Sequence stratigraphic intelligent interpretation method and device and storage medium

Abstract

Embodiments of the present application provide a layer sequence stratigraphic intelligent interpretation method, device, processor, storage medium and computer program product. The method comprises: acquiring seismic data of a target work area; inputting the seismic data into a pre-trained knowledge-guided deep learning network model to obtain a preliminary sequence division result, wherein the pre-trained knowledge-guided deep learning network model is a deep learning model obtained by supervised training based on sequence label data marked according to prior geological knowledge, and a comprehensive loss function that fuses geological rule constraints is used in the training process; using a sedimentary pattern rule base to constrain and optimize the preliminary sequence division result in terms of geological rationality to obtain a final sequence division result enhanced in terms of geological rationality, wherein the sedimentary pattern rule base at least contains a system domain vertical superposition rule constraint and a seismic reflection geometry shape matching constraint; and generating sequence stratigraphic interpretation results based on the final sequence division result.

Description

Technical Field

[0001] This application relates to the field of oil and gas geological exploration and development technology, specifically to a sequence stratigraphic intelligent interpretation method, device, storage medium, and computer program product. Background Technology

[0002] Sequence stratigraphy is the core theoretical foundation for the detailed description and prediction of hydrocarbon reservoirs. Traditional sequence stratigraphy methods heavily rely on the experience of geological experts, employing a "top-down" approach: ① Applying a standard sequence model → ② Identifying key interfaces → ③ Dividing the sequence → ④ Calibration and filling with data (well and seismic data). Standard sedimentary models (such as the Vail classic model) are applied to seismic, well logging, and core data for manual interpretation. This method has significant drawbacks: ① High subjectivity: Differences in experience among interpreters can lead to drastically different partitioning schemes, resulting in poor consistency; ② Low efficiency: Manual interpretation is time-consuming and difficult to handle the massive amounts of data from large-scale 3D seismic surveys; ③ Limited resolution: The human eye struggles to identify subtle but important patterns in the data, easily overlooking complex or hidden stratigraphic interfaces.

[0003] Existing technologies have also attempted to apply machine learning (ML) and deep learning (DL) techniques to the automatic segmentation of sequence stratigraphy. For example, the U-Net network is used for semantic segmentation of seismic profiles. However, purely data-driven machine learning methods suffer from a "black box" problem; their segmentation results may lack geological rationality and are difficult for geologists to trust. Furthermore, model training requires a large amount of high-quality labeled data, resulting in limited generalization ability. Summary of the Invention

[0004] The purpose of this application is to provide a sequence stratigraphic intelligent interpretation method, a sequence stratigraphic intelligent interpretation device, a storage medium, and a processor.

[0005] To achieve the above objectives, the first aspect of this application provides a sequence stratigraphic intelligent interpretation method, comprising: Acquire seismic data for the target work area; The seismic data is input into a pre-trained knowledge-guided deep learning network model to obtain preliminary sequence division results. The preliminary sequence division results include at least the system domain category of each voxel in three-dimensional space and the automatically identified sequence interface position. The pre-trained knowledge-guided deep learning network model is a deep learning model obtained by supervised training based on sequence label data annotated with prior geological knowledge, and a comprehensive loss function that integrates geological rule constraints is used during the training process. The preliminary sequence stratigraphy results are subjected to geological rationality constraints and optimization using a sedimentary model rule base to obtain a final sequence stratigraphy result with enhanced geological rationality. The sedimentary model rule base includes at least constraints on the vertical stacking regularity of system tracts and constraints on the matching of seismic reflection geometry. Sequence stratigraphic interpretation results are generated based on the final sequence stratigraphic division results.

[0006] A second aspect of this application provides a sequence stratigraphic intelligent interpretation device, comprising: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the above-described hierarchical intelligent interpretation method.

[0007] A third aspect of this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the above-described sequential stratigraphic intelligent interpretation method.

[0008] The fourth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described intelligent interpretation method of stratigraphy.

[0009] This method deeply integrates geologists' prior knowledge (sedimentary models) with data-driven machine learning algorithms to form a "dual-drive" workflow, enabling rapid, objective, and geologically sound sequence stratigraphy modeling.

[0010] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0011] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 The schematic diagram illustrates a flowchart of the sequence stratigraphic intelligent interpretation method according to an embodiment of this application; Figure 2 The diagram illustrates the internal structure of a computer device according to an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] Figure 1 The illustration shows a schematic flowchart of a sequence stratigraphic intelligent interpretation method according to an embodiment of this application. Figure 1 As shown, in one embodiment of this application, a sequence stratigraphic intelligent interpretation method is provided. This embodiment mainly illustrates the application of this method to a processor, and includes the following steps: Step 102: Obtain seismic data for the target work area; Step 104: Input the seismic data into the pre-trained knowledge-guided deep learning network model to obtain preliminary sequence division results. The preliminary sequence division results include at least the system domain category of each voxel in three-dimensional space and the automatically identified sequence interface location. The pre-trained knowledge-guided deep learning network model is a deep learning model obtained by supervised training based on sequence label data annotated with prior geological knowledge, and a comprehensive loss function that integrates geological rule constraints is used during the training process. Step 106: Use the sedimentary model rule base to constrain and optimize the geological rationality of the preliminary sequence division results to obtain the final sequence division results with enhanced geological rationality. The sedimentary model rule base includes at least the vertical stacking regularity constraint of the system tract and the seismic reflection geometry matching constraint. Step 108: Generate sequence stratigraphic interpretation results based on the final sequence stratigraphic division results.

[0014] First, the processor can acquire seismic data and related auxiliary data for the target work area. Seismic data may include amplitude, phase, and instantaneous attributes. Related auxiliary data may include well logging, core samples, and structural interpretation results. Optionally, after acquiring the seismic data for the target work area, preprocessing operations such as standardization, normalization, and spatial alignment can be performed to construct a multi-channel input dataset suitable for deep learning models. Each channel corresponds to a seismic attribute or geological feature, constituting a sedimentary pattern-sensitive input set. Then, the processor can input the seismic data into a pre-trained knowledge-guided deep learning network model to obtain preliminary sequence division results output by the pre-trained knowledge-guided deep learning network model. The preliminary sequence division results include at least the system domain category of each voxel in three-dimensional space and the automatically identified sequence boundary location. A system domain refers to a stratigraphic unit with specific sedimentary characteristics and spatial structure formed within a sequence under the control of sedimentary environment, provenance supply, and sea-level changes. Seismic systems tracts can be categorized into lowstand systems tracts (LST), transgressive systems tracts (TST), and highstand systems tracts (HST). Lowstand systems tracts are sedimentary bodies formed during periods of sea-level decline, transgressive systems tracts are formed during periods of rapid sea-level rise, and highstand systems tracts are formed during periods of high sea-level stability or slow sea-level decline. A voxel is the smallest unit of volume in three-dimensional space, equivalent to the extension of a two-dimensional pixel in three dimensions. In a seismic data volume, each voxel represents a seismic attribute value at a spatial location (x, y, z). The pre-trained knowledge-guided deep learning network model is a deep learning model trained under supervised conditions using stratigraphic label data annotated with prior geological knowledge. Furthermore, a comprehensive loss function incorporating geological rule constraints is used during training. In other words, the knowledge-guided deep learning network model not only uses labeled data for supervised learning during training but also explicitly introduces geological rule constraints into the deep learning model through a comprehensive loss function. These geological rule constraints include systems tract order, thickness range, and reflection structure matching. The model learns geological patterns during the training phase and possesses predictive capabilities guided by geological rationality. The comprehensive loss function refers to a loss function composed of multiple geological rule constraint terms added to the traditional cross-entropy loss.

