Method for calibrating pre-stack seismic inversion using fully connected neural network

Abstract

A method for calibrating pre-stack seismic inversion is provided. The method includes: selecting multiple features from the inverted elastic properties to generate reservoir properties; using a fully connected neural network model to learn the mapping relationship between these features and real data at the well location; and applying the predictions to the reservoir description across the entire exploration area to generate one or more final models.

Description

Technical Field

[0001] This disclosure generally relates to the fields of seismic exploration and geophysical reservoir characterization. More specifically, the present invention is a method for calibrating pre-stack seismic inversion using a fully connected neural network. Background Technology

[0002] Seismic inversion is widely used in the oil and gas industry to predict subsurface geological features and properties that cannot be directly observed. This technique involves collecting seismic and well logging data through seismic exploration, and then performing mathematical and computational analysis on these data to derive information about lithology, fluid content, and other physical properties. Pre-stack inversion uses seismic data before stacking (stacking is the process of integrating and averaging data to improve the signal-to-noise ratio). This can leverage the amplitude-with-migration (AVO) or amplitude-with-angle (AVA) effects to extract more comprehensive subsurface geological information than post-stack inversion.

[0003] Seismic data offers broad coverage and provides information on large-volume subsurface structures; however, its vertical resolution is relatively low compared to well logging data. Conversely, while well logging data offers high vertical resolution and precise measurements of rock and fluid properties at specific depths, it lacks spatial coverage. Pre-stack seismic inversion aims to combine seismic and well logging data to generate high-resolution reservoir properties with broad spatial coverage.

[0004] Applying pre-stack inversion in complex environments requiring high accuracy and high vertical resolution presents several challenges. One challenge is that certain properties, such as density, are only sensitive to large-offset seismic data, which are typically noisier and less reliable than near-offset data. Another challenge is that seismic data and wavelets are band-limited, leading to the loss of high-frequency details crucial for discerning fine geological features, thin layers, and small reservoir heterogeneity. A third challenge is that certain geological properties, such as fluid indicators, cannot often be obtained directly through inversion but require the combination of P-wave velocity, S-wave velocity, and density. The inherent uncertainties associated with these parameters can accumulate, resulting in an overall uncertainty far greater than that associated with any single parameter. Therefore, addressing these challenges is essential for improving the accuracy and resolution of pre-stack seismic inversion to enhance lithological identification, fluid identification, and reservoir characterization in seismic exploration and reservoir characterization. Summary of the Invention

[0005] One embodiment of this disclosure provides a method for calibrating pre-stack seismic inversion using a fully connected neural network to provide more accurate and higher-resolution predictions of reservoir properties across a seismic survey area. The method includes the following steps: acquiring seismic and well logging data; deriving seismic horizons, estimated wavelets, and an initial geological model; obtaining a pre-stack inversion geological model by integrating the seismic data, well logging data, horizons, estimated wavelets, and the initial geological model; calculating the geological properties of a target reservoir by integrating the inversion geological model; generating features from the target reservoir properties and optionally from other geological properties and data types; selecting the number of layers and parameters for each layer in the neural network to optimize the calibration process; training the neural network model and rigorously evaluating the training results at well locations and across the entire survey area; and implementing the calibrated model's predictions across the entire survey area, optionally generating additional rock and fluid property volumes.

[0006] Examples of pre-stack seismic inversion methods in this approach may include elastic wave equation inversion, model-based inversion, stochastic inversion, and sparse spike inversion.

[0007] In this method, the inverted geological model consists of three components and is parameterized as P-wave velocity, S-wave velocity and density, or P-wave impedance, S-wave impedance and density, or bulk modulus, shear modulus and density.

[0008] In this method, the features extracted from the target geological attributes include amplitude, amplitude envelope, amplitude-weighted frequency, amplitude-weighted phase, average frequency, apparent polarity, cosine instantaneous phase, seismic data derivative, instantaneous amplitude derivative, dominant frequency, instantaneous frequency, instantaneous phase, and integral absolute amplitude.

[0009] Examples of other attributes and data types in this method include initial model, master survey line, connecting line, two-way travel time, depth, strata, topography, and tidal data.

