Deep learning seismic lithology prediction method based on geological statistical characteristic rule constraint
By combining deep learning methods with seismic data and geological statistical characteristics, the problems of insufficient resolution and insufficient geological prior knowledge in traditional seismic lithology prediction have been solved, achieving more accurate lithology prediction, especially showing significant advantages in complex geological bodies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2024-12-31
- Publication Date
- 2026-06-23
Smart Images
Figure CN119758452B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic exploration, and in particular to a deep learning-based seismic lithology prediction method based on constraints of geological statistical characteristics. Background Technology
[0002] Traditional model-driven pre-stack inversion and post-stack seismic reservoir prediction techniques are limited by seismic resolution, making it difficult to effectively predict thin layers. Traditional frequency-division seismic techniques can only provide qualitative and semi-quantitative descriptions of thin layers and lack integration with pre-stack seismic gathers and multiple seismic attributes. These methods largely rely on physical models, often failing to effectively identify complex, concealed reservoirs or those disturbed by coal seams. While deep learning-based lithology prediction methods introduce the advantages of data-driven approaches, they typically rely solely on seismic data and limited well logging information, neglecting the comprehensive utilization of prior geological knowledge. In actual geological exploration, lithology prediction tasks often involve multiple strata with significant differences between them, and the geostatistical characteristics within different strata exhibit complex variations.
[0003] Existing earthquake lithology prediction methods rely solely on training sample sets constructed from multiple earthquake data, resulting in low prediction accuracy and reliability. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning method for earthquake lithology prediction based on the constraints of geostatistical characteristics in order to improve the accuracy and reliability of earthquake lithology prediction.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] A deep learning-based seismic lithology prediction method based on constraints from geostatistical characteristics includes the following steps:
[0007] S1. Acquire post-stack seismic data and pre-stack gather data, perform feature extraction and seismic inversion on the post-stack seismic data and pre-stack gather data, and obtain seismic attribute and elastic parameter inversion results. Seismic attributes, elastic parameter inversion results, post-stack seismic data and pre-stack gather data constitute multiple seismic information.
[0008] S2. Acquire geological data and well logging data, statistically analyze the characteristics of geological data and well logging data, and quantify the characteristics. The specific steps for quantifying the characteristics are as follows: For each stratigraphic unit, perform unique thermal coding to obtain a coded representation, obtain the lithology type prior probability vector of each stratigraphic unit, obtain the conversion probability vector between different lithologies, obtain the thickness of mudstone, sandstone and coal seam in each stratigraphic unit, and obtain the thickness distribution vector. The coded representation of stratigraphic unit, prior probability vector, conversion probability vector and thickness distribution vector serve as multiple geological statistical features.
[0009] S3. By using temporal depth relationships, multiple seismic information and multiple geological statistical features are associated, corresponding lithological labels are set for multiple seismic information, and multiple geological statistical features are used to constrain the multiple seismic information to obtain the labeled dataset required for training.
[0010] S4. Based on the labeled dataset required for training, the deep convolutional neural network model is trained, validated, and tested to obtain a deep learning seismic lithology prediction model.
[0011] S5. Obtain actual earthquake data and input the actual earthquake data into the deep learning earthquake lithology prediction model to obtain the actual earthquake lithology prediction results.
[0012] Furthermore, the encoding is represented as:
[0013] For the six stratigraphic units, unit 1 is encoded as [1,0,0,0,0,0], unit 2 is encoded as [0,1,0,0,0,0], unit 3 is encoded as [0,0,1,0,0,0], unit 4 is encoded as [0,0,0,1,0,0], unit 5 is encoded as [0,0,0,0,1,0], and unit 6 is encoded as [0,0,0,0,0,1].
[0014] Furthermore, the prior probability vector for lithology type is [p M ,p S ,p C ], p M ,p S ,p C These represent the prior probabilities of mudstone, sandstone, and coal seam, respectively.
[0015] Furthermore, the conversion probability vector between different lithologies is:
[0016] [T M,M ,T M,S ,T M,C ,T S,M ,T S,S ,T S,C ,T C,M ,T C,S ,T C,C ]
[0017] Among them, T M,M ,T M,S ,T M,C ,T S,M ,T S,S ,T S,C ,T C,M ,T C,S ,T C,CThese represent the conversion probabilities of mudstone to mudstone, mudstone to sandstone, mudstone to coal seam, sandstone to mudstone, sandstone to sandstone, sandstone to coal seam, coal seam to mudstone, coal seam to sandstone, and coal seam to coal seam, respectively.
