological data to determine the three-dimensional output geological model in each three-dimensional generated geological model; wherein, the relative error between the output geological data corresponding to each three-dimensional output geological model and the target geological data is less than a preset error threshold; the trained geological model generation module is obtained by training with real three-dimensional geological models and training noise vectors; By inputting each low-dimensional noise vector (for example, 1000 different low-dimensional noise vectors can be obtained) into the trained geological model generation module, a "probable" three-dimensional generated geological model of the geological laws corresponding to each low-dimensional noise vector can be obtained. Based on the target geological data of the target area, each three-dimensional generated geological model is screened. The target geological data corresponding to each voxel in the target three-dimensional voxel grid is compared with the geological data corresponding to each voxel in the three-dimensional generated geological model. If the relative error between the geological data corresponding to the three-dimensional generated geological model and the target geological data corresponding to the target three-dimensional voxel grid is less than the preset error threshold (which can be set to 5%), then the three-dimensional generated geological model is used as the output of the trained geological model generation module to obtain the three-dimensional output geological model.

[0023] Step S300: Perform uncertainty quantification and synthesis on each three-dimensional output geological model to obtain the uncertainty geological model corresponding to the target area; In the above steps, a three-dimensional output geological model dataset is constructed based on each three-dimensional output geological model. The output geological data corresponding to each three-dimensional output geological model are then synthesized by uncertainty quantification to obtain the uncertainty geological model corresponding to the target area. The uncertainty geological model intuitively shows which areas in the uncertainty geological model are reliable and which areas have greater variables, providing key risk assessment basis for subsequent engineering decisions.

[0024] The method provided in this embodiment of the invention obtains target geological data for a target area. Since the trained geological model generation module has been trained using a real 3D geological model and training noise vectors, inputting multiple low-dimensional noise vectors into the trained geological model generation module can directly obtain multiple 3D generated geological models corresponding to the target area. The geological data corresponding to each 3D generated geological model is compared with the target geological data of the target area. The 3D generated geological model corresponding to the relative error between the geological data and the target geological data is less than a preset error threshold is used as the 3D output geological model. Uncertainty quantification and synthesis are performed on each 3D output geological model to obtain the uncertain geological model corresponding to the target area. This allows the geological data in the uncertain geological model to combine the output geological data corresponding to multiple 3D output geological models, avoiding the uniformity of geological model data. Furthermore, the uncertain geological model contains all possible real geological conditions in the target area. The geological data corresponding to the highest probability is taken as the most likely real geological condition in the target area, effectively avoiding safety accidents during mine production.

[0025] In one embodiment, the specific implementation method provided in this embodiment for synthesizing uncertainties from various three-dimensional output geological models to obtain an uncertain geological model corresponding to the target area includes: Step S301: Determine uncertainty parameters and uncertainty indices based on the output geological data corresponding to each three-dimensional output geological model; wherein, uncertainty parameters include variance, mean, and confidence interval; uncertainty indices include coefficient of variation; The uncertainty quantification process in the above steps refers to the process of quantitatively evaluating the uncertainty of the results of each predicted three-dimensional output geological model (i.e., evaluating the probability of the existence of a certain geological structure (such as a fault) in the target area through multiple three-dimensional output geological models). Specifically, the output geological data corresponding to each three-dimensional output geological model are combined to determine the variance and mean of the output geological data corresponding to the three-dimensional output geological model, and a confidence interval is constructed based on the variance and mean of the output geological data (for example, it can be set to a 95% confidence interval). The uncertainty index (i.e., coefficient of variation) is determined based on the variance and mean of the output geological data. Among them, the above uncertainty parameters also include the standard deviation and expectation of the output geological data corresponding to the three-dimensional output geological model.

[0026] Step S303: Map the uncertainty parameters and uncertainty indices to the preset geological grid corresponding to the target area to obtain the uncertainty geological model corresponding to the target area; In the above steps, the half-width of the confidence interval is determined based on the confidence interval, the variance (or standard deviation, or half-width of the confidence interval) corresponding to the output geological data is mapped to the preset geological raster corresponding to the target area, and the coefficient of variation is mapped to the preset geological raster to obtain the uncertain geological model corresponding to the target area.

[0027] In one embodiment, if the number of samples of the three-dimensional output geological model is 1000, and 800 of the three-dimensional output geological models show that the target area is a coal seam and 200 of the three-dimensional output geological models show that the target area is a fault, then the probability that there is a fault in the target area can be quantified as 20%.

