A lithology prediction method based on geological statistical characteristic rule constraint

Abstract

This invention relates to the field of geophysical exploration technology and proposes a lithology prediction method based on the constraints of geostatistical feature regularities. This method uses a large-scale exploration model to extract geostatistical feature regularities for lithology prediction, including: extracting seismic data; constructing virtual wells, simulating and generating virtual well data, and generating synthetic seismic records based on the seismic data; extracting geostatistical features to obtain their regularities; and constructing a multi-channel deep neural network model using seismic data and geostatistical features as inputs for lithology prediction. This invention expands the virtual wells using a Markov chain model and a sequential Gaussian simulation algorithm, and optimizes the model using K-Fold cross-validation, improving the generalization ability and stability of the seismic dataset. By combining seismic data with geostatistical features, multi-channel input, and a feature fusion layer, multi-source information is integrated. The accuracy is improved from 85.34% to 95.48%, providing a scientific basis for oil and gas reservoir identification and well location deployment.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and proposes a lithology prediction method based on the constraints of geological statistical characteristics. Background Technology

[0002] Large-scale exploration models are models built using massive datasets and advanced machine learning techniques. These models are typically used to process and analyze complex datasets to discover patterns and correlations within the data. In the exploration field, these models can be applied to geological data analysis, mineral resource prediction, seismic data processing, and other areas to improve the accuracy and efficiency of exploration.

[0003] Large-scale exploration models include, but are not limited to, the following: 1. The GPT series models, such as GPT-2 and GPT-3, were developed by OpenAI and are used to generate text and understand language.

[0004] 2. BERT and its variants, such as RoBERTa and ALBERT, were developed by Google for natural language processing tasks.

[0005] 3. The T5 model, developed by Google's research team, is used for text-to-text conversion tasks.

[0006] 4. The XLNet model, jointly developed by Carnegie Mellon University and Google research teams, is used for natural language understanding.

[0007] 5. The Transformer model, proposed by Vaswani et al. in 2017, is the foundational architecture for many subsequent models.

[0008] 6. The Megatron-LM model, developed by NVIDIA, is used to train large-scale language models.

[0009] 7. The Switch Transformer model, developed by Google's research team, is designed to improve the scalability and efficiency of models.

[0010] Conventional intelligent lithology prediction methods include seismic attribute analysis, machine learning algorithms, statistical analysis, neural networks, fuzzy logic, support vector machines, and ensemble learning methods such as random forests and gradient boosting trees. In addition, there are geostatistical methods, such as Kriging interpolation, used to consider spatial correlations in lithology prediction. Stochastic simulation methods in geological modeling software, such as sequential indicator simulation, can also generate a variety of possible lithology distribution scenarios. Deep learning algorithms, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), perform excellently in processing seismic images and time-series data and can be used for lithology identification and prediction. There are also hybrid methods that combine the advantages of multiple technologies to improve the accuracy and robustness of predictions. These methods typically require large amounts of geological, geophysical, and geochemical data as training sets to build reliable prediction models.

[0011] In large-scale exploration models, the mapping relationship between complex heterogeneous lithological characteristics and seismic attributes is highly nonlinear. Conventional intelligent lithology prediction methods mostly rely solely on model training methods between well logging and seismic data, which is overly dependent on model quality, resulting in low training efficiency and high uncertainty in prediction results.

[0012] Developing lithology prediction guided by geology and physics is a technological trend to further improve prediction accuracy. By guiding learning and accelerating learning, better and faster lithology prediction results can be obtained.

[0013] Chinese patent application CN202311498385.4 discloses a method for predicting formation lithology. The method includes: real-time acquisition of multiple sets of logging data and actual drilling formation data from the wellbore; extraction of key parameters from the multiple sets of logging data; data filtering of the key parameters to obtain filtered key parameters; steady-state testing of the filtered key parameters; selection of logging data groups corresponding to key parameters whose steady-state test results meet preset requirements; merging the selected multiple sets of logging data with the actual drilling formation data; grouping the merged data according to different formations; training formation lithology prediction models based on the multiple merged data sets; obtaining trained formation lithology prediction models; the formation lithology prediction models are trained using deep learning algorithms; inputting the logging data to be tested into the trained formation lithology prediction models; and outputting the formation lithology prediction results corresponding to the logging data to be tested.