[0018] Furthermore, the thickness distribution vector is:
[0019]
[0020] in These represent the category distributions of mudstone (1-5ms), mudstone (5-10ms), mudstone (10-20ms), and mudstone (20-70ms), respectively; and the category distributions of sandstone (1-5ms), sandstone (5-10ms), sandstone (10-20ms), and sandstone (20-70ms), respectively. These represent the category distribution of coal seams from 1 to 2 ms and from 2 to 3 ms, respectively.
[0021] Furthermore, the deep convolutional neural network model is trained in a sequence-to-point manner, and the input label sequence consists of multiple sampling points near the point to be predicted.
[0022] Furthermore, the deep convolutional neural network model is trained using a cross-blind testing method.
[0023] Furthermore, the specific steps for training using the cross-blind testing method are as follows:
[0024] The labeled dataset is divided into training well data and one blind logging well data. The training well data is used to input the deep convolutional neural network model for training and validation, and the blind logging well data is used to input the trained deep convolutional neural network model for testing. The above steps are repeated to obtain the blind logging results for each well. Based on the blind logging results of each well, well data with prediction accuracy less than a threshold are selected, and the selected well data is input into the deep convolutional neural network model again for training.
[0025] Furthermore, the deep convolutional neural network model employs an early stopping mechanism during training.
[0026] Furthermore, the learning rate is dynamically adjusted during the training process of the deep convolutional neural network model.
[0027] Compared with the prior art, the present invention has the following beneficial effects:
[0028] This invention encodes and quantifies four geostatistical features (stratigraphic units, lithological prior probabilities, lithological transformation matrices, and lithological thickness distribution), transforming complex geological information into numerical features that can be understood and processed by deep neural networks. Stratigraphic units are represented using one-hot encoding, while lithological prior probabilities, lithological transformation matrices, and lithological thickness distribution are processed through probabilistic numerical representations, providing explicit constraints for lithology prediction. By using these four geostatistical features as input to a deep convolutional neural network, the model can effectively constrain itself using input geological prior information while simultaneously making lithology predictions based on seismic data. This method enhances the model's stability and prediction accuracy, while fully utilizing geological background knowledge to optimize the lithology prediction process. By combining multiple seismic information sources with constraints from various geostatistical features, it achieves more comprehensive and accurate seismic lithology prediction. Through this synergistic effect, the model can more accurately capture complex lithological variations, showing a significant advantage, especially in the prediction of complex geological bodies such as coal seams. Attached Figure Description
[0029] Figure 1 This is a flowchart of the present invention;
[0030] Figure 2 This invention relates to a deep convolutional neural network architecture with multiple seismic information and multiple geological constraints.
[0031] Figure 3 The results are from a cross-blind test of two wells;
[0032] Figure 4 The predicted results are for seismic lithology slices, among which... Figure 4 (a) shows the prediction results using only multiple seismic information. Figure 4 (b) Prediction results based on the synergistic effect of multiple seismic information and multiple geological statistical characteristics. Detailed Implementation
[0033] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0034] This invention proposes a deep learning-based seismic lithology prediction method based on constraints from geostatistical features. The flowchart of the method is as follows: Figure 1As shown, a training sample set with multiple geostatistical features and seismic information was constructed by comprehensively considering stratigraphic information, lithological prior probabilities, lithological succession relationships, and lithological thickness distribution. Blind test results show that introducing geological prior constraints can significantly improve the prediction accuracy and reliability of sandstone and thin coal seams. Compared with methods that only construct training sample sets based on multiple seismic information, this method provides prediction results that are more consistent with actual geological prior knowledge in complex reservoirs with significant differences in lithological distribution, and the predicted lithological slices are more closely aligned with the sedimentary environment characteristics of each geological unit.
[0035] The method of the present invention includes the following steps:
[0036] S1. Acquire post-stack seismic data and pre-stack gather data, perform feature extraction and seismic inversion on the post-stack seismic data and pre-stack gather data, and obtain seismic attribute and elastic parameter inversion results. The seismic attribute and elastic parameter inversion results, post-stack seismic data and pre-stack gather data constitute multiple seismic information.
