Method for processing well sonic dispersion data

Abstract

Machine learning-enhanced interpretation of wellbore sonic data. The disclosure in question refers to the interpretation of wellbore sonic data using machine learning. In an example of a method according to aspects of the present disclosure, wellbore sonic data is received and machine learning is used to interpret the wellbore sonic data.

Description

1 / 32 METHOD FOR PROCESSING WELL SONIC DISPERSION DATA CROSS-REFERENCES

[001] This application claims the benefit of: a) U.S. Provisional Patent Application 62 / 861,756, “Machine learning enhanced borehole sonic data interpretation,” filed June 14, 2019; and b) U.S. Provisional Patent Application 62 / 902,009, “Neural network enhanced borehole data,” filed September 18, 2019, the full disclosures of which are incorporated herein by reference. BACKGROUND

[002] Several methods have been developed to interpret sonic borehole scattering data to invert formation elastic properties, such as the classical model-based inversion of the parametric shear inversion (SPI) method. See U.S. Patent 7,120,541, hereinafter referred to as Wang, which is incorporated herein by reference in its entirety. Such techniques are computationally expensive and assume that scattering follows a certain parametric mathematical function and, consequently, cannot accurately obtain the formation shear slowness when the formation is strongly anisotropic. Furthermore, this inversion method only works for the dipole scattering mode (i.e., flexural mode) and cannot be easily extended to other modes such as Stoneley, Quadrupole, and Pseudo-Rayleigh modes.

[003] Furthermore, for many applications, physics-based analytical, semi-analytical, or numerical solutions for direct modeling are time-consuming. This prevents their use for urgent tasks. In the interpretation of well sonic scattering, the labeling and extraction of different scattering modes depend on signal processing methods that completely ignore physics or are based on the linear scaling of simplified physical models (e.g., model Petition 870260065747, dated 03 / 07 / 2026, page 10 / 93 2 / 32 isotropic homogeneous), which fail to capture nonlinear effects, especially when the formation is strongly anisotropic. All existing approaches are not fully automated and rely on parameter adjustment for challenging cases. Furthermore, there is no automatic algorithm to analyze and identify the different anisotropy mechanism presented by the sonic signals measured in the well.

[004] Therefore, this application discloses various methods and techniques for enhancing and interpreting well data. Furthermore, those skilled in the art will appreciate that many of the methods and techniques disclosed herein can be used with ultrasonic scattering modes. The methods and techniques disclosed in this document can improve and enhance the understanding of scattering modes, uncertainty, anisotropy, and other well data interpretation issues. SUMMARY

[005] This summary is provided to present a selection of concepts that are described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used as an aid in limiting the scope of the claimed subject matter.

[006] In one embodiment, the inversion and interpretation of well sonic scatter data using machine learning-based approaches are described. The method includes generating a training dataset (i.e., lookup tables) from a mode search or any other analytical or numerical method and using the training dataset to train neural networks to approximate solutions. These neural network proxies can be used for sensitivity analysis and performing inversion on the measured sonic data. Parameters of Petition 870260065747, dated 03 / 07 / 2026, page 11 / 93 3 / 32 Relevant models can be estimated with associated uncertainties. Furthermore, various machine learning methods can also be developed based on the generated training dataset, which can be used to infer relevant model parameters with uncertainties from the field data directly.

[007] Other features and advantages of the disclosure in question will become more easily apparent from the following detailed description, when considered together with the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[008] The disclosure in question is further described in the detailed description that follows, with reference to the plurality of drawings noted by means of non-limiting examples of the disclosure in question, in which similar reference numbers represent similar parts along the various views of the drawings.

[009] Fig. 1 represents a fully connected conceptual network with a hidden layer, where x is the vector of model parameters and y is the vector of flexural dipole scattering.

[0010] Fig. 2 represents a comparison of dipole scattering curves generated by neural networks and a lookup table that is equivalent to a mode search.

[0011] Fig. 3 shows a comparison of dipole scattering curves generated by neural networks and a mode search.

[0012] Fig. 4 represents a distribution of calculated relative mean absolute differences (RMAD) between the neural network results and the training dataset.

[0013] Fig. 5 represents the synthetic scatter plot data (points) versus the reconstructed data (overlaid curves) from 100 inversion realizations. Petition 870260065747, dated 03 / 07 / 2026, page 12 / 93 4 / 32

[0014] Fig. 6 represents a violin plot of the inverted parameter distribution of 100 inversion realizations.

[0015] Fig. 6A is an enlargement of the central area of ​​Fig. 6.

[0016] Fig. 7 represents a violin plot of the inverted parameter distribution of 100 inversion realizations.

[0017] Fig. 8 represents a graph for analysis method one.

[0018] Fig. 9 represents a graph for analysis method one below.

[0019] Fig. 10 represents a graph for analysis method one below.

[0020] Fig. 11 represents a graph for analysis method two below, which uses a neural network mode and performs inversion on raw data.

[0021] Fig. 12 represents a graph for analysis method two below.

[0022] Fig. 13 represents a graph for analysis method two below.

[0023] Fig. 14 represents a graph for analysis method two below.

[0024] Fig. 15 represents a graph for analysis method two below.

[0025] Fig. 16 represents a graph for analysis method three below, which shows the combination of methods one and two below.

[0026] Figs. 17 to 22 illustrate various methods according to some modalities.

[0027] Fig. 23 illustrates an example computing system according to some embodiments. Petition 870260065747, dated 03 / 07 / 2026, page 13 / 93 5 / 32 DETAILED DESCRIPTION

[0028] The details shown in this document are by way of example and for the purpose of illustrating the examples of the disclosed matter only and are presented in order to provide what is considered to be the most useful and readily understood description of the principles and conceptual aspects of the present disclosure. In this respect, no attempt is made to show structural details in more detail than is necessary, the description taken with the drawings making it apparent to those skilled in the art how the various forms of the disclosure in question can be incorporated in practice. Furthermore, similar reference numbers and designations in the various drawings indicate similar elements.