[0014] The above methods are highly dependent on high-quality logging and actual drilling data, and rely too much on model quality, resulting in low training efficiency and high uncertainty in prediction results.

[0015] Chinese patent application CN201910562055.4 discloses a lithology prediction method based on elastic parameters within a fuzzy logic framework. The method includes: establishing a membership function for each of multiple elastic parameters to each of multiple lithologies; substituting the elastic parameters from multiple sets of training data into the membership function to obtain the membership degree of each elastic parameter in each set of training data to each lithology, where each set of training data includes multiple elastic parameters and corresponding real lithologies; constructing a lithology prediction model based on the membership degree; and inputting the elastic parameter to be predicted into the lithology prediction model to obtain the corresponding predicted lithology.

[0016] The above methods are highly complex and rely on the selection and combination of elastic parameters, which limits their stability and generalizability under complex geological conditions.

[0017] The paper "Research on Semi-Supervised Learning Algorithm for Lithology Prediction Based on Label Propagation," published in the journal *Progress in Geophysics*, Vol. 36, No. 2, 2021 (pp. 540-548), employs a semi-supervised learning algorithm to address the issue of well logging interpretation accuracy. A lithology prediction framework of "clustering—manual annotation—pseudo-annotation—classification" is proposed. First, a clustering algorithm is used to select samples to be labeled. Then, based on the similarity of data in the feature space and geographic space, a semi-supervised learning method is used to propagate the label from manually labeled samples to unlabeled samples. Finally, based on the confidence of the pseudo-labels, a weighted support vector machine algorithm is used to design and train the classification model.

[0018] The aforementioned algorithms have certain limitations in relying on manual annotation and geospatial similarity. Summary of the Invention

[0019] The purpose of this invention is to address the shortcomings of existing technologies by proposing a lithology prediction method based on the constraints of geostatistical characteristics, thereby solving the following technical problems: Conventional intelligent lithology prediction methods mostly rely solely on model training between well logging and seismic data, which is overly dependent on model quality, resulting in low training efficiency and high uncertainty in prediction results. They also lack stability and generalization under complex geological conditions. Furthermore, the accuracy of methods based on manual annotation is difficult to guarantee, and their reliance on geographic spatial similarity has certain limitations.

[0020] To achieve the above objectives, the present invention adopts the following technical solution: A lithology prediction method based on geostatistical feature constraints, which uses a large-scale exploration model to extract geostatistical feature patterns for lithology prediction, including: Extracting earthquake data; Virtual wells are constructed, virtual well data is generated through simulation, and synthetic seismic records are generated based on seismic data; Extract geostatistical features to obtain the regularity of geostatistical features; Using seismic data and geological statistical features as input, a multi-channel deep neural network model is constructed to predict lithology.

[0021] Furthermore, step one specifically includes: The large-scale exploration model was analyzed to extract seismic data related to time and depth. Extract well logging information and obtain lithofacies calibration of the well logging information; By converting seismic time to depth, the depth calibration of seismic data and well logging data is aligned, ensuring consistency between the two data in the depth domain.

[0022] Furthermore, step two specifically includes: Virtual well lithology type simulation; Virtual well physical property parameter simulation; Seismic data synthesis: Extracting wavelets from actual seismic data; By combining the simulated P-wave velocity, S-wave velocity, and density data of the virtual well with the extracted wavelet, a synthetic seismic channel conforming to the geological characteristics of the specified working area is generated using the co-convolution technique.

[0023] Furthermore, step two also includes dataset expansion and the division of training and validation wells: Add the generated virtual well data and the corresponding synthetic seismic channel to the existing real well dataset; The expanded dataset is divided into training wells and validation wells proportionally.