[0010] In this method, a fully connected neural network is selected and trained under a regression architecture, where the real data consists of a set of continuous measurements.

[0011] In this method, the target reservoir geological properties include P-wave velocity, S-wave velocity, density, Poisson impedance, and the P-wave to S-wave velocity ratio. And fluid indicator factors. Attached Figure Description

[0012] The teachings of the invention can be readily understood by taking into account the accompanying drawings and the following description: Figure 1 This is a schematic top view of a survey area containing incident points of various seismic sources according to an embodiment of the present disclosure; Figure 2 It is a schematic diagram showing a cross-sectional view of an environment according to one embodiment, including the incident point of the seismic source, seismic data recording sensors, well location, well casing, various propagating rays and various incident angles; Figure 3 This is a schematic cross-sectional view of an environment including a wellbore and a logging tool according to one embodiment, the logging tool including one or more acoustic generators and one or more logging data recording sensors; Figure 4 This is a schematic diagram illustrating a high-performance computing system according to one embodiment; Figure 5 This is a flowchart illustrating a method for calibrating pre-stack seismic inversion using a fully connected neural network according to an embodiment of the present disclosure; Figure 6 This is a schematic diagram showing an environment mapping that includes the training data on the left and the prediction results of the survey area on the right. Figure 7 This is a flowchart illustrating the process of using pre-stack seismic inversion to predict target reservoir properties; Figure 8 This demonstrates a fully connected neural network; Figure 9 This is a flowchart illustrating the process of calibrating pre-stack seismic inversion results; Figure 10 It is a graph showing the true P-wave velocity, S-wave velocity, and density of a thin layer with high contrast and lateral heterogeneity; Figure 11 It shows the curves of P-wave velocity, S-wave velocity, density, and generated fluid factor retrieved for seismic trace No. 220; Figure 12 It shows Figure 10 Curves of fluid factors retrieved at seismic traces 75, 175, 220, and 400; Figure 13 It shows Figure 10 A graph showing the fluid factor predicted using the methods described in this disclosure at seismic traces 75, 175, 220, and 400; and Figure 14 It shows the predicted fluid factor and compares it with Figure 10 A graph comparing the fluid factors inverted from all seismic traces. Detailed Implementation

[0013] Reference will now be made in detail to several embodiments of this disclosure, examples of which are illustrated in the accompanying drawings. It should be noted that, where possible, similar or identical reference numerals are used in the drawings, and may represent similar or identical functions.

[0014] The accompanying drawings depict embodiments of the present disclosure and are for illustrative purposes only. Those skilled in the art will readily recognize from the following description that alternative embodiments of the structures, systems, and methods shown herein exist without departing from the principles described herein.

[0015] Throughout this specification, the terms “means,” “method,” and “technique” are used interchangeably and have the same meaning.

[0016] In this specification, the terms "loss function", "cost function" and "error function" are used interchangeably and have the same meaning.

[0017] In this specification, the terms "fluid factor" and "fluid indicator factor" have the same meaning.

[0018] Pre-stack seismic inversion can be achieved by iteratively updating the geological model and generating synthetic seismograms to match observed seismic data. The forward modeling problem can be formulated as follows: ,(1); in, This represents observed seismic data at different offsets or angles; Representing the elastic model (P-wave velocity) S-wave velocity or formation density This model can be directly measured at the well site as real data; As a forward modeling operator, it can be simplified to a linearized operator. If an elastic wave propagation operator is used, it can be nonlinear.

[0019] An example of a general cost function is the quadratic cost function. Mathematically, the quadratic cost function for seismic inversion... This can be expressed as: (2).

[0020] Several limiting factors affect the accuracy and vertical resolution of the inversion results. One factor is that seismic data recorded at the surface typically attenuates, with deeper segments attenuating much faster than shallower segments, leading to band-limiting of seismic data at reservoir targets. Another factor is that wavelets are usually required in seismic inversion. Wavelets can be non-stationary, meaning they can vary not only vertically in depth or time but also spatially, making accurate estimation of wavelets challenging. Furthermore, when using linearized Zoeppritz equations or even full-wave elastic equations for pre-stack seismic inversion, density is only sensitive to high-angle reflections, which are difficult to obtain.