[0028] In one embodiment, the specific implementation method for determining uncertainty parameters and uncertainty indices based on the output geological data corresponding to each three-dimensional output geological model provided in this embodiment includes: Step S3011: Determine the uncertainty parameters corresponding to each volume element position based on the output geological data corresponding to the volume elements at the same volume element position in each three-dimensional output geological model. Determine the size of the target area (optional: 1km × 1km × 0.5km). Based on the target area size, set the voxel size of the preset geological raster (i.e., detail resolution, which can be set to 15.6m × 15.6m × 7.8m) to ensure the voxels capture the geological features of the target area (e.g., minimum fault width). Since voxels are the basic units in 3D raster data, similar to pixels in a 2D image, obtain the output geological data corresponding to each voxel in each 3D output geological model. The output geological data corresponding to each voxel reflects the geological attributes at the voxel's location in the 3D output geological model (e.g., lithology (encoded by integers), coal seam thickness (meters), gas content (m³ / t), etc.). Since each 3D output geological model corresponds to the geological conditions of the target area, the structures of each 3D output geological model are identical. An output geological data set can be constructed based on the output geological data corresponding to voxels at the same voxel location in each 3D output geological model. ; in, This is the set of output geological data corresponding to the voxel at position i in all three-dimensional output geological models. This refers to the output geological data corresponding to the volume element at position i in the j-th three-dimensional output geological model. Based on the output geological data set, the variance (which can be used as a measure of uncertainty of the output geological data), mean, standard deviation and expectation (which can be used as the best estimate of the output geological data, representing the most likely geological properties of the target area at the location of the volume element) of the output geological data corresponding to each volume element location are obtained. The central limit theorem in statistics states that when the sample size is large enough (N>30), the sample mean approximately follows a normal distribution. Since multiple three-dimensional output geological models are obtained by using 1000 low-dimensional noise vectors as input, the number of samples of the generated three-dimensional output geological models is sufficient to ensure that the output geological data corresponding to the three-dimensional output geological models approximately follows a normal distribution. Based on the variance and mean of the output geological data corresponding to each voxel position, a confidence interval (e.g., a 95% confidence interval) is constructed for the output geological data corresponding to each voxel position. The confidence interval gives the range of possible values ​​of the geological attributes corresponding to the target area at that voxel position. A confidence level of 95% means that the true value has a 95% probability of falling within the confidence interval.

[0029] Step S3012: Calculate the ratio of variance to mean for each voxel position to obtain the coefficient of variation for each voxel position. Calculate the ratio of variance to mean of the output geological data corresponding to each volume element location: ; in, is the coefficient of variation corresponding to position i of the volume element; Let be the variance corresponding to position i of the volume element; Let be the mean value corresponding to position i of the volume element.

[0030] In one embodiment, the specific implementation method provided in this embodiment for mapping uncertainty parameters and uncertainty indices to a preset geological grid corresponding to a target area to obtain an uncertainty geological model corresponding to the target area includes: Step S3031: Map the uncertainty parameters and uncertainty indices corresponding to each voxel position to the voxels at the voxel positions corresponding to the preset geological grid to obtain the uncertainty geological model; Determine the half-width of the confidence interval, and map the variance (or standard deviation, or half-width of the confidence interval) and uncertainty index (i.e., coefficient of variation) corresponding to each voxel location as geological data to the voxels corresponding to the preset geological raster to obtain an uncertainty geological model (i.e., a three-dimensional uncertainty distribution map). This uncertainty geological model intuitively shows which areas in the model have reliable predictions (small variance) and which areas have huge uncertainties (large variance). Based on the threshold segmentation algorithm, the voxels corresponding to the coefficient of variation greater than the preset coefficient of variation threshold (which can be set to 0.3) are marked as high uncertainty regions, and the voxels corresponding to the coefficient of variation less than or equal to the preset coefficient of variation threshold (which can be set to 0.3) are marked as low uncertainty regions. Based on the Moving Cubes (MC) algorithm, different levels of uncertainty isosurfaces (e.g., high uncertainty isosurfaces composed of high uncertainty regions; low uncertainty isosurfaces composed of low uncertainty regions) are extracted from the labeled uncertainty geological model. By using different color codes to characterize isosurfaces at different levels in the uncertain geological model (for example, characterizing low-uncertainty isosurfaces as blue and high-uncertainty isosurfaces as red), the final uncertain geological model corresponding to the target area is obtained. Since the high-uncertainty area is clearly identified in the uncertain geological model, the model can be used as a probabilistic basis model for providing safety decisions. At the same time, the model can realize interactive three-dimensional visualization, allowing personnel to explore the spatial distribution of uncertainty corresponding to the target area.

[0031] In one embodiment, the specific implementation method provided in this embodiment for inputting each low-dimensional noise vector into the trained geological model generation module to obtain multiple three-dimensional generated geological models includes: Step S201: Based on the trained geological model generation module, each low-dimensional noise vector is mapped to the corresponding feature tensor. Step S203: Upsample each feature tensor to obtain the three-dimensional generated geological model corresponding to each low-dimensional noise vector; Obtain a low-dimensional noise vector (typically a 128-dimensional random low-dimensional noise vector). Specifically, the low-dimensional noise vector is a d-dimensional vector z ∈ [0,1] randomly sampled from a standard normal distribution P(0,1) or a uniform distribution U[-1,1]. Where z is a low-dimensional noise vector, and d is usually set to 100-200. The low-dimensional noise vector (i.e., low-dimensional Gaussian noise vector) is used as the initial input of the generator in the geological model generation module (3D generative adversarial network). Its mathematical representation is z~M(0,I), where I is a 128×128 identity matrix. The low-dimensional vector carries randomness and can be used as a seed when generating diverse 3D geological models. The generator based on the trained geological model generation module (which can be set to an 8-layer deconvolutional neural network) maps low-dimensional noise vectors into feature tensors and performs multi-step upsampling on the feature tensors to finally obtain a three-dimensional generated geological model with the same dimensions as the target geological data sample (the voxel size of the voxel in this model is the same as the voxel size of the target three-dimensional voxel grid). Specifically, the 128-dimensional low-dimensional noise vector is mapped to a 4×4×4×512 feature tensor through the fully connected layer in the generator. The feature tensor is upsampled by a 3D deconvolution layer (kernel size 4×4×4, stride 2, tanh activation) to obtain a first upsampled model of size 8×8×8×256. The first upsampled model is then processed by batch normalization layer (BatchNorm layer) and Leaky Rectified Linear Unit (LeakyReLU) activation function (the slope of which can be set to 0.2) to obtain the processed first upsampled model. The first upsampled model after processing is upsampled by a 3D deconvolution layer to obtain a second upsampled model of size 16×16×16×128. The second upsampled model is then processed by batch normalization and linear rectification by a batch normalization layer and a leaky linear rectified activation function to obtain the processed second upsampled model. The processed second upsampling model is upsampled by a 3D deconvolution layer to obtain a third upsampling model of size 32×32×32×64. The third upsampling model is then processed by batch normalization and linear rectification using a batch normalization layer and a leaky linear rectified activation function to obtain the processed third upsampling model. A 3D generated geological model of size 64×64×64×1 is obtained by upsampling the processed third upsampled model through a 3D deconvolution layer.