[0037] S2. Acquire geological data and well logging data, statistically analyze the characteristics of geological data and well logging data, and quantify the characteristics. The specific steps for quantifying the characteristics are as follows: For each stratigraphic unit, perform unique thermal coding to obtain a coded representation, obtain the lithology type prior probability vector of each stratigraphic unit, obtain the conversion probability vector between different lithologies, obtain the thickness of mudstone, sandstone and coal seam in each stratigraphic unit, and obtain the thickness distribution vector. The coded representation of stratigraphic unit, prior probability vector, conversion probability vector and thickness distribution vector serve as multiple geological statistical features.
[0038] S3. By using temporal depth relationships, multiple seismic information and multiple geological statistical features are associated, corresponding lithological labels are set for multiple seismic information, and multiple geological statistical features are used to constrain the multiple seismic information to obtain the labeled dataset required for training.
[0039] S4. Based on the labeled dataset required for training, the deep convolutional neural network model is trained, validated, and tested to obtain a deep learning seismic lithology prediction model.
[0040] S5. Obtain actual earthquake data and input the actual earthquake data into the deep learning earthquake lithology prediction model to obtain the actual earthquake lithology prediction results.
[0041] In S1, feature extraction and seismic inversion are performed on post-stack seismic data and pre-stack gathers to obtain inversion results of various seismic attributes and elastic parameters.
[0042] In S2, the geological statistical characteristics hidden within the geological data and well logging data are analyzed and processed into a numerical format suitable for deep neural network algorithms. First, each stratigraphic unit is represented using one-hot encoding. The six stratigraphic units are encoded separately: unit 1 is encoded as [1,0,0,0,0,0], unit 2 as [0,1,0,0,0,0], and so on. For each stratigraphic unit, the prior probabilities of the three lithological types (mudstone, sandstone, and coal seam) are represented as [p...]. M ,p S ,p C The 3×3 transfer matrix T is flattened into a 9-dimensional vector [T]. M,M ,T M,S ,T M,C ,T S,M ,T S,S ,T S,C ,T C,M ,T C,S ,T C,C The vector [] represents the conversion probability between different lithologies. Here, M, S, and C represent mudstone, sandstone, and coal seam, respectively. The thickness of mudstone and sandstone in each stratigraphic unit is divided into four different thickness categories: 1-5 ms, 5-10 ms, 10-20 ms, and 20-70 ms, while the thickness of coal seam is divided into two different thickness categories. The distribution of these thickness categories is represented as a 10-dimensional thickness distribution vector, in the form:
[0043]
[0044] In S3, multiple seismic information at the well location (such as post-stack earthquakes, pre-stack earthquakes, seismic attributes, and inverted elastic parameters) are associated with various geostatistical features (stratigraphic units, lithological probabilities, transformation matrices, and thickness distributions) through time-depth relationships. Combined with well logging lithological features constructed based on core, logging, and drilling data, a labeled dataset required for training is generated.
[0045] In S4, based on the constructed labeled learning samples, we trained, validated, and tested a deep convolutional neural network model, obtaining a model constrained by multiple seismic information and geostatistical features. The deep neural network model was trained sequentially point-to-point, with the input label sequence consisting of multiple sampling points near the point to be predicted, providing more information compared to a single-point input. To ensure effective model training, appropriate time windows were selected for different branches. A cross-blind testing method was adopted, meaning that blind test wells did not participate in the model optimization process. Multiple wells were used as experimental samples; each time, a subset of wells was selected as the training dataset, and the remainder as the blind test set, repeating the experiment to obtain the blind test results for each well. The model's loss function was classification cross-entropy, and the optimizer was an optimization algorithm with adaptive learning rate and efficient convergence. During training, an early stopping mechanism was introduced to prevent overfitting, and the learning rate was dynamically adjusted to further improve model performance.
[0046] In S5, the trained neural network model constrained by multiple seismic information and geological statistical features is finally applied to the entire three-dimensional seismic data to complete the prediction of seismic lithology spatial distribution based on deep neural networks.
[0047] The core innovation of this invention mainly consists of the following parts:
[0048] 1) Encoding and Numericalization of Four Geostatistical Features: This invention encodes and quantifies four geostatistical features (stratigraphic units, lithological prior probabilities, lithological transformation matrices, and lithological thickness distribution), transforming complex geological information into numerical features that can be understood and processed by deep neural networks. Specifically, stratigraphic units are represented using unique thermal encoding, while lithological prior probabilities, lithological transformation matrices, and lithological thickness distribution are processed using probabilistic numerical representations, providing clear constraints for lithological prediction.