[0015] Furthermore, after obtaining the preliminary sequence stratigraphic results corresponding to the seismic data for the target work area, the geological rationality constraints and optimizations of the preliminary sequence stratigraphic results can be applied using a sedimentary model rule base to obtain a final sequence stratigraphic result with enhanced geological rationality. The sedimentary model rule base includes at least constraints on the vertical stacking regularity of system tracts and constraints on the matching of seismic reflection geometry. The sedimentary model rule base is a structured knowledge base containing constraint rules extracted from geological sedimentary models. These can include: stacking regularity constraints, geometric morphology constraints, and thickness range constraints. The stacking regularity constraint requires that the vertical sequence must be LST→TST→HST, prohibiting reverse order. Geometric morphology constraints can require matching relationships between reflection structures (pre-lap, top-lap, bottom-lap) and system tracts, such as top-lap→TST, pre-lap→LST / HST. Thickness range constraints can require reasonable thickness ranges for each system tract under a specific geological background. The sedimentary model rule base can also include contact relationship constraints to ensure natural transitions at system tract boundaries and avoid abrupt changes. Therefore, the final sequence stratigraphic division result with enhanced geological rationality retains both data-driven details and conforms to geological laws. Sequence stratigraphic interpretation results can then be generated based on this final result. Specifically, based on the optimized sequence stratigraphic division result, structural interpretation results from well data can be combined. For example, two-dimensional sequence stratigraphic profiles, three-dimensional sequence volumes, and three-dimensional sequence interface mesh models can be constructed.

[0016] In one embodiment, the expression for the comprehensive loss function Loss is as follows: Loss = L_CE + λ1·L_order + λ2·L_geometry + λ3·L_thick + λ4·L_contact Wherein, L_CE is the cross-entropy loss, used to constrain the consistency between the prediction and the manual label; L_order is the constraint loss for the vertical stacking pattern of the system domains; L_geometry is the constraint loss for the matching of the reflection geometry with the system domains; L_thick is the constraint loss for the thickness of the system domains; L_contact is the constraint loss for the rationality of the contact relationship between the system domains; λ1, λ2, λ3, and λ4 are weighting coefficients used to balance the importance of each constraint.

[0017] Specifically, cross-entropy loss is a loss function used to measure the difference between the model's predicted probability distribution and the true label distribution. In this application, cross-entropy loss is essentially the standard loss term in supervised learning, ensuring that the model learns the mapping relationship from seismic attributes to geological categories. True labels are manually labeled by geological experts based on sedimentary patterns and sequence stratigraphy principles, including systems tract classification (LST, TST, HST) and sequence boundary locations. Cross-entropy loss constrains model predictions to be consistent with this expert knowledge and is the basis for model fitting known data. The systems tract vertical stacking regularity constraint loss L_order is used to force the systems tracts predicted by the model to follow the normal evolutionary order in sedimentology in the vertical direction, i.e., LST→TST→HST. This order reflects the complete cycle of sea-level decline, rise, and then high-level stability. The basic principle of sequence stratigraphy requires that within a complete sequence, there should be no vertical reversal phenomenon where TST is below LST or HST is below TST (unless entering the next cycle). The systems tract vertical stacking regularity constraint loss transforms this geological regularity into an optimizable mathematical form.

[0018] In a specific embodiment, for the system domain prediction sequence S(z) of the same vertical profile, denoted by 1, 2, and 3 as LST, TST, and HST respectively, the system domain vertical stacking constraint loss L_order can be calculated using the following formula:

[0019] in, This represents the system domain category code value at depth z, which increases in the order of low-order system domain, transgressive system domain, and high-order system domain; z is the depth sampling point index. This is the system domain category encoding value at depth z+1. When a reverse order occurs... The penalty is applied at certain times to prevent the model from generating geologically impossible sequences from the training source, thereby improving the vertical logical rationality of the prediction results.

[0020] The constraint loss L_geometry, which matches the reflection geometry to the system tract, is used to force a geologically empirical correspondence between the predicted system tract type and the seismic reflection geometry (such as onlap, onlap, and downlap). Specifically, seismic reflection geometry is a direct response to sedimentary processes. For example, onlap typically indicates marine transgression and preferentially corresponds to TST. Progradation is common as sediments advance towards the basin and often corresponds to LST or HST. Downlap often occurs at the end of LST or HST. If a location is identified as a typical onlap but predicted as HST, it is geologically unreasonable. Therefore, in this embodiment, the geometry is constructed by calculating the matching degree between geometric properties (such as dip angle and coherence) and system tract type. For example, the geometry type is defined as... The system domain prediction probability is Then, the constraint loss L_geometry for matching the reflection geometry with the system domain can be calculated using the following formula:

[0021] in, This represents the system domain most likely corresponding to this geometric type. It can be seen that the constraint loss L_geometry, which matches the reflection geometry with the system domain, allows the model to not only rely on amplitude characteristics but also utilize the important geological clue of seismic reflection morphology, enhancing the spatial structural rationality of the prediction results.

[0022] In one embodiment, the constraint loss for matching the reflection geometry with the system tract includes: calculating seismic geometric properties based on seismic data, including the dip and azimuth of the reflecting layer; identifying the reflection termination type of the stratigraphic reflection phase axis based on the seismic geometric properties, including at least one of top overflow, bottom overflow, and progradation; establishing a conditional probability relationship between the reflection termination type and the system tract category based on geological experience, the conditional probability relationship representing the probability distribution of different system tract categories when a specific reflection termination type occurs; and calculating the constraint loss for matching the reflection geometry with the system tract based on the difference between the predicted probability of the system tract category in the preliminary sequence stratigraphy results and the conditional probability relationship.

[0023] Among them, reflection geometry refers to the macroscopic morphology and spatial combination characteristics of in-phase axes on seismic profiles, reflecting the original occurrence, sedimentary processes, and subsequent tectonic alterations of sedimentary strata. Common reflection geometries include parallel, progradational, top-overlap, and under-overlap structures, which are direct seismic responses to sedimentary environments and processes. Different geometries correspond to different sedimentary dynamic backgrounds. For example, progradational structures indicate sediment advance towards the basin; top-overlap structures represent newer strata overlapping older strata during marine transgression; and under-overlap structures are commonly seen at the end of sediment advance towards the basin center. A summary is as follows: 1. Progradation: Commonly seen in the stage where sediments advance towards the basin, usually corresponding to LST or HST; 2. Onlap: A typical feature of marine transgression, primarily corresponding to TST; 3. Downlap: This often occurs at the end of the LST or HST. 4. Offlap / Retrogradation: This is commonly seen in the later stages of HST during the retrogradation process.

[0024] Therefore, if a certain region is identified as a “typical top superstructure”, its system domain is likely to be TST, and should not be predicted as HST.

[0025] A systems tract is a three-dimensional sedimentary unit with unified sedimentary dynamics formed within a sequence stratigraphy during a specific stage of sea-level change. The systems tracts in this application include three types: lowstand systems tracts, transgressive systems tracts, and highstand systems tracts. Seismic geometric attributes refer to quantitative parameters extracted from seismic data that describe the geometry of reflection interfaces, and may include dip and azimuth. Dip refers to the maximum tilt angle of the reflection interface in three-dimensional space, reflecting the original dip degree of the strata. Azimuth is the direction of the dip angle, indicating the orientation of the strata's dip. Reflection termination type refers to the lateral termination of the stratigraphic reflection phase axis, a key indicator for identifying sequence boundaries. Its types may include top-overlap, down-overlap, and progradation. Top-overlap refers to younger strata overlapping older strata against a sedimentary slope, indicating a transgressive process. Down-over refers to sedimentary layers terminating downwards along a sedimentary slope at an older interface, commonly found at the bottom of progradational sequences. Progradation refers to sedimentary bodies advancing towards the basin, with reflection phase axes superimposed in an S-shape or oblique pattern. Specifically, the identification method for reflection termination type can be achieved by comparing the dip angle differences between adjacent reflecting layers. For example, if the dip angle of the strata is less than that of the overlying interface, it can be considered top overflow. If the dip angle of the strata is greater than that of the underlying interface, it can be considered bottom overflow. If the dip angle along the depositional strike is greater than a certain threshold, it can be considered progradation.

