An intelligent forward modeling method for induction logging response in layered media

Abstract

The present invention discloses an intelligent forward modeling method for induction logging responses in layered media, relating to the technical field of petroleum exploration and development. The method comprises the following steps: inputting parameters; randomly generating N one-dimensional horizontal layered formation models and generating a pixelated grid for each model along the vertical depth direction; using a rigorous forward modeling method to obtain the logging responses of subarray 1 of the N models; approximating and calculating the logging responses of subarray 1 of the N models; establishing a deep neural network model for subarray 1 for training, with the input set to the pixelated grid resistivity of the N models and the output set to the residual between the rigorous forward modeling response and the approximate forward modeling response of the N models; obtaining the trained deep neural network model for all subarrays; and randomly establishing a horizontal layered model. Using the deep neural network model, the method achieves rapid calculation of the array resistivity logging response of the model. The method effectively utilizes the propagation laws of low-frequency electromagnetic fields, reduces the difficulty of intelligent forward modeling, and reduces the range of residual responses.

Description

Technical Field

[0001] The present invention relates to the technical field of petroleum exploration and development, and in particular to an intelligent forward modeling method for layered medium induction logging response. Background Art

[0002] Array resistivity logging is a core method for evaluating formation electrical properties. The accuracy and efficiency of its forward modeling directly impact the reliability of reservoir parameter inversion and oil and gas evaluation. While existing one-dimensional pseudo-analytical algorithms can achieve strict calculations of array induction logging responses for one-dimensional layered models, they require sequentially calculating the raw subarray responses for multiple subarrays and multiple frequencies, then applying skin effect correction to obtain the corrected subarray responses. This results in a significant time-consuming full-subarray forward modeling for formation models with numerous layers, making it difficult to meet the requirements of real-time imaging and large-scale inversion iterations. While artificial intelligence has been introduced into the field of logging forward modeling in recent years, purely data-driven models suffer from a lack of physical constraints. These include black-box network predictions that violate electromagnetic field propagation mechanisms (e.g., negative resistivity), poor extrapolation of training samples (errors increase dramatically when formation parameters exceed the training range), and poor model generalization.

[0003] Therefore, it is urgent to study an intelligent forward modeling method for well logging response, in order to provide ideas for the rapid calculation of array resistivity logging response in horizontal layered models. Summary of the Invention

[0004] To solve the above technical problems, the present invention discloses an intelligent forward modeling method for induction logging responses in layered media. By integrating the Doll geometry factor theory in induction logging with deep learning technology, this method improves computational efficiency by more than three orders of magnitude while maintaining physical rationality. The error can be controlled within 3%, providing technical support for high-precision real-time reservoir evaluation.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An intelligent forward modeling method for induction logging response in layered media includes the following steps:

[0007] s1. Input the array resistivity logging instrument parameters, including the number of subarrays, the source distance between each subarray, and the transmitting coil frequency;

[0008] s2. Fixed the number of formation layers and well inclination, set the formation thickness and resistivity range, randomly generate N one-dimensional horizontal layered formation models, and generate a pixelated grid along the vertical depth direction for each model;

[0009] s3. Using the existing one-dimensional horizontal layered formation model pseudo-analytical forward modeling method (ie, rigorous forward modeling method), obtain the logging response of the subarray 1 of the N models generated in step s2;

[0010] s4. Set the calculation window range based on the detection performance of subarray 1, convolve the resistivity of each pixelated grid within the window with the low-order geometric factors at different vertical depths, and approximate the logging response of subarray 1 for the N models calculated in step s2;

[0011] s5. Establish a deep neural network model for subarray 1 for training, with the input set to the pixelated grid resistivity of N models and the output set to the residual between the rigorous forward modeling response in step s3 and the approximate forward modeling response in step s4 corresponding to the N models;

[0012] s6. Repeat steps s3 to s5 for all subarrays to obtain the trained deep neural network models for all subarrays;

[0013] s7. Randomly establish a horizontal layered model and use the deep neural network model in step s6 to quickly calculate the array resistivity logging response of the model.