[0049] 2) Four geostatistical features as input features for the deep neural network: This invention uses four geostatistical features as input to a deep convolutional neural network, enabling the model to effectively constrain itself using prior geological information while making lithology predictions based on seismic data. This method enhances the model's stability and prediction accuracy, while fully utilizing geological background knowledge to optimize the lithology prediction process.
[0050] 3) Synergistic effect of multiple seismic information and multiple geostatistical feature constraints: Another innovation of this invention lies in combining multiple seismic information with multiple geostatistical feature constraints to achieve more comprehensive and accurate seismic lithology prediction. Through this synergistic effect, the model can more accurately capture complex lithological changes, showing a significant advantage, especially in the prediction of complex geological bodies such as coal seams.
[0051] The purpose of this invention is to provide a deep learning-based lithology prediction method that comprehensively utilizes prior geological information to fully explore the geostatistical features hidden in well logging information and geological data. Conventional deep learning-based methods, under a data-driven supervised deep learning framework, use seismic data for lithofacies prediction, often suffering from problems such as inconsistencies in geological priors, poor generalization, and weak interpretability. This invention proposes a novel workflow that comprehensively integrates seismic data, well logging information, and prior geological knowledge to fully explore the geostatistical features hidden in the data, leading to a deeper understanding of the characteristics and patterns of different strata. This method utilizes multi-source data, including seismic data, inversion results, seismic attributes, and geostatistical patterns, to design a deep learning model adapted to complex reservoirs with multiple strata, better capturing the differences and complexities between different strata. This invention considers prior geological knowledge and predicts lithofacies distribution by integrating well logging, seismic, and geological data. Based on geostatistical characteristics and considering factors such as stratigraphic position, lithological prior probability, lithological continuity, and lithological thickness distribution, the constructed feature set outperforms the baseline set composed of seismic records, inversion results, and seismic attributes. This successfully achieves a more accurate simulation of the correlation between geological attributes, demonstrating innovation in model consistency.
[0052] Figure 1 The specific process consists of five main steps: First, attribute extraction and seismic inversion are performed; next, the geological statistical characteristics are analyzed and numerically processed; then, based on the seismic information at the well location, lithology labels, and four geological statistical characteristics, reasonable lithology label samples are established; subsequently, a deep convolutional neural network model is trained using the constructed sample labels to obtain a prediction model with multiple seismic information and multiple geological constraints; finally, the trained model is used to predict two-dimensional or three-dimensional seismic lithology distribution.
[0053] Figure 2 It is a network architecture constrained by multiple seismic information and multiple geostatistical features. The network consists of four branches: Branch 1: takes pre-stack gathers as input, providing waveform information on amplitude variation with offset distance; Branch 2: inputs time spectrum maps in the range of 10-80Hz with 10Hz intervals, solving the problem of limited frequency bandwidth of seismic data and enabling the prediction of thin layers with higher resolution; Branch 3: integrates multiple seismic attribute information; Branch 4: introduces four geostatistical features (such as stratigraphic units, lithological prior probabilities, lithological transformation matrices, and lithological thickness distribution).
[0054] Figure 3The invention was applied to a real coal-bearing clastic rock work area, and the results of cross-blind well testing were compared and analyzed in eight sets of experiments to highlight its effectiveness. The method of this invention was then applied to a coal-bearing clastic rock work area. Figure 3 The results of cross-blind testing of two wells are presented. For comparative analysis, a total of eight sets of experiments were conducted. From right to left, they represent: lithology classification using only inversion results as input features to the deep neural network; lithology classification using only pre-stack angle gathers, frequency division volumes, inversion results, and multiple seismic attributes, without considering geostatistical constraints; lithology classification using only stratigraphic constraints; lithology classification using only prior probability constraints; lithology classification using only transformation matrix constraints; lithology classification using only thickness distribution constraints; lithology classification using all geostatistical constraints; and lithology classification using only true lithology labels. Analysis of the statistical and prediction results demonstrates the effectiveness of the proposed lithology prediction process, namely, the synergistic effect of multiple seismic information and the four geostatistical constraints provides better lithological differentiation than relying solely on multiple seismic information.
