WO2024045285A1 - 基于多源异构数据的地质构造建模方法 - Google Patents

基于多源异构数据的地质构造建模方法 Download PDF

Info

Publication number
WO2024045285A1
WO2024045285A1 PCT/CN2022/125275 CN2022125275W WO2024045285A1 WO 2024045285 A1 WO2024045285 A1 WO 2024045285A1 CN 2022125275 W CN2022125275 W CN 2022125275W WO 2024045285 A1 WO2024045285 A1 WO 2024045285A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
geological
fault
model
structural
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/125275
Other languages
English (en)
French (fr)
Inventor
伍新明
毕钲发
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to US18/546,360 priority Critical patent/US20250029330A1/en
Publication of WO2024045285A1 publication Critical patent/WO2024045285A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three-dimensional [3D] modelling for computer graphics
    • G06T17/05—Geographic models

Definitions

  • the present disclosure relates to the technical fields of geological exploration and artificial intelligence, and more specifically to a geological structure modeling method based on multi-source heterogeneous data.
  • Geological modeling aims to establish numerical models of geological conditions by using various mathematical methods to accurately represent their spatial positions and geometric relationships.
  • implicit methods are generally used for geological modeling.
  • the implicit method uses structural data to calculate a spatial scalar function to represent all geological structural features in the entire region.
  • geological interfaces are represented as isosurfaces, while faults and unconformities are represented as discontinuous changes in the value of the function, also known as the implicit tectonic model.
  • the present disclosure provides a geological structure modeling method based on multi-source heterogeneous data.
  • a geological structure modeling method based on multi-source heterogeneous data including:
  • multi-source heterogeneous data about the target area, where the above-mentioned multi-source heterogeneous data includes fault interpretation data and layer interpretation data, and the above-mentioned multi-source heterogeneous data includes at least one of the following: geological outcrop observation data, well logging data, A variety of geophysical data and empirical knowledge data;
  • the above converted fault interpretation data and the above converted layer interpretation data are input into the pre-trained neural network, and a complete geological structure model corresponding to the above target area is output.
  • Another aspect of the present disclosure provides a geological structure modeling device based on multi-source heterogeneous data, including:
  • the acquisition module is configured to acquire multi-source heterogeneous data of the target area, where the above-mentioned multi-source heterogeneous data includes fault interpretation data and horizon interpretation data, and the above-mentioned multi-source heterogeneous data includes at least one of the following: geological outcrop observation Data, well logging data, various geophysical data, empirical knowledge data;
  • a conversion module configured to convert the above-mentioned fault interpretation data and the above-mentioned horizon interpretation data based on a preset model grid to obtain converted fault interpretation data and converted horizon interpretation data;
  • the input and output module is configured to input the above-mentioned converted fault interpretation data and the above-mentioned converted horizon interpretation data into a pre-trained neural network, and output a complete geological structure model corresponding to the above-mentioned target area.
  • Another aspect of the present disclosure provides an electronic device, including: one or more processors; a memory configured to store one or more programs, wherein when the above one or more programs are used by the above one or more When the processor executes, one or more processors are caused to execute the above method.
  • Another aspect of the present disclosure also provides a computer-readable storage medium on which executable instructions are stored, which when executed by a processor causes the processor to perform the above method.
  • Another aspect of the present disclosure also provides a computer program product, including a computer program that implements the above method when executed by a processor.
  • Figure 1 schematically shows a flow chart of a geological structure modeling method based on multi-source heterogeneous data according to an embodiment of the present disclosure
  • Figure 2 schematically shows a flow chart of a neural network training method according to an embodiment of the present disclosure
  • Figure 3A schematically shows an architectural diagram of a neural network according to an embodiment of the present disclosure
  • Figure 3B schematically shows a structural diagram of a reverse residual module according to an embodiment of the present disclosure
  • Figure 3C schematically shows a structural diagram of a channel attention module according to an embodiment of the present disclosure
  • Figure 4 schematically shows a flow chart of an implicit construction modeling method according to an embodiment of the present disclosure
  • Figure 5 schematically illustrates a construction method according to an embodiment of the present disclosure that generates a spatial scalar function from input sparse structural data as an implicit representation of all geological structures;
  • Figure 6 schematically shows the application of two-dimensional structural modeling based on field outcrop observation data
  • Figure 7 schematically shows a real-time interactive structural modeling application based on well logging data
  • Figure 8 schematically shows an application case of three-dimensional geological structure modeling based on actual seismic interpretation data
  • Figure 9 schematically shows a structural block diagram of a geological structure modeling device based on multi-source heterogeneous data according to an embodiment of the present disclosure
  • Figure 10 schematically shows a structural block diagram of a neural network training device for constructing a geological structure model based on multi-source heterogeneous data according to an embodiment of the present disclosure
  • Figure 11 schematically shows a block diagram of an electronic device suitable for implementing the above method according to an embodiment of the present disclosure.
  • Geological modeling aims to establish numerical models of geological conditions by using various mathematical methods to accurately represent their spatial positions and geometric relationships.
  • Structural modeling is mainly explicit or surface modeling methods, which directly use geometric elements such as points, lines, and surfaces to express geological structures. Since there are various geological interfaces in the geological body, when stratigraphic unconformity, faults, stratigraphic pinch-out, etc. occur, modeling will encounter complex geological interface intersection relationship processing problems, and there are problems between the various structures. Closely related, any error will have an impact on the modeling results.
  • modeling techniques in related technologies usually require strict quality control to ensure the geological rationality and numerical stability of the model. This makes this method inefficient when dealing with more complex structural modeling problems, and model update is very difficult.
  • the modeling process is affected by the subjective factors and experience of geologists, making it difficult to reproduce the same results even if the same structural data is input.
  • Implicit modeling methods use structural data to calculate a spatial scalar function to represent all geological structural features throughout the region.
  • geological interfaces are represented as isosurfaces, while faults and unconformities are represented as discontinuous changes in the value of the function, which is also called an implicit tectonic model.
  • the implicit modeling method integrates all available geological information into the model by integrating structural observation data and geological empirical knowledge to reconstruct underground structural characteristics from a global perspective.
  • the input data of implicit modeling methods usually include various types of modeling objects, such as spatial points, vectors, polylines and surfaces through field observations and expert interpretations.
  • Geologists and geophysicists combine empirical knowledge to infer possible structural combinations from observational data to constrain the possible geometric relationships between geological interfaces, thereby guiding the modeling behavior of implicit methods.
  • the structural characteristics of the model need to conform to geological understanding and follow the structural characteristics known in the input data. Since the local distribution of underground space is only observed in limited areas (such as highly developed mines and oil fields), it is almost impossible to directly obtain the true situation of the complete geological structure. Therefore, additional geological rules and assumptions need to be introduced into the method to constrain the modeling process. For example, implicit interpolation in related technologies uses explicit smoothing criteria to simplify local changes in the model, thereby obtaining stable and reliable results.
  • DSI discrete smooth interpolation
  • the modeling process is formulated as solving a least squares minimization problem with smoothness constraints to estimate a scalar field compatible with the input.
  • This smoothness constraint introduces empirical knowledge into the modeling process, and its basic assumption is that the expected model should be as smooth as possible, that is, the changes in the local structure should be as small as possible.
  • the scalar function is always continuous across grid cells, model grid cells cannot span discontinuous structures, and this method cannot correctly estimate gradient fields near faults or unconformities. To solve this problem, it is necessary to calculate the irregular grid by modeling the discontinuous structures independently to constrain the modeling behavior, so that the method is still valid in these cases.
  • the potential field method is another type of implicit method.
  • PFM formulates structural modeling as a double cokriging interpolation or radial basis function interpolation problem. Compared with DSI, although the models are defined on volume meshes, PFM does not need to define any mesh when calculating scalar functions. The modeling results depend entirely on the distribution of the observed data, and the influence range of each data point is determined by the selected interpolation operator. However, PFM typically solves complex mathematical equations to adjust the impact of the interpolation operator throughout the region of interest to obtain geologically meaningful results. However, this also causes the computational cost to increase rapidly with the size of the input data.
  • geological modeling is formulated as a least squares or spatial interpolation problem, and the optimal solution is solved based on various mathematical formulas under the assumption of smooth regularization.
  • known data is too sparse or unevenly distributed, it is difficult for methods in related technologies to reasonably represent complex geological geometric features and relationships, and iteratively solving large sets of equations will be very inefficient when dealing with three-dimensional models.
  • the present disclosure proposes an interactive implicit modeling method (also known as DeepISMNet) using a convolutional neural network (CNN) to address the above technical problems.
  • Convolutional neural networks are used to construct a complete geological model from a small amount of structural data.
  • the network can flexibly integrate a variety of geological constraints and prior knowledge during the training process, effectively breaking the geological construction gap in related technologies. Die technology performance bottleneck.
  • the trained network can use a small amount of data to quickly generate geological understanding and compatibility with known knowledge in multiple practical modeling applications. Construct a model of the feature.
  • the method provided by the embodiments of the present disclosure belongs to a data-driven statistical algorithm and can estimate a mapping function by minimizing a given quality evaluation criterion based on empirical knowledge and sample examples, which maps input data to a desired output.
  • the advantage of deep learning is that there is no need to solve large systems of linear or nonlinear equations at a computationally expensive cost during the prediction process.
  • CNN has shown excellent capabilities in analyzing geometric features and capturing complex nonlinear mapping relationships between input and output.
  • learning units that can be used to build CNN architectures, such as depthwise separable convolutions, attention mechanisms, and residual learning modules.
  • Deep learning provides a more flexible framework for introducing multiple structural constraints and empirical understandings into the modeling process, thereby showing excellent performance when processing implicit and discrete data.
  • the input sparse structural data can come from a variety of data sources, including geological outcrop observations, well logging data, various geophysical data and expert experience knowledge, etc.
  • the method provided by the present disclosure uniformly transforms these multi-source heterogeneous information. Taking data in the form of matrices or tensors as input, the network can efficiently calculate implicit structural models that fit various geological data.
  • embodiments of the present disclosure provide a geological structure modeling method based on multi-source heterogeneous data, including:
  • multi-source heterogeneous data about the target area, where the above-mentioned multi-source heterogeneous data includes fault interpretation data and layer interpretation data, and the above-mentioned multi-source heterogeneous data includes at least one of the following: geological outcrop observation data, well logging data, A variety of geophysical data and empirical knowledge data;
  • the above converted fault interpretation data and the above converted layer interpretation data are input into the pre-trained neural network, and a complete geological structure model corresponding to the above target area is output.
  • Figure 1 schematically shows a flow chart of a geological structure modeling method based on multi-source heterogeneous data according to an embodiment of the present disclosure.
  • the geological structure modeling method based on multi-source heterogeneous data in this embodiment includes operations S110 to S130.
  • multi-source heterogeneous data of the target area is obtained, where the multi-source heterogeneous data includes fault interpretation data and horizon interpretation data, and the multi-source heterogeneous data includes at least one of the following: geological outcrop observation data, survey data Well data, various geophysical data, empirical knowledge data.
  • the above method further includes: obtaining multi-source heterogeneous data corresponding to the target area from the geological construction database by calling a data interface.
  • the fault interpretation data and horizon interpretation data may be fault interpretation data and horizon interpretation data obtained by manually converting geological outcrop observation data, well logging data, various geophysical data, and empirical knowledge data.
  • the above-mentioned fault interpretation data and the above-mentioned horizon interpretation data are converted based on a preset model grid to obtain converted fault interpretation data and converted horizon interpretation data.
  • the above-mentioned conversion of the above-mentioned fault interpretation data and the above-mentioned horizon interpretation data using a preset model grid, and obtaining the converted fault interpretation data and the converted horizon interpretation data includes: targeting the fault interpretation data in the above-mentioned fault interpretation data For each fault, mark the point within one grid from the above-mentioned fault as the first value to obtain the above-converted fault interpretation data; for each layer in the above-mentioned layer interpretation data, mark the point one grid away from the above-mentioned layer. The points within are set as the second value to obtain the above-mentioned converted layer interpretation data, where the above-mentioned second value is the ordinate average of all scattered points on the above-mentioned layer.
  • the acquired multi-source heterogeneous data is transmitted to a preset model grid to facilitate the conversion of fault interpretation data and horizon interpretation data.
  • the network is applied to actual structural interpretation data in different areas to obtain corresponding complete geological structural models. Since structural interpretation data are usually not necessarily located on a regular grid, it is necessary to convert the horizon interpretation data and fault interpretation data to the modeling grid and obtain structural data for input into the network before practical application.
  • the output of the network is an implicit structural model that is compatible with the input structural features and is geologically sound. By extracting model isosurfaces and discontinuous jump features, complete geological horizons and fault planes can be obtained respectively.
  • the above-described converted fault interpretation data and the above-described converted layer interpretation data are input into the pre-trained neural network, and a complete geological structure model corresponding to the above-mentioned target area is output.
  • underground structural features can be restored based on a complete geological structural model corresponding to the target area, and the structural features can be visually displayed, so as to facilitate the exploration of underground natural resources and the prevention and control of geological disasters based on the underground structural characteristics.
  • the structural parameters in the geological structural model can also be adjusted to adjust the structural characteristics, and the structural characteristics can be displayed visually in real time to realize the geological structure. Real-time interactivity of the model.
  • FIG. 2 schematically shows a flow chart of a training method for a neural network according to an embodiment of the present disclosure.
  • the training method in this embodiment includes operations S210 to S230.
  • geological simulation structural models In operation S210, data simulation technology is used to generate multiple geological simulation structural models to obtain a geological simulation structural model library.
  • the above geological simulation structural models have diverse fold and fault characteristics.
  • the above-mentioned use of data simulation technology to generate a geological simulation structural model and obtain a geological simulation structural model library includes: constructing an initial geological simulation structural model, wherein the above-mentioned initial geological simulation structural model is constant in the horizontal direction and monotonically increasing in the vertical direction.
  • the layered model, the above-mentioned initial geological simulation structural model includes at least one structural parameter; wherein the above-mentioned structural parameters include at least one of fold parameters, tilt parameters, and fault structural parameters; for each structural parameter in the above-mentioned at least one structural parameter, Adjust the above structural parameters within the preset range to obtain a geological simulation structural model library.
  • the convolutional neural network CNN can learn representative knowledge to achieve various applications in the field of geological modeling generalization on.
  • a workflow is proposed that uses a set of random parameters to generate a series of typical fold and fault structures and obtains a large number of training samples through data simulation.
  • the workflow starts with a layered model with constant horizontal direction and monotonically increasing vertical values as the initial structural model, and then sequentially adds folds, dips, and fault structures to the model, gradually complicating the geometric features of the model.
  • By randomly selecting parameters within a reasonable range a variety of structural models that are consistent with geological understanding can be generated and are not limited to a specific geometric model.
  • the data simulation method parametrically represents the deformation process of stratigraphic folds, dips, and fault structures based on geological assumptions about typical fold and fault patterns.
  • the preset range corresponding to each of the above construction parameters can be determined based on data from different data sources.
  • data sources can include geological outcrop observations, well log data, various geophysical data, and expert experience knowledge.
  • each training sample data in the above-mentioned training sample data set includes a geological simulation structural model and the above-mentioned geological fault data and the above-mentioned geological layer data corresponding to the above-mentioned geological simulation structural model, and the above-mentioned geological simulation structural model and Fault and horizon data together serve as training labels.
  • geological layer data and geological fault data are used as input data
  • the neural network outputs a complete implicit geological structural model
  • the geological simulation structural model itself and the input structural data are jointly As labels for neural network training to provide supervision of prediction results.
  • the above-mentioned acquisition of geological fault data and unevenly distributed geological layer data of the geological simulation structural model, and obtaining the training sample data set includes: extracting discontinuous features in the above-mentioned geological simulation structural model, and obtaining the above-mentioned geological faults data; select a preset number of isosurfaces from the isosurfaces of the above-mentioned geological simulation structural model to obtain a target isosurface set; for each target isosurface in the above target isosurface set, convert the above target isosurfaces The scatter point data on the surface is divided into multiple data intervals according to the ordinate value range; the second preset number of data intervals among the above multiple data intervals are randomly discarded, and the scatter point data in the remaining data intervals are used as the above-mentioned geological layers. bit data; generate the above training sample data based on the above geological fault data, the above geological layer data and the above geological simulation structural model, and finally obtain the above training sample
  • a method for extracting geological layer data includes: arranging all scattered points on a single layer in descending order according to the ordinate, and dividing them into multiple evenly spaced intervals according to the ordinate value range, and then randomly discarding one or All scatter points within multiple intervals, thereby generating horizon data similar to human interpretation.
  • portions are randomly removed from the horizon scatter data to simulate the sparse and unevenly distributed horizon interpretation in real situations.