[0029] Aspects of this disclosure relate to the interpretation of well sonic data. In the oil and gas industry, sonic devices are placed in a well to measure the travel time of sound, which can be excited by different types of sources. Several characteristics of the formation and well environment can be interpreted from the recorded sonic waves using a variety of methods. Interpreted formation properties, including compression and shear velocities, density, and elastic anisotropy, are widely used in the mechanical characterization and petrophysical evaluation of the formation.

[0030] Although different wave modes can be initiated from different sources and converted along their travel paths, some of these wave modes exhibit dispersive characteristics, i.e., the velocity varies with frequency, for example, Stoneley dispersion, pseudo-Rayleigh, flexural, and quadrupole. Traditional ways of estimating relevant model parameters from these measurements include methods based on Petition 870260065747, dated 03 / 07 / 2026, page 14 / 93 6 / 32 appearance, such as Slowness Time Coherence (STC) analysis and the model-based inversion method, minimizing the differences between measured and modeled dispersion data, such as parametric shear inversion (SPI). See Wang and Kimball et al. “Semblance processing of borehole acoustic array data”, Geophysics, 49, 3, 274-281, 1984.

[0031] Modeled dispersion data can be computed by analytical or numerical algorithms or interpolated from precomputed lookup tables. These measured modal dispersion data can be extracted from sonic waveforms measured by different algorithms, such as the Prony method. See Lang et al., “Estimating slowness dispersion from arrays of sonic logging waveforms”, Geophysics, 52, 4, 530-544, 1987. The model-based inversion process is generally expensive in terms of computational cost. The lookup table can be used for fast inversion, but is usually large in size and not accurate enough in some regions of the parameter space where higher nonlinearity is present.

[0032] In one embodiment, machine learning-based methods for improving the interpretation of well sonic data are described. Lookup tables are generated from a mode search or other analytical or numerical solvers and are used to train neural network models as accurate and efficient proxies that can be used for field data inversion. Machine learning-based methods are also used to automate the labeling and extraction of different modal dispersion data from recorded waveforms with appropriate preprocessing. In one embodiment, the methods include training machine learning models from field sonic data to infer relevant model parameters. Petition 870260065747, dated 03 / 07 / 2026, page 15 / 93 7 / 32 with associated uncertainties. In some embodiments, the methods according to this disclosure may also include training machine learning models from the generated synthetic data to directly infer relevant model parameters with associated uncertainties from field data.

[0033] Data preparation

[0034] In one embodiment, a mode-search algorithm is used to generate synthetic scattering curves for a large volume of parameter combinations. As a non-limiting example, for an ANNIE model [Schoenberg, M., Muir, F. and Sayers, CM, 1996, Introducing ANNIE: A simple tripameter anisotropic model for schists, J. Six. Expl., 5, 35-49] where three Thomson parameters [Thomsen, L., 1986, Weak elastic anisotropy, Geophysics, 51, 1954-1966, hereinafter Thomsen] are reduced to 1, we assume the parameters with assigned values ​​listed as below: • bhr: [3.50, 3.75, 4.00, 4.25, 4.50] • dtm: [190, 196, 202, 208, 214, 220] • rhom: [800, 1000, 1200, 1400] • dts: [80, 85, 90, 95, 100, 105, 110, 115, 120, 125, 130, 135, 140, 145, 150, 155, 160, 165, 170, 175, 180] • pr: [0,1, 0,2, 0,3, 0,4] • rhob: [1800, 2000, 2200, 2400, 2600] • faixa: [0, 0.05, 0.10, 0.15, 0.200, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60] • freq: [1.0, 1.2, 1.4, 1.6, 1.8, 2.0, 2.2,2.4, 2.6, 2.8, 3.0, 3.2, 3.4, 3.6, 3.8, 4.0, 4.2, 4.4, 4.6, 4.8 5.0, 5.2, 5.4, 5.6, 5.8, 6.0, 6.2, 6.4, 6.6, 6.8, 7.0, 7.2 7.4, 7.6, 7.8, 8.0

[0035] Here, bhr, dtm, rhom, dts, pr, rhob, gamma, and freq represent the well radius (inches), mud slowness (ps / ft), mud density (kg / m3), slowness of Petition 870260065747, dated 03 / 07 / 2026, p. 16 / 93 8 / 32 formation shear (ps / ft), Poisson's ratio, formation density (kg / m3), Thomsen anisotropic parameter γ (unitless) [See Thomsen] and frequency (kHz), respectively. The ranges of these parameters can be extended further if necessary. The ANNIE model can be extended to a three-parameter Thomsen model if required. Those skilled in the art will also appreciate that other models, parameter combinations, and related approaches can be practiced according to the methods disclosed in this document.

[0036] The mode search algorithm is executed for any combination of the above parameters within the defined ranges. The inputs (e.g., seven model parameters including bhr, dtm, rhom, dts, pr, rhob, and gamma) and corresponding outputs (different slowness values ​​corresponding to different frequencies within the sonic tool acquisition range) are stored in a data file.

[0037] Training and validation

[0038] According to some embodiments, the machine learning problem can then be formulated as follows: given a set of model parameters, find the corresponding slowness at a set of predefined frequencies. To solve this problem, different neural network structures can be used. In one embodiment, a standard fully connected neural network is considered, as conceptually represented in Figure 1, with ReLU activation functions. The number of layers and hidden neurons in the network can be optimized through the training process. To train, optimize, and test the solution, we divided the samples into three subsets for training, validation, and testing, respectively. We trained the network with an Adam optimizer. The objective function was an error. Petition 870260065747, dated 03 / 07 / 2026, page 17 / 93 9 / 32 standard mean square. The training process converges well without overfitting.