[0026] In other words, in this embodiment, for the input 3D seismic data volume, gradient or structural tensor algorithms can be used to calculate the dip and azimuth of the reflection layer for each voxel. These attributes form the geometric basis for identifying the reflection termination type. For example: Dip(x,y,z) = arctan( T / x, T / y) Different termination types can be identified by comparing the dip angle of the deposition layer (Dip_layer) with the dip angle of the interface (Dip_surface). Onlap: If the dip angle of the sedimentary layer tends to face the overlying surface, it can be approximated by the following relationship: Dip_layer <Dip_surface; Downlap: If the dip angle of the deposited layer points to the underlying interface, it can be represented as follows: Dip_layer > Dip_surface; Progradation: If the dip angle Dip_x in the horizontal direction towards the basin is stable and greater than a certain threshold θ_prog, it can be expressed as follows: Dip_x>θ_prog.

[0027] In addition, properties such as coherence volume and amplitude trend can be combined to help determine the geometric shape.

[0028] Based on the calculated dip and azimuth angles, reflection termination types such as top wax, bottom wax, and progradation are automatically detected by setting thresholds or rules (e.g., comparing the dip angles of adjacent reflecting layers). This method discretizes termination types with clear sedimentological significance from continuous geometric attributes, establishing a preliminary correlation between seismic response and sedimentary processes. Furthermore, based on geological experience, a conditional probability relationship can be established between reflection termination types and systems tract categories. This conditional probability relationship represents the probability distribution of different systems tract categories corresponding to a specific reflection termination type, and can be denoted as P(system tract | reflection termination type). That is, a prior probability distribution of a systems tract category can be assigned to each reflection termination type based on regional sedimentary models and expert experience. In a specific embodiment, based on geological experience, systems tract preference probabilities can be assigned to different geometric types, for example: 1. The probability of a top-tier team transitioning to TST is extremely high: P(TST | Onlap) is close to 1, P(LST | Onlap) ≈ 0, P(HST | Onlap) ≈ 0; 2. Front product → More biased towards LST or HST: P(LST | Progradation)>0.5, P(HST | Progradation)>0.3; 3. Underdog → More inclined towards LST or HST: P(HST | Downlap) > a certain threshold.

[0029] This method quantifies the qualitative experience of geological experts, providing prior knowledge constraints for machine learning models.

[0030] Then, the processor can calculate the constraint loss for matching the reflection geometry with the system domain based on the difference between the predicted probability and conditional probability relationship of the system domain categories in the preliminary hierarchical partitioning results. Specifically, during model training, for each training sample location, the prior conditional probability distribution P_geo(c|G) is obtained based on its identified reflection termination type G. Simultaneously, the system domain probability distribution currently predicted by the model is also obtained. (c) The difference between the two is calculated as the constraint loss L_geometry for matching the reflection geometry with the system domain. Cross-entropy is typically used as the difference metric to encourage model predictions to approximate the prior distribution. Thus, by incorporating geological rules as soft constraints into the model's optimization objective, the model's predictions are consistent with the sedimentary environment implied by the seismic reflection geometry while fitting the data. For example, the probability of geometric information formation is denoted as P_geo, and the class probability predicted by the model is denoted as P_ML; the two can be fused by weights: P_final = α·P_ML + (1 α)·P_geo Where 0 ≤ α ≤ 1. In this way, the final system domain prediction results simultaneously satisfy data fitting and geometric rationality.

[0031] The system tract thickness constraint loss L_thick is used to constrain the predicted thickness of each system tract to be within a reasonable range according to regional geological statistics, avoiding non-geological phenomena such as abnormally thin or thick tracts. In specific basins and tectonic settings, the thicknesses of different sequence stratigraphic levels and system tracts exhibit statistical regularities; for example, the total thickness of a third-order sequence stratigraphic unit ranges from 30m to 200m. If the model predicts a certain TST thickness of only 2m, while the minimum thickness in the region is 10m, then the result has significant geological inconsistencies and should be corrected.

[0032] In one embodiment, the method for determining the thickness constraint loss for different system domains within a geologically reasonable range includes: statistically analyzing the predicted thickness of each system domain category in spatial location based on the preliminary sequence stratigraphy results; obtaining the geologically reasonable range corresponding to each system domain category; for each system domain category, determining a first difference and a second difference based on the comparison between the predicted thickness of the system domain category and its corresponding geologically reasonable range; wherein, when the predicted thickness is less than the lower limit of the geologically reasonable range, the first difference is the difference between the lower limit and the predicted thickness; otherwise, the first difference is zero; when the predicted thickness is greater than the upper limit of the geologically reasonable range, the second difference is the difference between the predicted thickness and the upper limit; otherwise, the second difference is zero; and calculating the system domain thickness constraint loss based on the first and second differences for all system domain categories. Specifically, the calculation formula for the thickness constraint loss of each system domain category is as follows:

[0033] Where Loss_thick,k is the thickness constraint loss for the k-th system domain category, k is the system domain category, and hk is the predicted thickness of the k-th system domain. This represents the lower limit of the geologically reasonable range corresponding to the k-th system domain. This represents the upper limit of the geologically reasonable range corresponding to the k-th system tract. The system tract thickness constraint loss is the sum of the thickness constraint losses for all system tract categories. Introducing this constraint can prevent the model from producing stratification with thickness anomalies due to data noise or local features, ensuring that the prediction results conform to the regional sedimentary scale.

[0034] The constraint loss L_contact for the rationality of system tract contact relationships is used to constrain the contact relationships (e.g., conformable, unconformable) between different system tracts to conform to geological laws and ensure natural transitions at boundaries. Sequence boundaries are typically unconformities or their corresponding conformable surfaces, and there should be continuous sedimentation or specific transitions between system tracts (e.g., a transgression surface often exists between LST and TST). Abrupt changes or discontinuities in contact relationships may indicate prediction errors. Therefore, this can be designed by analyzing the matching degree between the gradient changes of the predicted body at the boundary and the geological contact type. For example, for unconformities, reflection termination or angular differences are expected; for conformable contacts, continuous and gradual reflections are expected. In this way, the naturalness and continuity of sequence boundaries and system tract boundaries can be improved, making the delineation results topologically reasonable in three-dimensional space.

[0035] In one embodiment, the geological rationality constraint and optimization of the preliminary sequence stratigraphy results using a sedimentary model rule base includes executing vertical superposition law constraints on system tracts through a finite state machine algorithm, Viterbi decoding, or dynamic programming algorithm. Specifically, the finite state machine algorithm detects and corrects the vertical sequence of the preliminary sequence stratigraphy results based on the rule that only the low-stand systems tract (LST) is allowed to remain or transition to the transgressive systems tract (TST), and the transgressive systems tract (TST) is allowed to remain or transition to the high-stand systems tract (HST). The Viterbi decoding or dynamic programming algorithm is used to search for the optimal sequence path that conforms to the vertical superposition law based on the system tract category probabilities provided by the preliminary sequence stratigraphy results.