[0014] Optionally, step s4 specifically includes:

[0015] s41. Using the subarray information input in step s1, based on Doll geometry factor theory, obtain the longitudinal differential geometry factor function of the composite coil system corresponding to subarray 1 , which reflects the contribution weight of the stratigraphic unit at depth z to the receiving coil signal, and defines the longitudinal range corresponding to 90% contribution in the longitudinal differential geometric factor as the calculation window range of subarray 1 ; For the longitudinal differential geometry factor function Integrate to obtain the longitudinal integral geometric factor ,use Equivalence is performed on the strata on both sides of the current calculation window to obtain the equivalent conductivity of the strata on both sides of the calculation window;

[0016] s42. Function The weight sequence is discretized into a pixelated grid, and discrete weight values ​​are generated within the calculation window of subarray 1. The formula is: ,i=-k,…,0,…,+k, , is the longitudinal position of the receiving coil, is the resolution of the pixelated grid;

[0017] s43. Select the j-th model and its pixelated grid generated in step s2. Assuming that the model has a total of D depth points, select the d-th depth point of the current model, d = 1, ..., D, and use this depth point as the origin to draw a calculation window range upward and downward. Extract the conductivity sequence within the calculation window from the pixelated grid of the current model. , load pre-generated geometry factor weights , weighted summation of the conductivity series and geometric factor weights within the calculation window , get the sub-array 1 approximation response at the d-th depth point; repeat the above operation for all depth points to get the sub-array 1 approximation response of the j-th model ;

[0018] s44. Execute step s43 cyclically for the N models to obtain the approximate responses of subarray 1 corresponding to the N models.

[0019] Optionally, step s5 specifically includes:

[0020] s51. Define the residual response of subarray 1 of the j-th model at the d-th depth point as: ,j=1,2,...,N (model index), d=1,2,...,D (depth point index), where, The rigorous forward response of subarray 1 at depth d for the j-th model calculated in step s3 is: The approximate response of subarray 1 at depth d for the j-th model calculated in step s4;

[0021] s52. Construct input data and rearrange the pixelated grids of the N models generated in step s2 into a three-dimensional tensor , where K is the total number of pixelated grids in the window corresponding to a single depth point, and each sample Represents the conductivity sequence of the j-th model in the window at depth point d, and the output label is the residual tensor , obtained from step s51 constitute;

[0022] s53. A two-level stacked ensemble learning architecture is used to enhance residual prediction accuracy. The first-level base model layer consists of a bagging ensemble tree, a gradient boosted regression tree, and a Gaussian process regression model, each trained independently. The second-level meta-model uses the LSBoost (Least Squares Boosting) framework. Cross-validation predictions are used during training to generate base model outputs, which are then fused with the original features and fed into the meta-model. Electromagnetic constraints are embedded in the meta-model training, and model fusion is optimized through step-by-step training and a dynamic weight adjustment mechanism. This architecture effectively improves the accuracy and generalization of residual predictions by leveraging the synergy between multi-model complementarity and physical constraints.

[0023] s54. Split the N models into training, validation, and test sets in a ratio of 7:2:1. Use mean absolute error (MAE) or mean square error (MSE) as the loss function. Use an early stopping strategy during training, stopping training when the validation set loss stops decreasing for several consecutive rounds.

[0024] After training, the residual prediction accuracy is evaluated on the test set, requiring the mean absolute error (MAE) to be less than 0.1% × And the coefficient of determination μ>0.95, if it does not meet the standard, return to step s5.3 to adjust the network structure and retrain; if it meets the standard, save the trained subarray 1 residual prediction model , which is called by step s7. Since each depth point is independent, samples from different models at different depth points can be mixed. This will increase the amount of training data and help train models with strong generalization capabilities.