[0021] To distinguish between various lithologies and rock types, or to identify the presence and distribution of fluids within a reservoir, an attribute is typically selected by cross-plotting with lithology or fluid content. This attribute can be directly derived from elastic parameters. However, more often, it involves constructing a new parameter by combining one or more parameters among P-wave velocity, S-wave velocity, and density. In an exemplary embodiment, a fluid factor based on the following formula is used (Russell B. et al., 2003, "Fluid Property Discrimination Using AVO: A Biot-Gassmann Perspective," Geophysics, Vol. 68, No. 1, pp. 29-39): , (3); in, Indicates the fluid factor. Indicates density, Is with The relevant constants, and as an application example, .

[0022] Equation (3) shows that if the inversion uses P-wave velocity, S-wave velocity, and density as parameters, the fluid factor is obtained indirectly. The error of the fluid factor can be approximated as follows (similar approximations are obtained by using P-wave impedance, S-wave impedance, and density, or by using bulk modulus, shear modulus, and density as parameters): (4).

[0023] Equation (4) shows that the error in the target attribute (such as the fluid factor) may be greater than that of any single parameter obtained directly from pre-stack seismic inversion (such as... , or ).

[0024] To alleviate the physical and numerical limitations associated with pre-stack seismic inversion, a method utilizing fully connected neural networks is presented, where the inverted results serve as the basis and input for training and testing the neural network. Fully connected neural networks can model highly nonlinear and complex relationships in data without relying on physical assumptions and constraints. They can efficiently integrate data from various sources, including seismic data, well logging data, and stratigraphic layers. These networks adaptively fuse these multi-source data to achieve a comprehensive subsurface characterization, thereby improving the quality and reliability of predictions.

[0025] From a mathematical perspective, the quadratic loss function of a fully connected neural network can be written as: , (5); in, These are the weights of the neural network, and the predicted value y. (θ) is a function of these weights. Equation (5) is similar in form to Equation (2); the difference is that Equation (2) uses a forward simulation operator G based on physical assumptions. In contrast, Equation (5) uses customized layers and parameters of a neural network to learn the mapping relationships between different datasets and generate prediction data.

[0026] Figures 1 to 14 The apparatus, process, method, and results of this disclosure are illustrated. Specifically, Figures 1 to 4 Exemplary embodiments of methods, apparatus, and media for acquiring and storing seismic data are illustrated. The seismic data, after processing, can generate one or more high-resolution geological models for high-resolution imaging of complex subsurface structures in an exploration area, enabling lithological identification, fluid identification, and reservoir description. The exploration area can be subsurface structures beneath land or beneath the seabed.

[0027] Figure 1 This is a schematic diagram showing a top-down view of a survey area containing incident points of different seismic sources according to one embodiment. More specifically, Figure 1 A seismic survey area (exploration area) 101 is shown, which is a land-based area indicated by reference numeral 102. Reference numeral 102 indicates the top strata 102 of the land-based area. Those skilled in the art will recognize that seismic survey areas can generate detailed images of the local geology to determine the location and size of potential hydrocarbon (oil and gas) reservoirs, thereby establishing well sites 103. In these survey areas, seismic waves are reflected back from the subsurface rock strata when emitted from one or more seismic sources located at different incident points 104. A blast is an example of a seismic source generated by a seismic device. The seismic waves reflected back to the surface are captured by a seismic data recording sensor 105, transmitted from the seismic data recording sensor 105 via one or more data transmission systems (typically wireless), and stored for post-processing and analysis by a high-performance computing system. While this example shows the top strata 102 of a land-based area, it should be understood that this is merely an example, and the method and system can also be applied to survey areas on the ocean floor.

[0028] Figure 2 It is shown according to one embodiment Figure 1 This is a schematic diagram of a cross-sectional view of the seismic survey area 101, including the incident point of the seismic source, seismic data recording sensors (seismographs), well location, well casing, various propagation rays, and various incident angles. More specifically, in Figure 2In the figures, the cross-sectional view of the subsurface portion above the seismic survey area is indicated by reference numeral 201, and different types of strata are shown by reference numerals 102, 203, and 204. Although the seismic survey area in this example is based on land, it should be understood that this is only an example, and the method and system can also be applied to survey areas on the ocean floor. Figure 2 A common-center gather is shown, where seismic data is ordered according to surface geometry to simulate individual reflection points on Earth. Exploration seismic data can also be called traces, gathers, or image gathers. Figure 2 In this example, data from one or more shot points or blast points and detectors can be combined into a single image gather, or used individually depending on the type of analysis to be performed.