[0032] In one embodiment, the specific implementation of the training geological model generation module provided in this embodiment includes: Step S400: Obtain a real 3D geological model and multiple training noise vectors; Since a three-dimensional geological model is a digital model that expresses the spatial distribution and attributes of geological objects such as underground rock strata, coal seams, and structures in the form of three-dimensional voxels (i.e., three-dimensional voxels) or grids, a real three-dimensional geological model (one or more can be selected) is constructed based on real geological data. In the real three-dimensional geological model, voxels that correspond to real geological data are marked as measured voxels; voxels that do not correspond to real geological data are marked as missing voxels. Multiple training noise vectors are obtained from a standard normal distribution or a uniform distribution (a 128-dimensional random low-dimensional noise vector can be used).

[0033] Step S500: Establish a training dataset based on the real 3D geological model and each training noise vector; A training dataset is established based on a real 3D geological model and its corresponding training noise vector, which is used to train the geological model generation module.

[0034] Step S600: Input the training dataset into the three-dimensional generative adversarial network for training to obtain the trained geological model generation module; The 3D Generative Adversarial Network (3D-GAN) mentioned in the above steps is a deep learning model consisting of a generator and a discriminator. It learns the distribution of real data through adversarial competition between the two, thereby generating new, similar data. Specifically, the 3D-GAN is trained based on real 3D geological models and their corresponding training noise vectors in the training dataset to obtain the trained geological model generation module. The discriminator can be configured as an 8-layer 3D convolutional neural network (each convolutional layer is followed by a batch filter). The chNorm layer and LeakyReLU activation function are used to downsample the input 3D geological model through each layer of the convolutional neural network. Finally, the scalar value of the discriminator when it judges the training 3D geological model as a real 3D geological model is obtained. The scalar value is normalized based on the Sigmoid function to obtain the true probability (probability interval is [0,1]) corresponding to each training 3D geological model. The generator can be set to an 8-layer 3D deconvolutional neural network. The above-mentioned 3D generative adversarial network can also be replaced by other generative models such as 3D variational autoencoder (3D-VAE) or normalizing flow model.

[0035] In one embodiment, the specific implementation method of inputting the training dataset into a three-dimensional generative adversarial network for training to obtain a trained geological model generation module includes: Step S601: Train the generator in the three-dimensional generative adversarial network based on each training noise vector input to obtain the training three-dimensional geological model and generator loss function corresponding to each training noise vector; In the above steps, when training the generator in the 3D generative adversarial network, it is necessary to fix the learning parameters of the discriminator, input each training noise vector into the generator in the 3D generative adversarial network for training, generate diverse training 3D geological models, and determine the loss function corresponding to the generator based on the training geological data corresponding to the training 3D geological models.

[0036] Step S603: Train the discriminator in the three-dimensional generative adversarial network based on each trained three-dimensional geological model and the real three-dimensional geological model to obtain the discriminator loss function; In the above steps, when training the discriminator in the 3D generative adversarial network, it is necessary to fix the learning parameters of the generator, input the real 3D geological model and its corresponding generated training 3D geological model into the discriminator in the 3D generative adversarial network for training, and obtain the discriminator loss function based on the probability that each geological model is a real 3D geological model.

[0037] Step S605: Based on the generator loss function and the discriminator loss function, determine whether the trained generator and the trained discriminator have reached Nash equilibrium. If they have not reached Nash equilibrium, repeat the training of the generator and the discriminator until they reach Nash equilibrium. If they have reached Nash equilibrium, complete the training of the three-dimensional generative adversarial network and obtain the trained geological model generation module. Since training a 3D generative adversarial network is an adversarial game between the generator and the discriminator, the generator continuously learns to generate a 3D geological model that can "deceive" the discriminator. The discriminator, on the other hand, continuously learns to more accurately distinguish between the real 3D geological model and the training 3D geological model, while also supervising whether the training 3D geological model conforms to geostatistical laws. Therefore, based on the generator loss function and the discriminator loss function, it is determined whether the trained generator and the trained discriminator have reached Nash equilibrium (i.e., dynamic equilibrium). If the trained generator and the trained discriminator have not reached Nash equilibrium, it is determined that the trained generator and the trained discriminator have not yet reached a stable state. Then, the above training steps for the generator and the trainer are repeated, and the generator and the discriminator are repeatedly trained using the training dataset until the trained generator and the trained discriminator reach Nash equilibrium. If the trained generator and the trained discriminator reach Nash equilibrium, it can be determined that the trained generator and the trained discriminator have reached a stable state. At this time, the generator has the ability to generate a three-dimensional geological model that is both highly realistic and strictly conforms to the laws of geological spatial structure, and the trained geological model generation module is obtained.