  • selecting a preset number of isosurfaces from the isosurfaces of the above-mentioned geological simulation structural model to obtain a target isosurface set includes: combining the value ranges of all isosurfaces in the above-mentioned geological simulation structural model The range is divided into at least one numerical interval in descending order; for each numerical interval in the above-mentioned at least one numerical interval, a value is randomly selected from the above-mentioned numerical interval to obtain the target value; the isosurface corresponding to the above-mentioned target value is determined as the target, etc. value surface to obtain the above target isovalue surface set.
  • the geological simulation structure model represents the corresponding geological sequence interface, and the jump change of its local scalar value represents discontinuous structures such as faults or unconformity surfaces, the geological simulation structure
  • the model itself can be viewed as an implicit scalar function.
  • horizon scatter data can be obtained through an isosurface extraction method.
  • this disclosure divides the value range of the geological simulation structural model into several groups of evenly spaced intervals in descending order, and then randomly selects within each interval A value and extract its corresponding isosurface.
  • the above method further includes: the input of the neural network is scatter data representing layer and fault positions.
  • the layer data and fault data need to be converted to a regular model network, and the input data only has valid values on the known structural features, and the rest are 0 .
  • a method for converting horizon data and fault data to a regular model network includes: for each fault in the fault data, marking points within one grid from the fault as a first value , to obtain the converted fault data; for each horizon in the horizon data, set the points within one grid from the horizon to the corresponding scalar value in the geological simulation structural model to ensure that the same horizon The scatter points have the same value or label, and the converted horizon data is obtained.
  • the neural network is trained using the above training sample data set to obtain a neural network that can be used for geological modeling of multi-source heterogeneous data.
  • the above-mentioned training of the neural network using the above-mentioned training sample data set to obtain a neural network that can be used for multi-source heterogeneous data geological structure modeling includes: inputting the above-mentioned fault data and the above-mentioned layer data into the above-mentioned neural network , output the geological structure model; input the above-mentioned output geological modeling results and the above-mentioned simulation structural model into the hybrid loss function, and output the loss result, where the above-mentioned hybrid loss function combines the regression loss based on a single point and the multi-scale structure based on the local sliding Gaussian window The similarity is combined; the network parameters of the above-mentioned neural network are adjusted according to the above-mentioned loss results until the above-mentioned hybrid loss function meets the preset conditions (for example, the preset number of iterations is reached); the neural network obtained when the above-mentioned hybrid loss function satisfie
  • the geological simulation structural model generated using data simulation technology has complex faults and fold structures
  • the richness of training samples is improved and can effectively support the initial neural network to learn representative structural patterns. and geogeometric features that are difficult to describe using mathematical means in related technologies.
  • the training method provided by the present disclosure can flexibly introduce a variety of geological constraints and prior knowledge in the form of loss functions during the training process, effectively breaking the performance bottleneck of geological modeling technology and improving the accuracy of geological structure modeling.
  • the above training method further includes: using an Adam optimizer with adaptive learning step size to accelerate network convergence, and the initial learning rate is set to 0.01.
  • the loss function stops decreasing between two iterations, set the current learning rate to half of its original value.
  • Network training stops after 120 epochs, and the training and validation loss curves gradually converge (less than 0.1). This shows that CNN has successfully learned the geometric characteristics and combination rules of different geological structures from training data samples.
  • the neural network iteratively optimizes network parameters by minimizing the error between the prediction model and the reference model, and ultimately achieves its generalization on actual data.
  • regression loss is combined with multi-scale structural similarity to define a hybrid loss function for network training.
  • the hybrid loss function is defined as follows:
  • x be the reference structural model
  • m be its binary mask
  • the sampling points corresponding to the input layer data in the mask are set to 1, and the rest are set to 0.
  • the dimension sizes of x and m are consistent with the samples in the training data set.
  • MSE mean square error
  • MAE mean absolute error
  • p represents the local square area cropped from the same position in the two models being compared
  • N represents the total number of sampling points in the area p.
  • MAE is better than MSE in geological modeling scenarios, the results are still not optimal. CNNs trained using MAE alone may not correctly extract geometric features represented by the distribution of nearby sampling points, and blur high-frequency and sharp structural discontinuities. Therefore, models with similar MAE may have significantly different constructions, negatively affecting the optimization of the network.
  • MAE is combined with structural similarity (SSIM) to define a hybrid loss function. By adaptively assigning higher weights at structural boundaries that differ greatly, SSIM can better preserve high-frequency geometric features.
  • SSIM loss used to measure the reference model and the prediction model in the region p can be expressed as follows:
  • ⁇ x and ⁇ y represent the average values of model x and y in the local area, respectively.
  • ⁇ x and ⁇ y represent the variance
  • ⁇ xy represents the covariance of the two.
  • the mean, variance and covariance are obtained by using an isotropic Gaussian filter with standard deviation ⁇ and zero mean. calculation to improve calculation efficiency.
  • ⁇ x and ⁇ y measure the changes in stratigraphic sequence and structure in the local area of the model, while ⁇ x , ⁇ y and ⁇ xy measure the relative change trend of the structure in the local area of the model being compared. Together they indicate that the structures of the two geological models are similar. sex.
  • the parameters ⁇ and ⁇ define the relative importance of the mean and variance terms in the loss function. According to the preliminary experimental results, both parameters are set to 1.
  • formula (2) uses smaller constant factors C 1 and C 2 to avoid numerical instability caused by division by zero.
  • Gaussian filter The standard deviation ⁇ g is a hyperparameter that needs to be determined before training the network. The choice of ⁇ g will affect the effect of network training. Using a larger ⁇ g for training may make the network overemphasize local changes and produce false features at the edges of the structure. Training with smaller ⁇ g may blur structural and stratigraphic boundaries. To avoid adjusting parameters, this paper uses multi-scale structural similarity (MS-SSIM) and expresses it as follows:
  • the parameter ⁇ j defines the relative importance of the variance terms of different scales in formula (3).
  • MS-SSIM creates a series of local Gaussian windows from large to small by using different standard deviations ⁇ g , and simultaneously estimates the structural similarity of the model within multiple windows. In order to improve calculation efficiency, this paper avoids directly defining the Gaussian standard deviation ⁇ g , but performs five consecutive 28282 average pooling operations on the model in the local area p to obtain five equivalent spatial scales.
  • MS-SSIM emphasizes changes in structural characteristics near a single sampling point in the model (the range of which is defined by the Gaussian standard deviation ⁇ g ), it may still introduce artifacts into the predictions. This is because during the network training process, it is difficult to correctly calculate the derivatives at the boundaries within the local area.
  • MS-SSIM is insensitive to mean deviations in predicted values, which may lead to changes in stratigraphic sequences or disturbances in geological interfaces in modeling results.
  • MAE can minimize the average error at each sampling point in the model, so it can better reconstruct stratigraphic sequences and eliminate boundary artifacts.
  • it is proposed to use a weighted hybrid loss function:
  • ⁇ is the weighting factor used to balance different loss terms
  • K represents the total number of local areas p within the model.
  • modeling performance is improved by using a hybrid function based on single point accuracy and structural similarity when updating a neural network.
  • Table 1 shows a quantitative analysis of networks trained with different loss functions using multiple quantitative evaluation metrics and lists the average scores of these metrics on the validation dataset.
  • the CNN trained using the MS-SSIM and MAE hybrid loss function (denoted by MS-SSIM and L1) outperforms the other loss functions in Table 1 on all quality metrics, even when used as network training itself quality indicators.
  • the excellent performance of this loss function is attributed to better reconstruction of structures related to faults and unconformity surfaces in the model by assigning high weights to areas with large structural local differences, and the accurate identification of faults helps to constrain Horizontal trends in stratigraphic interfaces across structural discontinuities.
  • a closed loop of knowledge is formed through data simulation and model training.
  • the proposed method can effectively support the network to learn representative structural patterns and geological geometric features that are difficult to describe by mathematical means in related technologies, thereby promising Break through the performance bottleneck of current modeling methods in complex structural environments.
  • the present disclosure implements the construction interpolation process through multi-layer spatial convolution with trainable kernel parameters.
  • the convolutional layer of CNN can be regarded as an implicit interpolation operator in related technologies, but the difference is that the convolution kernel parameters of the network can be optimized through training.
  • the complex geological knowledge and structural patterns involved in the structural forward modeling process are implicitly embedded into the neural network, effectively ensuring the generalization ability of the method in actual data applications.
  • the trained network is not limited to the specific format of input data in various geological modeling applications. The input data is flexible and can be changed in real time, achieving real-time interactive Modeling of geological structures.
  • the above-mentioned training method further includes: normalizing the above-mentioned geological simulation structural model using a preset normalization method to obtain a normalized geological simulation structural model; wherein, the above-mentioned acquisition of the above-mentioned geological simulation structural model The geological fault data and unevenly distributed geological layer data of the structural model are used to obtain the training sample data set, which includes: extracting the geological fault data and geological layer data of the above-mentioned normalized geological simulation structural model to obtain the above-mentioned training sample data set.
  • the structural data needs to be scaled during the geological structure model training stage to obtain normalized geology with a value range of 0 to 1 Simulation construction model.
  • scatter points on the same geological interface are assigned corresponding values to the normalized model.
  • the above-mentioned neural network includes an encoder branch and a decoder branch, the above-mentioned encoder branch includes at least one encoding feature processing unit, and the above-mentioned decoder branch includes at least one decoding feature processing unit; wherein, the above-mentioned use of the above-mentioned training samples
  • the data set training neural network includes: inputting the above-mentioned geological fault data and the above-mentioned geological layer data into the above-mentioned encoder branch, and after processing by at least one of the above-mentioned encoding feature processing units in the above-mentioned encoder branch, outputting the final encoded hidden feature;
  • the encoded hidden features are input to the above-mentioned decoder branch, and after being processed by at least one of the above-mentioned decoding feature processing units in the above-mentioned decoder branch, a prediction construction model is output; the above-mentioned prediction construction model and
  • the encoder branch may include five encoding feature processing single units connected in sequence.
  • the above-mentioned encoder branch includes a first coding feature processing unit, a second coding feature processing unit, a third coding feature processing unit, a fourth coding feature processing unit and a fifth coding feature processing unit connected in sequence;
  • the above-mentioned input of the above-mentioned geological fault data and the above-mentioned geological layer data into the above-mentioned encoder branch, and after processing by at least one of the above-mentioned encoding feature processing units in the above-mentioned encoder branch, the output of the final encoded hidden features includes:
  • the fourth encoded hidden feature is input to the fifth encoded feature processing unit to obtain the final encoded hidden feature.
  • the above-mentioned encoding feature processing unit includes an inverse residual module and a maximum pooling layer.
  • the above-mentioned inverse residual module includes an extended convolution layer, a depth-separable convolution layer and a projection convolution layer; the above-mentioned will
  • the above-mentioned geological fault data and the above-mentioned geological layer data are input to the above-mentioned encoder branch.
  • outputting the final encoded hidden feature includes: combining the above-mentioned geological fault data and the above-mentioned geological layer.
  • the data is input into the above-mentioned extended convolution layer to expand the feature dimension and output the extended features; the above-mentioned extended features are input into the above-mentioned depth-separable convolution layer for feature extraction and the extracted features are output; the above-mentioned extracted features are input into the above-mentioned projection convolution layer to output Perform dimensionality reduction processing on the above-mentioned extracted features, and output the first coded hidden features; input the above-mentioned first coded hidden features into the above-mentioned maximum pooling layer, and output the above-mentioned final coded hidden features.
  • the above-mentioned depth-separable convolution layer includes a single-channel feature extraction layer and a cross-channel feature fusion layer; the above-mentioned input of the above-mentioned extended features into the above-mentioned depth-separable convolution layer for feature extraction, and the output extraction feature includes: The above-mentioned extended features are input into the above-mentioned single-channel feature extraction layer and the single-channel features are output; the above-mentioned single-channel features are input into the above-mentioned cross-channel feature fusion layer to fuse the above-mentioned single-channel features and the above-mentioned extraction features are output.
  • the depthwise separable convolutional layer includes 18181 single-channel feature extraction layer and 18181 cross-channel feature fusion layer for extending the input data to a higher-dimensional feature space and mapping it back to the output channel.
  • Figure 3A schematically shows an architectural diagram of a neural network according to an embodiment of the present disclosure.
  • FIG. 3B schematically shows a structural diagram of a reverse residual module according to an embodiment of the present disclosure.
  • Figure 3C schematically shows a structural diagram of a channel attention module according to an embodiment of the present disclosure.
  • the neural network consists of an encoder branch and a decoder branch.
  • the input sparse structural data (such as geological fault data and geological horizon data) are first extracted in the encoder branch and then restored into a complete implicit structural model in the decoder branch.
  • the input sparse structural data passes through 5 coding feature processing units (ie, E0 to E4) in sequence.
  • Each coding feature processing unit unit corresponds to a spatial scale. Except for the E0 unit, other coding units
  • the input of the feature processing unit is the output of the previous encoding feature processing unit.
  • Each encoding feature processing unit uses a maximum pooling layer internally to downsample the input features and simultaneously expand its feature dimensions to extract corresponding spatial scale information. Based on the low-rank nature of the input data, the encoder unit adopts a linear bottleneck and an inverse residual structure.
  • the reverse residual module includes a 1 ⁇ 1 ⁇ 1 extended convolution layer, a 3 ⁇ 3 ⁇ 3 depth-separable convolution layer and a 1 ⁇ 1 ⁇ 1 projected convolution layer, and after each layer of convolution Including batch normalization (BN) and nonlinear activation function (ReLu).
  • the two 1 ⁇ 1 ⁇ 1 convolutional layers at the end of the depthwise separable convolutional layer are used to expand the input to a higher dimensional feature space and map it back to the output channel.
  • the unit is represented as a residual learning function using local skip connections on the dilated convolutional layer and the projected convolutional layer, which accelerates the backpropagation of the gradient response in the network.
  • the encoder aggregates rich information, not all features are beneficial for prediction. Due to the sparsity and heterogeneity of the input data, there are many features that are not related to geological structures, and most of the values in the feature channels are zero.
  • a lightweight channel attention module is inserted into the last three encoding feature extraction units of the encoder, that is, a channel attention module is inserted into the reverse residual module.
  • the channel attention module consists of a squeeze module and an excitation secondary module.
  • the input features are reconstructed through a remake operation, compressed into a low-dimensional feature space in the squeeze module, and then converted into channel attention weights by firing the secondary module.
  • This enables the network to adaptively learn the dependencies between various feature dimensions, thereby re-adjusting their importance in the prediction process by suppressing irrelevant features and improving network prediction performance.
  • the decoder branch includes five decoding feature processing units (from D4 to D0) corresponding to the encoder branch to obtain the full model space receptive field of the network.
  • Each decoding feature processing unit in the decoder branch is responsible for aggregating hidden features from the previous decoding feature processing unit and the encoder skip layer, while compensating for the mismatch in spatial resolution between different features.
  • the decoder features are concatenated with the corresponding encoder features (such as the hidden features of E3 and D4) and input into two consecutive depth-separable convolutional layers to further extract effective information.
  • the depthwise separable convolutional layer decomposes the convolution operation into two separate steps, including a 3 ⁇ 3 ⁇ 3 convolutional layer for single-channel feature extraction and a 1 ⁇ 1 ⁇ 1 for fusion of cross-channel features. Convolutional layer. By decomposing the convolution operation into two steps, depthwise separable convolution can significantly reduce the computational complexity and memory footprint, thus helping to build lightweight networks.
  • the output layer after the decoder branch uses a 1 ⁇ 1 ⁇ 1 convolutional layer to implement a simple linear transformation to fuse all decoding features and output modeling results.
  • an encoder-decoder architecture design similar to UNet and its related variants is used on the CNN architecture and is improved to find the best between geological modeling accuracy and computational efficiency. compromise.
  • the data input to the network is first down-sampled at multiple spatial resolutions in the encoder, and then recombined with the corresponding up-sampled features through the skip connection mechanism in the decoder. Therefore, local features of the input data are usually extracted in the shallow layers of CNN, and gradually transform into high-level global features when the receptive fields in the deep convolutional layers become larger and larger. Since hidden features with different spatial resolutions usually have different geometric information, systematically aggregating multi-scale hidden features helps the network produce reliable and stable prediction results. In addition, because structural information may be gradually lost in multiple downsampling, low-level features extracted by shallow layers can usually follow the input data better than deep features, and the use of skip connection mechanism helps the network to generate structures consistent with known structures. Compatible models.
  • the modeling experience and knowledge learned from the training sample data set will be implicitly embedded into the network parameters.
  • the trained geological structure model was applied to the geological simulation structure model that was not included in the training sample data set.
  • These geological simulation structural models have complex fault and fold structures, and the geological interface is modified by multiple high-angle normal faults.
  • the incomplete layer data and fault data of these geological simulation structural models were extracted and input into the geological structural model. Although there is no global information to constrain the structural pattern, the geological structural model can still output a model with similar input structural characteristics and preserve the changing trend of the geological interface.
  • the modeling ability of the network is quantitatively evaluated by using a variety of indicators.
  • quantitative evaluation indicators include SSIM, MSE, MAE, explained variance score (EVS), mean square log error (MSLE), median absolute error (MDAE), and goodness of fit (R2S).
  • EVS explained variance score
  • MSLE mean square log error
  • MDAE median absolute error
  • R2S goodness of fit
  • HFA horizon fitting accuracy
  • Table 2 shows a quantitative comparison between the construction method according to the embodiment of the present disclosure and the current mainstream network on geological modeling problems.