[0029] like Figure 2 As shown, one or more shot points or blast points represent seismic sources located on the Earth's surface, indicated by reference numeral 104, at various incident points or sites where one or more sources are activated. Seismic energy or seismic sources from multiple incident points 104 will be reflected from interfaces between different strata. These reflections will be captured by multiple seismic data recording sensors 105, each placed at a different offset distance 210 and at well site 103. Since all incident points 104 and all seismic data recording sensors 105 are placed at different offset distances 210, reconnaissance seismic data or sets (also referred to in the art as "gathers" or "image gathers") will be recorded at different incident angles 208. Incident points 104 generate downward propagating rays 205, which are captured at the surface by upward propagation reflections from the seismic data recording sensors 105. In this example, well site 103 connected to an existing drilled well 209 is shown, with multiple measurements taken along well 209 using techniques known in the art. The wellbore 209 is used to acquire logging data, which may include P-wave velocity, S-wave velocity, density, etc. (Not in...) Figure 2 Other sensors shown can be placed within the survey area to capture seismic data. Seismic data can be used to examine the dependence of amplitude, signal-to-noise ratio, time difference, frequency content, phase, and other seismic properties on the incident angle 208°, offset 210°, azimuth, and other geometric properties that are crucial for data processing and imaging in the seismic survey area.

[0030] Figure 3This is a schematic diagram illustrating a cross-sectional view of a seismic exploration area comprising a wellbore and logging tools according to one embodiment, wherein the logging tools include one or more acoustic wave generators and one or more logging data recording sensors. An acoustic wave generator is an example of a device that generates one or more acoustic waves (acoustic waves). An acoustic wave generator can be referred to as a sound source because it generates or produces one or more acoustic waves (acoustic waves), which are also called seismic waves. The one or more logging data recording sensors are examples of one or more seismic data recording sensors (seismic detectors or seismic data recorders), and are the same seismic data recording sensor as seismic data recording sensor 105. In embodiments of the invention, oil and / or gas production is suspended in order to generate seismic waves and record seismic data, including reflections of seismic waves as they move through one or more subsurface strata in the seismic exploration area.

[0031] Figure 3 An oil drilling system 300 on land 305 is shown, including a drilling rig 310. The drilling rig 310 supports the insertion of a logging tool 315 into a wellbore 320. The logging tool 315 includes one or more acoustic generators (sound sources) for generating one or more acoustic waves, which are transmitted to one or more formations to generate reflected or reflected waves in the formations. Although this example shows one or more formations in a land-based exploration area, it should be understood that this is only an example, and the method and system can also be applied to exploration areas on the surface or bottom of a body of water, such as the ocean. The logging tool 315 also includes one or more logging data recording sensors. As described above, the one or more logging data recording sensors receive and record logging data, which includes reflected data received by the one or more logging data recording sensors in response to acoustic waves emitted into the one or more formations by the one or more acoustic generators. The logging data is an example of seismic data. The logging data includes compressive wave velocity or P-wave velocity (Vp), S-wave velocity (Vs), and density as an indicator of porosity. This logging process used to record logging data can also be called sonic logging. The logging vehicle 325 can be coupled to the logging tool 315 to assist in its lowering and raising, and to communicate with the logging tool 315 to obtain logging data. Alternatively, in methods and systems targeting exploration areas on the surface or bottom of water bodies (such as oceans), other equipment or systems can be used to assist in the lowering and raising of the logging tool 315 and to communicate with the logging tool 315 to obtain logging data.