[0039] The trained neural network model can be validated against a mode search. Figure 2 shows a comparison of results for the case defined in the table node (i.e., a case existing in the training dataset, and in this case, the lookup table was generated from a mode search). Skilled practitioners will understand that the two curves represented in the graph are almost exactly the same because this case is present in the training dataset (bhr = 4 inches, dtm = 202 ft, rhom = 1200 kg / m3, dts = 90 ft, pr = 0.1, rhob = 1800 kg / m3, faixa = 0.5).

[0040] Figure 3 shows a comparison for the case outside the table node, i.e., a case that is not present in the training dataset (This case is not present in the training dataset (bhr = 4.26, dtm = 215.4, rhom = 1.016, dts = 138.8, pr = 0.34, rhob = 1.954.6, gamma = 0.17)). The trained neural network model generated perfect results compared to the mode search. We performed the comparison with all cases in the training dataset and plotted the histogram of the relative mean absolute differences (RMAD) between the results of the neural network models and the mode search. As shown in Figure 4, we obtained very good accuracy, except for a small volume of cases with RMAD close to 0.001. Further investigation shows that all these cases are located within the ranges of the intervals defined for the parameters.This problem can be overcome by extending the definition intervals, or by making the grids denser near the edges, or by increasing the weights of these boxes along the edges.

[0041] A benchmark shows that the trained neural network models achieved a significant increase in speed. Petition 870260065747, dated 03 / 07 / 2026, page 18 / 93 10 / 32 compared to mode search. A workstation test with a Titan Xp GPU demonstrated that neural network models are 30,000 times faster than the mode search method.

[0042] Inversion based on the trained neural network

[0043] The trained neural network model that has high accuracy and efficiency can be used as a direct solver for the inversion of sonic well scatter data.

[0044] First we tested on a synthetic case 1 with parameters bhr = 0.5, dtm = 0.5, rhom = 0.5, dts = 0.5, pr = 0.5, rhob = 0.5 and gama = 0.5. Note that here these values ​​have no unit after normalization with respect to the range of each parameter defined previously, i.e., bhr = 0.5 can be converted to bhr = 4 inches based on the defined range of 3.5 ~ 4.5. We performed 100 inversion realizations separately with each one starting from a random initial model.

[0045] As shown in Figure 5, all inversions converge to match the data well, but yield very different solutions, as shown in Figure 6, which represents a violin plot of the inverted parameter distribution of 100 inversion realizations, showing that, in this case, the median of each inverted parameter corresponds well to the true value. The true model is defined as bhr = 0.5, dtm = 0.5, rhom = 0.5, dts = 0.5, pr = 0.5, rhob = 0.5, and gamma = 0.5, which are normalized values ​​for the interval [0, 1] based on the ranges of these parameters. The central points in the respective graphs represent the true values. To elaborate and make this clearer, an enlargement of the area of ​​the violin plots with the points in Figure 6 is provided in Figure 6A to better show the median vs. true values. true values ​​(see Fig. 6A, internal points 602, 604, 606, 608, 610, 612, 614 vs. larger points 602A, 604A, 606A, 608A, 610A, 612A, Petition 870260065747, dated 03 / 07 / 2026, page 19 / 93 11 / 32 614A, respectively.) Shear slowness can be accurately inverted due to high sensitivity. Well radius and mud slowness show relatively less uncertainty compared to densities, Poisson's ratio, and Thomsen parameter γ.

[0046] We tested in synthetic case 2 with parameters bhr = 0.9, dtm = 0.8, rhom = 0.6, dts = 0.5, pr = 0.4, rhob = 0.3 and gamma = 0.2. All 100 inversion realizations converge to match the data well, but obtain different solutions, as shown in Figure 7. As in Figure 6, the central points in the respective graphs represent the true values. In this case, we still obtain accurate shear slowness, good well radius, but the median of other inverted parameters deviates from the true values. In terms of uncertainties of the inverted parameters, we observe a similar trend with case 1.

[0047] From the inversion study, we confirm that this inverse problem is highly non-unique. The uncertainties of the estimated parameters can be quantified.

[0048] By applying the neural network model trained for field data inversion, different modal scatter data are labeled and extracted from Prony's method. A machine learning-based method can also be developed for this application to make the process automatic and efficient.

[0049] In one embodiment, a machine learning framework is described to improve the interpretation of sonic well data. The machine learning technique can be used to train accurate and efficient neural network models to replace computationally expensive solvers such as mode search and finite difference method. The trained models can be used as proxies for direct solvers for inversion. Petition 870260065747, dated 03 / 07 / 2026, page 20 / 93 12 / 32

[0050] The example used here is a seven-parameter ANNIE model in vertical wells, but the same concept and structure can be applied to other scenarios, such as the nine-parameter TIV model defined by bhr, dtm, rhom, dts, pr, rhob and 3 Thomsen parameters epsilon, gamma and delta [See Thomsen], or cases that have more parameters, such as orthorhombic anisotropy and in deviated, high-angle or horizontal wells where well deviation, well azimuth and transmitter firing direction are extra input parameters, etc.

[0051] In addition to training a neural network as a routing proxy, it is also possible to train advanced neural networks to infer model parameters with uncertainties from sonic field data.

[0052] In some modalities, a predefined parametric function, for example, Bessel's function, can be used to approximate the scatter data, i.e., dimensionality reduction, to train neural networks to map from these function parameters to model parameters.

[0053] In some embodiments, the variational automatic encoder and other neural networks are combined to map the scatter data to the model parameters.

[0054] In some modalities, neural networks are trained to automatically label and extract different modal scatter data from the results of the Prony method.