  • the method provided by the present disclosure not only has better performance on all indicators, but also has a more lightweight network architecture, that is, fewer training parameters and computational complexity. This enables the method to be applied to large-scale 3D modeling tasks.
  • the use of the attention module allows the network to learn the interdependencies between different feature dimensions, thereby adaptively emphasizing features that help improve prediction performance while suppressing irrelevant features. Therefore, although the dimensions of the features are smaller than the CNN structure in related technologies (such as UNet), the network still achieves stable modeling results and reliable generalization capabilities on the validation data set.
  • a CNN trained using only simulation data can also show excellent modeling capabilities.
  • the CNN-based intelligent structural modeling process is driven by spatial convolution with trainable parameters and loss functions that provide various geological constraints.
  • Spatial convolution in the network can be equivalent to the local interpolation operator used in related technologies. The only difference is that spatial convolution can be optimized through training, providing more flexibility and therefore able to handle mathematical models that are difficult to solve in related technologies.
  • the data simulation method used in this disclosure can generate a large number of realistic structural models, and fully considers the common local missing situations of actual data when constructing the training data set.
  • This method of combining structural forward modeling and network training can embed various structural patterns and geological laws in the training samples into the neural network in a targeted manner, thereby forming a closed loop of knowledge. This is also an important reason why the trained network can be successfully applied to real examples with different construction patterns. Due to the full automation of the modeling process and the high-efficiency implementation based on the parallel computing platform, the network can perform real-time modeling of input construction data under manual interaction. For example, the same set of input data is randomly perturbed to generate multiple modeling results to analyze the geological uncertainty and structural consistency in structural interpretation.
  • Figure 4 schematically shows a flow chart of an implicit construction modeling method according to an embodiment of the present disclosure.
  • this method is mainly divided into the training sample library preparation stage, the training stage and the inference stage.
  • the preparation stage of the training sample library parameters of faults, folds, and tilt structures are randomly selected; a simulation structural model generator is used to generate a simulation structural model; then the fault and layer scatter points of the simulation structural model are extracted to form scatter data as input data .
  • the incomplete layer and scattered point data of all faults are used as input data and input into the neural network CNN for training, and the simulation construction model is used to monitor the network learning process, thereby outputting a complete implicit construction model.
  • the network is applied to actual structural interpretation data in different areas to obtain the corresponding complete geological structural model.
  • the structural interpretation data is usually not necessarily located on a regular grid, before practical application, it is necessary to convert the layer and fault interpretation data of the structural interpretation data of the actual work area to the modeling grid, and obtain the structural data for input into the network ; Then input the structural data into the above-trained implicit structural model, and output an implicit structural model that is compatible with the input structural characteristics and is geologically reasonable; then by extracting the model isosurface and discontinuous jump features, complete models can be obtained respectively. Geological horizons and fault planes.
  • Figure 5 schematically illustrates a construction method according to an embodiment of the present disclosure that generates a spatial scalar function from input sparse structural data as an implicit representation of all geological structures.
  • implicit modeling is regarded as a deep learning image restoration problem.
  • the network can reconstruct a complete geological structure model from sparse or non-uniform sampling data. Deep learning methods allow the differences between compared models to be reasonably measured by defining appropriate loss functions and can flexibly introduce geological understanding and structural constraints.
  • the input of this network is fault and horizon data, and the output is a full-space three-dimensional scalar field as an implicit representation of all geological structures. Parameterize fault and fold simulations and automatically create simulation models with diverse structural characteristics by randomly selecting parameters within a reasonable range, thereby building a complete training sample library.
  • horizon and fault scatter points are randomly extracted from these models to generate sparse or unevenly distributed data as network input to predict the complete geological structure model, while the simulation model itself is used for supervision.
  • a hybrid loss function is defined that combines single-point-based regression loss with multi-scale structural similarity based on local sliding Gaussian windows to ensure that geological prediction results are compatible with known structures.
  • the network outputs the implicit model, by extracting its isosurface and detecting rapid changes in its local values, any complete geological layer plane and fault can be extracted respectively (Figure 5).
  • the trained network can effectively create geologically reasonable and structurally consistent models, demonstrating the potential of using deep learning technology to improve the quality of geological modeling.
  • the modeling results of this method are reproducible because there is no need to randomly initialize network parameters in each calculation process.
  • structural modeling was performed in three different geological application scenarios, including respectively based on field outcrop observation data.
  • Figure 6 schematically shows the application of two-dimensional structural modeling based on field outcrop observation data.
  • Panel a in Figure 6 shows the geological interpretation data obtained through field outcrop observations of a basin.
  • This outcrop profile contains a series of medium- and high-angle faults (indicated by dashed lines) due to local subsidence caused by synsedimentary dissolution of gypsum in the deposit. These synsedimentary faults control lateral thickness variations at the stratigraphic interface.
  • these field observations are integrated into a regular model grid to be used as input to a CNN to obtain a complete structural model of the outcrop.
  • Panel b in Figure 6 shows the predicted structural model using discrete color bars
  • panel d in Figure 6 shows the corresponding modeling results using continuous color bars. It can be found that the different geological layers where dislocations occur in the model (discontinuities in scalar values) match well with the fault data shown in panel a in Figure 6 as input to the network.
  • Figure c in Figure 6 shows the fitting situation of the four horizons extracted from the modeling results (Panels b and d in Figure 6) and the input horizon data respectively, proving that the modeling results have a good influence on the input Compatibility of structural data and their geological plausibility.
  • the network was trained using simulated data, it still achieved good modeling performance on real data sets with complex geological formations.
  • Figure 7 schematically illustrates a real-time interactive structural modeling application based on well log data.
  • Panel a in Figure 7 was modeled using horizons and faults interpreted from borehole and outcrop observations under different geological conditions to study the geological uncertainties associated with changes in structural position.
  • the first data example shows a simple continuous layer model.
  • the modeling situation becomes more complex and diverse when considering additional geometric objects (such as faults or unconformities) that disrupt continuous geological layers.
  • additional geometric objects such as faults or unconformities
  • random perturbations of the interpreted horizon positions can produce variations in stratigraphic thickness.
  • the network takes multiple structural combinations as input to simulate the impact of multiple possible structural forms and relationships on the modeling process in Figure 7 a. All results shown in Figure 7, panel c, are calculated using a single GPU (NVIDIA Tesla V100).
  • the average time to generate each model using the trained network is about 2 milliseconds (model size is 128 ⁇ 256).
  • the trained network is not limited to a specific format of the input data.
  • the input data is flexible and can be changed in real time, so this method has the ability to model interactively in real time.
  • panel d in Figure 7 shows a good fit between the input horizon data and the modeled horizons, once again proving that the method can output a geological structural model that is compatible with given structural characteristics.
  • Figure 8 schematically shows an application case of three-dimensional geological structure modeling based on actual seismic interpretation data.
  • the trained CNN in order to verify the modeling capabilities of the network, is applied to the three-dimensional structural modeling problem, and it is planned to extract unevenly sampled scatter data from seismic interpretation data as network input to construct a complete Geological structure model.
  • Three-dimensional seismic data collected in structurally complex deformation areas have relatively low resolution and signal-to-noise ratio.
  • Some seismic reflectors are difficult to track continuously throughout the data space due to noise. Closely spaced and intersecting faults, as well as noise and stratigraphic features similar to structural discontinuities, further complicate and obscure the geometric relationships and spatial organization of seismic structures.
  • Panel d in Figure 8 shows a reconstructed complete horizon interface colored by vertical coordinate values, where there is a vertical jump across the fault.
  • both the structural discontinuities and geological interfaces output by the network are compatible with the input structures, and the predicted model can even maintain changes in fold structures without additional provision of Global information that constrains modeled behavior.
  • the input horizon data on the same geological interface can be well matched with the corresponding isosurface of the prediction model, which proves that the network can generate a model that is compatible with the input structural characteristics.
  • the present disclosure also provides a geological structure modeling device based on multi-source heterogeneous data.
  • the device will be described in detail below with reference to FIG. 9 .
  • Figure 9 schematically shows a structural block diagram of a geological structure modeling device based on multi-source heterogeneous data according to an embodiment of the present disclosure.
  • the geological structure modeling device 900 based on multi-source heterogeneous data in this embodiment includes an acquisition module 910 , a conversion module 920 and an input and output module 930 .
  • the acquisition module 910 is configured to acquire multi-source heterogeneous data of the target area, where the multi-source heterogeneous data includes fault interpretation data and horizon interpretation data, and the multi-source heterogeneous data includes at least one of the following: geological outcrop. Observation data, well logging data, various geophysical data, and empirical knowledge data. In an embodiment, the acquisition module 910 may be configured to perform the operation S110 described above, which will not be described again here.
  • the conversion module 920 is configured to convert the above-mentioned fault interpretation data and the above-mentioned horizon interpretation data based on a preset model grid to obtain converted fault interpretation data and converted horizon interpretation data.
  • the conversion module 920 may be configured to perform the operation S120 described above, which will not be described again here.
  • the input and output module 930 is configured to input the above-mentioned converted fault interpretation data and the above-mentioned converted horizon interpretation data into a pre-trained neural network, and output a complete geological structure model corresponding to the above-mentioned target area.
  • the input and output module 930 may be configured to perform the operation S130 described above, which will not be described again here.
  • Figure 10 schematically shows a structural block diagram of a neural network training device that can be used for multi-source heterogeneous data modeling according to an embodiment of the present disclosure.
  • the neural network training device 1000 of this embodiment includes a generation module 1010, an extraction module 1020 and a training module 1030.
  • the generation module 1010 is configured to generate a geological simulation structural model using data simulation technology to obtain a geological simulation structural model library.
  • the above geological simulation structural model has diverse fold and fault characteristics.
  • the generation module 1010 may be configured to perform the operation S210 described above, which will not be described again here.
  • the extraction module 1020 is configured to obtain the geological fault data and geological layer data of the above-mentioned geological simulation structural model for each geological simulation structural model in the above-mentioned geological simulation structural model library, and finally obtain a training sample data set.
  • the extraction module 1020 may be configured to perform the operation S220 described above, which will not be described again here.
  • the training module 1030 is configured to train a neural network using the above training sample data set to obtain a neural network that can be used for geological modeling of multi-source heterogeneous data.
  • the training module 1030 may be configured to perform the operation S230 described above, which will not be described again here.
  • modules, sub-modules, units, sub-units may be implemented in one module. Any one or more of the modules, sub-modules, units, and sub-units according to the embodiments of the present disclosure may be split into multiple modules for implementation.
  • any one or more of the modules, sub-modules, units, and sub-units according to embodiments of the present disclosure may be at least partially implemented as hardware circuits, such as field programmable gate arrays (FPGAs), programmable logic arrays (PLA), System-on-a-chip, system-on-substrate, system-on-package, application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuits that can be implemented in hardware or firmware, or in a combination of software, hardware, and firmware Any one of these implementation methods or an appropriate combination of any of them.
  • FPGAs field programmable gate arrays
  • PLA programmable logic arrays
  • ASIC application-specific integrated circuit
  • one or more of the modules, sub-modules, units, and sub-units according to the embodiments of the present disclosure may be at least partially implemented as a computer program module, and when the computer program module is executed, corresponding functions may be performed.
  • any multiple modules among the acquisition module 910, the conversion module 920, the input and output module 930, the generation module 1010, the extraction module 1020 and the training module 1030 can be combined and implemented in one module, or any one of them Modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in one module. According to an embodiment of the present disclosure, at least one of the acquisition module 910, the conversion module 920, the input-output module 930, the generation module 1010, the extraction module 1020 and the training module 1030 may be at least partially implemented as a hardware circuit, such as a field programmable gate.
  • At least one of the acquisition module 910, the conversion module 920, the input and output module 930, the generation module 1010, the extraction module 1020 and the training module 1030 may be at least partially implemented as a computer program module, and when the computer program module is run, Can perform corresponding functions.
  • the geological structure modeling device based on multi-source heterogeneous data in the embodiment of the present disclosure corresponds to the geological structure modeling method based on multi-source heterogeneous data in the embodiment of the present disclosure. Based on For a description of the geological structure modeling device of multi-source heterogeneous data, please refer to the geological structure modeling method based on multi-source heterogeneous data, and will not be described again here.
  • Figure 11 schematically shows a block diagram of an electronic device suitable for implementing the above method according to an embodiment of the present disclosure.
  • an electronic device 1100 includes a processor 1101 that can be loaded into a random access memory (RAM) 1103 according to a program stored in a read-only memory (ROM) 1102 or from a storage part 1108 program to perform various appropriate actions and processes.
  • processor 1101 may include, for example, a general purpose microprocessor (eg, CPU), an instruction set processor and/or related chipset, and/or a special purpose microprocessor (eg, application specific integrated circuit (ASIC)), or the like.
  • Processor 1101 may also include onboard memory for caching purposes.
  • the processor 1101 may include a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of the present disclosure.
  • the processor 1101, ROM 1102 and RAM 1103 are connected to each other through a bus 1104.
  • the processor 1101 performs various operations according to the method flow of the embodiment of the present disclosure by executing programs in the ROM 1102 and/or RAM 1103. It should be noted that the program may also be stored in one or more memories other than ROM 1102 and RAM 1103.
  • the processor 1101 may also perform various operations according to the method flow of embodiments of the present disclosure by executing programs stored in the one or more memories.
  • the electronic device 1100 may further include an input/output (I/O) interface 1105 that is also connected to the bus 1104 .
  • Electronic device 1100 may also include one or more of the following components coupled to I/O interface 1105: an input portion 1106 including a keyboard, mouse, etc.; including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and an output section 1107 of a speaker and the like; a storage section 1108 including a hard disk and the like; and a communication section 1109 including a network interface card such as a LAN card, a modem and the like.
  • the communication section 1109 performs communication processing via a network such as the Internet.
  • Driver 1110 is also connected to I/O interface 1105 as needed.
  • Removable media 1111 such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive 1110 as needed, so that a computer program read therefrom is installed into the storage portion 1108 as needed.
  • the present disclosure also provides a computer-readable storage medium.
  • the computer-readable storage medium may be included in the device/device/system described in the above embodiments; it may also exist independently without being assembled into the device/system. in the device/system.
  • the above computer-readable storage medium carries one or more programs. When the above one or more programs are executed, the method according to the embodiment of the present disclosure is implemented.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium, which may include, but is not limited to, portable computer disks, hard disks, random access memory (RAM), and read-only memory (ROM). , erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device.
  • the computer-readable storage medium may include one or more memories other than ROM 1102 and/or RAM 1103 and/or ROM 1102 and RAM 1103 described above.
  • Embodiments of the present disclosure also include a computer program product including a computer program containing program code for performing the method illustrated in the flowchart.
  • the program code is used to cause the computer system to implement the training method or construction method of the geological structure model based on multi-source heterogeneous data provided by the embodiments of the present disclosure.
  • the computer program may rely on tangible storage media such as optical storage devices and magnetic storage devices.
  • the computer program can also be transmitted and distributed in the form of a signal on a network medium, and downloaded and installed through the communication part 1109, and/or installed from the removable medium 1111.
  • the program code contained in the computer program can be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.
  • the computer program may be downloaded and installed from the network via communication portion 1109 and/or installed from removable media 1111 .
  • the computer program is executed by the processor 1101, the above-described functions defined in the system of the embodiment of the present disclosure are performed.
  • the systems, devices, devices, modules, units, etc. described above may be implemented by computer program modules.
  • the program code for executing the computer program provided by the embodiments of the present disclosure may be written in any combination of one or more programming languages. Specifically, high-level procedural and/or object-oriented programming may be utilized. programming language, and/or assembly/machine language to implement these computational procedures. Programming languages include, but are not limited to, programming languages such as Java, C++, python, "C" language or similar programming languages.
  • the program code may execute entirely on the user's computing device, partly on the user's device, partly on a remote computing device, or entirely on the remote computing device or server.
  • the remote computing device may be connected to the user computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device, such as provided by an Internet service. (business comes via Internet connection).
  • LAN local area network
  • WAN wide area network
  • Internet service business comes via Internet connection
  • each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logic functions that implement the specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown one after another may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved.
  • each block in the block diagram or flowchart illustration, and combinations of blocks in the block diagram or flowchart illustration can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or may be implemented by special purpose hardware-based systems that perform the specified functions or operations. Achieved by a combination of specialized hardware and computer instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Geometry (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Remote Sensing (AREA)
  • Computer Graphics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