[0032] Figure 4 This is a schematic diagram illustrating a high-performance computing system according to one embodiment, which receives (typically wirelessly) data from... Figure 1 and Figure 2 Earthquake data recording sensor 105 and / or Figure 3Earthquake data recording sensors (in) Figure 3 Seismic data of seismic waves (also known as well logging data recording sensors). Figure 4 A high-performance computer system stores the seismic data in at least one memory for post-processing and analysis via a computer-implemented method and apparatus according to one or more embodiments. The analyzed or processed seismic data can be accessed via a personal computer system. More specifically, Figure 4 A data transmission system 400 is shown for wirelessly transmitting seismic data from a seismic data recording sensor to a system computer 405 coupled to one or more storage devices 410 for storing the seismic data in a database. The data transmission system can also wirelessly transmit seismic data directly from the seismic data recording sensor 405 to one or more storage devices 410 for storing the seismic data in a database, which can be accessed by the system computer 405. Wireless transmission is indicated by reference numeral 402. The one or more storage devices 410 may also store other computer software instructions or programs to implement the apparatus and methods described in the embodiments. The system computer 405 may be coupled (e.g., wirelessly) to one or more output storage devices 420, which can receive the results of computer-implemented processes or methods executed by the system computer 405. A personal computer system 425 may be coupled (e.g., wirelessly) to one or more output storage devices 420 and / or the system computer 405 so that a user can use the user interface of the personal computer system 425 to input information or obtain the results of computer-implemented processor methods executed by the system computer 405. One or more storage devices 420 may also store other computer software instructions or programs to implement the apparatus and methods described in the embodiments.

[0033] The user interface of the personal computer system 425 may include, for example, one or more of the following: a keyboard, mouse, joystick, button, switch, electronic pen or stylus, gesture recognition sensor (e.g., for recognizing gestures of the user including body part movements), input sound device or voice recognition sensor (e.g., microphone for receiving voice commands), output sound device (e.g., speaker), trackball, remote control, portable (e.g., cellular or smartphone) phone, tablet computer, pedal or foot switch, virtual reality device, etc. The user interface may also include a haptic device to provide haptic feedback to the user. The user interface may also include, for example, a touchscreen. Furthermore, the personal computer system 425 may be a desktop computer, laptop computer, tablet computer, mobile phone, or any other personal computing system.

[0034] The processes, functions, methods, and / or computer software instructions or programs in the apparatus and methods described in the embodiments herein may be recorded, stored, or fixed in one or more non-transitory computer-readable media (computer-readable storage (recording) media) including program instructions (computer-readable instructions) executable by a computer to cause one or more processors to execute (implement or implement) the program instructions. The media may also be included alone or in combination with program instructions, data files, data structures, etc. The media and program instructions may be specially designed and constructed, or may be well-known and usable by those skilled in the art of computer software. Examples of non-transitory computer-readable media include magnetic media, such as hard disks, floppy disks, and magnetic tapes; optical media, such as CD-ROMs and DVDs; magneto-optical media, such as optical discs; and hardware devices specifically configured for storing and executing program instructions, such as read-only memory (ROM), random access memory (RAM), flash memory, etc. Examples of program instructions include machine code (e.g., generated by a compiler) and files containing higher-level code that can be executed by a computer using an interpreter. The program instructions may be executed by one or more processors. The described hardware device can be configured as one or more software modules that are recorded, stored, or embedded in one or more non-transitory computer-readable media to perform the above-described operations and methods, and vice versa. Furthermore, the non-transitory computer-readable media can be distributed among computer systems connected via a network, and program instructions can be stored and executed in a distributed manner. In addition, the computer-readable media can also be embodied in at least one application-specific integrated circuit (ASIC) or field-programmable logic array (FPGA).

[0035] One or more databases may include a collection of data and supporting data structures that can be stored, for example, in one or more storage devices 410 and 420. For example, one or more storage devices 410 and 420 may be embodied as one or more non-transitory computer-readable storage media, such as non-volatile memory devices, such as read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), and flash memory, USB drives, volatile memory devices (e.g., random access memory (RAM)), hard disks, floppy disks, Blu-ray discs, or optical media (e.g., CD-ROMs and DVDs), or combinations thereof. However, examples of storage devices 410 and 420 are not limited to those described above, and the storage can be implemented by a variety of other devices and structures understood by those skilled in the art.