[0024] Furthermore, the division between training wells and validation wells is performed using a random partitioning method or a stratified sampling method.

[0025] Furthermore, step three includes combining the geological structures and boundary conditions contained in the stratigraphic model within the large exploration model to statistically analyze the probability and thickness distribution of various lithologies for each well.

[0026] Furthermore, the lithological probability statistics include: The lithology probability of each well is calculated based on the thickness of different lithologies within the stratigraphic layer using a stratified statistical method. Calculate the average lithology probability of all wells in the stratigraphic unit based on the lithology probability of each well. The mean lithological probability of each stratum is organized into a transformation matrix, where the rows of the matrix represent strata, the columns represent lithology, and the matrix elements are the mean lithological probability of that stratum.

[0027] Furthermore, the thickness distribution statistics are obtained by measuring the thickness distribution of each lithology in each layer to describe the spatial variation of the thickness of the corresponding lithology.

[0028] Furthermore, the aforementioned geostatistical characteristics are as follows: —(1).

[0029] Furthermore, step four specifically includes: The input data is constructed and lithological information is used as a label. The input data includes seismic data and geological statistical features. Design corresponding multi-channel input structures based on earthquake attributes and geological statistical characteristics; The data from each channel is processed by an adaptive convolutional network to extract high-dimensional features, and then the features are deeply fused through a feature fusion layer. The fused features are further processed through a multi-layer nonlinear network to extract the complex mapping relationship between the data and lithological prediction. By generating lithology prediction results in the output layer using a normalization function, a multi-channel deep neural network architecture that integrates seismic attributes and geostatistical features is constructed.

[0030] The beneficial effects of this invention are as follows: Compared with existing technologies, this invention expands the virtual well using a Markov chain model and synthesizes seismic channels through sequential Gaussian simulation and co-convolution, thereby expanding the dataset and effectively alleviating the problem of insufficient training data, improving the model's generalization ability and stability. The model employs multi-channel input, integrating seismic data and geostatistical features, and achieves effective combination of multi-source information through feature extraction and fusion layers. Furthermore, K-Fold cross-validation technology is used to optimize the model, ensuring high prediction accuracy across different datasets. Taking the lithology prediction of a block in the Yanjia area as an example, the prediction accuracy of traditional methods is only 85.34%, and often results in situations that do not conform to basic geological understanding (such as incorrect identification of sandstone-mudstone interbedded layers). The model of this invention not only successfully avoids these problems but also improves the prediction accuracy to 95.48%. This improvement not only enhances the accuracy of lithology prediction but also better conforms to geological laws, providing a more scientific reference for oil and gas reservoir identification and well location deployment, meeting the application needs of practical engineering. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the process of this invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0033] A lithology prediction method based on the constraints of geostatistical characteristics mainly includes the following steps: Analyze the exploration model data, convert the seismic data extracted from the exploration model to time and depth, and align it with the lithological information in the actual well logging data in the depth domain; The lithology of virtual wells is constructed based on the Markov chain model, and the P-wave velocity and density are generated by sequential Gaussian co-simulation. Combined with wavelets in real seismic data, the generated synthetic seismic record conforms to the geological characteristics of the specified working area through co-convolution. Combining the geological structures and boundary conditions contained in the stratigraphic model within the large exploration model, a layered statistical method is used to calculate the probability distribution of different lithologies in each well, and further generate a lithology transformation matrix. At the same time, a thickness distribution model is constructed by using the actual thickness of lithology in the strata, thereby extracting the geological statistical characteristics and regularities. A multi-channel deep neural network architecture integrating seismic attributes and geostatistical features is constructed. Seismic attributes and geostatistical features are input into the network in a coordinated manner. High-dimensional features are extracted through an adaptive convolutional network, and combined with a feature fusion layer, multi-dimensional information is deeply fused. Based on statistical feature constraints, the accuracy of lithology prediction is improved, the optimization space is reduced, and the prediction efficiency is accelerated. K-Fold cross-validation technology is employed to continuously optimize and evaluate model performance, ensuring the model possesses good generalization ability and stability. This enables rapid prediction of target blocks, yielding lithological models that better align with geological understanding. Example 2