[0036] A finite state machine (FSM) is a mathematical model used to describe the behavior and state transitions of a system. It consists of a set of states, a set of inputs, a set of transition rules (which specify how to transition from one state to another given inputs), and an initial state. In this application, the states are system domain categories (LST, TST, HST). The state transition rules are strictly limited to allowing only transitions that conform to sedimentary evolution patterns: LST can remain LST or transition to TST; TST can remain TST or transition to HST; HST can remain HST or transition to the LST of the next cycle (indicating the start of a new sequence). The role of the FSM is to define geologically impossible reverse sequence transitions (such as HST → TST, TST → LST) as "illegal," thereby forcing corrections during optimization. The Viterbi algorithm is a dynamic programming algorithm used to find the hidden state sequence most likely to have generated a noisy observation sequence. In this application, the hidden state refers to the true system domain category (LST, TST, HST). The observation sequence refers to the probability distribution of the predicted systems tracts given by the deep learning model for each depth point. Transition probabilities are defined by geological laws and are typically set to 1 for only forward transitions (LST→TST→HST), and 0 or a minimum for reverse transitions. The Viterbi algorithm searches for a systems tract sequence that runs through the entire vertical profile. This sequence should match the high-probability categories predicted by the model as closely as possible while fully adhering to geological state transition rules, thus obtaining a globally optimal and geologically sound stratigraphic sequence. In sequence stratigraphy, the problem is decomposed into: for each depth point, calculating the maximum cumulative probability of reaching that point and being in a specific systems tract state (such as LST). The Viterbi algorithm recursively calculates from top to bottom, eventually backtracking to obtain the optimal sequence for the entire vertical direction. In this way, a unique optimal solution can be quickly found from a vast number of possible sequence combinations (3^N, where N is the number of depth points). Vertical superposition refers to the basic geological law that, during sedimentation, due to sea-level changes and the periodicity of sediment supply, different systems tracts follow a fixed sequential order in the vertical direction. For example, within a complete sea-level rise and fall cycle, the vertical sequence must be LST → TST → HST. This reflects the complete process of sea level rising from low (falling phase) through rapid rise (transgression phase) to high (stable or slow falling phase). Therefore, the core constraint is that within the same sequence, inverted sequences such as TST below LST or HST below TST are prohibited. Preliminary sequence delineation results refer to the direct output of preliminary predictions made by a knowledge-guided deep learning model on 3D seismic data. The preliminary sequence delineation result is a 3D data volume, where each voxel contains: the predicted probability of the system domain category, and may also include preliminary category determination.

[0037] Specifically, construct a finite state machine (allowing only geologically reasonable state transitions): 1. LST → LST or TST; 2. TST → TST or HST; 3. HST → HST or a new cycle of LST. The predicted sequence is scanned depth by depth. Illegal transitions (such as HST→TST, TST→LST) are marked as unreasonable and corrected using rules (e.g., local reclassification or truncation). Furthermore, when the model outputs probabilities P(LST), P(TST), and P(HST) for each category at each depth point, a transition matrix A that only allows "forward" transitions can be constructed. The Viterbi algorithm is then used to search for the optimal path S*(z) that satisfies both probability and order constraints, thus obtaining a geologically reasonable vertical systems domain sequence. Through this method, the superposition rules are transformed from "empirical" to explicit constraints within the model.

[0038] By employing finite state machine algorithms, Viterbi decoding, or dynamic programming algorithms to enforce vertical stacking constraints within the system domain, it is ensured that the final sequence stratigraphy conforms to the sedimentary evolution order of "LST→TST→HST" in the vertical direction. This eliminates any reverse-order errors that might arise from artificial intelligence models, providing an indisputable geological basis for the results. Furthermore, by performing independent sequence optimization on each spatial column, the optimized 3D model is logically self-consistent in every local vertical direction. This avoids the contradictions and patchwork errors that are unavoidable in manual section-by-section correction, significantly improving the intrinsic working quality of the 3D geological model.

[0039] In one embodiment, the implementation of seismic reflection geometry matching constraints includes: calculating seismic geometric attributes based on seismic data, the seismic geometric attributes including the dip angle and azimuth angle of the reflecting layer; identifying the reflection termination type of the stratigraphic reflection phase axis according to the seismic geometric attributes, the reflection termination type including at least one of top overflow, bottom overflow, and pre-accretion; obtaining a predefined system tract category conditional probability distribution associated with the reflection termination type; and weightedly fusing the conditional probability distribution with the system tract category prediction probability of the preliminary sequence stratigraphy result at the corresponding location to obtain a corrected prediction result.

[0040] Seismic reflection geometry refers to the macroscopic features of the spatial morphology, contact relationships, and termination patterns of in-phase axes (strata reflections) on seismic profiles. Seismic reflection geometry is a comprehensive seismic response reflecting the original occurrence of sedimentary strata, the sedimentary process, and subsequent tectonic deformation. Different geometries directly correspond to different sedimentary dynamic backgrounds and stratigraphic contact relationships, serving as key indicators for identifying sequence boundaries and systems tracts. For example, progradational structures indicate the propagation of sediments towards the basin; top wax represents stratigraphic overlap during marine transgression; and underwagging is commonly found at the end of progradational sequences. Seismic geometric attributes are parameters quantitatively calculated from 3D seismic data volumes to describe the spatial morphology of reflection interfaces. Core attributes include dip and azimuth. The dip angle is the maximum angle between the reflection interface and the horizontal plane, measured in degrees (°). It quantitatively describes the degree of dip of the strata. The azimuth is the direction of the dip angle, i.e., the orientation of the strata's dip, usually expressed as an angle rotating clockwise from true north (0°-360°).

[0041] First, the dip and azimuth angles can be calculated from the input raw 3D seismic data volume using gradient calculation or structural tensor analysis algorithms. This operation transforms continuous seismic amplitude signals into a continuous attribute field describing the spatial morphology of stratigraphic interfaces. Based on the calculated dip volume and azimuth angles, a set of rule-based logical discrimination algorithms is designed and applied. This algorithm automatically classifies each voxel or local region into a specific reflection termination type (top supremacy, bottom supremacy, anterior accretion, etc.) by scanning and comparing the dip and azimuth angle relationships of adjacent spatial locations. Then, a pre-established reflection termination type-systems tract conditional probability lookup table stored in the sedimentary model rule base can be obtained. After identifying the reflection termination type at a certain location, the system automatically retrieves the corresponding prior probability distribution from the lookup table. For the same spatial location, the prior probability distribution obtained from the above query (i.e., the conditional probability distribution) is linearly weighted and fused with the preliminary predicted probability distribution (i.e., the predicted probability of the systems tract category at the corresponding location based on the preliminary sequence division results) given by the knowledge-guided deep learning model. The weighting coefficients can be adjusted globally or locally based on data quality, regional credibility, or expert experience to obtain corrected prediction results.

[0042] In this embodiment, the macroscopic morphology of seismic reflections—a core geological piece of evidence—is forcibly incorporated into the interpretation process, ensuring that the final system tract division is physically consistent with the sedimentary processes revealed by the earthquake. This transforms the output of artificial intelligence from an incomprehensible "black box" result into one that geologists can understand, verify, and trust. Furthermore, seismic geometric properties (such as dip angle and curvature) are extremely sensitive to minute undulations and discontinuities at stratigraphic interfaces. Through this process, the model can effectively identify key interface markers such as weak top and bottom overshoots that are difficult to detect with amplitude data alone, as well as the fine internal stacking patterns of pre-accretionary structures, thereby improving the resolution and completeness of sequence stratigraphy. Moreover, the weighting coefficients in the weighted fusion in this embodiment provide a flexible adjustment knob. In areas with high model prediction confidence, larger weighting coefficients can be set, relying more on data-driven results; in areas with particularly clear geological patterns or poor data quality, the weighting coefficients of prior knowledge can be increased. This flexibility enhances the method's adaptability to different geological conditions and data situations in different work areas.

[0043] In one embodiment, the method further includes preprocessing the seismic data before inputting it into a pre-trained knowledge-guided deep learning network model. The preprocessing includes at least one of standardization, normalization, and alignment with sequence label data. Preprocessing operations can be common processing operations, which will not be elaborated here.

[0044] In one embodiment, the knowledge-guided deep learning network model is a U-Net convolutional neural network.

[0045] U-Net is a fully convolutional neural network with a symmetrical encoder-decoder structure, originally designed for biomedical image segmentation. In this application, U-Net is not simply applied, but rather used as a deeply customizable and knowledge-injected infrastructure, transformed into a "knowledge-guided" model through the following multi-layered methods: 1. Architecture level: Adapt to geological data and tasks.