[0025] Optionally, step s7 specifically includes:

[0026] s71. Randomly generate a new horizontal layered stratum model. The number of layers is consistent with the number of layers in the training model. The thickness and resistivity of each layer are randomly selected within the range of the stratum parameters set in step s2. The resolution of the pixelated grid is the same as that in step s2. Discretize the stratum into pixels and generate resistivity sequence vectors along the depth direction;

[0027] s72. For the response of subarray 1, load the pre-stored geometric factor weight sequence G1 for the subarray and determine the calculation window range Win1; traverse each depth point of the model, extract the resistivity sequence within the window centered at the current depth point, perform a weighted summation of the resistivity sequence and the geometric factor, and obtain the geometric factor approximation response curve of subarray 1. ;

[0028] s73. Call the subarray 1 deep neural network model trained in step s6 , the resistivity of the pixelated grid of the entire model is input into the network, and the network outputs the predicted residual curve , the residual contains high-frequency response components that are not captured by the geometric factor approximation; the geometric factor approximation response and the neural network residual are added at each depth point to obtain the intelligent forward response of subarray 1 ;

[0029] Repeat steps s72 and s73 for all subarrays to obtain the array resistivity logging instrument response in the model of step s71. Calculate the reference response for the same model using the rigorous forward modeling method of step s3. Calculate the relative error metrics for each subarray: mean absolute percentage error and maximum local deviation. Verify that the relative error in the main reservoir interval is <3%. If the error exceeds this limit, trigger the model retraining mechanism.

[0030] The beneficial effect of the present invention is that it provides an intelligent forward modeling method for induction logging response in layered media. By integrating the Doll geometric factor theory in induction logging with deep learning technology, the computational efficiency is improved by more than three orders of magnitude while maintaining physical rationality, and the error can be controlled within 3%, providing technical support for high-precision real-time reservoir evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of an intelligent forward modeling method for induction logging response in layered media in the present invention;

[0032] Figure 2A It is a one-dimensional horizontal layered model;

[0033] Figure 2B It is a pixelated grid model of a one-dimensional horizontal layered model;

[0034] Figure 3 The strict array resistivity logging response for the one-dimensional horizontal layered model;

[0035] Figure 4A is the vertical differential geometric factor diagram of array resistivity logging tool;

[0036] Figure 4B It is the vertical integration geometric factor diagram of array resistivity logging tool;

[0037] Figure 5 Approximating array resistivity logging responses for a one-dimensional horizontal layered model;

[0038] Figure 6 for Figure 3 Strict response and Figure 5 Comparison of approximation responses;

[0039] Figure 7 for Figure 3 Strict response and Figure 5 residuals between approximated responses;

[0040] Figure 8 Schematic diagram of the deep neural network model and training process;

[0041] Figure 9A A pixelated grid of a randomly generated one-dimensional horizontal layered model;

[0042] Figure 9B For vertical wells Figure 9A Comparison of the model's intelligent forward response and rigorous forward response results;

[0043] Figure 9C For vertical wells Figure 9A The relative error between the model's intelligent forward modeling response and the rigorous forward modeling response;

[0044] Figure 10A For inclined well Figure 9A Comparison of the model's intelligent forward response and rigorous forward response results;

[0045] Figure 10B For inclined well Figure 9A The relative error between the model's intelligent forward modeling response and the rigorous forward modeling response;

[0046] Figure 11 For different well inclination angles, Figure 9A Comparison of single-point calculation time between intelligent forward modeling and strict forward modeling of the model, as well as the forward modeling speed improvement effect. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] An intelligent forward modeling method for induction logging response in layered media, the process is as follows Figure 1 As shown, the following steps are included:

[0049] s1. Enter the array resistivity logging instrument parameters, including the number of subarrays, the source spacing between each subarray, and the transmit coil frequency. This example uses the Baker Atlas HDIL array induction logging instrument as an example. This instrument includes one transmit coil, eight operating frequencies, and seven subarray receive coils. Each subarray includes a shielded receive coil and a main receive coil.

[0050] s2. Fixed the number of formation layers and well inclination, set the formation thickness and resistivity range, where the number of formation layers is 20, the well inclination is 60°, the formation thickness is [0.5, 5] m, and the resistivity is [0.2, 200] Ω·m. Randomly generate 10,000 one-dimensional horizontal layered formation models within the parameter range, and generate a pixelated grid along the vertical depth direction for each model. The resolution of the pixelated grid is =0.05 m. Figure 2A A one-dimensional horizontal layered model is given. Figure 2B Given Figure 2A Illustration of the pixelated mesh generated by the model.