[0034] A lithology prediction method based on the constraints of geostatistical characteristics includes: In step 1, the large-scale exploration model is first analyzed to extract seismic data related to time-depth relationships, such as seismic wave amplitude, reflection coefficient, and velocity model. Then, well logging information is extracted, with a focus on obtaining the lithofacies calibration of the well logging information. Finally, through seismic time-depth conversion (converting the seismic data from the time domain to the depth domain), the depth calibration of the seismic data and the well logging data is aligned to ensure consistency between the seismic data and the well logging data in the depth domain.

[0035] In step 2, the first step is to expand the virtual well. Statistical analysis is performed based on existing real well logging data to extract the statistical regularities of lithofacies types and elastic coefficients of the subsurface reservoir. Subsequently, based on these statistical results, a Markov chain model is applied to simulate the spatial variation characteristics of the geological body. The transformation moments of the Markov chain are used to simulate the lithology of each layer in the virtual well using the transformation matrix. Next, the physical property parameter simulation stage begins. First, a sequential Gaussian simulation algorithm is used to generate the P-wave velocity distribution. Based on the principles of seismic exploration, and combined with velocity data from known wells, a P-wave velocity simulation array for the virtual well is completed to describe the spatial transformation probability between different lithofacies, thereby capturing the spatial correlation of the subsurface geological body. Then, using the simulated P-wave velocity as a secondary variable, a sequential Gaussian co-simulation algorithm is used to further simulate the shear wave velocity and density distribution. Finally, seismic data synthesis is performed: wavelets are extracted from actual seismic data and statistical analysis is conducted to obtain a wavelet model suitable for the regional stratigraphic characteristics. By combining the simulated P-wave and S-wave velocity and density data of the virtual well with the extracted wavelets, a synthetic seismic channel conforming to the geological characteristics of the specified working area is generated using the co-convolution technique. In step 2, the second step involves expanding the dataset and dividing it into training and validation wells. First, the generated virtual well data and corresponding synthetic seismic channels are added to the existing real-world well dataset. Next, the expanded dataset is divided into training and validation wells at an 8:2 ratio. To ensure that the training and validation sets have similar geological feature distributions, a random partitioning method is typically used. Furthermore, depending on actual needs, stratified sampling can also be used to further ensure consistency in the geological feature distributions between the training and validation sets. In step 3, combining the geological structures and boundary conditions contained in the stratigraphic model within the large-scale exploration model, the probability and thickness distribution of various lithologies for each well are statistically analyzed using a stratified statistical method. For lithology probability, the probability within each stratigraphic level refers to the proportion of different lithologies within that level. This proportion is calculated based on Formula 1, which represents the thickness of different lithologies within that level. After calculating the lithology probability for each well, the average lithology probability of all wells within each stratigraphic level is calculated. Then, the average lithology probability of each stratigraphic level is organized into a transformation matrix. The rows of the matrix represent stratigraphic levels, the columns represent lithologies, and the matrix elements are the average lithology probability of that stratigraphic level. For thickness distribution, the thickness distribution of each lithology within each stratigraphic level is measured to describe the spatial variation of the corresponding lithology thickness. Thus, two types of geological statistical characteristics are obtained. —(1) In step 4, the input data is first constructed, mainly consisting of two parts: seismic data (such as amplitude, waveform, and velocity) and geostatistical features (including lithology probability and thickness distribution), with lithology information used as a label. A corresponding multi-channel input structure is designed for the seismic attributes and geostatistical features. The data in each channel first undergoes high-dimensional feature extraction through an adaptive convolutional network, followed by deep feature fusion through a feature fusion layer. Next, the fused features are further processed through a multi-layer nonlinear network to extract the complex mapping relationship between the data and lithology prediction. Finally, the lithology prediction result is generated in the output layer using the Softmax activation function, constructing a multi-channel deep neural network architecture that integrates seismic attributes and geostatistical features.