本公开提供了一种基于多源异构数据的地质构造建模方法，可以应用于地质勘查和人工智能技术领域。该方法包括：利用数据仿真技术生成多个地质仿真构造模型，构建地质仿真构造模型库，地质仿真构造模型具有多样化褶皱和断层特征；针对地质仿真构造模型库中的每个地质仿真构造模型，获取地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集；以及利用训练样本数据集训练神经网络；将目标区域的多源异构数据输入到训练好的神经网络中，输出与多源异构数据对应的地质构造模型，其中，多源异构数据包括断层解释数据和层位解释数据，多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据。

Description

基于多源异构数据的地质构造建模方法
相关申请的引用
1、本申请要求于2022年8月31日向中国国家知识产权局递交的发明名称为“基于深度学习的可交互式智能三维隐式构造建模方法”的中国专利申请202211064255.5的优先权，该申请的全部内容通过引用加入至本申请中。
2、本申请要求于2022年10月11日向中国国家知识产权局递交的发明名称为“基于多源异构数据的地质构造建模方法”的中国专利申请202211240817.7的优先权，该申请的全部内容通过引用加入至本申请中。
技术领域
本公开涉及地质勘查和人工智能技术领域，更具体地涉及一种基于多源异构数据的地质构造建模方法。
背景技术
可靠的地质构造模型对充分了解地下空间组织和定量模拟各种地质过程等一系列科学研究问题至关重要。地质建模旨在通过使用各种数学方法，建立关于地质情况的数值模型，以精确表示其空间位置和几何关系。相关技术中，一般使用隐式方法进行地质建模，隐式方法利用构造数据计算一个空间标量函数，以表示整个区域内所有的地质构造特征。在该函数中，地质界面被表示为等值面，而断层和不整合面被表示为函数值的不连续变化，该函数也称为隐式构造模型。
公开内容
本公开提供了一种基于多源异构数据的地质构造建模方法。
根据本公开的一个方面，提供了一种基于多源异构数据的地质构造建模方法，包括：
获取关于目标区域的多源异构数据，其中，上述多源异构数据包括断层解释数据和层位解释数据，上述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据；
基于预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据；以及
将上述转换后的断层解释数据和上述转换后的层位解释数据输入到预先训练好的神经网络中，输出与上述目标区域对应的完整地质构造模型。
本公开的另一个方面提供了一种基于多源异构数据的地质构造建模装置，包括：
获取模块，被配置用于获取目标区域的多源异构数据，其中，上述多源异构数据包括断层解释数据和层位解释数据，上述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据；
转换模块，被配置用于基于预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据；以及
输入输出模块，被配置用于将上述转换后的断层解释数据和上述转换后的层位解释数据输入到预先训练好的神经网络中，输出与上述目标区域对应的完整地质构造模型。
本公开的另一个方面提供了一种电子设备，包括：一个或多个处理器；存储器，被配置用于存储一个或多个程序，其中，当上述一个或多个程序被上述一个或多个处理器执行时，使得一个或多个处理器执行上述方法。
本公开的另一个方面还提供了一种计算机可读存储介质，其上存储有可执行指令，该指令被处理器执行时使处理器执行上述方法。
本公开的另一个方面还提供了一种计算机程序产品，包括计算机程序，该计算机程序被处理器执行时实现上述方法。
附图说明
通过以下参照附图对本公开实施例的描述，本公开的上述内容以及其他目的、特征和优点将更为清楚，在附图中：
图1示意性示出了根据本公开实施例的基于多源异构数据的地质构造建模方法的流程图；
图2示意性示出了根据本公开实施例的神经网络的训练方法的流程图；
图3A示意性示出了根据本公开实施例的神经网络的架构图；
图3B示意性示出了根据本公开实施例的反向残差模块的结构示意图；
图3C示意性示出了根据本公开实施例的通道注意力模块的结构示意图；
图4示意性示出了根据本公开实施例的隐式构造建模方法的流程图；
图5示意性示出了根据本公开实施例的构建方法从输入的稀疏构造数据生成空间标量函数，作为对所有地质构造的隐式表示；
图6示意性示出了基于野外露头观测资料的二维构造建模的应用；
图7示意性示出了基于测井资料的实时交互式构造建模应用；
图8示意性示出了基于实际地震解释数据的三维地质构造建模应用案例；
图9示意性示出了根据本公开实施例的基于多源异构数据的地质构造建模装置的结构框图；
图10示意性示出了根据本公开实施例的用于基于多源异构数据构建地质构造模型的神经网络训练装置的结构框图；以及
图11示意性示出了根据本公开实施例的适于实现上述方法的电子设备的方框图。
具体实施方式
以下，将参照附图来描述本公开的实施例。但是应该理解，这些描述只是示例性的，而并非要限制本公开的范围。在下面的详细描述中，为便于解释，阐述了许多具体的细节以提供对本公开实施例的全面理解。然而，明显地，一个或多个实施例在没有这些具体细节的情况下也可以被实施。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要地混淆本公开的概念。
在此使用的术语仅仅是为了描述具体实施例，而并非意在限制本公开。在此使用的术语“包括”、“包含”等表明了所述特征、步骤、操作和/或部件的存在，但是并不排除存在或添加一个或多个其他特征、步骤、操作或部件。
在此使用的所有术语(包括技术和科学术语)具有本领域技术人员通常所理解的含义，除非另外定义。应注意，这里使用的术语应解释为具有与本说明书的上下文相一致的含义，而不应以理想化或过于刻板的方式来解释。
在使用类似于“A、B和C等中至少一个”这样的表述的情况下，一般来说应该按照本领域技术人员通常理解该表述的含义来予以解释(例如，“具有A、B和C中至少一个的系统”应包括但不限于单独具有A、单独具有B、单独具有C、具有A和B、具有A和C、具有B和C、和/或具有A、B、C的系统等)。
构建可靠的地质构造模型对充分了解地下空间组织和定量模拟各种地质过程等一系列科学研究问题至关重要。地质建模旨在通过使用各种数学方法，建立关于地质情况的数值模型，以精确表示其空间位置和几何关系。构造建模主要为显式或曲面建模方法，这类方法直接使用点、线、面等几何要素对地质构造进行表达。由于地质体中存在各种地质界面，当出现地层不整合、断层错断地层、地层尖灭等情形时，建模就会遇到复杂的地质界面交切关系处理问题，其各个构造之间存在紧密联系，任何误差都会对建模结果造成影响。因此，相关技术中建模技术通常需要严格的质量控制，以保证模型的地质合理性和数值稳定性。这使得该方法在处理较复杂构造建模问题时效率低下，模型更新十分困难。此外，建模过程受地质人员主观因素和经验的影响，导致即使输入相同构造数据，也很难复现相同的结果。
相对于显式方法，隐式建模方法具有高效、可更新和可复现的特征。隐式建模方法利用构造数据计算一个空间标量函数，以表示整个区域内所有的地质构造特征。在该函数中，地 质界面被表示为等值面，而断层和不整合面被表示为函数值的不连续变化，该函数也称为隐式构造模型。隐式建模方法通过整合构造观测数据和地质经验知识，将所有可用的地质信息融合到模型中，以从全局角度重建地下构造特征。
隐式建模方法的输入数据通常包括各种类型的建模对象，如通过野外观测和专家解释的空间点、向量、多段线和曲面。地质学家和地球物理学家结合经验知识从观测数据中推断可能的构造组合，用以约束地质界面之间可能的几何关系，进而引导隐式方法的建模行为。模型的构造特征需要符合地质学认识，同时遵循输入数据中已知的构造特征。由于地下空间的局部分布仅在有限的区域(如高度开发的矿井和油田)被观测，几乎不可能直接获得完整地质构造的真实情况。因此，需要额外在方法中引入地质规则和假设来约束建模过程。例如，相关技术中的隐式插值采用显式平滑准则来简化模型的局部变化，从而获得稳定且可靠的结果。
例如，离散平滑插值(DSI)是一类通过在空间网格上离散标量函数来计算地质模型的隐式方法。在DSI及其变体方法中，建模过程被表述为求解带有平滑约束的最小二乘最小化问题来估计与输入兼容的标量场。该平滑约束将经验知识引入建模过程，其基本假设是预期的模型应尽可能平滑，即局部构造的改变应尽可能小。然而，因为标量函数在网格单元上始终是连续的，所以模型网格单元不能跨越不连续构造，且该方法无法正确估计断层或不整合面附近的梯度场。为解决这个问题，需要通过对不连续构造独立建模来计算非规则网格以约束建模行为，从而保证在这些情况下方法也依然有效。除DSI外，势场法(PFM)是另一类隐式方法。
PFM将构造建模表述为双重协克里格插值或径向基函数插值问题。与DSI相比，虽然模型都在体网格上定义，但PFM在计算标量函数时不需要定义任何网格。建模结果完全取决于观测数据的分布，每个数据点的影响范围由所选插值算子所确定。然而，PFM通常会求解复杂数学方程来调整插值算子在整个感兴趣区域的影响，以获得符合地质意义的结果。然而，这也会导致计算成本随着输入数据的大小而迅速增加。
虽然相关技术中的方法具有很多优点，但在保持模型和输入数据地质一致性方面，处理构造复杂区域的地质模型仍然是一项具有挑战性的任务，建模的可靠性取决于观测数据的可用性和质量。在某些复杂的地质环境中，如果数据采样稀疏或不均匀分布，完全由数学方程指导的建模行为可能并不总能产生地质上有效的模型。相应的模型可能产生与地质认识或与相关构造空间关系不兼容的错误的几何特征。该问题主要归因于插值算子所允许的几何特征相对有限。在相关技术中，所有输入数据和先验知识在数学上被表示为线性约束条件，以尽可能平滑地计算连续的标量场。虽然该假设有助于推导出特定的地质模型，但施加这种平滑 度准则可能会影响局部地质变化的影响，并对构造高度复杂区域的建模精度产生负面影响。由于建模灵活性仅限于特定插值算子能够表达的几何关系，因此隐式方法通常会受到负面影响，从而生成在地质上不合理的几何特征或假象。
因此，将地质建模表述为最小二乘或空间插值问题，在平滑正则化假设下基于各种数学公式求解最优解。当已知数据过于稀疏或分布不均时，相关技术中的方法难以合理表示复杂的地质几何特征和关系，且迭代求解大型方程组在处理三维模型时会十分低效。
有鉴于此，本公开针对以上技术问题，提出一种使用卷积神经网络(CNN)的可交互隐式建模方法(又名DeepISMNet)。利用卷积神经网络从少量构造数据中构建完整的地质模型，基于自动生成的仿真模型样本库，网络能够在训练过程中灵活地整合多种地质约束和先验知识，有效打破相关技术中地质建模技术性能瓶颈。此外，通过将可用的多源异构信息转化成矩阵或张量的数据形式作为输入，训练好的网络能够在多个实际建模应用中利用少量的数据快速地生成符合地质认识和兼容已知构造特征的模型。
本公开实施例提供的方法属于数据驱动的统计算法，能基于经验知识和样本示例，通过最小化给定的质量评价标准估计一个映射函数，该函数将输入数据映射到期望的输出。与相关技术相比，深度学习的优势是在预测过程中无需以昂贵的计算代价求解大型线性或非线性方程系统。在计算机图像领域，CNN在分析几何特征和捕捉输入与输出之间复杂的非线性映射关系方面显示出卓越能力。为了在准确度和效率之间找到最佳折衷，有许多学习单元可用于构建CNN的架构，如深度可分离卷积、注意力机制和残差学习模块等。深度学习为建模过程中引入多种构造约束和经验认识提供了更为灵活的框架，从而在处理隐式和离散数据时表现出优异的性能。此外，输入的稀疏构造数据可以来自于多种数据源，包括地质露头观测、测井数据、多种地球物理数据和专家经验知识等，本公开提供的方法通过将这些多源异构信息统一转化成矩阵或张量的数据形式作为输入，网络能够高效地计算出拟合各种地质数据的隐式构造模型。
具体地，本公开的实施例提供了一种基于多源异构数据的地质构造建模方法，包括：
获取关于目标区域的多源异构数据，其中，上述多源异构数据包括断层解释数据和层位解释数据，上述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据；
基于预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据；以及
将上述转换后的断层解释数据和上述转换后的层位解释数据输入到预先训练好的神经网络中，输出与上述目标区域对应的完整地质构造模型。
图1示意性示出了根据本公开实施例的基于多源异构数据的地质构造建模方法的流程图。
如图1所示，该实施例的基于多源异构数据的地质构造建模方法包括操作S110～操作S130。
在操作S110，获取目标区域的多源异构数据，其中，上述多源异构数据包括断层解释数据和层位解释数据，上述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据。
根据本公开的实施例，上述方法还包括：通过调用数据接口从地质构建数据库中获取与目标区域对应的多源异构数据。
根据本公开的实施例，断层解释数据和层位解释数据可以是由人工将地质露头观测数据、测井数据、多种地球物理数据、经验知识数据转换得到的断层解释数据和层位解释数据。
在操作S120，基于预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据。
根据本公开的实施例，上述利用预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据包括：针对上述断层解释数据中的每个断层，将距离上述断层一个网格以内的点标为第一值，得到上述转换后的断层解释数据；针对上述层位解释数据中的每个层位，将离上述层位一个网格以内的点设为第二值，得到上述转换后的层位解释数据，其中，上述第二值为上述层位上所有散点的纵坐标平均值。
根据本公开的实施例，将获取的多源异构数据传输至预设模型网格中，以便于对断层解释数据和层位解释数据进行转换。