[0038] In one embodiment, the specific implementation method provided in this embodiment for training a generator in a three-dimensional generative adversarial network based on each training noise vector input to obtain a trained three-dimensional geological model and a generator loss function corresponding to each training noise vector includes: Step S6011: Input each training noise vector into the generator in the three-dimensional generative adversarial network for training to obtain each training three-dimensional geological model. With the learning parameters of the discriminator fixed, each training noise vector is input into the generator in the 3D generative adversarial network for training. Each training noise vector is then upsampled to obtain the corresponding 3D geological model.

[0039] Step S6012: Calculate the first average relative error between the training geological data corresponding to each training 3D geological model and the real geological data corresponding to the real 3D geological model as the generator loss function. Obtain the corresponding training geological data from each training 3D geological model, calculate the relative error between the training geological data corresponding to each training 3D geological model and the real geological data corresponding to the real 3D geological model, and calculate the mean of the relative error to obtain the first average relative error as the generator loss function.

[0040] In one embodiment, the specific implementation method provided in this embodiment for training the discriminator in a three-dimensional generative adversarial network based on various trained three-dimensional geological models and real three-dimensional geological models to obtain the discriminator loss function includes: Step S6031: Determine the true variation function corresponding to the true three-dimensional geological model based on real geological data; The realistic 3D geological model is divided according to the three-dimensional principal directions of its corresponding mine strata (X-axis: mine strike; Y-axis: mine dip; Z-axis: vertical). Sixteen azimuth intervals are set in each direction to ensure comprehensive capture of the anisotropic characteristics of the realistic 3D geological model. Valid data point pairs are determined in each direction, and these pairs are grouped according to the distance between the valid data points. The grouping interval is set to the range prediction value. Each set of data points should contain no fewer than 30 pairs to avoid statistical errors; The experimental variation function was calculated based on the actual geological data corresponding to the measured volume elements in the real three-dimensional geological model: ; in, Let be the discrete experimental variation function (i.e., semivariance), h be the spatial separation vector (which can be along the X, Y, or Z axes), and N(h) be the number of valid data pairs spaced h apart. Let represent the three-dimensional coordinates (i.e., spatial position) of the a-th data point in the real three-dimensional geological model. Divide the real three-dimensional geological model into multiple distance intervals along the principal directions (e.g., X, Y, Z axes). Calculate the variance of all valid data pairs within each interval, and obtain the average variance value of the valid data pairs corresponding to each interval. This yields the values ​​corresponding to different lag distances h. value; Based on the experimental variogram, the experimental variogram curves of the real 3D geological model in various directions are obtained. A theoretical variogram model (such as a spherical model, exponential model, or Gaussian model) is selected to fit the experimental variogram, so that the discrete experimental variogram is fitted into a continuous and deterministic variogram. The squared error between the experimental variogram and the spherical model is minimized by an optimization method (such as the least squares method). Specifically, the spherical model is used as the theoretical variogram model, and the squared error between the experimental variogram and the spherical model is minimized by an optimization method (least squares method). ; in, The minimum squared error between the experimental variation function and the spherical model. The theoretical variation function model (a spherical model) is used, where θ is the model parameter vector; based on the coefficient of determination. Determine the goodness of fit between the experimental variation function and the spherical model, and select the coefficient of determination. The theoretical variation function model corresponding to the maximum value is used as the final theoretical variation function model; Key parameters of the theoretical variation function model (spherical model) were derived through regression: ; in, Let be the fitted experimental variation function. denoted as nugget value, representing microscale variation and / or measurement error; C is the partial sill value, representing the structural variation caused by spatial autocorrelation; d is the range, representing the maximum distance at which spatial autocorrelation exists. The total sill value represents the total spatial variation; The key parameters of the theoretical variogram model are iteratively adjusted based on the Levenberg-Marquardt algorithm until each key parameter converges to the optimal solution. The fitted experimental variogram is then used as the "prior knowledge" (i.e., the real variogram) corresponding to the real geological model to constrain the generation of subsequent geological models.

[0041] Step S6032: Input each training 3D geological model and the real 3D geological model into the discriminator in the 3D generative adversarial network to obtain the true probability and training variation function corresponding to each training 3D geological model; wherein, the true probability is the probability value corresponding to the discriminator judging the training 3D geological model as the real 3D geological model; The sliding window algorithm is used to extract multiple sub-regions (e.g., each sub-region has a 64×64×64 voxel mesh) from the real 3D geological model as real samples. Data augmentation processing is performed on each real sample (including random rotation (90°, 180°, 270°), mirror flip (X / Y / Z axes), and slight affine transformation) to increase the data diversity of the real samples. Finally, conditional information (such as the boundary of the work area and the location of the main fault lines) is added to each real sample to form a sample 3D geological model corresponding to the real 3D geological model. With the generator's learning parameters fixed, each sample 3D geological model and each training 3D geological model are input into the discriminator in the 3D generative adversarial network. The true probabilities of the discriminator judging each sample 3D geological model as a real 3D geological model and the true probabilities of the discriminator judging each training 3D geological model as a real 3D geological model are obtained. The true probabilities are in the range [0,1]. When the true probability is 1, it indicates that the discriminator judges the input training 3D geological model or sample 3D geological model as a real 3D geological model. When the true probability is 0, it indicates that the discriminator judges the input training 3D geological model or sample 3D geological model as a training 3D geological model generated by the generator. The closer the true probability is to 1, the closer the 3D geological model is to a real 3D geological model; the closer the true probability is to 0, the closer the 3D geological model is to the training 3D geological model generated by the generator. The discriminator in the three-dimensional distributed generative adversarial network provided in this embodiment of the invention integrates a variogram calculation module. Based on the variogram calculation module, the training variogram corresponding to each training three-dimensional geological model is determined. Specifically, the training variogram is calculated along the three main directions X, Y, and Z of the training three-dimensional geological model. The spatial lag distances corresponding to the three main directions are divided into T intervals (e.g., ...). That is, the spatial lag distance of the e-th trained 3D geological model in the X principal direction is divided into ten intervals; for each lag distance interval (e.g., [ , +Δ ],in, Let Δ be the lag distance of the e-th trained 3D geological model in the b-th segment along the X principal direction; Let be the length of each lag distance of the e-th trained 3D geological model in the X principal direction. Calculate the squared average of the training geological data differences for all valid data point pairs that satisfy this distance range. Construct the training variogram (i.e., variogram curve) of the e-th trained 3D geological model in the three principal directions. , , .