[0055] Another aspect of this disclosure concerns the labeling, extraction, and interpretation of well sonic scattering data. In the oil and gas industry, well sonic measurements have been widely used for petrophysical formation evaluation and geomechanical characterization. Well sonic waves can be excited by different sources, such as monopoles, dipoles, and quadrupoles. Guided waves exhibit different modes, such as Stoneley, flexural, and screw modes, which present Petition 870260065747, dated 03 / 07 / 2026, page 21 / 93 13 / 32 Dispersive characteristics. Dispersion analysis often produces sounding-hole modes in the form of disconnected, scattered points in the frequency-slowness plane. Labeling, extraction, and interpretation of the different modes from these scattered points are essential steps in sonic interpretation. Traditional methods for dealing with this problem rely on signal processing [Aeron et al., “Automatic dispersion extraction using continuous wavelet transform”, 2008 IEEE International Conference on Acoustics, Speech and Signal Processing, Las Vegas, NV, 2008, pp. 2405-2408] or are based on a simplified model [See Wang and Endo et al., “Shear Parametric Inversion SPI-3 and VKSPI”, Sep 28, 2005 at MSIP Seminar]. Signal processing-based methods ignore physics; therefore, the solution is sometimes wild when measurements are noisy or incomplete.Some methods are based on the linear scaling of a simplified model, for example, a homogeneous isotropic model. This method works well in many cases, but often fails to account for nonlinear effects caused by complicated sonic physics, such as formation anisotropy and formation alteration near the wellbore, within a given interval. More complex and accurate physics-based models have not been used for dispersion and extraction labeling, mainly due to the intractable computational resource requirement.

[0056] In one embodiment, a method is described for improving the labeling, extraction, and interpretation of well sonic data using neural networks. The results of synthetic direct modeling are pre-calculated for different model scenarios and stored. This synthetic data is then used to train neural network models as an efficient approximation of the real physical model that can be used for effective labeling, extraction, and interpretation of field data. By combining different models, Petition 870260065747, dated 03 / 07 / 2026, page 22 / 93 14 / 32 we can efficiently identify a different anisotropy mechanism. The algorithms and workflows do not depend on user input, are faster than conventional methods, and allow for real-time well interpretation.

[0057] Generation of proxy models to approximate expensive direct modeling

[0058] First, analytical, semi-analytical, or numerical solutions are employed to generate synthetic dispersion data for combinations of parameters selected within the physical range of the parameter space. As an example, for borehole sonic dipole dispersion caused by intrinsic TI anisotropy, an ANNIE model [Schoenberg et al., “Introducing ANNIE: A simple three-parameter anisotropic velocity model for shales”, J. Seis. Expl., 5, 35-49, 1996] is used, and we assume that the parameters are varying within a given range, as shown below: • bhr: [bhr lb, bhr ub] • dtm: [dtm lb, dtm ub] • rhom: [rhom_lb, rhom_ub] • dts: [dts_lb, dts_ub] • pr: [pr_lb, pr_ub] • rhob: [rhob_lb, rhob_ub] • gamma: [gamma_lb, gamma_ub] • freq: [freq_lb, freq_ub]

[0059] Here, bhr, dtm, rhom, dts, pr, rhob, gamma, and freq represent the well radius (inches), mud slowness (ps / ft), mud density (kg / m3), formation shear slowness (ps / ft), Poisson's ratio, formation density (kg / m3), Thomsen anisotropic parameter γ (unitless) [Thomsen, L., 1986, Weak elastic anisotropy, Geophysics, 51, 1954-1966], and frequency (kHz), respectively. The symbol “lb” indicates a lower limit. Petition 870260065747, dated 03 / 07 / 2026, p. 23 / 93 15 / 32 of the corresponding parameter and “ub” is for the upper limit. The sampling density within the physical range depends on the nonlinearity of the problem along each individual direction.

[0060] A mode search algorithm is executed to select combinations of the above parameters within the defined ranges; in some embodiments, the mode search algorithm is executed for each combination of the above parameters within the defined ranges. The inputs (e.g., seven model parameters including bhr, dtm, rhom, dts, pr, rhob, and gamma) and corresponding outputs (different slowness values ​​corresponding to different frequencies within the sonic tool acquisition range) are stored in a data file.

[0061] A neural network can be trained using the inputs and outputs in synthetic data. The machine learning problem can be expressed as follows: given a set of model parameters, find the corresponding slowness at a set of predefined frequencies. To solve this problem, although different neural network structures can be explored, here we consider a standard fully connected neural network, as conceptually represented in Fig. 1, with ReLU activation functions. The number of layers and hidden neurons in the network can be optimized through the training process. An Adam optimizer is used to train the network. The objective function was a standard mean squared error. The training process converges well without overfitting. The trained neural network model can be validated against mode search. A benchmark shows that the trained neural network achieved a four-order speed increase compared to mode search.

[0062] Labeling, extraction and interpretation using the trained neural network Petition 870260065747, dated 03 / 07 / 2026, page 24 / 93 16 / 32

[0063] In the interpretation of sonic data, an important method is to convert the waveforms received by the receiver array into the frequency domain and process the slowness and dispersion estimation. Classical approaches, for example, the Prony method, are used for this processing. The resulting scatter plot includes different modes and is usually noisy. An essential step is to label and extract different dispersion modes from the scatter plot and then invert the extracted mode for the relevant parameters.

[0064] Once we trained an efficient neural network model for modeling the dipole scattering mode, we developed a plurality of analysis methods that utilize the trained neural network model for automatic labeling, extraction, and inversion of dipole scattering data.

[0065] Analysis method 1: clustering with neural network model fitting

[0066] In this analysis method, we perform a clustering operation for the normalized scatter plot, i.e., the slowness and frequency are normalized to an interval of 0 to 1, defining a minimum and maximum slowness and a minimum and maximum frequency. The discrete slowness frequency points are grouped into a limited number of classes, see Fig. 8. In this work, we use the DBSCAN method [Ester et al., “A density-based algorithm for discovering clusters in large spatial databases with noise. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), AAAI Press. pp. 226-231, 1996] but other methods are considered.