根据本公开的实施例，将网络应用于不同区域的实际构造解释数据上，以获得对应的完整地质构造模型。由于构造解释数据通常不一定位于规则网格上，在实际应用前需要将层位解释数据和断层解释数据转换到建模网格，并获得用于输入网络的构造数据。网络的输出是与输入构造特征兼容且地质上合理的隐式构造模型。通过提取模型等值面和不连续跳跃特征，可以分别得到完整的地质层位面和断层面。
在操作S130，将上述转换后的断层解释数据和上述转换后的层位解释数据输入到预先训练好的神经网络中，输出与上述目标区域对应的完整地质构造模型。
根据本公开的实施例，根据与目标区域对应的完整地质构造模型可以还原地下构造特征，并可视化展示构造特征，以便于根据地下构造特征勘察地下自然资源，以及防治地质灾害。
根据本公开的实施例，在可视化展示的构造特征不符合地质学认识的情况下，还可以通过调整地质构造模型中的构造参数，以调整构造特征，并实时通过可视化展示构造特征，实现地质构造模型的实时可交互能力。
图2示意性示出了根据本公开实施例的神经网络的训练方法的流程图。
如图2所示，该实施例的训练方法包括操作S210～操作S230。
在操作S210，利用数据仿真技术生成多个地质仿真构造模型，得到地质仿真构造模型库，上述地质仿真构造模型具有多样化褶皱和断层特征。
根据本公开的实施例，上述利用数据仿真技术生成地质仿真构造模型，得到地质仿真构造模型库包括：构建初始地质仿真构造模型，其中，上述初始地质仿真构造模型为水平方向恒定、垂向单调递增的层状模型，上述初始地质仿真构造模型包括至少一个构造参数；其中，上述构造参数包括褶皱参数、倾斜参数、断层构造参数中的至少一个；针对上述至少一个构造参数中的每个构造参数，在预设范围内调整上述构造参数，得到地质仿真构造模型库。
根据本公开的实施例，考虑到训练数据集应尽可能包含具有各种构造模式的模型，从而使卷积神经网络CNN能够学习到具代表性的知识，以实现在地质建模领域各种应用上的泛化。然而，由于无法获得地下空间的真实情况，很难通过人工解释的方式了解某区域所有地质构造。为解决这一问题，提出使用一组随机参数来生成一系列典型的褶皱和断层构造，以数据仿真的方式获得大量的训练样本的工作流程。该工作流程首先以一个水平方向恒定、垂向单调递增值的层状模型作为初始构造模型，然后在该模型中依次添加褶皱、倾斜和断层构造，逐渐使该模型的几何特征复杂化。通过在合理范围内随机选择参数，可以生成多种且符合地质学认识的构造模型，且不限于某种特定的几何模式。
根据本公开的实施例，基于对典型褶皱和断层模式的地质假设，数据仿真方法将地层褶曲、倾斜和断层构造的变形过程进行参数化表示。只需随机设定一组在预定义合理范围内的参数，就能创建彼此不同且逼真的构造模型，从而丰富训练样本数据，提高地质构造模型的泛华能力。
需要说明的是，对上述构造参数中的每个构造参数对应的预设范围，可以基于来自不同数据源的数据进行确定。例如，数据源可以包括地质露头观测、测井数据、多种地球物理数据和专家经验知识等。
在操作S220，针对上述地质仿真构造模型库中的每个地质仿真构造模型，获取上述地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集。
根据本公开的实施例，上述训练样本数据集中的每个训练样本数据包括一个地质仿真构造模型和与上述地质仿真构造模型对应的上述地质断层数据和上述地质层位数据，上述地质仿真构造模型和断层和层位数据共同作为训练标签。
根据本公开的实施例，在进行神经网络训练阶段，地质层位数据和地质断层数据被用作输入数据，神经网络输出完整的隐式地质构造模型，而地质仿真构造模型本身和输入构造数据共同作为对神经网络训练的标签，以提供对预测结果的监督。
根据本公开的实施例，上述获取地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集包括：提取上述地质仿真构造模型中的不连续特征，得到上述地质断层数据；从上述地质仿真构造模型的等值面中选取预设数量的等值面，得到目标等值面集合；针对上述目标等值面集合中的每个目标等值面，将上述目标等值面上的散点数据，按照纵坐标值域范围划分为多个数据区间；随机丢弃上述多个数据区间中第二预设数量的数据区间，剩余的数据区间内的散点数据作为上述地质层位数据；根据上述地质断层数据、上述地质层位数据和上述地质仿真构造模型，生成上述训练样本数据，最终得到上述训练样本数据集。
根据本公开的实施例，提取地质层位数据的方法包括：根据纵坐标将单个层位上所有散点降序排列，同时根据纵坐标值域划分为均匀间隔的多个区间，之后随机丢弃一个或多个区间内的所有散点，从而生成类似于人工解释的层位数据。此外，在每次输入数据生成过程中，从层位散点数据中随机删除部分，以模拟实际情况中稀疏且不均匀分布的层位解释。
根据本公开的实施例，从上述地质仿真构造模型的等值面中选取预设数量的等值面，得到目标等值面集合包括：将上述地质仿真构造模型中的所有等值面的值域范围按照降序划分为至少一个数值区间；针对上述至少一个数值区间中的每个数值区间，从上述数值区间中随机选取一个数值，得到目标数值；确定与上述目标数值对应的等值面为目标等值面，得到上述目标等值面集合。
根据本公开的实施例，由于地质仿真构造模型中的等值面代表相应的地质层序界面，而其局部标量值的跳跃变化则表示断层或不整合面等不连续构造，因此地质仿真构造模型本身可视为隐式标量函数。
根据本公开的实施例，由于地质层位面被隐式嵌入到仿真构造模型中，层位散点数据可通过等值面提取方法获得。为避免在训练数据中引入特定的采样模式或生成紧密相邻的不同层位，本公开将地质仿真构造模型的值域按降序划分为均匀间隔的几组区间，之后在每个区间内随机选取一个值并提取其对应的等值面。
根据本公开的实施例，上述方法还包括：神经网络的输入为表示层位和断层位置的散点数据。在将层位数据和断层数据输入神经网络之前，层位数据和断层数据需要被转换到规则的模型网络上，并且使得输入数据仅在已知的构造特征上具有效的值，其余则为0。
根据本公开的实施例，将层位数据和断层数据需要被转换到规则的模型网络上的方法包括：针对断层数据中的每个断层，将距离断层一个网格以内的点标为第一值，得到转换后的断层数据；针对层位数据中的每个层位，将离层位一个网格以内的点设为在地质仿真构造模型中所对应的标量值，以确保相同层位上的散点具有相同的值或标注，得到转换后的层位数 据。
在操作S230，利用上述训练样本数据集训练神经网络，得到可用于多源异构数据地质建模的神经网络。
根据本公开的实施例，其中，上述利用上述训练样本数据集训练神经网络，得到可用于多源异构数据地质构造建模的神经网络包括：将上述断层数据和上述层位数据输入上述神经网络，输出地质构造模型；将上述输出地质建模结果和上述仿真构造模型输入混合损失函数，输出损失结果，其中，上述混合损失函数将基于单点的回归损失与基于局部滑动高斯窗的多尺度结构相似性进行结合；根据上述损失结果调整上述神经网络的网络参数，直至上述混合损失函数满足预设条件(例如达到预设的迭代次数)；将上述混合损失函数满足上述预设条件时得到的神经网络作为上述训练好的神经网络。
根据本公开的实施例，由于利用数据仿真技术生成的地质仿真构造模型具有复杂的断层和褶皱构造，从而提高了训练样本的丰富性，能够有效地支持初神经网络学习到具有代表性的构造模式以及相关技术中数学手段难以描述的地质几何特征。同时，本公开提供的训练方法能够在训练过程中灵活以损失函数的方式引入多种地质约束和先验知识，有效打破地质建模技术性能瓶颈，提高地质构造建模的精度。
根据本公开的实施例，上述训练方法还包括：利用自适应学习步长的Adam优化器来加速网络收敛，且初始学习率设为0.01。当损失函数在两次迭代中停止继续降低时，将当前学习率设为原来的一半。网络训练在120个周期后停止，训练和验证损失曲线逐渐收敛(小于0.1)。这表明CNN已从训练数据样本成功学习了不同地质构造的几何特征和组合规律。
根据本公开的实施例，神经网络通过最小化预测模型和参考模型之间的误差，以迭代优化网络参数，并最终实现其在实际数据上的泛化。为使网络做出地质意义上有效的预测，将回归损失与多尺度结构相似性相结合，来定义用于网络训练的混合损失函数。
在其中一个实施例中，混合损失函数的定义如下：
设x为参考构造模型，m为其二进制掩码，掩码中对应于输入层位数据上的采样点设为1，其余设置为0。x和m的维度大小与训练数据集中的样本一致。对于每个参考模型x，带有可训练参数的CNN网络将层位数据h＝x⊙m和断层数据f作为输入，以生成预测模型
Figure PCTCN2022125275-appb-000001
其中，在已知层位附近被替换成输入数据的预测模型可以表示为
Figure PCTCN2022125275-appb-000002
在许多回归问题中，均方误差(MSE)和平均绝对误差(MAE)通常用于在结果中逐点地衡量方法精度。MSE强调误差较大的采样点，更能容忍误差较小的采样点，且不考虑数据的潜在空间分布模式。与MSE相比，MAE对局部数据的变化更敏感，减少了因过度惩罚较 大误差的采样点而在结果中引入的假象。我们采用加权MAE作为混合损失函数中的一项，其公式如下：
Figure PCTCN2022125275-appb-000003
其中，p表示从被比较的两个模型中的相同位置裁剪的局部方形区域，N表示区域p内采样点总数。
尽管MAE在地质建模场景中优于MSE，但结果仍然不是最优。单独使用MAE训练的CNN可能无法正确提取由临近采样点分布表示的几何特征，且模糊了高频和尖锐的构造不连续。因此，具有相似MAE的模型可能具有明显不同的构造，从而对网络的优化产生负面影响。为避免MAE的平滑效应，将MAE与结构相似性(SSIM)结合，定义一种混合损失函数。通过在差异较大的构造边界处自适应分配更高权重，SSIM可以更好地保留高频几何特征。用于衡量参考模型和预测模型在区域p内的SSIM损失可以表示如下：
Figure PCTCN2022125275-appb-000004
其中，μ x和μ y分别表示局部区域内模型x和y的平均值。σ x和σ y代表方差，σ xy代表二者的协方差。平均值、方差和协方差通过使用一个标准差为σ g均值为零的各向同性高斯滤波器
Figure PCTCN2022125275-appb-000005
计算，以提高计算效率。μ x和μ y衡量模型局部区域中地层层序和构造的变化，而σ x、σ y和σ xy衡量被比较的模型局部区域构造的相对变化趋势，它们共同表示两个地质模型的构造相似性。参数β和γ定义了损失函数中均值项和方差项的相对重要性。根据预实验结果，这两个参数均设为1。此外，公式(2)使用较小的常数因子C 1和C 2来避免被零除而导致数值不稳定的情况。高斯滤波器
Figure PCTCN2022125275-appb-000006
的标准差σ g是一个超参数，需要在训练网络前确定。σ g的选择会影响网络训练的效果，使用具有较大σ g进行训练可能使网络过度强调局部变化，并在构造边缘处产生虚假特征。而使用较小σ g进行训练可能模糊构造和地层边界。为避免调整参数，本文使用多尺度结构相似性(MS-SSIM)，并将其表示如下：
Figure PCTCN2022125275-appb-000007
其中，参数γ j定义公式(3)中不同尺度方差项的相对重要性。MS-SSIM通过使用不同标准差σ g创建一系列从大到小的局部高斯窗，并在多个窗口内同时估计模型的结构相似性。为提升计算效率，本文避免直接定义高斯标准差σ g，而是对局部区域p内的模型进行连续5次28282平均池化操作，从而获得等效的5种空间尺度，此外，采用权重系数γ j＝ {0.05，0.29，0.3，0.24，0.12}调整从不同空间尺度估计的结构相似性损失，以确保它们的总和等于1。
虽然MS-SSIM强调了模型中单个采样点附近的构造特征变化(其范围由高斯标准差σ g定义)，但仍可能会在预测中引入假象。这是因为在网络训练过程中，很难在局部区域内正确计算边界处的导数。
此外，MS-SSIM对预测值的平均偏差不敏感，这可能导致建模结果中地层层序的变化或地质界面的扰动。与之相比，虽然无法保留对比度高的构造特征，MAE却能将模型内每个采样点的平均误差降至最低，因此能更好地重建地层层序和消除边界假象。为同时获得两个损失函数的优势，提出使用加权混合损失函数：
Figure PCTCN2022125275-appb-000008
其中，λ是用于平衡不同损失项的加权因子，K表示模型内的局部区域p总数。在训练CNN时，从被比较的两个模型中逐点地裁剪局部区域，并根据公式(4)计算每个区域内的损失。根据预实验结果，将局部区域的尺寸设为7，λ设为1.25。最终的损失为所有局部区域内损失函数的平均值。
根据本公开的实施例，通过在更新神经网络时使用基于单点精度和结构相似性的混合函数，提高了建模性能。
表1示出了使用多种量化评价指标对使用不同损失函数训练的网络进行了定量分析，并列出了验证数据集上这些指标的平均得分。
表1
Figure PCTCN2022125275-appb-000009
如表1所示，使用MS-SSIM和MAE混合损失函数(由MS-SSIM和L1表示)训练的CNN在所有质量指标上都优于表1中的其他损失函数，甚至包括本身用作网络训练的质量指标。该损失函数的优异性能归因于通过将高权重分配给构造局部差异较大的区域，更好地重建了模型中与断层和不整合面等相关的构造，而断层的准确识别有助于约束跨越构造不连续处地层界面的水平变化趋势。
根据本公开的实施例，通过数据仿真和模型训练的方式形成知识闭环，提出的方法能有效地支持网络学习到具有代表性的构造模式以及相关技术中数学手段难以描述的地质几何特征，从而有望突破当前建模方法在复杂构造环境下的性能瓶颈。
根据本公开的实施例，本公开通过具可训练核参数的多层空间卷积来实现构造插值过程。CNN的卷积层可以看作是相关技术中的隐式插值算子，而区别在于网络的卷积核参数可通过训练被优化。通过使用大量仿真样本数据进行训练，构造正演过程所涉及的复杂地质知识和构造模式被隐式地嵌入到了神经网络中，有效保证了方法在实际数据应用上的泛化能力。此外，依托图形处理器GPU并行计算平台，经过训练的网络在各种地质建模的应用中，不局限于输入数据的特定格式，输入数据灵活，可实时更改输入数据，实现实时可交互式的地质构造建模。