[0042] Step S6033: Obtain the image quality index between each trained 3D geological model and the real 3D geological model, and obtain the average image quality index based on each image quality index; wherein, the image quality index includes the Fraser initial distance score; Extract the training image features corresponding to each training 3D geological model and the real image features corresponding to the real 3D geological model; Determine the training mean vector and training covariance matrix corresponding to each training image feature, and determine the true mean vector and true covariance matrix corresponding to the true image features; Each training mean vector, each training covariance matrix, the true mean vector, and the true covariance matrix are input into the Fréchet distance formula to obtain the image quality index (Fréchet Inception Distance, FID) between each trained 3D geological model and the true 3D geological model. Based on each image quality index, the average image quality index (i.e., the average Fréchet Inception Distance) is obtained.

[0043] Step S6034: Determine the average true probability based on each true probability; The average true probability of the training 3D geological model is calculated based on the true probability of each training 3D geological model.

[0044] Step S6035: Calculate the function similarity loss between each training variation function and the true variation function; wherein, the function similarity loss includes the average relative error, mean square error or L2 norm error between each training variation function and the true variation function. The above steps calculate the second average relative error (or mean square error, or L2 norm error) between each training variogram and the true variogram. True variation function (i.e., the fitted experimental variogram) is decomposed along the principal direction to determine the true variogram of the real 3D geological model in the principal direction X. The true variation function in the principal direction Y The true variation function in the principal direction Z ; The average training variation function of the trained 3D geological model in each principal direction is determined based on the training variation function of each trained 3D geological model in each principal direction. The function similarity loss is determined based on the true variation function of the real 3D geological model in each principal direction and the average training variation function of the trained 3D geological model in each principal direction: ; in, For functional similarity loss; The average training variation function for training the 3D geological model in the principal direction X; The average training variation function for training the 3D geological model in the principal direction Y; The average training variation function for training the three-dimensional geological model in the principal direction Z; , , These are the weights in the main direction X, the main direction Y, and the main direction Z, respectively. Represents the L2 norm; The function similarity loss ensures that the generated training 3D geological model conforms to geostatistical laws in terms of spatial structure; the final calculated complete function similarity loss is: ; in, It is the complete functional similarity loss; is the geological constraint weight coefficient (usually set to 0.5); G(z,c) is the generated training 3D geological model; It is a true three-dimensional geological model.

[0045] Step S6036: Use the average image quality index, average true probability, and function similarity loss as the discriminator loss function; The average image quality metric (i.e., the Fraser initial distance score), the average true probability, and the function similarity loss (e.g., the second average relative error, mean square error, and L2 norm error) are used as the discriminator loss function.

[0046] In one embodiment, the method provided in this embodiment determines whether the trained generator and discriminator have reached Nash equilibrium based on the generator loss function and the discriminator loss function. If Nash equilibrium has not been reached, the generator and discriminator are trained repeatedly. If Nash equilibrium has been reached, the training of the 3D generative adversarial network is completed, and the trained geological model generation module is obtained. The specific implementation method includes: Step S6051: Determine training completion requirements based on the generator loss function and the discriminator loss function; wherein, the training completion requirements include a first completion requirement, a second completion requirement, a third completion requirement, a fourth completion requirement, and a fifth completion requirement; the completion condition for the first completion requirement is that the fluctuation amplitude of the first average relative error in a preset training batch is less than a first preset fluctuation amplitude threshold; the completion condition for the second completion requirement is that the fluctuation amplitude of the function similarity loss in a preset training batch is less than a second preset fluctuation amplitude threshold; the completion condition for the third completion requirement is that the fluctuation amplitude of the average image quality index in a preset training batch is less than a third preset fluctuation amplitude threshold; the completion condition for the fourth completion requirement is that the value of the average true probability is within a preset average true probability range; the completion condition for the fifth completion requirement is that the function similarity loss is less than a preset similarity loss threshold; The first completion requirement in the above steps is that the fluctuation range of the first average relative error in the preset training batches (which can be set to 50 consecutive training batches) is less than the first preset fluctuation range threshold (which can be set to 0.01); the second completion requirement is that the fluctuation range of the function similarity loss in the preset training batches (which can be set to 50 consecutive training batches) is less than the second preset fluctuation range threshold (which can be set to 0.01); the third completion requirement is that the fluctuation range of the average image quality index in the preset training batches (which can be set to 50 consecutive training batches) is less than the third preset fluctuation range threshold (which can be set to 10); the fourth completion requirement is that the value of the average true probability is within the preset average true probability range (which can be set to 50±5%); and the fifth completion requirement is that the function similarity loss is less than the preset similarity loss threshold (which can be set to 10%).