[0067] In some modes, for each class, an inversion is performed using the neural network model to combine the data points belonging to that class (see Petition 870260065747, dated 03 / 07 / 2026, page 25 / 93 17 / 32 Fig. 9 shows each class inverted using a neural network model. The solid points are reconstructed scattering modes obtained from the inversion. Class 0 is well matched, as shown by the good fit between the raw data ('x' and circles) and the modeled data (solid points). Other classes are not fitted because they are outside a predefined physical dipole scattering region. Classes that are far from the predefined physical dipole scattering region will show large mismatch [e.g., the trend of class 2 in Fig. 9 is obviously different from any dipole scattering mode, therefore the mismatch is very large (e.g., greater than 1E5) if we use the dipole scattering model to fit the data points belonging to this class], while classes located in the region will be well matched showing a small residual error (e.g., class 0 in Fig. 9, with residual error / misfit less than 0.01).

[0068] In some embodiments, an inversion is performed for only one or more classes using the neural network model to combine the data points belonging to the respective classes into one or more classes.

[0069] Classes with large misfit errors are removed from the data. Classes associated with small misfit errors are merged and inverted again as a whole. In most cases, this will lead to a well-reconstructed dipole scattering curve, and the data points can be filtered based on the distance from this curve (see Fig. 10, where all classes with small misfit are gathered and then inverted with additional iterations. All points close to the reconstructed dipole scattering curve are marked and extracted. It was also inverted and interpreted simultaneously during the process).

[0070] Analysis method 2: model-based progressive fitting and filtering Petition 870260065747, dated 03 / 07 / 2026, page 26 / 93 18 / 32

[0071] In this analysis method, we perform inversion on raw data in the scatter plot to reconstruct a best-fit dipole scatter curve. Obviously, this reconstructed curve will be offset by out-of-the-curve points and data belonging to other modes. However, it should be reasonably not too far from the actual dipole scatter curve. Therefore, we can use a distance-based filter [e.g., filter | iteration0 = 100, filter | iteration1 = 70, filter | iteration2 = 50, filter | iteration3 = 30, filter | iteration4 = 10, filter | iteration5 = 10, ...] to remove out-of-the-curve points and some data points that are too far from the dipole scatter mode. After filtering, we repeat the inversion for the remaining data and expect to obtain a better reconstructed dipole scatter curve, and then we can apply the distance-based filter again to further clean up the data.We repeat this process and choose an adaptive filter that decreases the distance with increasing iterations. The iteration will be interrupted if we reach the predefined maximum iterations (e.g., 3, 4, 5, 6, 7, 8, 9, 10 or more when appropriate) or the residual mismatch error is less than a predefined threshold (e.g., 5E-5). Different adaptive filters can be designed, for example, they can be predefined or automatically calculated based on the variation of the data (e.g., filter = F(std(Y)), where Y are the scatter points and F is a projected function). This progressive adaptation and filtering process is demonstrated in Figs. 11, 12, 13, 14 and 15.

[0072] In Fig. 11, which uses a neural network mode and performs inversion on raw data, a dipole scattering curve (black dots) will be reconstructed. Filtering is performed with a relatively large distance, for example, 100 us / ft. Petition 870260065747, dated 03 / 07 / 2026, page 27 / 93 19 / 32

[0073] In Fig. 12, an inversion is performed again for the remaining data (after the filtering of the previous step) to obtain a new reconstructed curve. The filtering is performed with a reduced distance, for example, 70 us / ft.

[0074] In Fig. 13, an inversion is performed again for the remaining data (after the filtering of the previous step) to obtain a new reconstructed curve. The filtering is performed with reduced distance.

[0075] In Fig. 14, the inversion is continuously performed again for the remaining data (after the filtering of the previous step) to obtain a new reconstructed curve. The filtering is performed with reduced distance.

[0076] In Fig. 15, the last step is to retrieve all the raw data and filter the raw data based on the reconstructed dipole scattering curve. The inversion can be performed again for the remaining data or the inversion results from the last step can be used.

[0077] Analysis method 3: combination of method 1 and method 2

[0078] This analysis method is a combination of analysis methods 1 and 2, i.e., a clustering is performed first and some classes with large residual misfit error (e.g., greater than 0.01, although under varying circumstances other error thresholds may be used, e.g., greater than 0.05, greater than 0.005, etc.) are removed from the data. For the remaining data, we apply method 2 for progressive fitting and filtering. In the last step, we take all the data points back and filter them based on a predefined distance from the final reconstructed dipole scattering curve. The result is shown in Fig. 16.

[0079] To increase the robustness of the algorithm, in the clustering step, we can divide the data into different frequency bands and perform clustering for each band. Petition 870260065747, dated 03 / 07 / 2026, page 28 / 93 20 / 32 separately. After removing the classes with large residual errors, we reassembled all remaining classes and proceeded with progressive adjustment and filtering.

[0080] In the three analysis methods above, labeling, extraction, and inversion are achieved in one procedure, i.e., we successfully label the dipole scattering data from the raw data and perform inversion on the identified data to obtain the model parameters.

[0081] Automatically identify the anisotropy mechanism

[0082] In the above modalities, a workflow is described for labeling, extracting, and inverting dipole scattering data using a neural network that honors the intrinsic anisotropy effect of TI. The same procedure can be applied to other scattering modes from the same source or from different sources. The same procedure can also be applied using other neural network models that respect other mechanisms, such as different types of anisotropy (e.g., stress-induced anisotropy, fracture-induced anisotropy). Once we have established neural network models of different physical mechanisms for each scattering mode, it is efficient to apply the models to a given dataset. Based on the data fitting, we can identify which physical mechanism is dominating the sonic data.

[0083] In one embodiment, a workflow is developed with different methods to enable fully automated labeling, extraction, and inversion of well sonic dipole scattering data, which can be extended to other modes. After developing models for all different scattering modes, we can easily and efficiently apply the workflow with all models to the measured data. Based on the residual data fitting error, we can estimate and identify different Petition 870260065747, dated 03 / 07 / 2026, page 29 / 93 21 / 32 physical phenomena presented by the data, including different mechanisms of anisotropy.

[0084] In one embodiment, a general workflow for using trained neural network proxy models to improve the interpretation of well sonic scatter data is described.

[0085] In one embodiment, a plurality of methods can be used for automated labeling, extraction and inversion of scattering modes.