[0046] Input Adaptation: The model's input channels are expanded to receive a "sedimentation pattern-sensitive input set." This is not just the raw seismic amplitude, but a multi-channel 3D tensor including dip angle, curvature, coherence volume, and transient properties. U-Net's powerful multi-channel feature fusion capability allows it to automatically learn and weight the contribution of these properties to system domain identification. Output Adaptation: The output is a probability map of each voxel belonging to categories such as LST, TST, and HST. This directly corresponds to pixel-level (voxel-level) semantic segmentation tasks, perfectly meeting the need to partition system domains on a voxel-by-voxel basis in 3D space. 2. Training mechanism level: Injecting geological rules through a comprehensive loss function. Standard U-Net uses cross-entropy loss. This application uses the following combined loss function for training: Loss = L_CE + λ1·L_order + λ2·L_geometry + λ3·L_thick + λ4·L_contact During backpropagation, the gradient is calculated not only from the difference between the prediction and the artificial label (L_CE), but also modulated by the degree of violation of geological rules (L_order, etc.).

[0047] U-Net's end-to-end differentiability allows this complex loss function, which incorporates geological patterns, to be used smoothly for training. All network parameters (convolutional kernel weights) are adjusted during the optimization of this composite objective, thus allowing the network to encode geological patterns internally.

[0048] 3. Structural enhancement level: Introduce a geological prior attention module.

[0049] A geologically prior-guided attention module can be inserted into the encoder or skip connections of U-Net. This module can accept additional guiding signals derived from the rule base (e.g., a probability map of the existence of a "topsupreme" structure in a certain area) and generate an attention weight map to weight the intermediate feature maps of U-Net. This is equivalent to telling the network, "In these areas where marine transgression may occur, pay more attention to features related to TST." This enables real-time, dynamic guidance of knowledge during the feature extraction stage, improving the network's ability to focus on key geological features.

[0050] 4. Process Integration Level: As the core predictive engine for dual-drive processes The U-Net model is positioned at the center of a dual-drive process: "depositing prior knowledge of patterns → initial prediction by machine learning → optimization of rule base constraints → final model". The probabilistic results (rather than hard classification) produced by this model provide rich and actionable inputs for subsequent optimization of the rule base (such as Viterbi decoding). At the same time, the constraints of the rule base are fed back to U-Net during training through the loss function, forming a closed loop.

[0051] In one embodiment, the sequence label data includes at least one of the following: system domain classification labels, sequence interface location labels, sequence element number labels, and typical seismic reflection structure labels.

[0052] Sequence boundary location labels refer to the spatial markings of key unconformities or corresponding conformable surfaces separating different levels of sequences (such as third-order and fourth-order sequences) within a data volume. In a two-dimensional profile, sequence boundary location labels can be a series of continuous polylines or curves. In a three-dimensional data volume, sequence boundary location labels can be a binary mask, where voxels with a value of 1 represent the boundary location, or a series of three-dimensional point sets or bedding grids. Sequence boundaries form the skeleton of sequence stratigraphy analysis, defining the spatiotemporal range of sequence units and are crucial for isochronous stratigraphic correlation and tectonic-sedimentary evolution analysis. Sequence unit numbering labels refer to unique identifiers for sequence units of different levels with specific time spans (such as "third-order sequence A" or "supersequence I"), used to spatially organize and associate strata belonging to the same sedimentary cycle. They can be viewed as instance segmentation labels, assigning a unique ID number to each independent sequence unit volume (whether a two-dimensional region or a three-dimensional volume). In the mask, different numbers represent different sequence units. Typical seismic reflection structure labels refer to the annotations on seismic profiles that characterize the termination mode and external morphology of seismic reflections, representing specific sedimentary processes or geometric configurations. The main types include: pre-echo, top hyperbole, and bottom hyperbole. The annotation format can be a pixel-level mask (marking areas with specific structures), a bounding box or polygon annotation, or even vector data with type attributes.

[0053] Specifically, knowledge-guided deep learning networks (U-Net) can use these multi-type labels for end-to-end supervised training. Labels guide model learning in various ways. For example, system domain classification labels serve as direct supervision signals for the primary task, driving the model to learn the mapping relationship between seismic attributes and system domain categories through cross-entropy loss. Sequence boundary location labels and sequence unit number labels can serve as supervision signals for auxiliary tasks. For instance, the model can simultaneously output system domain classification maps and boundary probability maps, learning shared features through multi-task learning, making the features extracted by the model beneficial for both boundary detection and in vivo classification. Typical seismic reflection structure labels can serve as feature guidance or constraints, acting as an independent segmentation task or used to calculate geometric matching constraint loss (L_geometry), forcing the model to focus on structural features with clear geological significance. This multi-label supervision mechanism essentially deconstructs and encodes the systematic, multi-factor comprehensive reasoning logic used by geological experts in interpretation into the training objective, enabling the model to mimic the comprehensive judgment ability of experts, rather than simply performing image classification.

[0054] In one embodiment, the input to the knowledge-guided deep learning network model is a multi-channel sedimentation pattern sensitive input set; the sedimentation pattern sensitive input set includes at least two of the following: seismic amplitude attributes, seismic geometric attributes, seismic reflection characteristic attributes, and sedimentation pattern response attributes.

[0055] A sedimentary pattern sensitive input set refers to a multi-channel, multi-dimensional data cube specifically designed and constructed for training and running knowledge-guided deep learning network models. Its core design idea is to not only provide raw seismic data, but also systematically integrate various geophysical and geological attributes that are highly indicative of sedimentary patterns, sequence boundaries, and systems tracts, forming a comprehensive input that is complementary and sensitive to changes in the sedimentary environment. Seismic amplitude attributes refer to various attributes related to the intensity (energy) of reflected waves in seismic records, and are the most direct reflection of differences in stratigraphic impedance. Seismic amplitude attributes can include raw amplitude, envelope / instantaneous amplitude, and root-mean-square amplitude. Raw amplitude is the basic reflection intensity, reflecting differences in lithology, porosity, and fluid properties. Envelope / instantaneous amplitude reflects the total reflected energy and can be used to trace strong reflection axes and identify unconformities. Root-mean-square amplitude is the average amplitude calculated within a time window, which helps to depict the spatial distribution of lithological bodies such as sand bodies and channels. Seismic geometric attributes are derived attributes describing the spatial morphology, attitude, and curvature variations of seismic reflection phase axes. They are calculated from raw seismic data using mathematical methods such as gradient and structural tensor. Seismic geometric attributes can include dip and azimuth, and curvature. Curvature describes the degree of bending of the reflecting surface. High curvature regions are often associated with geological phenomena such as fracture zones, channel edges, and reef boundaries. In this embodiment, converting the "morphology" of seismic reflections into quantitative data is the core basis for identifying sedimentary structures, tectonic deformations, and stratigraphic contact relationships. Seismic reflection characteristic attributes describe the characteristics of seismic reflected waves in terms of continuity, similarity, texture, and frequency, reflecting the lateral homogeneity and vertical stacking pattern of strata. Seismic reflection characteristic attributes can include coherence / similarity and texture attributes. Coherence / similarity measures the degree of similarity of waveforms between adjacent seismic traces. Low coherence zones indicate faults, channels, unconformities, or zones of drastic stratigraphic change, and are a powerful tool for identifying sequence boundaries and sedimentary unit boundaries. Texture attributes, calculated using methods such as gray-level co-occurrence matrices, quantify the texture features of seismic images, distinguishing between different sedimentary structures such as blocky, layered, and disordered structures. Sedimentary pattern response attributes are a higher-order category of attributes, directly calculated or derived from specific sedimentary pattern theories or geological concepts. Their design aims to "encode" geologists' prior knowledge into data features.