[0036] In step 5, K-Fold cross-validation is used to optimize model performance. First, the data (including well data and seismic attribute data) in the training and validation sets are divided into K non-overlapping subsets. During each validation iteration, one subset is selected as the validation set, and the remaining K-1 subsets are used for training. This process is repeated K times. In this way, each subset is used as a validation set, ensuring the model can be comprehensively evaluated across different data subsets. Through multiple training and validation iterations, accuracy is used as the evaluation metric to select the best-performing model. After training, the optimized model is applied to the target block for rapid lithology prediction. Example 3

[0037] Taking a block in the Yanjia area of ​​Shengli Oilfield as an example, a lithology prediction method based on the constraints of geostatistical characteristics includes: Step 1: Select a block in the Yanjia area of ​​Shengli Oilfield as the research object. The block covers an area of ​​49 square kilometers and has a depth of about 6501.81 meters. 17 well logging data were obtained. The seismic data includes raw seismic data, instantaneous phase, instantaneous frequency and instantaneous amplitude attribute information. Time-depth conversion was performed on the seismic information and it was aligned with the lithological information in the depth domain. Step 2: Expand the total number of wells to 50 by expanding the virtual wells and seismic channels. Preprocess the dataset and divide the dataset into two parts: a training set containing 40 wells and a validation set containing 10 wells. The preprocessing mainly includes data format conversion and outlier removal. Step 3: Statistically analyze the lithological probability and thickness distribution of each well at different strata. The results show that the lithological probability of mudstone at a certain stratum in this block is 34.01%, sandstone is 26.29%, and conglomerate is 36.85%, with average thickness distribution values ​​of 68.5m, 43.2m, and 70.2m, respectively. Step 4: Input the seismic data, lithology probability, and thickness distribution as inputs, and lithology as a label into the lithology prediction network to construct a multi-channel deep neural network architecture that integrates seismic attributes and geological statistical characteristics of the region. Step 5: Model training was performed using K-Fold cross-validation, with K set to 5. The model performance was evaluated using the validation set. During the cross-validation process, the model achieved an average accuracy of 93.67%. Finally, the best-performing model was selected to predict lithology in other locations of the target block to guide subsequent exploration and development work. Example 4

[0038] Taking a certain block in Shengli Oilfield as an example, a lithology prediction method based on the constraints of geostatistical characteristics includes: Step 1: A block in the Shengli Oilfield was selected as the research object. The block covers an area of ​​65 square kilometers and has a depth of approximately 5800 meters. 22 well logging data were acquired. Seismic data included raw seismic data, instantaneous amplitude, phase, and velocity attribute information. These seismic data underwent time-depth conversion and were aligned with the lithological information in the well logging data in the depth domain to ensure data consistency and accuracy. Step 2 involves expanding the number of virtual wells and seismic channels to 60, thereby increasing the diversity and richness of the training data. Preprocessing of this well data primarily includes data format conversion, outlier handling, and normalization. The dataset is then divided into a training set and a validation set, with the training set containing 45 wells and the validation set containing 15 wells. Step 3 involves statistical analysis of the lithological probability and thickness distribution of each well at different strata. Taking a specific stratum as an example, the statistical results show that the lithological probability of mudstone is 30.5%, sandstone is 29.7%, and limestone is 38.2%. The average thickness distribution of this stratum is as follows: mudstone 60.4 meters, sandstone 55.6 meters, and limestone 72.8 meters. Based on these statistical data, a lithological transformation matrix and thickness distribution model for this block were further constructed, and geological statistical characteristics were extracted. Step 4: Seismic data, lithology probability, and thickness distribution data are input as input, and lithology type is used as a label, which are then fed into a multi-channel deep neural network. This network architecture integrates seismic attributes and geostatistical features, and uses a convolutional neural network (CNN) for high-dimensional feature extraction. Combined with a feature fusion layer, multi-source data is deeply fused to improve the accuracy of lithology prediction. Step 5: Train the model using K-Fold cross-validation, setting K to 5. Divide the dataset into 5 non-overlapping subsets, selecting one subset as the validation set and the remaining 4 subsets as the training set each time. Evaluate the model performance using the validation set, and finally select the model with the best performance on the validation set for lithology prediction. After training, use the model to predict lithology at other undrilled geological locations in the block.