根据本公开的实施例，上述训练方法还包括：利用预设归一化方式对上述地质仿真构造模型进行归一化处理，得到归一化后的地质仿真构造模型；其中，上述获取上述地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集包括：提取上述归一化后的地质仿真构造模型的地质断层数据和地质层位数据，得到上述训练样本数据集。
根据本公开的实施例，考虑到实际地质数据集的坐标范围可能彼此相差很大，因此在地质构造模型训练阶段需要对构造数据进行缩放，以获得值域范围为0到1的归一化地质仿真构造模型。在归一化训练数据时，将同一地质界面上的散点赋予归一化模型相应的值。
根据本公开的实施例，上述神经网络包括编码器分支和解码器分支，上述编码器分支包括至少一个编码特征处理单元，上述解码器分支包括至少一个解码特征处理单元；其中，上述利用上述训练样本数据集训练神经网络包括：将上述地质断层数据和上述地质层位数据输入上述编码器分支，经过上述编码器分支中的至少一个上述编码特征处理单元处理之后，输出最终编码隐藏特征；将上述最终编码隐藏特征输入上述解码器分支，经过上述解码器分支中的至少一个上述解码特征处理单元处理之后，输出预测构造模型；将上述预测构造模型和上述仿真构造模型输入混合损失函数，输出损失结果，其中，上述混合损失函数将基于单点的回归损失与基于局部滑动高斯窗的多尺度结构相似性进行结合；根据上述损失结果调整上述神经网络的网络参数，直至上述混合损失函数满足预设条件；将上述混合损失函数满足上述预设条件时得到的预测构造模型作为上述训练好的神经网络。
根据本公开的实施例，编码器分支可以包括五个编码特征处理单单元依次连接。
根据本公开的实施例，上述编码器分支包括依次连接的第一编码特征处理单元、第二编码特征处理单元、第三编码特征处理单元、第四编码特征处理单元和第五编码特征处理单元；
上述将上述地质断层数据和上述地质层位数据输入上述编码器分支，经过上述编码器分 支中的至少一个上述编码特征处理单元处理之后，输出最终编码隐藏特征包括：
将上述地质断层数据和上述地质层位数据输入上述第一编码特征处理单元，输出第一编码隐藏特征；
将上述第一编码隐藏特征输入上述第二编码特征处理单元，输出第二编码隐藏特征；
将上述第二编码隐藏特征输入上述第三编码特征处理单元，得到第三编码隐藏特征；
将上述第三编码隐藏特征输入上述第四编码特征处理单元，得到第四编码隐藏特征；
将上述第四编码隐藏特征输入上述第五编码特征处理单元，得到上述最终编码隐藏特征。
根据本公开的实施例，上述编码特征处理单元包括反向残差模块和最大池化层，上述反向残差模块包括扩展卷积层、深度可分离卷积层和投影卷积层；上述将上述地质断层数据和上述地质层位数据输入上述编码器分支，经过上述编码器分支中的至少一个上述编码特征处理单元处理之后，输出最终编码隐藏特征包括：将上述地质断层数据和上述地质层位数据输入上述扩展卷积层，以扩展特征维数，输出扩展特征；将上述扩展特征输入上述深度可分离卷积层进行特征提取，输出提取特征；将上述提取特征输入上述投影卷积层，以对上述提取特征进行降维处理，输出第一编码隐藏特征；将上述第一编码隐藏特征输入上述最大池化层，输出上述最终编码隐藏特征。
根据本公开的实施例，上述深度可分离卷积层包括单通道特征提取层和跨通道特征融合层；上述将上述扩展特征输入上述深度可分离卷积层进行特征提取，输出提取特征包括：将上述扩展特征输入上述单通道特征提取层，输出单通道特征；将上述单通道特征输入上述跨通道特征融合层，以对上述单通道特征进行融合，输出上述提取特征。
根据本公开的实施例，深度可分离卷积层包括18181单通道特征提取层和18181跨通道特征融合层，用于将输入数据扩展到更高维特征空间，并将其映射回输出通道。
图3A示意性示出了根据本公开实施例的神经网络的架构图。
图3B示意性示出了根据本公开实施例的反向残差模块的结构示意图。
图3C示意性示出了根据本公开实施例的通道注意力模块的结构示意图。
如图3A所示，神经网络由编码器分支和解码器分支组成。输入的稀疏构造数据(例如地质断层数据和地质层位数据)首先在编码器分支进行特征提取，之后在解码器分支恢复成完整的隐式构造模型。在神经网络的编码器分支部分，输入的稀疏构造数据按顺序经过5个编码特征处理单元(即E0到E4)，每个编码特征处理单元单元对应一种空间尺度，除E0单元外，其他编码特征处理单元的输入均为前一个编码特征处理单元的输出。
每个编码特征处理单元内部均使用最大池化层对输入特征进行下采样，并同时扩展其特征维数，以提取相应的空间尺度信息。基于输入数据的低秩性，编码器单元采用了线性瓶颈 和反向残差结构。
如图3B所示，反向残差模块包括1×1×1扩展卷积层、3×3×3深度可分离卷积层和1×1×1投影卷积层，且每层卷积之后包括批量归一化(BN)和非线性激活函数(ReLu)。深度可分离卷积层末端的两个1×1×1卷积层用于将输入扩展到更高维特征空间，并将其映射回输出通道。利用扩展卷积层和投影卷积层上的局部跳跃连接将该单元表示为残差学习函数，加速了梯度响应在网络中的反向传播。
虽然编码器能聚合丰富信息，但并非所有特征都对预测有利。由于输入数据的稀疏性和异质性，存在许多和地质构造不相关的特征，且特征通道中的大部分值为零。为增强编码器分支的辨别学习能力，在编码器的最后三个编码特征提取单元中插入轻量级的通道注意力模块，即在反向残差模块中插入通道注意力模块。
如图3C所示，通道注意力模块由挤压模块和激发次级模块组成。输入特征通过重拍操作重拍，在挤压模块中被压缩到低维特征空间，然后通过激发次级模块转换为通道注意力权重。这使得网络能自适应学习各个特征维度之间的依赖关系，从而通过抑制不相关特征来重新调整它们在预测过程中的重要性，提升了网络预测性能。
如图3A所示，解码器分支包括与编码器分支对应的五个解码特征处理单元(从D4到D0)，以获得网络的全模型空间感受野。解码器分支中每个解码特征处理单元负责聚合来自前一解码特征处理单元和编码器跳连层的隐藏特征，同时补偿不同特征之间的空间分辨率的不匹配。在每个空间尺度层级上，解码器特征与对应的编码器特征拼接(如E3和D4的隐藏特征)，并输入到两个连续的深度可分离卷积层中以进一步提取有效信息。
深度可分离卷积层将卷积运算拆解为两个单独的两个步骤，包括用于单通道特征提取的3×3×3卷积层和用于融合跨通道特征的1×1×1卷积层。通过将卷积操作分解为两个步骤，深度可分离卷积可以显著降低计算复杂度和内存占用，从而有助于构建轻量级网络。解码器分支之后的输出层采用1×1×1卷积层实现简单的线性变换，以融合所有解码特征并输出建模结果。
根据本公开的实施例，CNN架构上使用了类似于UNet及其相关变体的编码器-解码器架构设计，并进行了改进，以求在地质建模准确性和计算效率之间找到最佳折衷。在编码器-解码器结构中，输入到网络的数据先在编码器中以多个空间分辨率进行下采样，然后通过解码器中的跳连机制与对应的上采样特征重新组合。因此，输入数据的局部特征通常在CNN的浅层被提取，而当深层卷积层中的感受野越来越大时，逐渐转变为高层次的全局特征。由于具有不同空间分辨率的隐藏特征通常具有不同的几何信息，系统地聚集多尺度隐藏特征有助于网络产生可靠且稳定的预测结果。此外，因为构造信息可能会在多次下采样中逐渐丢失， 浅层提取的低层次特征通常比深层特征能更好地遵循输入数据，而跳连机制的使用有助于网络生成与已知构造兼容的模型。
根据本公开的实施例，当CNN训练完成后，从训练样本数据集中学习的建模经验和知识将隐式地嵌入到网络参数中。为验证得到的地址构造模型的建模性能，将训练得到的地质构造模型应用于未包含在训练样本数据集的地质仿真构造模型上。这些地质仿真构造模型具有复杂的断层和褶皱构造，且地质分界面被多个高角度正断层改造。为测试方法的有效性，提取这些地质仿真构造模型的不完整的层位数据和断层数据，并输入到地质构造模型。尽管没有用于约束构造模式的全局信息，但地质构造模型仍然能输出与输入构造特征相似的模型，并保留地质界面的变化趋势。
同时，除定性分析外，通过使用多种指标定量评估网络的建模能力。这些量化评价指标包括SSIM、MSE、MAE、可解释方差得分(EVS)、均方对数误差(MSLE)、中值绝对误差(MDAE)和拟合优度(R2S)。此外，还通过计算层位拟合精度(HFA)衡量与输入层位相关的建模精度。
表2示出了根据本公开的实施例的构建方法与目前主流网络在地质建模问题上的定量对比。
如表2所示，与其他方法相比，本公开提供的方法不仅在所有指标上均有更好的表现，而且具有更轻量级的网络架构，即更少的训练参数和计算复杂度。这使得该方法能够应用于大型三维建模任务。此外，注意力模块的使用让网络学习到不同特征维度间的相互依赖关系，从而自适应地强调有助于提升预测性能的特征，同时抑制不相关的特征。因此，尽管特征的维度均少于相关技术中的CNN结构(如UNet)，但该网络仍然实现了在验证数据集上稳定的建模结果和可靠的泛化能力。
表2
Figure PCTCN2022125275-appb-000010
根据本公开的实施例，在不同的复杂实际数据应用中，仅使用仿真数据训练的CNN也能 表现出优异的建模能力。与隐式方法中的任何显式数学约束不同，基于CNN的智能构造建模过程是通过具有可训练参数的空间卷积和提供各种地质约束的损失函数来驱动的。网络中的空间卷积可以等效于相关技术中使用的局部插值算子，唯一的区别在于空间卷积能通过训练优化，提供了更多的灵活性，因此能处理相关技术难以在数学模型中准确表达的几何特征。由于构造建模依赖于对观测数据空间分布和几何关系的分析，提供具代表性的数据样本对训练CNN以及实现其可靠的泛化能力至关重要。本公开使用的数据仿真方法能生成大量且逼真的构造模型，且在构建训练数据集时充分考虑实际数据常见的局部缺失情况。这种构造正演和网络训练相结合的方式能有针对性地将训练样本中的各种构造模式和地质规律嵌入到神经网络中，从而形成知识闭环。这也是经过训练的网络能成功应用于具有不同构造模式的实际例子的重要原因。由于建模过程的全自动化和基于并行计算平台的高效率实现，网络能对输入构造数据进行人工交互下的实时建模。例如，对同一套输入数据加以随机扰动以生成多个的建模结果，进而分析构造解释中存在的地质不确定性和构造一致性。
图4示意性示出了根据本公开实施例的隐式构造建模方法的流程图。
如图4所示，该方法主要分为训练样本库准备阶段、训练阶段和推理阶段。在训练样本库准备阶段，随机选择断层、褶皱、倾斜构造的参数；利用仿真构造模型生成器生成仿真构造模型；然后提取仿真构造模型的断层和层位散点，构成作为输入数据的散点数据。在训练阶段，将不完整的层位和所有断层的散点数据作为输入数据，输入神经网络CNN中进行训练，并利用仿真构造模型监控网络学习过程，从而输出完整的隐式构造模型。在推理阶段，将网络应用于不同区域的实际构造解释数据上，以获得对应的完整地质构造模型。首先，由于构造解释数据通常不一定位于规则网格上，在实际应用前，需要对实际工区构造解释资料的层位和断层解释数据转换到建模网格，并获得用于输入网络的构造数据；然后将构造数据输入上述训练好的隐式构造模型中，输出与输入构造特征兼容且地质上合理的隐式构造模型；之后通过提取模型等值面和不连续跳跃特征，可以分别得到完整的地质层位面和断层面。
图5示意性示出了根据本公开实施例的构建方法从输入的稀疏构造数据生成空间标量函数，作为对所有地质构造的隐式表示。
如图5所示，将隐式建模视为深度学习图像修复问题，基于从训练数据集学习的知识，网络能从稀疏或非均匀采样数据中重建完整的地质构造模型。深度学习方法允许通过定义适当的损失函数来合理衡量所比较模型间的差异，且能灵活地引入地质认识和构造约束。该网络输入为断层和层位数据，输出为全空间三维标量场，作为对所有地质构造的隐式表示。将断层和褶皱模拟参数化，通过在合理范围内随机选择参数，自动创建具有多样化构造特征的仿真模型，从而构建完备的训练样本库。在训练网络时，从这些模型中随机提取层位和断层 散点，以生成稀疏或不均匀分布的数据作为网络输入，以预测完整的地质构造模型，而仿真模型本身被用于监督。在网络训练的过程中，定义了一个混合损失函数，该函数将基于单点的回归损失与基于局部滑动高斯窗的多尺度结构相似性结合，以确保地质预测结果与已知构造兼容。当网络输出隐式模型后，通过提取其等值面，以及检测其局部值的快速变化，可分别提取出任意完整的地质层位面和断层(如图5)。通过实际数据测试，训练后的网络能有效地创建地质合理且构造一致的模型，显示出利用深度学习技术提高地质建模质量的潜力。此外，相对于其他方法，由于不需要在每个计算过程中随机初始化网络参数，该方法的建模结果是可复现的。
根据本公开的实施例，为综合评估训练后网络的性能表现以及提出构建方法的通用性和有效性，分别在三种不同的地质应用场景中进行构造建模，分别包括基于野外露头观测资料的二维构造建模、基于测井资料的实时交互式构造建模和基于实际地震解释资料的三维构造建模。
图6示意性示出了基于野外露头观测资料的二维构造建模的应用。
图6中的图a显示了通过对某盆地进行野外露头观测获取的地质解释资料。由于矿床中石膏的同沉积溶解引起的局部沉降，该露头剖面存在一系列中、高角度断层(虚线表示)。这些同沉积断层控制了地层界面的横向厚度变化。为进行构造建模，这些野外观测资料被整合到规则的模型网格中，以便用作CNN的输入，进而获得该露头完整的构造模型。
图6中的图b展示了使用离散色棒显示的预测构造模型，图6中的图d则显示了对应的使用连续色棒显示的建模结果。可以发现，模型中发生位错的不同地质层位(标量值的不连续)与图6中的图a展示的作为网络输入的断层数据有较好的匹配。此外，图6中的图c分别显示了从建模结果(图6中的图b和图d)中提取的四个层位与输入层位数据的拟合情况，证明了建模结果对输入构造数据的兼容性和其地质上的合理性。此外，尽管网络是使用仿真数据进行训练的，但它在具有复杂地质构造的实际数据集上仍具有良好的建模性能。