[0055] Figure 4 This involves a spatial distribution comparison of predicted two-dimensional seismic lithology slices. Finally, the method is applied to the entire three-dimensional work area. Figure 4 The results of seismic lithology slice predictions along the stratigraphic horizon are presented: (a) using only multi-seismic information, and (b) the synergistic effect of constraints from multi-seismic information and multiple geostatistical features. The small black circles highlight the predicted lithology at the well location. Without geological constraints, it was incorrectly predicted as sandstone, but the actual well logging label was mudstone. The method of this invention accurately predicted mudstone. This further demonstrates that deep learning network models applying geological constraints can obtain lithology prediction results that better reflect the sedimentary environment.
[0056] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A deep learning-based seismic lithology prediction method based on constraints from geostatistical characteristics, characterized in that, The method includes the following steps: S1. Acquire post-stack seismic data and pre-stack gather data, perform feature extraction and seismic inversion on the post-stack seismic data and pre-stack gather data, and obtain seismic attribute and elastic parameter inversion results. Seismic attributes, elastic parameter inversion results, post-stack seismic data and pre-stack gather data constitute multiple seismic information. S2. Acquire geological data and well logging data, statistically analyze the characteristics of geological data and well logging data, and quantify the characteristics. The specific steps for quantifying the characteristics are as follows: For each stratigraphic unit, perform unique thermal coding to obtain a coded representation, obtain the lithology type prior probability vector of each stratigraphic unit, obtain the conversion probability vector between different lithologies, obtain the thickness of mudstone, sandstone and coal seam in each stratigraphic unit, and obtain the thickness distribution vector. The coded representation of stratigraphic unit, prior probability vector, conversion probability vector and thickness distribution vector serve as multiple geological statistical features. S3. By using temporal depth relationships, multiple seismic information and multiple geological statistical features are associated, corresponding lithological labels are set for multiple seismic information, and multiple geological statistical features are used to constrain the multiple seismic information to obtain the labeled dataset required for training. S4. Based on the labeled dataset required for training, the deep convolutional neural network model is trained, validated, and tested to obtain a deep learning seismic lithology prediction model. S5. Obtain actual earthquake data, input the actual earthquake data into the deep learning earthquake lithology prediction model, and obtain the actual earthquake lithology prediction results. The conversion probability vector between different lithologies is: in, These represent the conversion probabilities of mudstone to mudstone, mudstone to sandstone, mudstone to coal seam, sandstone to mudstone, sandstone to sandstone, sandstone to coal seam, coal seam to mudstone, coal seam to sandstone, and coal seam to coal seam, respectively.
2. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The encoding is represented as follows: For the six stratigraphic units, unit 1 is coded as [1, 0, 0, 0, 0, 0], unit 2 is coded as [0, 1, 0, 0, 0, 0], unit 3 is coded as [0, 0, 1, 0, 0, 0], unit 4 is coded as [0, 0, 0, 1, 0, 0], unit 5 is coded as [0, 0, 0, 0, 1, 0], and unit 6 is coded as [0, 0, 0, 0, 0, 1].
3. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The prior probability vector of lithology type is , These represent the prior probabilities of mudstone, sandstone, and coal seam, respectively.
4. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The thickness distribution vector is: in These represent the category distributions of mudstone (1-5 ms), mudstone (5-10 ms), mudstone (10-20 ms), mudstone (20-70 ms), sandstone (1-5 ms), sandstone (5-10 ms), sandstone (10-20 ms), and sandstone (20-40 ms), respectively. These represent the category distribution of coal seams from 1 to 2 ms and from 2 to 3 ms, respectively.
5. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The deep convolutional neural network model is trained in a sequence-to-point manner, and the input label sequence consists of multiple sampling points near the point to be predicted.
6. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The deep convolutional neural network model is trained using a cross-blind testing method.
7. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 6, characterized in that, The specific steps for training using the cross-blind testing method are as follows: The labeled dataset is divided into training well data and one blind logging well data. The training well data is used to input the deep convolutional neural network model for training and validation, and the blind logging well data is used to input the trained deep convolutional neural network model for testing. The above steps are repeated to obtain the blind logging results for each well. Based on the blind logging results of each well, well data with prediction accuracy less than a threshold are selected, and the selected well data is input into the deep convolutional neural network model again for training.
8. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The deep convolutional neural network model employs an early stopping mechanism during training.
9. The deep learning-based seismic lithology prediction method based on geostatistical feature constraints according to claim 1, characterized in that, The learning rate is dynamically adjusted during the training of the deep convolutional neural network model.