[0056] In one embodiment, the geological rationality constraint and optimization of the preliminary sequence stratigraphy results using a sedimentary model rule base includes: inputting the preliminary sequence stratigraphy results into the sedimentary model rule base, identifying suspicious areas that violate superposition rule constraints and / or geometric shape matching constraints; and adjusting the classification results of the suspicious areas using local reclassification, Markov random field optimization, or Viterbi path repair algorithm.

[0057] Suspicious areas refer to the set of spatial locations identified by the sedimentary model rule base as violating one or more geological constraints in the preliminary delineation results. The geological interpretation of these areas is clearly unreasonable. Suspicious areas can be detected by scanning each constraint condition in the rule base: 1. Sequence Violation Detection: Mark depth intervals where vertical order is reversed (e.g., HST below TST). 2. Geometric Conflict Detection: In regions identified as specific reflective structures (e.g., supermassive tops), if the predicted system domain type does not match the rule (e.g., predicted as HST instead of TST), then mark the region. 3. Thickness Anomaly Detection: Calculate the local system domain thickness; regions exceeding a reasonable range are marked. The detection result for suspicious regions can be a three-dimensional binary mask, where "1" represents a suspicious voxel and "0" represents a reliable voxel.

[0058] Local reclassification is a fine-grained correction method for suspicious regions. It uses the suspicious region as its core, combining information from surrounding reliable regions and multi-source data features to redetermine the system domain category of each voxel within that region. Multi-source features, including seismic attributes, geometric features, and original prediction probabilities, are extracted within the suspicious region and its neighborhood. Then, a lightweight classifier (such as random forest or support vector machine) or a rule-based classifier is used for reclassification locally. The classification process considers consistency with adjacent reliable regions to ensure a natural transition of the corrected result to the surrounding geological environment. Markov random field optimization is a global optimization method based on probabilistic graphical models. It treats the category of each voxel as a random variable and models the spatial dependencies and data fit between variables by defining an energy function. Common algorithms include graph cut, belief propagation, and iterative conditional patterns. The goal is to simultaneously optimize the classification of all voxels globally, eliminating noise and isolated errors, and ensuring spatial continuity and reasonable boundaries. The Viterbi path repair algorithm is a dynamic programming-based sequence optimal path search algorithm specifically designed for optimizing one-dimensional sequence problems with order constraints.

[0059] Specifically, the preliminary sequence stratigraphy results can be used as input, loading all constraint rules from the sedimentary model rule base for a comprehensive and automated geological rationality check. For each spatial location (x, y), the vertical sequence S0(z) is scanned using finite state machine rules. Once an illegal state transition (e.g., TST→LST) is detected, the relevant depth interval is immediately marked as suspicious. The preliminary system domain map is compared with a reflection termination type map independently calculated from seismic data. In areas where geological rules define "should match" (e.g., the top-super region should primarily be TST), if a severe mismatch occurs, it is marked. A location may violate multiple rules simultaneously. Finally, a comprehensive 3D mask of suspicious areas is generated, quantifying the "suspicion level" of each location (e.g., the number and severity of rule violations). Based on the distribution characteristics, type, and scale of suspicious areas, the system intelligently selects or combines the following optimization algorithms: local reclassification, Markov random field optimization, or Viterbi path repair. The implementation process of local reclassification is as follows: A 3D neighborhood window of appropriate size is extracted centered on each suspicious region; within this window, reliable regions (non-suspicious) are used as training samples, combined with multi-source features, to train a local classification model. This model is then applied to reclassify suspicious voxels, ensuring a smooth transition with reliable results at the window boundaries. The advantage of local reclassification is that it preserves the results of reliable regions to the greatest extent possible, only making minimal necessary corrections to problem regions. The implementation process of Markov Random Field (MRF) optimization is as follows: The entire 3D data volume or a large sub-volume is constructed as an MRF model; an energy function is defined: data terms are based on the original predicted probabilities, and a smoothing term encourages spatial consistency; optional rule penalty terms are added to apply stronger correction forces to known suspicious regions; algorithms such as graph cut are used to minimize the energy, resulting in an optimized classification volume. The advantage of MRF optimization is that it achieves global consistency optimization, making it particularly suitable for handling large-area, complex spatial inconsistency problems. The implementation process of Viterbi path repair is as follows: For each vertical sequence requiring optimization (such as a well trajectory or a grid column), its predicted probability sequence is extracted; a transition matrix conforming to geological laws is defined (e.g., reverse transitions are prohibited); the Viterbi algorithm is run to output a new, compliant system domain sequence; and all optimized sequences are recombined into a three-dimensional volume. The advantage of Viterbi path repair lies in ensuring the correctness of the vertical geological sequence.

[0060] After the above optimization, the suspicious areas were corrected, resulting in a final sequence stratigraphy result with significantly enhanced geological rationality. This result has passed rigorous automated quality control. The final sequence stratigraphy result can be input into the rule base again for verification to ensure that all constraints are met, forming a closed loop.

[0061] In one embodiment, the method further includes a training step for a knowledge-guided deep learning network model. The training step includes: acquiring sample seismic data and corresponding sample stratigraphic label data of the training area, wherein the sample stratigraphic label data is labeled by geological experts based on prior geological knowledge; constructing a deep learning network; using the sample seismic data as input and the sample stratigraphic label data as supervision, iteratively optimizing the deep learning network using a comprehensive loss function that incorporates geological rule constraints to obtain a pre-trained knowledge-guided deep learning network model; wherein the comprehensive loss function includes at least one geological rule constraint term used to penalize prediction results that violate geological rules.

[0062] This process involves collecting sample seismic data and corresponding sequence label data for training work areas. The training work area is essentially a work area, distinguished only from the target work area in terms of description. The sample seismic data is multi-source, including 2D / 3D seismic data, well logging data, core description data, and prior geological knowledge. The seismic and well logging data can then be standardized, normalized, and aligned to form an input dataset usable by the machine learning model. Geologists can use this prior geological knowledge and standard sedimentary patterns to perform artificial sequence divisions on key profiles or well sections, clearly labeling sequence boundaries and systems tracts (such as lowstand series (LST), transgressive series (TST), and highstand series (HST). The artificial interpretation results are then converted into labeled data (such as masked images) for supervising the training of the learning model, ensuring the geological reliability of the training data. The input data for the training phase consists of preprocessed multi-channel two-dimensional / three-dimensional tensors, primarily including: seismic amplitude, phase, and instantaneous attributes (such as envelope and instantaneous frequency), dip angle, curvature, and other seismic geometric attributes, as well as reflection characteristics such as coherence volume, azimuth energy, and anisotropy attributes. These multi-source attributes collectively constitute the "Depositional-Pattern Enhanced Input," denoted as: ·, …, A n [x,y,z)], where A i This represents different seismic attribute channels. Training labels, on the other hand, are manually interpreted by geological experts from a few representative wells and profiles based on sedimentary patterns and sequence stratigraphy principles, and then converted into label data recognizable by the model. These include: System domain classification labels (LST, TST, HST); Location of the layer sequence interface; Numbering of hierarchical units at each level (unit hierarchical sequence, superhierarchical sequence, etc.); Annotation of typical earthquake reflection structures (forelap, top overshoot, bottom overshoot).

[0063] Tags can be represented as semantic segmentation masks or sequence annotation matrices, denoted as: Label-Data(x,y,z) ∈ {LST, TST, HST, …}.