[0036] like Figure 5As shown, this workflow diagram illustrates the process of training and predicting target reservoir properties using a fully connected neural network. First, pre-stack seismic data, well logging data, and seismic horizons are inverted to generate target geological properties. Features are extracted from the target properties, and optionally from other image and data volumes or historical data. During the training phase, a variety of optional features can be utilized. These features include, but are not limited to, inverted elastic parameters and properties derived from these inversions, horizons, spatial locations, and two-way travel times. High-resolution well logging data, as well as optional enhanced synthetic well logging data, can be used as real data for training. The predicted geological properties are evaluated against those obtained at the well location. If, according to predetermined criteria, the predicted properties match those obtained at the well location, the target geological properties can be constructed using the predicted properties and optional additional rock and fluid properties data volumes. To produce high-fidelity results that can be used to determine the location and size of potential oil and gas reservoirs, model performance is typically predicted and evaluated within the seismic exploration area. The criteria for model evaluation may include the R² score at the well location, Pearson correlation coefficient, root mean square error (RMSE), and mean absolute error (MAE); as well as the vertical resolution, horizontal continuity, and feedback from human experts across the entire survey area.

[0037] For reference Figure 6 The diagram illustrates a schematic top-down view of actual data (left) and the entire survey area (right). The actual data, derived from well logging and synthetic well logging data, is often sparsely and irregularly distributed within the survey area. In contrast, seismic survey data points are regularly spaced and evenly distributed. This model is applicable to both onshore and offshore oil fields, and is suitable for both the exploration and development phases of oil fields.

[0038] Figure 7 This is a flowchart illustrating the process of predicting target reservoir properties using pre-stack seismic inversion. Seismic observation data is collected through seismic exploration, and well logging data is collected by performing logging operations at one or more well locations, such as... Figures 1 to 4 As shown, strata are specific geological events exhibiting distinct seismic response characteristics. Strata are obtained by interpreting geological interfaces within seismic data. After seismic data processing, key seismic events representing subsurface variations are interpreted using both manual and automated picking methods. Quality control checks, integration with other geological data, and iterative optimization contribute to the accurate identification of strata.

[0039] Seismic wavelets are derived to represent the waveforms characterizing seismic data. These wavelets are estimated using techniques such as deconvolution and least-squares matching based on synthetic seismic traces.

[0040] The initial model for seismic inversion can be derived from well logging data, rock physics models, regional geological knowledge, and prior research. Interpolation methods such as Kriging interpolation are employed to combine well logging data with seismic data to optimize the estimation of the initial model.

[0041] In Operation 701, i.e., forward modeling, seismic data, well logging data, seismic horizons, estimated wavelets, and an initial model are processed using one of the pre-stack seismic inversion methods to generate a geological model of the entire survey area, such as P-wave velocity, S-wave velocity, and density. Examples of pre-stack seismic inversion methods include elastic wave equation inversion, model-based inversion, stochastic inversion, and sparse spike inversion.

[0042] In operation 702, a synthetic seismogram is generated using linearized equations (such as the Zoeppritz equation) or nonlinear full-wave elastic equations based on P-wave velocity, S-wave velocity, and density. Operation 703, the model update, involves comparing the data residuals between the seismic data and the synthetic seismogram and iteratively adjusting the model until the data residuals are within acceptable limits or the number of iterations exceeds a preset maximum number of iterations.

[0043] In operation 705, the properties of the target reservoir (such as P-wave velocity, S-wave velocity, density, Poisson impedance, and the P-wave to S-wave velocity ratio) and fluid indicator factors are calculated by combining the geological model obtained in operation 704.

[0044] Figure 8 A fully connected neural network is illustrated, comprising an input layer 801, an intermediate layer 802, and an output layer 803. The input layer 801 includes inputs 811 to 815, which can be any combination of seismic features. The intermediate layer 802 is a hidden layer and may include one or more layers. The processors (neurons) in each intermediate layer can be regressive architectures. The output from the processors (neurons) in the previous layer serves as the input to the next layer. The final output 831 of the output layer 803 is one of the geological properties to be predicted.