[0039] This patent utilizes a large-scale exploration model example to introduce geological statistical characteristics into the intelligent lithology prediction process in a data-driven manner. It fully leverages the feature extraction and multi-dimensional fusion capabilities of deep learning to ensure the physical rationality of the prediction results, thereby reducing the optimization space and improving prediction accuracy.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A lithology prediction method based on the constraints of geostatistical characteristics, wherein the method extracts geostatistical characteristics from a large-scale exploration model for lithology prediction, characterized in that, include: Extracting earthquake data; Virtual wells are constructed, virtual well data is generated through simulation, and synthetic seismic records are generated based on seismic data; Extract geostatistical features to obtain the regularity of geostatistical features; Using seismic data and geological statistical features as input, a multi-channel deep neural network model is constructed to predict lithology.

2. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 1, characterized in that, Step one specifically includes: The large-scale exploration model was analyzed to extract seismic data related to time and depth. Extract well logging information and obtain lithofacies calibration of the well logging information; By converting seismic time to depth, the depth calibration of seismic data and well logging data is aligned, ensuring consistency between the two data in the depth domain.

3. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 1, characterized in that, Step two specifically includes: Virtual well lithology type simulation; Virtual well physical property parameter simulation; Seismic data synthesis: Extracting wavelets from actual seismic data; By combining the simulated P-wave velocity, S-wave velocity, and density data of the virtual well with the extracted wavelet, a synthetic seismic channel conforming to the geological characteristics of the specified working area is generated using the co-convolution technique.

4. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 3, characterized in that, Step two also includes dataset expansion and the division of training and validation wells: Add the generated virtual well data and the corresponding synthetic seismic channel to the existing real well dataset; The expanded dataset is divided into training wells and validation wells proportionally.

5. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 4, characterized in that, The division of training wells and validation wells is performed using a random partitioning method or a stratified sampling method.

6. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 1, characterized in that, Step three includes combining the geological structures and boundary conditions contained in the stratigraphic model in the exploration model to statistically analyze the probability and thickness distribution of various lithologies for each well.

7. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 6, characterized in that, The lithological probability statistics include: The lithology probability of each well is calculated based on the thickness of different lithologies within the stratigraphic layer using a stratified statistical method. Calculate the average lithology probability of all wells in the stratigraphic unit based on the lithology probability of each well. The mean lithological probability of each stratum is organized into a transformation matrix, where the rows of the matrix represent strata, the columns represent lithology, and the matrix elements are the mean lithological probability of that stratum.

8. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 7, characterized in that, The thickness distribution statistics are obtained by measuring the thickness distribution of each lithology in each layer to describe the spatial variation of the thickness of the corresponding lithology.

9. The lithology prediction method based on the constraints of geostatistical characteristics according to claim 8, characterized in that, The geological statistical characteristics are as follows: ——（1）。 10. A lithology prediction method based on the constraints of geostatistical characteristics according to any one of claims 1-9, characterized in that, Step four specifically includes: The input data is constructed and lithological information is used as a label. The input data includes seismic data and geological statistical features. Design corresponding multi-channel input structures based on earthquake attributes and geological statistical characteristics; The data from each channel is processed by an adaptive convolutional network to extract high-dimensional features, and then the features are deeply fused through a feature fusion layer. The fused features are further processed through a multi-layer nonlinear network to extract the complex mapping relationship between the data and lithological prediction. By generating lithology prediction results in the output layer using a normalization function, a multi-channel deep neural network architecture that integrates seismic attributes and geostatistical features is constructed.

Images