当对复杂地质构造建模时，隐式方法的可靠性在很大程度上取决于输入数据的质量以及是否较均匀分布。然而，在实际应用中构造数据通常是极度稀疏或局部聚集分布的。这导致建模变成不适定问题，即存在多个拟合输入数据的模型。因此，尤其是对于噪声和难以自洽的构造解释数据，地质不确定性分析对于寻找最佳的最符合地质认识的建模结果至关重要。虽然相关技术中的隐式方法可以通过扰动输入数据来描述由此产生的地质不确定性，并生成各种不同的模型。但它们无法通过随机模拟单一的几何模式来研究可能存在的多种复杂的地质构造组合规律和空间关系。相比之下，由于智能建模方法的全自动化和的高计算效率，网络能对输入构造数据进行人工交互下的实时建模，以生成不同的构造模型，从而有助于分析 地质不确定性。
图7示意性示出了基于测井资料的实时交互式构造建模应用。
图7中的图a利用在不同地质条件下从钻孔和露头观测中解释的层位和断层进行建模，以研究与构造位置变化相关的地质不确定性。第一个数据例子展示了简单的连续层状模型。而当考虑额外的几何对象(如断层或不整合面)使连续地质层位错断时，建模情况变得更加复杂化和多样化。此外，由于地层边界的过渡通常无法从垂直钻孔中准确测，随机扰动解释的层位位置可以产生地层厚度的变化。如图7中的图b所示，网络将多种构造组合作为输入，以模拟图7中的图a多种可能的构造形态和关系对建模过程的影响。图7中的图c所示的所有结果都是通过使用单块GPU(NVIDIA Tesla V100)的计算得到的。虽然网络训练需要花费几小时，但使用训练后的网络生成每个模型的平均时间约为2毫秒(模型尺寸为128×256)，同时，该训练后的网络不局限于输入数据的特定格式，输入数据灵活，可实时更改输入数据，因此该方法具有实时可交互式建模的能力。此外，图7中的图d显示与输入层位数据和建模层位之间良好的拟合情况，再次证明了方法能输出与给定构造特征兼容的地质构造模型。
图8示意性示出了基于实际地震解释数据的三维地质构造建模应用案例。
根据本公开的实施例，为验证网络的建模能力，将经过训练后的CNN应用于三维构造建模问题，拟从地震解释资料中提取不均匀采样散点数据作为网络输入，以构建完整的地质构造模型。在构造复杂变形区域采集的三维地震数据具有相对较低的分辨率和信噪比。由于受到噪声的影响，一些地震反射层很难在整个数据空间内被连续地追踪。紧密间隔以及交叉断层，以及与结构构造不连续性类似的噪声和地层特征时使得地震构造的几何关系和空间组织进一步复杂化和模糊化。
如图8中的图a所示，由于采样数据之间的距离变化较大，导致在地质界面周围的数据点在某些局部区域相对稀疏或聚集。图8中的图b显示的建模结果表明，提出的方法能够稳定生成地质上有效的三维构造模型。在图8中的图c中，使用等值面提取方法从生成的标量场中提取完整的地质界面，并去除断层附近的层位，以突出中由于断层存在而造成的构造不连续。
图8中的图d显示了一个重建的由垂直坐标值着色的完整层位界面，其中存在跨越断层的垂向跳跃。如图8中的图c和8中的图d所示，网络输出的构造不连续和地质界面均可与输入构造兼容，预测出的模型甚至可以保持褶皱构造的变化，而无需额外提供用于约束建模行为的全局信息。通过比较，在同一地质界面上的输入层位数据能与预测模型相应等值面较好匹配，证明了网络能够生成与输入构造特征兼容的模型。
需要说明的是，本公开实施例中的流程图所示的操作除非明确说明不同操作之间存在执 行的先后顺序，或者不同操作在技术实现上存在执行的先后顺序，否则，多个操作之间的执行顺序可以不分先后，多个操作也可以同时执行。
基于上述基于多源异构数据的地质构造建模方法，本公开还提供了一种基于多源异构数据的地质构造建模装置。以下将结合图9对该装置进行详细描述。
图9示意性示出了根据本公开实施例的基于多源异构数据的地质构造建模装置的结构框图。
如图9所示，该实施例的基于多源异构数据的地质构造建模装置900包括获取模块910、转换模块920和输入输出模块930。
获取模块910，被配置用于获取目标区域的多源异构数据，其中，上述多源异构数据包括断层解释数据和层位解释数据，上述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据。在一实施例中，获取模块910可以被配置用于执行前文描述的操作S110，在此不再赘述。
转换模块920，被配置用于基于预设模型网格转换上述断层解释数据和上述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据。在一实施例中，转换模块920可以被配置用于执行前文描述的操作S120，在此不再赘述。
输入输出模块930，被配置用于将上述转换后的断层解释数据和上述转换后的层位解释数据输入到预先训练好的神经网络中，输出与上述目标区域对应的完整地质构造模型。在一实施例中，输入输出模块930可以被配置用于执行前文描述的操作S130，在此不再赘述。
图10示意性示出了根据本公开实施例的可用于多源异构数据建模的神经网络训练装置的结构框图。
如图10所示，该实施例的神经网络的训练装置1000包括生成模块1010、提取模块1020和训练模块1030。
生成模块1010被配置用于利用数据仿真技术生成地质仿真构造模型，得到地质仿真构造模型库，上述地质仿真构造模型具有多样化褶皱和断层特征。在一实施例中，生成模块1010可以被配置用于执行前文描述的操作S210，在此不再赘述。
提取模块1020被配置用于针对上述地质仿真构造模型库中的每个地质仿真构造模型，获取上述地质仿真构造模型的地质断层数据和地质层位数据，最终得到训练样本数据集。在一实施例中，提取模块1020可以被配置用于执行前文描述的操作S220，在此不再赘述。
训练模块1030被配置用于利用上述训练样本数据集训练神经网络，得到可用于多源异构数据地质建模的神经网络。在一实施例中，训练模块1030可以被配置用于执行前文描述的操作S230，在此不再赘述。
根据本公开的实施例的模块、子模块、单元、子单元中的任意多个、或其中任意多个的至少部分功能可以在一个模块中实现。根据本公开实施例的模块、子模块、单元、子单元中的任意一个或多个可以被拆分成多个模块来实现。根据本公开实施例的模块、子模块、单元、子单元中的任意一个或多个可以至少被部分地实现为硬件电路，例如现场可编程门阵列(FPGA)、可编程逻辑阵列(PLA)、片上系统、基板上的系统、封装上的系统、专用集成电路(ASIC)，或可以通过对电路进行集成或封装的任何其他的合理方式的硬件或固件来实现，或以软件、硬件以及固件三种实现方式中任意一种或以其中任意几种的适当组合来实现。或者，根据本公开实施例的模块、子模块、单元、子单元中的一个或多个可以至少被部分地实现为计算机程序模块，当该计算机程序模块被运行时，可以执行相应的功能。
根据本公开的实施例，获取模块910、转换模块920、输入输出模块930、生成模块1010、提取模块1020和训练模块1030中的任意多个模块可以合并在一个模块中实现，或者其中的任意一个模块可以被拆分成多个模块。或者，这些模块中的一个或多个模块的至少部分功能可以与其他模块的至少部分功能相结合，并在一个模块中实现。根据本公开的实施例，获取模块910、转换模块920、输入输出模块930、生成模块1010、提取模块1020和训练模块1030中的至少一个可以至少被部分地实现为硬件电路，例如现场可编程门阵列(FPGA)、可编程逻辑阵列(PLA)、片上系统、基板上的系统、封装上的系统、专用集成电路(ASIC)，或可以通过对电路进行集成或封装的任何其他的合理方式等硬件或固件来实现，或以软件、硬件以及固件三种实现方式中任意一种或以其中任意几种的适当组合来实现。或者，获取模块910、转换模块920、输入输出模块930、生成模块1010、提取模块1020和训练模块1030中的至少一个可以至少被部分地实现为计算机程序模块，当该计算机程序模块被运行时，可以执行相应的功能。
需要说明的是，本公开的实施例中基于多源异构数据的地质构造建模装置部分与本公开的实施例中基于多源异构数据的地质构造建模方法部分是相对应的，基于多源异构数据的地质构造建模装置部分的描述具体参考基于多源异构数据的地质构造建模方法部分，在此不再赘述。
图11示意性示出了根据本公开实施例的适于实现上述方法的电子设备的方框图。
如图11所示，根据本公开实施例的电子设备1100包括处理器1101，其可以根据存储在只读存储器(ROM)1102中的程序或者从存储部分1108加载到随机访问存储器(RAM)1103中的程序而执行各种适当的动作和处理。处理器1101例如可以包括通用微处理器(例如CPU)、指令集处理器和/或相关芯片组和/或专用微处理器(例如，专用集成电路(ASIC))等等。处理器1101还可以包括用于缓存用途的板载存储器。处理器1101可以包括用于执行根据本公 开实施例的方法流程的不同动作的单一处理单元或者是多个处理单元。
在RAM 1103中，存储有电子设备1100操作所需的各种程序和数据。处理器1101、ROM 1102以及RAM 1103通过总线1104彼此相连。处理器1101通过执行ROM 1102和/或RAM 1103中的程序来执行根据本公开实施例的方法流程的各种操作。需要注意，所述程序也可以存储在除ROM 1102和RAM 1103以外的一个或多个存储器中。处理器1101也可以通过执行存储在所述一个或多个存储器中的程序来执行根据本公开实施例的方法流程的各种操作。
根据本公开的实施例，电子设备1100还可以包括输入/输出(I/O)接口1105，输入/输出(I/O)接口1105也连接至总线1104。电子设备1100还可以包括连接至I/O接口1105的以下部件中的一项或多项：包括键盘、鼠标等的输入部分1106；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分1107；包括硬盘等的存储部分1108；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分1109。通信部分1109经由诸如因特网的网络执行通信处理。驱动器1110也根据需要连接至I/O接口1105。可拆卸介质1111，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器1110上，以便于从其上读出的计算机程序根据需要被安装入存储部分1108。
本公开还提供了一种计算机可读存储介质，该计算机可读存储介质可以是上述实施例中描述的设备/装置/系统中所包含的；也可以是单独存在，而未装配入该设备/装置/系统中。上述计算机可读存储介质承载有一个或者多个程序，当上述一个或者多个程序被执行时，实现根据本公开实施例的方法。
根据本公开的实施例，计算机可读存储介质可以是非易失性的计算机可读存储介质，例如可以包括但不限于：便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。例如，根据本公开的实施例，计算机可读存储介质可以包括上文描述的ROM 1102和/或RAM 1103和/或ROM 1102和RAM 1103以外的一个或多个存储器。
本公开的实施例还包括一种计算机程序产品，其包括计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。当计算机程序产品在计算机系统中运行时，该程序代码用于使计算机系统实现本公开实施例所提供的基于多源异构数据的地质构造模型的训练方法或构建方法。
在该计算机程序被处理器1101执行时执行本公开实施例的系统/装置中限定的上述功能。根据本公开的实施例，上文描述的系统、装置、模块、单元等可以通过计算机程序模块来实 现。
在一种实施例中，该计算机程序可以依托于光存储器件、磁存储器件等有形存储介质。在另一种实施例中，该计算机程序也可以在网络介质上以信号的形式进行传输、分发，并通过通信部分1109被下载和安装，和/或从可拆卸介质1111被安装。该计算机程序包含的程序代码可以用任何适当的网络介质传输，包括但不限于：无线、有线等等，或者上述的任意合适的组合。
在这样的实施例中，该计算机程序可以通过通信部分1109从网络上被下载和安装，和/或从可拆卸介质1111被安装。在该计算机程序被处理器1101执行时，执行本公开实施例的系统中限定的上述功能。根据本公开的实施例，上文描述的系统、设备、装置、模块、单元等可以通过计算机程序模块来实现。
根据本公开的实施例，可以以一种或多种程序设计语言的任意组合来编写用于执行本公开实施例提供的计算机程序的程序代码，具体地，可以利用高级过程和/或面向对象的编程语言、和/或汇编/机器语言来实施这些计算程序。程序设计语言包括但不限于诸如Java，C++，python，“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算设备上执行、部分地在用户设备上执行、部分在远程计算设备上执行、或者完全在远程计算设备或服务器上执行。在涉及远程计算设备的情形中，远程计算设备可以通过任意种类的网络，包括局域网(LAN)或广域网(WAN)，连接到用户计算设备，或者，可以连接到外部计算设备(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，上述模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图或流程图中的每个方框、以及框图或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
本领域技术人员可以理解，本公开的各个实施例和/或权利要求中记载的特征可以进行多种组合或/或结合，即使这样的组合或结合没有明确记载于本公开中。特别地，在不脱离本公开精神和教导的情况下，本公开的各个实施例和/或权利要求中记载的特征可以进行多种组合和/或结合。所有这些组合和/或结合均落入本公开的范围。
以上对本公开的实施例进行了描述。但是，这些实施例仅仅是为了说明的目的，而并非 为了限制本公开的范围。尽管在以上分别描述了各实施例，但是这并不意味着各个实施例中的措施不能有利地结合使用。本公开的范围由所附权利要求及其等同物限定。不脱离本公开的范围，本领域技术人员可以做出多种替代和修改，这些替代和修改都应落在本公开的范围之内。