[0064] During the training phase, the U-Net model outputs a multi-class probability map with the same size as the input, where each voxel (x, y, z) corresponds to a probability P(x, y, z, c) of belonging to each class. Here, c represents the system domain or sequence unit class. To ensure the results both closely match the data and satisfy geological patterns, this application employs a comprehensive loss function of the following form: Loss = L_CE + λ1·L_order + λ2·L_geometry + λ3·L_thick + λ4·L_contact Where L_CE is the conventional cross-entropy loss, used to constrain the consistency between the prediction and the manual label; L_order is the constraint loss for the vertical stacking order of system domains (LST→TST→HST); L_geometry is the constraint loss for matching the reflection geometry (pre-product, top-overlap, bottom-overlap) with the system domains; L_thick is the thickness constraint loss for different system domain thicknesses to be within a geologically reasonable range; L_contact is the constraint loss for the reasonableness of the vertical contact relationship between system domains; λ1–λ4 are weighting coefficients used to balance the importance of each constraint. For example, for the system domain prediction sequence S(z) of the same vertical profile (represented by 1, 2, and 3 for LST, TST, and HST respectively), the order constraint loss can be constructed as: L_order = Σ_z max(0, S(z)). If S(z) > S(z+1) at some point, that is, the higher-order system domain is "subjected" to the lower-order system domain, then a penalty occurs.

[0065] Through the above loss design, the model not only "fits the data" during training, but also "learns geological rules" simultaneously. Moreover, the prediction stage only requires the input of the entire seismic data volume of the entire work area and its various attributes, in the same format as the training stage, namely: Input-Data(x,y,z) (but label data is no longer needed).

[0066] After obtaining a trained knowledge-guided deep learning network model, automatic prediction can be performed on the entire work area, with outputs including: 1. System domain category (LST, TST, HST, etc.) for each voxel in 3D space; 2. Automatically identified sequence boundary locations; 3. 3D sequence unit volume; 4. Optional output probability maps of sedimentary geometries (pre-accretion, top superposition, bottom superposition, etc.). This preliminary result is denoted as S0(x,y,z), and then enters the sedimentary model constraint and optimization module. After rule-based constraints and corrections, the final prediction result S*(x,y,z) conforming to geological laws is obtained.

[0067] In one embodiment, the knowledge-guided deep learning network model is a U-Net convolutional neural network. This invention does not simply adopt the existing U-Net model, but rather addresses the problem of layer-by-layer construction by making multifaceted improvements to the model structure, input features, and training mechanism, forming a "knowledge-guided deep learning network." The main innovations are as follows: 1) Knowledge-integrated U-Net constrained by sedimentary model: Based on the traditional U-Net, this application explicitly incorporates the following geological knowledge into the model: a. Vertical order priors of system domain evolution (LST → TST → HST); b. Correspondence between typical seismic reflection geometries (pre-product, top-overlap, bottom-overlap) and different system tracts; C. The geologically reasonable range of each system tract in terms of thickness, spatial distribution, and relative position.

[0068] These priors, through the combined action of the following functions, enable the network to "embed" sedimentary pattern knowledge throughout the entire training and prediction process: rule-constrained loss functions (such as L_order, L_geometry, L_thick, L_contact); a geological prior-guided attention module; and a geological-constrained filter (Geology-Filter / CRF) in the post-processing stage. 2) Multi-channel fusion input for sequence stratigraphy: Traditional methods mostly use only seismic amplitude, while this application designs a "sedimentary model sensitive input set" that introduces multiple channels including dip angle, curvature, coherence volume, azimuth energy, instantaneous attributes, and sedimentary model response attributes constructed in this invention, which fundamentally enhances the model's ability to identify sequence interfaces, system tract boundaries, and sedimentary geometries.

[0069] 3) A dual-driven process of "geological constraints + deep learning" Prior knowledge of sedimentary models includes preliminary predictions from machine learning (S0), rule base constraints and optimization, and the final sequence stratigraphic model (S*).

[0070] Unlike traditional "purely manual interpretation" or "purely deep learning automatic segmentation", this application achieves a deep integration of model efficiency and geological expert knowledge, which significantly improves the credibility and consistency of the results.

[0071] In one embodiment, the sequence stratigraphic interpretation results include at least one of: a two-dimensional sequence stratigraphic profile, a three-dimensional sequence volume, and a three-dimensional mesh model of the sequence interface.

[0072] The final output of this application is not a simple model pixel classification result, but a high-precision sequence stratigraphic interpretation result constructed by combining multiple data sources based on the optimized system domain prediction volume S*(x,y,z), including: Two-dimensional sequence stratigraphic profile; 3D Sequence Volume; Horizon Grid, a three-dimensional mesh model of a hierarchical interface.

[0073] In this embodiment, the data sources on which the high-precision sequence stratigraphic interpretation results are constructed mainly include: 1. The optimized system domain prediction volume S*(x,y,z); 2. Raw seismic data and various seismic attribute volumes (amplitude, dip, coherence, curvature, etc.); 3. Well data (well logging curves, sequence boundary well depths, synthetic seismic records, time-depth conversion relationships, etc.); 4. Structural interpretation results (faults, major stratigraphic interfaces, structural morphology). 5. Sedimentary model rule base (location of maximum flooding surface, spatial distribution rules of system tracts, etc.); 6. 3D modeling mesh and interpolation / smoothing parameters (such as minimum curvature, Kriging, RBF, etc.).

[0074] Specifically, the construction process of the two-dimensional sequence profile includes: extracting the system tract division results of S* from the selected profile; superimposing them onto the seismic amplitude profile as a background display; locally adjusting the position of the sequence interface using properties such as dip angle and coherence to ensure the interface matches the reflection events; accurately correcting the interface depth on the profile based on well data; and eliminating unnecessary noise through curve smoothing methods to ensure a smooth and continuous sequence interface. The final profile simultaneously reflects "model results + seismic characteristics + well constraints," facilitating expert review and interpretation. The construction process of the three-dimensional sequence volume includes: based on S*, performing voxel-level system tract division on the entire three-dimensional volume, and combining: coherence volume and structural interpretation to constrain the spatial boundaries between system tracts; well and structural surfaces to ensure the distribution of system tracts in three-dimensional space is consistent with the structural pattern; connectivity checks and morphological corrections to eliminate isolated small blocks and obvious noise volumes. The final three-dimensional sequence volume can directly serve subsequent reservoir prediction, sedimentary system analysis, etc. The construction process of the sequence interface grid model (HorizonGrid) includes: extracting transition zones between different system tracts from S* as the initial locations of the sequence interfaces; using the seismic dip body Dip(x,y,z) to tangentially constrain the interface morphology, making it as parallel as possible to the formation reflections; fine-tuning or projecting the interface points to locations with better reflection continuity, ensuring the interface closely follows the main reflection events; anchoring the interface at the well location based on well data to ensure the interpretation results are consistent with the measured well depths; constructing a smooth and continuous sequence interface in three-dimensional space using surface fitting and smoothing algorithms such as minimum curvature and RBF; and combining fault interpretation results to perform fault cutting and fitting processing on the interface, ensuring reasonable geometric relationships on both sides of the fault. The final three-dimensional sequence interface grid model is geometrically smooth and continuous, geologically conforms to sedimentary and tectonic patterns, and is highly consistent with seismic reflections and well data.

[0075] The core idea of ​​this application is to deeply integrate geologists’ prior knowledge (sedimentary models) with data-driven machine learning algorithms to form a “dual-drive” workflow, so as to achieve rapid, objective and geologically sound sequence stratigraphy modeling.

[0076] Figure 1 This is a flowchart illustrating a sequence stratigraphic intelligent interpretation method in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0077] In one embodiment, a sequence stratigraphic intelligent interpretation device (not shown in the figure) is provided, the device comprising: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the sequential stratigraphic intelligent interpretation method described above.

[0078] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and the intelligent interpretation method for sequence stratigraphy can be implemented by adjusting kernel parameters.

[0079] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0080] This application provides a storage medium storing a program that, when executed by a processor, implements the above-described sequence stratigraphic intelligent interpretation method.

[0081] This application provides a processor for running a program, wherein the program executes the above-described sequence stratigraphic intelligent interpretation method.

[0082] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 2 As shown, the computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with external terminals via a network connection. When executed by the processor A01, the computer program B02 implements a hierarchical intelligent interpretation method.

[0083] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0084] This application provides a computer (electronic) device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above-mentioned sequence stratigraphic intelligent interpretation methods.

[0085] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing the steps of an initialized hierarchical intelligent interpretation method.