[0086] In some forms, the clustering methods include an option to split the data into different frequency bands and cluster them separately.

[0087] The same workflow can be extended to other sonic and ultrasonic wellbore dispersion modes, for example, quadrupole, Stoneley, pseudo-Rayleigh, etc., in any LWD or wire rope logging, in any open hole or cased hole logging. Those skilled in the art will appreciate that the non-limiting methods 1700, 1800, 1900, 2000, 2100 and 2200 disclosed in this document and illustrated in Figs. 17-22 can be applied in any of these dispersion modes.

[0088] In one embodiment, for specific field data, we can perform this workflow using all the different models, so we can automatically label, extract, and invert all scatter modes. Those versed in the art will appreciate that the non-limiting 1700, 1800, 1900, 2000, 2100, and 2200 methods disclosed in this document and illustrated in Figs. 17-22 can be applied to any of these types of field data.

[0089] In some embodiments, the analytical methods disclosed and other techniques herein can be used to characterize any anisotropy mechanism. Based on data fitting, we can estimate and identify the potential mechanism that contributes to the anisotropy. Petition 870260065747, dated 03 / 07 / 2026, page 30 / 93 22 / 32 presented in any well data. Those skilled in the art will appreciate that the non-limiting methods 1700, 1800, 1900, 2000, 2100 and 2200 disclosed in this document and illustrated in Figs. 17-22 can be used to characterize any anisotropy mechanism.

[0090] Attention is now directed to Fig. 17 which discloses a computer-implemented method 1700 for interpreting well sonic data. In some embodiments of the 1700 method, the well sonic scatter data are in the ultrasound range.

[0091] Method 1700 involves receiving 1702, using one or more processors, sonic well data. Machine learning is used 1704 to interpret the sonic well data.

[0092] In some embodiments, the 1700 method includes generating 1706 a large volume of synthetic well sonic scatter data for defined parameter combinations within a certain range using analytical or numerical solvers. For example, in some embodiments, millions of scatter curves can be calculated to generate the volume of synthetic well sonic scatter data.

[0093] In some embodiments, the 1700 method includes using the synthetic data generated as input to neural networks and training the neural networks to predict borehole sonic scatter data from input model parameters.

[0094] In some modalities, the 1700 method includes performing 1710 global and local sensitivity analysis using the trained neural networks.

[0095] In some embodiments, the 1700 method includes inferring 1712 model parameters from synthetic well sonic scatter data.

[0096] In some embodiments, the 1700 method includes using 1714 of the trained neural networks as proxies Petition 870260065747, dated 03 / 07 / 2026, page 31 / 93 23 / 32 direct and inversion for the model parameters with uncertainties from the measured well sonic scatter data. In other embodiments, the inversion is simultaneous or sequential based on prior information or sensitivities (see Fig. 17, 1716). In other embodiments, the measured well sonic scatter data are inverted separately, sequentially, or simultaneously (see Fig. 17, 1718). In other embodiments, the measured well sonic is from a single depth or multiple depths (see Fig. 17, 1720).

[0097] Attention is now directed to Fig. 18 which discloses a computer-implemented method 1800 for interpreting well sonic data.

[0098] Method 1800 includes training 1802 one or more neural networks to infer model parameters and uncertainties directly from directly measured dispersion data. Method 1800 also includes determining 1804 one or more model parameters directly from at least one trained neural network. Finally, method 1800 also includes determining 1806 one or more uncertainties directly from at least one trained neural network.

[0099] Attention is now directed to Fig. 19 which discloses a 1900 method for processing borehole sonic scatter data.

[00100] The 1900 method includes generating 1902 synthetic direct modeling results corresponding to sonic scatter data for a plurality of model scenarios.

[00101] The 1900 method also includes training 1904 one or more neural network models based on the generated sonic scattering data to generate one or more physical sonic well scattering models.

[00102] The 1900 method also includes receiving 1906 the sonic scatter data from the well. In some embodiments of Petition 870260065747, dated 03 / 07 / 2026, page 32 / 93 24 / 32 method 1900, the sonic scattering data from the well are in the ultrasound range.

[00103] The 1900 method also includes interpreting 1908 well sonic scatter data based, at least in part, on inversion of well sonic scatter data using one or more trained neural network models.

[00104] In some embodiments, the interpretation of well sonic scatter data includes estimating the slowness, scatter, and / or other model parameters from well sonic scatter data (see Fig. 19, 1910).

[00105] In other modalities, the estimated dispersion is categorized in a plurality of ways (see Fig. 19, 1912). In other modalities, the relevant dispersion data are extracted from one or more of the plurality of ways (see Fig. 19, 1914).

[00106] In some embodiments, the interpretation of well sonic scattering data further involves combining a plurality of physical well sonic scattering models; and identifying well anisotropy parameters of the plurality of physical models (see Fig. 19, 1916).

[00107] Attention is now directed to Fig. 20 which discloses a 2000 method for processing borehole sonic scatter data.

[00108] The 2000 method includes defining 2002 frequency and slowness intervals. In some embodiments, the frequency and slowness intervals are normalized to a range of 0 to 1 (see Fig. 20, 2004).

[00109] Method 2000 also includes receiving 2006 sonic scatter data from the well. In some embodiments of Method 2000, the sonic scatter data from the well are in the ultrasound range.

[00110] The 2000 method also includes categorizing the 2008 sonic dispersion data from the well into a plurality of discrete slowness-frequency points based on, by Petition 870260065747, dated 03 / 07 / 2026, page 33 / 93 25 / 32 less in part, within the defined frequency and slowness intervals.

[00111] The 2000 method also includes classifying 2010 the plurality of discrete slowness-frequency points into a plurality of clustering classes.

[00112] The 2000 method also includes that, for each of the plurality of clustering classes, an inversion is performed in 2012 using a neural network model to combine the discrete slowness-frequency points belonging to the respective clustering classes.