[0047] Step S6052: Based on the training completion requirements, determine whether the trained generator and the trained discriminator have reached Nash equilibrium. If any one or more of the first, second, third, fourth, and fifth completion requirements are not met, then it is determined that the trained generator and the trained discriminator have not reached Nash equilibrium. Repeat the training of the generator and the discriminator until the trained generator and the trained discriminator reach Nash equilibrium. Step S6053: If the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement all meet the completion conditions, then it is determined that the trained generator and the trained discriminator have reached Nash equilibrium, completing the training of the three-dimensional generative adversarial network and obtaining the trained geological model generation module.

[0048] The specific implementation of the geological model generation method provided in this invention also includes: constructing an output geological model set based on the three-dimensional output geological model output by the generator in the trained geological model generation module; dividing the three-dimensional output geological models in the output geological model set into an AI model training set and an AI model test set in a ratio of 7:3, which are used to train and test intelligent mining AI models such as the adaptive cutting model of the coal mining machine and the dynamic prediction model of gas outburst, thereby solving the problem of scarce three-dimensional geological model data; By combining uncertain geological models with mining plans, the mining risks at different locations within the target area can be simulated (e.g., areas with variance ≥ 0.2 are defined as high-risk areas, requiring optimization of roadway layout or additional drilling verification), providing a quantitative risk basis for mine mining planning.

[0049] In the geological model generation method provided in this embodiment of the invention, when training the three-dimensional generative adversarial network, a trained geological knowledge graph can be used to guide the generation process of the generator to ensure that the generated stratigraphic sequence conforms to the regional geological laws; rock mechanics equations can be introduced as constraints to make the generated three-dimensional geological model mechanically balanced; and seismic interpretation data, gravity or magnetic exploration data can be introduced as additional input conditions for the generator to generate a three-dimensional geological model that matches more exploration evidence.

[0050] Based on the foregoing embodiments, this embodiment provides an example of generating a geological model of a target area using the aforementioned geological model generation method. See [link to example]. Figure 2 The diagram illustrates a process for generating a 3D model of a target area based on a geological model generation method. The specific steps are as follows: Step S701: Obtain a real 3D geological model and multiple training noise vectors, and construct a training dataset based on the real 3D geological model and each training noise vector. Acquire real geological data, construct a real three-dimensional geological model based on the real geological data, and determine the measured volume elements and missing volume elements in the real three-dimensional geological model; Multiple training noise vectors (128-dimensional) are randomly obtained based on a standard normal or uniform distribution. Step S702: Obtain the true variation function based on the real geological data corresponding to the measured volume elements in the real three-dimensional geological model; Determine the three-dimensional principal directions in the real three-dimensional geological model, set multiple azimuth intervals (up to 16 azimuth intervals) in each principal direction, and determine the effective data point pairs in each reverse direction in the real three-dimensional geological model based on the coordinates of the measured volume elements in the real three-dimensional geological model. Determine the discrete experimental variation function based on the corresponding real geological data for the effective data points. By fitting the discrete experimental variation function to the theoretical variation function model, a continuous experimental variation function can be obtained. The continuous experimental variation function is used as the true variation function; Step S703: Input the training dataset into the three-dimensional generative adversarial network for training to obtain the generator loss function and the discriminator loss function; By fixing the learning parameters of the discriminator in the 3D generative adversarial network, and inputting each training noise vector in the training dataset into the generator in the 3D generative adversarial network for training, various training 3D geological models are obtained. Obtain the corresponding training geological data in each training 3D geological model, calculate the relative error between the training geological data corresponding to each training 3D geological model and the real geological data corresponding to the real 3D geological model, and calculate the mean of the relative error to obtain the first average relative error as the generator loss function. By fixing the learning parameters of the discriminator in the 3D generative adversarial network, each training 3D geological model and the real 3D geological model are input into the discriminator in the 3D generative adversarial network to obtain the real probability and training variation function corresponding to each training 3D geological model. The image quality index (i.e., Fraser initial distance score) between each trained 3D geological model and the real 3D geological model is obtained, and the average image quality index (i.e., average Fraser initial distance score) is obtained based on each image quality index. The average true probability of the training three-dimensional geological model is calculated based on the true probability of each training three-dimensional geological model. The function similarity loss between each training variogram and the true variogram is calculated (including the average relative error, mean square error, or L2 norm error between each training variogram and the true variogram). The average image quality index, average true probability, and function similarity loss are used as the discriminator loss function. Step S704: Determine whether the 3D generative adversarial network has been trained based on the generator loss function and the discriminator loss function, and obtain the trained geological model generation module; Training completion requirements (including the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement) are determined based on the generator loss function and the discriminator loss function. Based on the training completion requirements, it is determined whether the trained generator and the trained discriminator have reached Nash equilibrium. If any one or more of the first, second, third, fourth, and fifth completion requirements are not met, it is determined that the trained generator and the trained discriminator have not reached Nash equilibrium. The generator and discriminator are trained repeatedly until they reach Nash equilibrium. If the first, second, third, fourth, and fifth completion requirements are all met, then the trained generator and the trained discriminator are judged to have reached Nash equilibrium, the training of the three-dimensional generative adversarial network is completed, and the trained geological model generation module is obtained. Step S705: Obtain the target geological data of the target area (usually a mine area) of the three-dimensional geological model to be generated; By collecting all available target geological data within the target area (usually discrete, limited-quantity borehole data obtained during geological exploration of the mine (including information such as coordinates, depth, lithology, coal thickness, and grade), and geological profile maps corresponding to the target area), the target geological data is standardized. The processed target geological data is then mapped into a target 3D volumetric raster through spatial interpolation (such as nearest neighbor interpolation to fill blank areas) or rasterization algorithms. Step S706: Obtain multiple low-dimensional noise vectors and input each low-dimensional noise vector into the trained geological model generation module to obtain multiple three-dimensional output geological models; By fixing the learning parameters of the generator, the low-dimensional noise vectors are upsampled based on the generator to obtain a three-dimensional generated geological model. The generated geological data corresponding to the 3D generated geological model is compared with the target geological data corresponding to each voxel in the target 3D voxel grid. If the relative error between the geological data corresponding to the 3D generated geological model and the target geological data corresponding to the target 3D voxel grid is less than the preset error threshold (which can be set to 5%), then the 3D generated geological model is used as the output of the trained geological model generation module to obtain each 3D output geological model. Step S707: Perform uncertainty quantification and synthesis on each three-dimensional output geological model to obtain the uncertainty geological model corresponding to the target area; Based on the output geological data corresponding to the voxels at the same voxel position in each 3D output geological model, determine the uncertainty parameters (variance, mean, and confidence interval) corresponding to each voxel position. The ratio of variance to mean at each position of the volume element is calculated to obtain the uncertainty index (i.e., coefficient of variation) at each position of the volume element. The uncertainty parameters and uncertainty indices corresponding to each voxel position are mapped to the voxels at the voxel positions corresponding to the preset geological raster to obtain the uncertainty geological model.