Claims (12)

  1. 一种基于多源异构数据的地质构造建模方法，包括：
    获取关于目标区域的多源异构数据，其中，所述多源异构数据包括断层解释数据和层位解释数据，所述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据；
    基于预设模型网格转换所述断层解释数据和所述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据；以及
    将所述转换后的断层解释数据和所述转换后的层位解释数据输入到预先训练好的神经网络中，输出与所述目标区域对应的完整地质构造模型。
  2. 根据权利要求1所述的方法，其中，所述利用预设模型网格转换所述断层解释数据和所述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据包括：
    针对所述断层解释数据中的每个断层，将距离所述断层一个网格以内的点标为第一值，得到所述转换后的断层解释数据；
    针对所述层位解释数据中的每个层位，将离所述层位一个网格以内的点设为第二值，得到所述转换后的层位解释数据，其中，所述第二值为所述层位上所有散点的纵坐标平均值。
  3. 根据权利要求1所述的方法，其中，所述训练好的神经网络的训练方法，包括：
    利用数据仿真技术生成多个地质仿真构造模型，得到地质仿真构造模型库，所述地质仿真构造模型具有多样化褶皱和断层特征；
    针对所述地质仿真构造模型库中的每个地质仿真构造模型，获取所述地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，最终得到训练样本数据集；以及
    利用所述训练样本数据集训练神经网络，得到可用于多源异构数据地质建模的神经网络。
  4. 根据权利要求3所述的方法，其中，所述利用数据仿真技术生成多个地质仿真构造模型，得到地质仿真构造模型库包括：
    构建初始地质仿真构造模型，其中，所述初始地质仿真构造模型为水平方向恒定、垂向单调递增的层状模型，所述初始地质仿真构造模型包括至少一个构造参数；其中，所述构造参数包括褶皱参数、倾斜参数、断层构造参数中的至少一个；
    针对所述至少一个构造参数中的每个构造参数，在预设范围内调整所述构造参数，得到所述地质仿真构造模型库。
  5. 根据权利要求3所述的方法，其中，所述获取所述地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集包括：
    提取所述地质仿真构造模型中的不连续特征，得到所述地质断层数据；
    从所述地质仿真构造模型的等值面中选取预设数量的等值面，得到目标等值面集合；
    针对所述目标等值面集合中的每个目标等值面，将所述目标等值面上的散点数据，按照纵坐标值域范围划分为多个数据区间；
    随机丢弃所述多个数据区间中第二预设数量的数据区间，剩余的数据区间内的散点数据作为所述地质层位数据；
    根据所述地质断层数据、所述地质层位数据和所述地质仿真构造模型，生成所述训练样本数据，得到所述训练样本数据集。
  6. 根据权利要求5所述的方法，其中，从所述地质仿真构造模型的等值面中选取预设数量的等值面，得到目标等值面集合包括：
    将所述地质仿真构造模型中的所有等值面的值域范围按照降序划分为至少一个数值区间；
    针对所述至少一个数值区间中的每个数值区间，从所述数值区间中随机选取一个数值，得到目标数值；
    确定与所述目标数值对应的等值面为目标等值面，得到所述目标等值面集合。
  7. 根据权利要求3所述的方法，还包括：
    利用预设归一化方式对所述地质仿真构造模型进行归一化处理，得到归一化后的地质仿真构造模型；
    其中，所述获取所述地质仿真构造模型的地质断层数据和不均匀分布的地质层位数据，得到训练样本数据集包括：
    提取所述归一化后的地质仿真构造模型的地质断层数据和地质层位数据，得到所述训练样本数据集。
  8. 根据权利要求3所述的方法，其中，所述神经网络包括编码器分支和解码器分支，所述编码器分支包括至少一个编码特征处理单元，所述解码器分支包括至少一个解码特征处理单元；
    其中，所述利用所述训练样本数据集训练神经网络包括：
    将所述地质断层数据和所述地质层位数据输入所述编码器分支，经过所述编码器分支中的至少一个所述编码特征处理单元处理之后，输出最终编码隐藏特征；
    将所述最终编码隐藏特征输入所述解码器分支，经过所述解码器分支中的至少一个所述解码特征处理单元处理之后，输出预测构造模型；
    将所述预测构造模型和所述仿真构造模型输入混合损失函数，输出损失结果，其中，所述混合损失函数将基于单点的回归损失与基于局部滑动高斯窗的多尺度结构相似性进行结合；
    根据所述损失结果调整所述神经网络的网络参数，直至所述混合损失函数满足预设条件；
    将所述混合损失函数满足所述预设条件时得到的神经网络作为所述训练好的神经网络。
  9. 根据权利要求8所述的方法，其中，所述编码特征处理单元包括反向残差模块和最大池化层，所述反向残差模块包括扩展卷积层、深度可分离卷积层和投影卷积层；
    所述将所述地质断层数据和所述地质层位数据输入所述编码器分支，经过所述编码器分支中的至少一个所述编码特征处理单元处理之后，输出最终编码隐藏特征包括：
    将所述地质断层数据和所述地质层位数据输入所述扩展卷积层，以扩展特征维数，输出扩展特征；
    将所述扩展特征输入所述深度可分离卷积层进行特征提取，输出提取特征；
    将所述提取特征输入所述投影卷积层，以对所述提取特征进行降维处理，输出第一编码隐藏特征；
    将所述第一编码隐藏特征输入所述最大池化层，输出所述最终编码隐藏特征。
  10. 一种基于多源异构数据的地质构造建模装置，包括：
    获取模块，被配置用于获取关于目标区域的多源异构数据，其中，所述多源异构数据包括断层解释数据和层位解释数据，所述多源异构数据包括以下至少一种：地质露头观测数据、测井数据、多种地球物理数据、经验知识数据；
    转换模块，被配置用于基于预设模型网格转换所述断层解释数据和所述层位解释数据，得到转换后的断层解释数据和转换后的层位解释数据；以及
    输入输出模块，被配置用于将所述转换后的断层解释数据和所述转换后的层位解释数据输入到预先训练好的神经网络中，输出与所述目标区域对应的完整地质构造模型。
  11. 一种电子设备，包括：
    一个或多个处理器；
    存储装置，被配置用于存储一个或多个程序，
    其中，当所述一个或多个程序被所述一个或多个处理器执行时，使得所述一个或多个处理器执行根据权利要求1～9中任一项所述的方法。
  12. 一种计算机可读存储介质，其上存储有可执行指令，该指令被处理器执行时使处理器执行根据权利要求1～9中任一项所述的方法。
PCT/CN2022/125275 2022-08-31 2022-10-14 基于多源异构数据的地质构造建模方法 Ceased WO2024045285A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/546,360 US20250029330A1 (en) 2022-08-31 2022-10-14 Method of modeling interactive intelligent three-dimensional implicit structure based on deep learning

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN202211064255.5 2022-08-31
CN202211064255 2022-08-31
CN202211240817.7A CN115587537A (zh) 2022-08-31 2022-10-11 基于多源异构数据的地质构造建模方法
CN202211240817.7 2022-10-11

Publications (1)

Publication Number Publication Date
WO2024045285A1 true WO2024045285A1 (zh) 2024-03-07

Family

ID=84779124

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/125275 Ceased WO2024045285A1 (zh) 2022-08-31 2022-10-14 基于多源异构数据的地质构造建模方法

Country Status (3)

Country Link
US (1) US20250029330A1 (zh)
CN (1) CN115587537A (zh)
WO (1) WO2024045285A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118626819A (zh) * 2024-08-15 2024-09-10 中南大学 一种地质特征识别方法、装置、终端设备及介质
CN118628756A (zh) * 2024-08-13 2024-09-10 大连健星建筑工程有限公司 基于多节点关联性的地质勘测预测方法及系统
CN118917555A (zh) * 2024-10-10 2024-11-08 山东能源数智云科技有限公司 基于工业互联的设备综合效率评估方法及装置
CN119006734A (zh) * 2024-10-18 2024-11-22 北京朗境创新技术有限公司 一种三维地质建模方法、装置及电子设备
CN119918122A (zh) * 2024-11-21 2025-05-02 宁波公路市政设计有限公司 一种隧道智能设计系统和方法
CN121071511A (zh) * 2025-09-03 2025-12-05 钜鑫电子技术（梅州）有限公司 一种基于特征空间映射的pcb产线工艺参数智能匹配方法

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117649529B (zh) * 2024-01-30 2024-07-09 中国科学技术大学 一种基于多维信号分析神经网络的测井数据解释方法
CN118469100B (zh) * 2024-07-11 2024-09-10 江苏省地质调查研究院 一种地热资源有利区预测方法、系统、设备及存储介质
CN118762144B (zh) * 2024-09-06 2024-11-22 中南大学 融合复杂断层网络的煤层地质体建模方法及系统
CN120067995B (zh) * 2025-03-04 2026-01-23 广东电网有限责任公司东莞供电局 多元负荷预测方法、设备、存储介质及程序产品
CN120370431B (zh) * 2025-04-11 2026-01-06 中国地质科学院 一种盆地区深层卤水成矿分析与评估方法
CN120354736A (zh) * 2025-04-16 2025-07-22 山东地矿开元工程科技有限公司 基于深度学习的岩土体参数智能反演系统
CN120707763B (zh) * 2025-06-27 2026-03-31 华能澜沧江水电股份有限公司 一种多源信息融合的岩石三维重构方法及系统
CN121235231B (zh) * 2025-12-04 2026-02-10 天津市地质工程勘测设计院有限公司 基于多源数据融合的煤矿采空区塌陷智能预测系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200183047A1 (en) * 2018-12-11 2020-06-11 Exxonmobil Upstream Research Company Automated Reservoir Modeling Using Deep Generative Networks
WO2022078514A1 (zh) * 2020-10-15 2022-04-21 山东大学 一种结构和波速随机布设的三维速度地质建模方法
CN114758085A (zh) * 2022-05-06 2022-07-15 煤炭科学研究总院有限公司 基于地质预测网络的三维地质模型生成方法及装置
CN114882187A (zh) * 2022-05-09 2022-08-09 吉林大学 一种区域三维地质建模数据的智能化生成方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110441820B (zh) * 2019-08-21 2020-06-16 中国矿业大学(北京) 地质构造的智能解释方法
CN112562075B (zh) * 2019-09-26 2024-08-27 中国石油天然气集团有限公司 多滑脱层构造三维地质建模方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200183047A1 (en) * 2018-12-11 2020-06-11 Exxonmobil Upstream Research Company Automated Reservoir Modeling Using Deep Generative Networks
WO2022078514A1 (zh) * 2020-10-15 2022-04-21 山东大学 一种结构和波速随机布设的三维速度地质建模方法
CN114758085A (zh) * 2022-05-06 2022-07-15 煤炭科学研究总院有限公司 基于地质预测网络的三维地质模型生成方法及装置
CN114882187A (zh) * 2022-05-09 2022-08-09 吉林大学 一种区域三维地质建模数据的智能化生成方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
BI ZHENGFA, WU XINMING, LI ZHAOLIANG, CHANG DEKUAN, YONG XUESHAN: "DeepISMNet: three-dimensional implicit structural modeling with convolutional neural network", GEOSCIENTIFIC MODEL DEVELOPMENT, vol. 15, no. 17, 8 September 2022 (2022-09-08), pages 6841 - 6861, XP093144899, ISSN: 1991-9603, DOI: 10.5194/gmd-15-6841-2022 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118628756A (zh) * 2024-08-13 2024-09-10 大连健星建筑工程有限公司 基于多节点关联性的地质勘测预测方法及系统
CN118626819A (zh) * 2024-08-15 2024-09-10 中南大学 一种地质特征识别方法、装置、终端设备及介质
CN118917555A (zh) * 2024-10-10 2024-11-08 山东能源数智云科技有限公司 基于工业互联的设备综合效率评估方法及装置
CN119006734A (zh) * 2024-10-18 2024-11-22 北京朗境创新技术有限公司 一种三维地质建模方法、装置及电子设备
CN119918122A (zh) * 2024-11-21 2025-05-02 宁波公路市政设计有限公司 一种隧道智能设计系统和方法
CN121071511A (zh) * 2025-09-03 2025-12-05 钜鑫电子技术（梅州）有限公司 一种基于特征空间映射的pcb产线工艺参数智能匹配方法

Also Published As

Publication number Publication date
CN115587537A (zh) 2023-01-10
US20250029330A1 (en) 2025-01-23

Similar Documents

Publication Publication Date Title
WO2024045285A1 (zh) 基于多源异构数据的地质构造建模方法
US11668853B2 (en) Petrophysical inversion with machine learning-based geologic priors
US11397272B2 (en) Data augmentation for seismic interpretation systems and methods
US11693139B2 (en) Automated seismic interpretation-guided inversion
EP3894902B1 (en) Subsurface models with uncertainty quantification
US12223420B2 (en) Artificial learning fracture system and method for predicting permeability of hydrocarbon reservoirs
US20180247227A1 (en) Machine learning systems and methods for data augmentation
CN108830421B (zh) 致密砂岩储层的含气性预测方法及装置
CN110276402B (zh) 一种基于深度学习语义边界增强的盐体识别方法
US20240045089A1 (en) Generating realistic synthetic seismic data items
Huang et al. Machine learning using U-net convolutional neural networks for the imaging of sparse seismic data
Wang et al. Seismic stratum segmentation using an encoder–decoder convolutional neural network
CN115639605A (zh) 基于深度学习的高分辨率断层的自动识别方法和装置
Lang et al. Semi-supervised seismic impedance inversion with convolutional neural network and lightweight transformer
Gao et al. Fault representation in structural modelling with implicit neural representations
CN115687887A (zh) 一种基于自监督生成对抗学习和重构物理约束的稀疏时频分析方法、设备及可读存储介质
CN118051793B (zh) 基于trd工法的地层数据实时分析与预测方法及系统
Liu et al. Foundation models for exploration geophysics
Wu et al. Seismic Impedance Inversion Using Conditional Generative Adversarial Networks Assisted by Spectral Reconstruction-Based Data Augmentation
Wu et al. Surrogate model for fracture propagation in heterogeneous reservoirs based on generative neural networks
CN120044591B (zh) 基于正交标注的断层识别模型的训练方法、断层识别方法、电子设备以及存储介质
US20250272916A1 (en) Geologically meaningful subsurface model generation based on a text description
US20260004398A1 (en) System and method for enhancing seismic images
Xie Impact of Deep Generative Models for solving non-linear Inverse Problems: application to Seismic Imaging
Xu et al. DiffSIM: Unconditional and conditional facies simulation based on denoising diffusion generative models

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 18546360

Country of ref document: US

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22957101

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 16.06.2025)

122 Ep: pct application non-entry in european phase

Ref document number: 22957101

Country of ref document: EP

Kind code of ref document: A1