[0045] Real-world data from well sites can be used when training neural networks. For example, to predict fluid factors, a set of input values ​​(seismic features) from the well site, including known fluid factors, is fed into the input layer. The output of the output layer is then compared to the known fluid factors at the well site. The matching degree between the output of each well site and the known fluid factors is improved by adjusting the weights and biases of each neuron, the number of layers, the number of neurons per layer, the activation function, the learning rate, the batch size, regularization techniques, optimization algorithms, and initialization methods. Training is complete when convergence is achieved, i.e., the incremental change in weights is less than a predetermined threshold.

[0046] Figure 9This is a flowchart illustrating the process of calibrating pre-stack seismic inversion results. According to this process, [the data will be calibrated based on...]. Figure 7 The method described above inputs a set of seismic features generated at the well site into a neural network. This neural network has been trained with a specific set of seismic features as input to predict the geological properties at the well site. The prediction results are then evaluated against real data at the well site to determine whether the predictions are satisfactory. Evaluation criteria may include R² score, Pearson correlation coefficient, root mean square error (RMSE), and mean absolute error (MAE).

[0047] If the prediction results are not ideal, you can choose to base your decision on... Figure 7 The method generates a second set of seismic features, which are then fed into a second neural network that has been trained using these second set of seismic features as input. This process iterates until the prediction results at the well location reach a satisfactory state.

[0048] Once the prediction results at the well location are satisfactory, the set of seismic features and the corresponding trained neural network can be selected for the survey area. Now, based on... Figure 7 The selected seismic feature set for the entire survey area generated by the method shown is input into a corresponding trained neural network to predict the geological properties of the entire survey area. The evaluation criteria for the predictions may include vertical resolution, horizontal continuity, and feedback from human experts.

[0049] Alternatively, in another embodiment of this disclosure, when the prediction results at the well site or the entire survey area are unsatisfactory, instead of selecting a new set of seismic features for the next iteration, the neural network can be retrained by selecting different layers and different parameters for the neural network.

[0050] Predictions from the calibration model can be applied to the entire survey area. Optionally, additional data volumes of rock and fluid properties can be generated, such as porosity, permeability, water saturation, sand layer thickness, resistivity, gamma rays, and shale volume.

[0051] Figures 10 to 14 A method is presented for calibrating the inverted fluid factor using a model with strongly interbedded thin layers and significant lateral heterogeneity at different scales, and comparing the results with a real model. The synthetic model is constructed to simulate the presence of heterogeneously distributed caves of various sizes and shapes in karst regions.

[0052] Figure 10A realistic elastic model was demonstrated. P-wave velocity, S-wave velocity, and density all exhibited similar anomaly patterns, but the background distribution differed slightly. Generally, the vertical resolution of these models varied between 3 and 15 meters. Synthetic seismic data was generated using one-dimensional elastic full-wave forward modeling to simulate the amplitude-variable-offset (AVO) effect. The wavelet used was the Ricker wavelet with a dominant frequency of 20 Hz. A quarter wavelength of approximately 57 meters and an eighth wavelength of approximately 26 meters both exceed the vertical resolution of realistic models. This synthetic test aims to challenge the detection limitations of current pre-stack seismic inversion methods in such complex geological environments.

[0053] Figure 11 The figure presents the P-wave velocity, S-wave velocity, density, and calculated fluid factor retrieved at seismic trace 220. This figure shows that while the presence and approximate location of subsurface reservoirs can be identified from the retrieval results, determining their size, actual reservoir properties, and precise location remains challenging due to the fact that the vertical scale of the actual model is far below the detection limits of seismic resolution.

[0054] Figure 12 It was also explained that the fluid factors calculated from the inversion results are insufficient to accurately depict the actual subsurface geology at different locations, especially in cases of significant scale differences and discontinuous distribution. For example, when there is only one cave in the real model (such as at seismic channel 75), the fluid factors can roughly capture its morphology. However, in scenarios with multiple caves (such as at seismic channels 175, 220, and 400), the accuracy of the fluid factors decreases, and identifying the morphology and number of caves becomes challenging.

[0055] Figure 13 The study revealed that the fluid factor results were significantly improved after employing a neural network-based calibration method. The real data used for training was distributed approximately uniformly at intervals of one-tenth of the model size. In all scenarios, regardless of the number and size of caves, the vertical position and true value of the fluid factor were successfully recovered.