[0051] The geological model generation method provided in this invention trains a three-dimensional generative adversarial network using a real three-dimensional geological model and training noise vectors. During the training process, geostatistical constraints (i.e., the true probability and function similarity loss (the difference between the true variogram and the training variogram)) are introduced as conditional constraints for the discriminator. Based on the adversarial relationship between the discriminator and the generator, the three-dimensional geological model generated by the trained generator follows the real geological spatial structure laws, ensuring that the generated three-dimensional geological model is not only a "data-driven" product but also a "physical law-driven" product. Its geological realism and reliability far exceed those of simple data augmentation or traditional generation methods based on three-dimensional generative adversarial networks. This results in the three-dimensional geological model generated by the trained generator having clear fault planes, irregular collapse column boundaries, complex fold morphology, and other fine structures, greatly improving the restoration and realism of complex geological structures. By inputting multiple low-dimensional noise vectors into the trained geological model generation module, multiple 3D generated geological models corresponding to the target area can be directly obtained. The target geological data corresponding to the target area is used as the output constraint condition of the generator in the trained geological model generation module. The 3D generated geological model corresponding to the geological data and the target geological data with a relative error of less than a preset error threshold is used as the 3D output geological model. Uncertainty quantification and synthesis are performed on each 3D output geological model to obtain the uncertain geological model corresponding to the target area. This makes the geological data in the uncertain geological model combine the output geological data corresponding to multiple 3D output geological models, avoiding the singleness of 3D geological model data. Moreover, the uncertain geological model contains all possible real geological conditions in the target area. The geological data corresponding to the highest probability is used as the most likely real geological condition in the target area. This allows personnel to clearly identify high-risk areas (such as areas with a high probability of faults) in the target area based on the uncertain geological model, thereby adaptively formulating more resilient mining plans and effectively preventing safety accidents. The geological model generation module of this invention can automatically, efficiently, and with low computational cost generate a massive amount of high-quality (i.e., geologically plausible) 3D geological models based on various random low-dimensional noise vectors. These generated 3D geological models can be constructed into an incredibly rich "virtual geological library." This "virtual geological library" can provide a vast amount of training samples for various AI algorithms (AI models), avoiding overfitting of AI models due to insufficient 3D geological model data. Furthermore, the 3D geological models in this "virtual geological library" include extreme geological scenarios (such as scenarios intentionally generated with complex conditions like large faults and thin coal seams). Inputting these 3D geological models, including those with extreme geological scenarios, into AI models for training can improve the robustness of the AI ​​models. The 3D geological models generated based on the trained geological model generation module can directly serve as the foundation for digital twins, solving the core bottleneck of scarce real 3D geological model data and greatly promoting the progress of intelligent mine construction.

[0052] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0053] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of generating a geological model, characterized by, The method comprises the following steps: obtaining target geological data of a target area; obtaining a plurality of low-dimensional noise vectors, inputting each low-dimensional noise vector into a trained geological model generation module to obtain a plurality of three-dimensional generated geological models, screening each three-dimensional generated geological model based on the target geological data to determine a three-dimensional output geological model in each three-dimensional generated geological model; wherein the relative error of the output geological data corresponding to each three-dimensional output geological model and the target geological data is less than a preset error threshold; the trained geological model generation module is trained by a real three-dimensional geological model and a training noise vector; quantifying and synthesizing the uncertainty of each three-dimensional output geological model to obtain an uncertainty geological model corresponding to the target area.

2. The geological model generation method of claim 1, wherein, The step of quantifying and synthesizing the uncertainty of each three-dimensional output geological model to obtain an uncertainty geological model corresponding to the target area comprises: determining an uncertainty parameter and an uncertainty index based on the output geological data corresponding to each three-dimensional output geological model; wherein the uncertainty parameter comprises variance, mean and confidence interval; the uncertainty index comprises a coefficient of variation; mapping the uncertainty parameter and the uncertainty index into a preset geological grid corresponding to the target area to obtain the uncertainty geological model corresponding to the target area.

3. The geological model generation method of claim 1, wherein, The step of determining an uncertainty parameter and an uncertainty index based on the output geological data corresponding to each three-dimensional output geological model comprises: determining an uncertainty parameter corresponding to each voxel position based on the output geological data corresponding to the voxels at the same voxel position in each three-dimensional output geological model; calculating the ratio of the variance and the mean corresponding to each voxel position to obtain the coefficient of variation corresponding to each voxel position.