[00113] When inverting the respective grouping classes in one or more of the plurality of grouping classes exhibits mismatch, method 2000 also includes removing the respective grouping classes to retain a subset of grouping classes.

[00114] In some embodiments, the mismatch is based, at least in part, on the distance of the respective discrete slowness-frequency points to a region of related dipole scattering physics (see Fig. 20, 2016).

[00115] The 2000 method also includes combining 2018 with the subset of clustering classes.

[00116] The 2000 method also includes inverting the 2020 combined clustering classes using the neural network model to combine the discrete combined slowness-frequency points.

[00117] Attention is now directed to Fig. 21 which discloses a 2100 method for processing borehole sonic scatter data.

[00118] Method 2100 involves receiving 2102 sonic scatter data from the well. In some embodiments of method 2100, the sonic scatter data from the well are in the ultrasound range. Petition 870260065747, dated 03 / 07 / 2026, page 34 / 93 26 / 32

[00119] Method 2100 also includes categorizing 2104 the sonic well scatter data into a plurality of discrete slowness-frequency points.

[00120] Method 2100 also includes inverting 2106 the categorized data to generate a dipole scattering curve and a refined dataset;

[00121] Method 2100 also includes performing 2108 iterative actions until a stopping criterion is met; specifically, applying an adaptive filter to the refined dataset to identify and remove points in the refined dataset that are outliers of the dipole scattering curve and inverting the refined dataset.

[00122] In some modes, the stopping criterion is a predefined number of iterations (see Fig. 21, 2110).

[00123] In some modes, the stopping criterion is based on a residual misfit error being less than a predefined threshold (see Fig. 21, 2112).

[00124] In some embodiments, the adaptive filter is a distance-based filter that decreases the distance in each iteration (see Fig. 21, 2114).

[00125] In some modes, the adaptive filter is calculated automatically based on variations in the categorized data or in the refined dataset (see Fig. 21, 2116).

[00126] Attention is now directed to Fig. 21 which discloses a 2100 method for processing borehole sonic scatter data.

[00127] The 2200 method includes defining 2202 frequency and slowness intervals.

[00128] Method 2200 also includes receiving 2206 sonic scatter data from the well. In some embodiments of method 2200, the sonic scatter data from the well are in the ultrasound range. Petition 870260065747, dated 03 / 07 / 2026, page 35 / 93 27 / 32

[00129] Method 2200 also includes categorizing 2208 the sonic scatter data from the well into a plurality of discrete slowness-frequency points based, at least in part, on the defined frequency and slowness intervals.

[00130] The 2200 method also includes classifying 2210 the plurality of discrete slowness-frequency points into a plurality of clustering classes.

[00131] Method 2200 also includes identifying 2212 each respective clustering class in the plurality of clustering classes that exhibit large residual mismatch error and removing the respective clustering class.

[00132] The 2200 method also includes combining 2214 the plurality of discrete slowness-frequency points in the remaining clustering classes in a combined dataset.

[00133] Method 2200 also includes inverting 2216 the combined dataset to generate a dipole scattering curve and a refined dataset.

[00134] Method 2200 also includes performing 2218 iterative actions until a stopping criterion is met; specifically, applying an adaptive filter to the refined dataset to identify and remove points in the refined dataset that are outliers of the dipole scattering curve and inverting the refined dataset.

[00135] The 2200 method also includes filtering the refined dataset based, at least in part, on a predefined distance from the dipole dispersion curve.

[00136] Attention is now directed to Figure 23. In some embodiments, the methods of the present disclosure can be performed by a computing system. Figure 23 illustrates an example of such a computing system 2300, according to some embodiments. The computing system 2300 may include a computer or computer system 2301A, Petition 870260065747, dated 03 / 07 / 2026, page 36 / 93 28 / 32 which may be an individual 2301A computer system or an arrangement of distributed computer systems. The 2301A computer system includes one or more 2302 analysis modules that are configured to perform various tasks according to some embodiments, such as one or more methods disclosed herein. In some embodiments, the 2302 analysis modules include machine learning, artificial intelligence, and / or neural network logic for supervised, unsupervised, and other forms of machine learning as those skilled in the art will appreciate.

[00137] To perform these various tasks, the analysis module 2302 runs independently, or in coordination with, one or more processors 2304, which are connected to one or more storage media 2306.The processor(s) 2304 is (or are) also connected to a network interface 2308 to enable the computer system 2301A to communicate via a data network 2310 with one or more additional computer systems and / or computing systems, such as 2301B, 2301C, and / or 2301D (note that computer systems 2301B, 2301C, and / or 2301D may or may not share the same architecture as computer system 2301A and may be located in different physical locations, for example, computer systems 2301A and 2301B may be located in one processing facility while communicating with one or more computer systems, such as 2301C and / or 2301D, which are located in one or more data centers and / or located in various countries on different continents).Those versed in the technical aspects will appreciate that the 2310 network can include cloud processing and / or storage infrastructure, so some or all 2300 computing systems may have cloud processing enabled to support Software as a Service offerings. Petition 870260065747, dated 03 / 07 / 2026, page 37 / 93 29 / 32

[00138] A processor may include a microprocessor, microcontroller, processor module or subsystem, programmable integrated circuit, programmable gate array, or other control or computing device.

[00139] Storage medium 2306 can be implemented as one or more computer-readable or machine-readable storage media. Note that, although in the example embodiment of Figure 23 the storage medium 2306 is represented as within the computer system 2301A, in some embodiments the storage medium 2306 can be distributed within and / or across multiple internal and / or external compartments of the computer system 2301A and / or additional computer systems.Storage media 2306 may include one or more different forms of memory, including semiconductor memory devices such as dynamic or static random access memories (DRAMs or SRAMs), erasable and programmable read-only memories (EPROMs), electrically erasable and programmable read-only memories (EEPROMs) and flash memories, magnetic disks such as hard disks, floppy disks and removable disks, other magnetic media including tapes, optical media such as compact discs (CDs) or digital video discs (DVDs), Blu-ray® discs or other types of optical storage, or other types of storage devices. Note that the instructions discussed above may be provided on a computer-readable or machine-readable storage medium, or may be provided on multiple computer-readable or machine-readable storage media distributed across a large system possibly having multiple nodes.This computer-readable or machine-readable storage medium(s) is / are considered part of an article (or article of manufacture). An article or article of manufacture may be... Petition 870260065747, dated 03 / 07 / 2026, page 38 / 93 30 / 32 refers to any single manufactured component or to multiple components. The storage medium or media may be located either on the machine running the machine-readable instructions or located at a remote location from which machine-readable instructions can be downloaded over a network for execution.