[0086] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0091] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0092] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0093] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0094] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A sequence stratigraphic intelligent interpretation method, characterized in that, The method includes: Acquire seismic data for the target work area; The seismic data is input into a pre-trained knowledge-guided deep learning network model to obtain preliminary sequence division results. The preliminary sequence division results include at least the system domain category of each voxel in three-dimensional space and the automatically identified sequence interface position. The pre-trained knowledge-guided deep learning network model is a deep learning model obtained by supervised training based on sequence label data annotated with prior geological knowledge, and a comprehensive loss function that integrates geological rule constraints is used during the training process. The preliminary sequence stratigraphy results are subjected to geological rationality constraints and optimization using a sedimentary model rule base to obtain a final sequence stratigraphy result with enhanced geological rationality. The sedimentary model rule base includes at least constraints on the vertical stacking regularity of system tracts and constraints on the matching of seismic reflection geometry. Sequence stratigraphic interpretation results are generated based on the final sequence stratigraphic division results. The expression for the comprehensive loss function Loss is as follows: Loss = L_CE + λ1·L_order + λ2·L_geometry + λ3·L_thick + λ4·L_contact Wherein, L_CE is the cross-entropy loss, used to constrain the consistency between the prediction and the artificial label; L_order is the constraint loss for the vertical stacking pattern of the system domain; L_geometry is the constraint loss for the matching of the reflection geometry with the system domain; L_thick is the constraint loss for the thickness of the system domain; L_contact is the constraint loss for the rationality of the contact relationship of the system domain; λ1, λ2, λ3, and λ4 are weighting coefficients used to balance the importance of each constraint.

2. The method according to claim 1, characterized in that, The constraint loss L_order of the vertical stacking law of the system domain is calculated by the following formula: in, This represents the system domain category code value at depth z, which increases in the order of low-order system domain, transgressive system domain, and high-order system domain; z is the depth sampling point index. This is the system domain category encoding value at depth z+1.

3. The method according to claim 1, characterized in that, The constraint loss for matching the reflection geometry with the system domain is determined by the following methods: The seismic geometric properties are calculated based on the seismic data, including the dip angle and azimuth angle of the reflecting layer; The reflection termination type of the stratigraphic reflection phase axis is identified based on the seismic geometric attributes, and the reflection termination type includes at least one of top overflow, bottom overflow, and forward accumulation; Based on geological experience, a conditional probability relationship is established between the reflection termination type and the system tract category. The conditional probability relationship is used to represent the probability distribution of different system tract categories when a specific reflection termination type occurs. Based on the difference between the predicted probability of the system domain category in the preliminary sequence partitioning results and the conditional probability relationship, the constraint loss for matching the reflection geometry with the system domain is calculated.

4. The method according to claim 1, characterized in that, The methods for determining the thickness constraint loss, which should be within a geologically reasonable range for the thickness of different system domains, include: Based on the preliminary hierarchical division results, the predicted thickness of each system domain category in spatial location is calculated. Obtain the geologically reasonable range corresponding to each system domain category; For each system domain category, a first difference and a second difference are determined based on the comparison between the predicted thickness of the system domain category and its corresponding geologically reasonable range; wherein, when the predicted thickness is less than the lower limit of the geologically reasonable range, the first difference is the difference between the lower limit and the predicted thickness; otherwise, the first difference is zero; when the predicted thickness is greater than the upper limit of the geologically reasonable range, the second difference is the difference between the predicted thickness and the upper limit; otherwise, the second difference is zero. The thickness constraint loss of the system domain is calculated based on the first and second differences of all system domain categories.

5. The method according to claim 4, characterized in that, The system domain thickness constraint loss is the sum of the thickness constraint losses of all system domain categories, wherein the thickness constraint loss of each system domain category is calculated according to the following formula: Where Loss_thick,k is the thickness constraint loss for the k-th system domain category, and k is the system domain category. h k The predicted thickness of the k-th system domain. This represents the lower limit of the geologically reasonable range corresponding to the k-th system domain. This represents the upper limit of the geologically reasonable range corresponding to the k-th system domain.

6. The method according to claim 1, characterized in that, Applying geologically sound constraints and optimizations to the preliminary sequence stratigraphy results using a sedimentary model rule base includes constraining the vertical stacking patterns of the system domains using finite state machine algorithms, Viterbi decoding, or dynamic programming algorithms. Among them, the finite state machine algorithm detects and corrects the vertical sequence of the preliminary sequence division results based on the rule that only the low-level system domain (LST) is allowed to remain or transition to the transgressive system domain (TST), and the transgressive system domain (TST) is allowed to remain or transition to the high-level system domain (HST). The Viterbi decoding or dynamic programming algorithm is used to search for the optimal sequence path that conforms to the vertical stacking rule, based on the system domain category probabilities provided by the preliminary sequence partitioning results.

7. The method according to claim 1, characterized in that, The implementation methods of the seismic reflection geometry matching constraint include: Based on the earthquake data, earthquake geometric properties are calculated, including the dip angle and azimuth angle of the reflecting layer; The reflection termination type of the stratigraphic reflection phase axis is identified based on the seismic geometric attributes, and the reflection termination type includes at least one of top overflow, bottom overflow, and forward accumulation; Obtain the predefined system domain category conditional probability distribution associated with the reflection termination type; The conditional probability distribution and the predicted probability of the system domain category at the corresponding position of the preliminary sequence division result are weighted and fused to obtain the corrected prediction result.

8. The method according to claim 1, characterized in that, The knowledge-guided deep learning network model is the U-Net convolutional neural network.

9. The method according to claim 1, characterized in that, The method further includes: Before inputting the seismic data into the pre-trained knowledge-guided deep learning network model, the seismic data is preprocessed, including at least one of standardization, normalization, and alignment with the sequence label data.

10. The method according to claim 1, characterized in that, The sequence label data includes at least one of the following: system domain classification labels, sequence interface location labels, sequence unit number labels, and typical seismic reflection structure labels.

11. The method according to claim 1, characterized in that, The input to a knowledge-guided deep learning network model is a multi-channel deposition pattern sensitive input set; The sedimentation model sensitive input set includes at least two of the following: seismic amplitude attributes, seismic geometric attributes, seismic reflection characteristic attributes, and sedimentation model response attributes.

12. The method according to claim 1, characterized in that, The process of applying geologically sound constraints and optimizations to the preliminary sequence stratigraphy results using a sedimentary model rule base includes: The preliminary sequence division results are input into the depositional pattern rule base to identify suspicious regions that violate the superposition rule constraints and / or the geometric shape matching constraints; The classification results of the suspicious regions are adjusted using local reclassification, Markov random field optimization, or the Viterbi path repair algorithm.

13. The method according to claim 1, characterized in that, The sequence stratigraphic interpretation results include at least one of the following: two-dimensional sequence stratigraphic profile, three-dimensional sequence volume, and three-dimensional mesh model of sequence interface.

14. The method according to claim 1, characterized in that, The method further includes a training step for a knowledge-guided deep learning network model, the training step comprising: Acquire sample seismic data and corresponding sample sequence label data for the training area. The sample sequence label data is labeled by geological experts based on prior geological knowledge. Build deep learning networks; Using the sample seismic data as input and the sample stratigraphic label data as supervision, the deep learning network is iteratively optimized using a comprehensive loss function that incorporates geological rule constraints to obtain a pre-trained knowledge-guided deep learning network model; wherein, the comprehensive loss function includes at least one geological rule constraint term used to penalize prediction results that violate geological rules.

15. A sequence stratigraphic intelligent interpretation device, characterized in that, include: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the sequence stratigraphic intelligent interpretation method according to any one of claims 1 to 14.

16. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the sequence stratigraphic intelligent interpretation method according to any one of claims 1 to 14.

17. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the sequence stratigraphic intelligent interpretation method according to any one of claims 1 to 14.

Images