[0056] Figure 14 The results of the entire model are presented, providing a comprehensive view of the achievements. The figure clearly demonstrates the method's ability to capture and characterize complex details of subsurface geological features. Notably, the method excels in capturing both vertical resolution and lateral heterogeneity in this complex and challenging geological environment. This method can serve as a powerful tool for improving the accuracy of reservoir characterization in complex geological environments, contributing to more informed decision-making in seismic exploration and reservoir management.

[0057] It will be apparent to those skilled in the art that various modifications can be made to the methods and systems of this disclosure without departing from the true scope of this disclosure as defined by the following claims.

Claims

1. A method for calibrating pre-stack seismic inversion using a fully connected neural network, the method comprising: (a) placing seismic data recording sensors at a plurality of locations within a survey area and placing a logging tool in a wellbore of the survey area, the logging tool comprising a sonic generator and one or more logging recording sensors; (b) performing a shot at an incidence point within the survey area to generate a seismic wave that travels through a subsurface formation; (c) observing the seismic wave using the seismic data recording sensors and recording seismic data from the seismic wave; (d) performing a logging operation using the logging tool and recording logging data using the one or more logging recording sensors; (e) transmitting the seismic data from the seismic data recording sensors and the logging data from the one or more logging recording sensors to a computer system comprising one or more memories and storing the seismic data and the logging data in the one or more memories; (f) processing the seismic data and the logging data to obtain an estimated wavelet and to identify one or more horizons; (g) obtaining one or more pre-stack inversion geologic models by combining the seismic data, the logging data, the one or more horizons, the estimated wavelet, and an initial model; (h) computing target reservoir properties by combining the plurality of inversion geologic models; (i) generating features from the target reservoir properties and / or from one or more additional properties and additional data types; (j) defining a neural network model, the neural network model comprising a plurality of layers, each layer comprising a plurality of parameters; (k) training the neural network model and evaluating the output of the neural network model at well locations and throughout the survey area; and (l) applying the predictions of the calibrated model to the entire survey area and optionally generating data volumes of additional rock and fluid properties.

2. The method of claim 1, in step (g), the one or more pre-stack inversion geologic models are obtained by a method selected from the group consisting of elastic wave equation inversion, model-based inversion, stochastic inversion, and sparse spike inversion.

3. The method of claim 1, wherein, The one or more inversion geologic models are selected from the group consisting of P-wave velocity model, S-wave velocity model, density model, and combinations thereof.

4. The method of claim 1, wherein, In step (i), the additional data types are selected from the group consisting of one or more of seismic data, logging data, horizons, estimated wavelet, and initial model within the same survey area.

5. The method of claim 1, wherein, In step (i), the additional properties are selected from the group consisting of one or more of main line locations, tie line locations, two-way travel times, depths, formations, topography, and tidal data.

6. The method of claim 1, wherein, Target reservoir properties are selected from one or more of P-wave velocity, S-wave velocity, density, Poisson's impedance, P-wave to S-wave velocity ratio and fluid-indicative factors.

7. The method of claim 1, wherein, The features extracted from the target geologic properties are selected from the group consisting of amplitude, amplitude envelope, amplitude weighted frequency, amplitude weighted phase, average frequency, apparent polarity, cosine instantaneous phase, seismic data derivative, instantaneous amplitude derivative, dominant frequency, instantaneous frequency, instantaneous phase, integrated absolute amplitude, and combinations thereof.

8. The method of claim 1, wherein, In step (k), the neural network is trained using a regression architecture, wherein the true data comprises a set of continuous measurements.

9. The method of claim 1, wherein, The parameters of the fully connected neural network include weights and biases, number of layers, number of neurons per layer, activation function, learning rate, batch size, regularization technique, optimization algorithm, and initialization method.

10. The method of claim 1, wherein, The criteria for the model evaluation are selected from one or more of R2 score, root mean square error (RMSE), mean absolute error (MAE), and vertical resolution and lateral continuity at the wellsite.

11. The method of claim 1, wherein, The optional additional data volumes of rock and fluid properties include porosity, permeability, water saturation, sand thickness, resistivity, gamma ray, and shale volume.

Images