[00140] In some embodiments, the computing system 2300 contains one or more well data analysis modules 2309. In the example of computing system 2300, computer system 2301A includes well data analysis module 2309. In some embodiments, a single well data analysis module may be used to perform some aspects of one or more embodiments of the methods disclosed herein. In other embodiments, a plurality of well data analysis modules may be used to perform some aspects of the methods in this document.

[00141] It should be appreciated that the 2300 computing system is merely an example of a computing system and that the 2300 computing system may have more or fewer components than shown, may combine additional components not represented in the example embodiment of Figure 23, and / or the 2300 computing system may have a different configuration or arrangement of the components represented in Figure 23. The various components shown in Figure 23 may be implemented in hardware, software, or a combination of both hardware and software, including one or more processing and / or application-specific integrated circuits.

[00142] In addition, the steps in the processing methods described in this document can be implemented by running one or more functional modules in information processing devices, such as general-purpose processors or application-specific chips, such as ASICs, FPGAs, PLDs, or other appropriate devices. These modules, combinations Petition 870260065747, dated 03 / 07 / 2026, page 39 / 93 31 / 32 of these modules and / or their combination with general hardware are included in the scope of this disclosure.

[00143] Computational interpretations, models, and / or other interpretation aids can be refined in an iterative manner; this concept is applicable to the methods discussed herein. This may include the use of feedback loops running on an algorithmic basis, such as in a computing device (e.g., 2300 computing system, Figure 23) and / or through manual control by a user who can make determinations as to whether a given step, action, model, template, or set of curves has become sufficiently accurate for the evaluation of the three-dimensional subsurface geological formation under consideration.

[00144] In addition, methods 1700, 1800, 1900, 2000, 2100 and 2200 are shown as including several blocks of 1.702m computer-readable storage media (CRM). 1,704m, 1,706m, 1,708m, 1,710m, 1,712m, 1,714m, 1,716m, 1,718m, 1,720m, 1,802m, 1,804m, 1,806m, 1,902m, 1,904m, 1,906m, 2,004m, 2,018m, 2,112m, 2,206m, 1,908m, 2,006m, 2,020m, 2,114m, 2,208m, 1,910m, 2,008m, 2,102m, 2,116m, 2,210m, 1,912m, 2,010m, 2,104m, 2,118m, 2,212m 1,914m, 2,012m, 2,106m, 2,120m, 2,214m, 1,916m, 2,014m, 2,108m, 2,202m 2,216m, 2,002m, 2,016m, 2,110m 2,204m, 2.218me 2.220m which may include processor-executable instructions that can instruct a computing system, such as the example computing system 2300 provided in Figure 23, to perform one or more of the actions described in relation to their respective methods.

[00145] Although only a few examples have been described in detail above, those skilled in the art will readily appreciate that many modifications to the examples are possible without materially departing from this disclosure. Consequently, all such modifications are intended to be included within the scope of this disclosure. Petition 870260065747, dated 03 / 07 / 2026, page 40 / 93 32 / 32 as defined in the following claims. In the claims, the terms "means plus function" are intended to cover the structures described herein as performing the stated function and not merely structural equivalents, but also equivalent structures. Thus, while a nail and a screw may not be structural equivalents in that a nail employs a cylindrical surface to fasten pieces of wood together, whereas a screw employs a helical surface, in the context of fastening pieces of wood together, a nail and a screw may be equivalent structures. It is the express intention of the applicant not to invoke 35 USC § 112, paragraph 6, for any limitations of any of the claims herein, except for those in which the claim expressly uses the words "means to" along with an associated function. Petition 870260065747, dated 03 / 07 / 2026, p. 41 / 93

Claims

1 / 2 CLAIMS 1. Method for processing well sonic scatter data, characterized in that it comprises: defining (2002) a frequency band; defining (2002) a slowness band; receiving (2102) well sonic scatter data; categorizing (2008, 2208) the well sonic scatter data into a plurality of discrete slowness-frequency points based, at least in part, on the defined frequency and slowness bands; classifying (2010, 2210) the plurality of discrete slowness-frequency points into a plurality of clustering classes; identifying (2212) each respective clustering class in the plurality of clustering classes that exhibit large residual mismatch error and removing the respective clustering class; combining (2214) the plurality of discrete slowness-frequency points into the remaining clustering classes into a combined dataset;invert (2216) the combined dataset to generate a dipole dispersion curve and a refined dataset; until a stopping criterion is met, iteratively (2218): apply an adaptive filter to the refined dataset to identify and remove points in the refined dataset that are out-of-bounds points of the dipole dispersion curve; and invert the refined dataset; and filter (2220) the refined dataset based, at least in part, on a predefined distance from the dipole dispersion curve. Petition 870260065747, dated 03 / 07 / 2026, p. 42 / 93 2 / 2; 2. Method, according to claim 1, characterized in that the stopping criterion (2110) is a predefined number of iterations.

3. Method, according to claim 1, characterized in that the stopping criterion (2112) is based on a residual mismatch error being less than a predefined threshold.

4. Method according to claim 1, characterized in that the adaptive filter (2114) is a distance-based filter that decreases the distance in each iteration.

5. Method according to claim 1, characterized in that the adaptive filter (2116) is automatically calculated based on variations in the categorized data or in the refined dataset. Petition 870260065747, dated 03 / 07 / 2026, p. 43 / 93

Images