4. The geological model generation method of claim 3, wherein, The step of mapping the uncertainty parameter and the uncertainty index into a preset geological grid corresponding to the target area to obtain the uncertainty geological model corresponding to the target area comprises: mapping the uncertainty parameter and the uncertainty index corresponding to each voxel position into the voxels at the voxel position corresponding to the preset geological grid to obtain the uncertainty geological model.

5. The geological model generation method of claim 1, wherein, The step of inputting each low-dimensional noise vector into a trained geological model generation module to obtain a plurality of three-dimensional generated geological models comprises: mapping each low-dimensional noise vector into a corresponding feature tensor based on the trained geological model generation module; performing up-sampling processing on each feature tensor to obtain the three-dimensional generated geological model corresponding to each low-dimensional noise vector.

6. The geological model generation method of claim 1, wherein, The training step of the geological model generation module comprises: obtaining a real three-dimensional geological model and a plurality of training noise vectors; establishing a training data set according to the real three-dimensional geological model and each training noise vector; inputting the training data set into a three-dimensional generative adversarial network for training to obtain the trained geological model generation module.

7. The geological model generation method of claim 6, wherein, The step of inputting the training data set into a three-dimensional generative adversarial network for training to obtain the trained geological model generation module comprises: training the generator in the three-dimensional generative adversarial network based on each training noise vector, to obtain a training three-dimensional geological model corresponding to each training noise vector and a generator loss function; training the discriminator in the three-dimensional generative adversarial network based on each training three-dimensional geological model and the real three-dimensional geological model, to obtain a discriminator loss function; determining whether the trained generator and the trained discriminator reach a Nash equilibrium based on the generator loss function and the discriminator loss function, and if not, repeatedly training the generator and the discriminator until the trained generator and the trained discriminator reach the Nash equilibrium; or if so, completing the training of the three-dimensional generative adversarial network to obtain the trained geological model generation module.

8. The geological model generation method of claim 7, wherein, The step of training the generator in the three-dimensional generative adversarial network based on each training noise vector, to obtain a training three-dimensional geological model corresponding to each training noise vector and a generator loss function, comprises: inputting each training noise vector into the generator in the three-dimensional generative adversarial network for training, to obtain each training three-dimensional geological model; calculating a first average relative error between the training geological data corresponding to each training three-dimensional geological model and the real geological data corresponding to the real three-dimensional geological model as the generator loss function.

9. The geological model generation method of claim 8, wherein, The step of training the discriminator in the three-dimensional generative adversarial network based on each training three-dimensional geological model and the real three-dimensional geological model, to obtain a discriminator loss function, comprises: determining a real variogram function corresponding to the real three-dimensional geological model based on the real geological data; inputting each training three-dimensional geological model and the real three-dimensional geological model into the discriminator in the three-dimensional generative adversarial network, to obtain a real probability corresponding to each training three-dimensional geological model and a training variogram function; wherein the real probability is a probability value corresponding to the judgment of the discriminator that the training three-dimensional geological model is a real three-dimensional geological model; obtaining an image quality index between each training three-dimensional geological model and the real three-dimensional geological model, and obtaining an average image quality index based on each image quality index; wherein the image quality index includes a Fréchet inception distance score; determining an average real probability based on each real probability; calculating a function similarity loss between each training variogram function and the real variogram function; wherein the function similarity loss includes an average relative error, a mean square error or an L2 norm error between each training variogram function and the real variogram function; taking the average image quality index, the average real probability and the function similarity loss as the discriminator loss function.

10. The geological model generation method of claim 9, wherein, determining whether the trained generator and the trained discriminator reach a Nash equilibrium based on the generator loss function and the discriminator loss function, and if not, repeatedly training the generator and the discriminator until the trained generator and the trained discriminator reach the Nash equilibrium; or if so, completing the training of the three-dimensional generative adversarial network to obtain the trained geological model generation module. The training of the three-dimensional generative adversarial network is completed if the Nash equilibrium is reached, and the trained geological model generation module is obtained, including: The training completion requirements are determined based on the generator loss function and the discriminator loss function, wherein the training completion requirements include a first completion requirement, a second completion requirement, a third completion requirement, a fourth completion requirement, and a fifth completion requirement; the first completion requirement has a requirement completion condition that the fluctuation amplitude of the first average relative error in a preset training batch is less than a first preset fluctuation amplitude threshold; the second completion requirement has a requirement completion condition that the fluctuation amplitude of the function similarity loss in a preset training batch is less than a second preset fluctuation amplitude threshold; the third completion requirement has a requirement completion condition that the fluctuation amplitude of the average image quality index in a preset training batch is less than a third preset fluctuation amplitude threshold; the fourth completion requirement has a requirement completion condition that the average real probability is within a preset average real probability range; and the fifth completion requirement has a requirement completion condition that the function similarity loss is less than a preset similarity loss threshold; It is determined whether the trained generator and the trained discriminator reach the Nash equilibrium based on the training completion requirements. If any one or more of the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement do not reach the requirement completion condition, it is determined that the trained generator and the trained discriminator do not reach the Nash equilibrium, and the generator and the discriminator are repeatedly trained until the trained generator and the trained discriminator reach the Nash equilibrium. If the first completion requirement, the second completion requirement, the third completion requirement, the fourth completion requirement, and the fifth completion requirement all reach the requirement completion condition, it is determined that the trained generator and the trained discriminator reach the Nash equilibrium, the training of the three-dimensional generative adversarial network is completed, and the trained geological model generation module is obtained.

Images