[0051] s3. Using the existing pseudo-analytical forward modeling method for the one-dimensional horizontal layered formation model, obtain the logging responses of subarray 1 for the 10,000 models generated in step s2. These responses are all responses after skin effect correction. In order to more intuitively demonstrate the accuracy and speed of the calculation of different subarray responses, Figure 3 Given Figure 2A The responses of all subarrays of the one-dimensional horizontal layered model are shown. Figure 3 As can be seen, different sub-arrays have different resolutions due to their different source distances. Based on this characteristic, the corresponding calculation window needs to be determined according to the detection performance of each sub-array in the subsequent steps.

[0052] s4. Set the calculation window range based on the detection performance of subarray 1. Convolve the resistivity of each pixelated grid within the window with low-order geometric factors at different vertical depths to approximate the logging response of subarray 1 for the 10,000 models in step s2.

[0053] Step s4 is described in further detail below:

[0054] s41. Using the subarray information input in step s1, based on Doll geometry factor theory, obtain the longitudinal differential geometry factor function of the composite coil system corresponding to subarray 1 , which reflects the contribution weight of the stratigraphic unit at depth z to the receiving coil signal, such as Figure 4A As shown; define the longitudinal range corresponding to 90% contribution in the longitudinal differential geometric factor as the calculation window range of subarray 1 =[-0.6,0.6]; for the longitudinal differential geometric factor function Integrate to obtain the longitudinal integral geometric factor ,like Figure 4B shown; using Equivalence is performed on the strata on both sides of the current calculation window to obtain the equivalent conductivity of the strata on both sides of the calculation window.

[0055] s42. Function Discretize the weight sequence into a pixelated grid. Assume that the resolution of the pixelated grid generated in step s2 is =0.05 m, and generate discrete weight values ​​within the calculation window of subarray 1. The formula is: ,i=-k,…,0,…,+k, , is the longitudinal position of the receiving coil, is the resolution of the pixelated grid.

[0056] s43. Select the j-th model and its pixelated grid generated in step s2, j = 1, ..., 10000. Assume that the model has a total of D depth points. Select the d-th depth point of the current model, d = 1, ..., D. Use this depth point as the origin to draw a calculation window upward and downward. Extract the conductivity sequence within the calculation window from the pixelated grid of the current model. , load pre-generated geometry factor weights , weighted summation of the conductivity series and geometric factor weights within the calculation window , get the approximate response of subarray 1 at the first depth point; repeat the above operation for all depth points, and get the approximate response of subarray 1 of the jth model by sliding window calculation ; Figure 2A The approximate forward response of the model is as follows Figure 5 As shown, Figure 3 The rigorous forward response comparison shown is Figure 5 The responses shown have similar patterns in different strata, but the resolution of each subarray is generally low. Figure 6 Subarray 1 and Subarray 7 were further compared. Figure 3 Strict forward response and Figure 5 The difference between the approximate forward modeling response and the strict forward modeling response shows that the law is consistent with that of the approximate forward modeling response, with slightly different results, and the difference is small. Therefore, by training the residual between the two, higher calculation accuracy can be achieved, verifying the feasibility of the method of this embodiment. Figure 7 Given Figure 3 Strict forward response and Figure 5 The specific residuals between the approximate forward responses are approximately 1 / 8 of the forward responses, which makes the model accuracy obtained by training the residuals higher with the same number of training sets.

[0057] s44. Repeat step s43 for 10,000 models to obtain approximate responses of subarray 1 corresponding to the 10,000 models.

[0058] s5. Build a deep neural network model for subarray 1 for training. The input is set to the pixelated grid resistivity of 10,000 models, and the output is set to the residual between the rigorous forward modeling response in step s3 and the approximate forward modeling response in step s4 corresponding to the 10,000 models.

[0059] Step s5 is described in further detail below:

[0060] s51. Define the residual response of subarray 1 of the j-th model at the d-th depth point as: ,j=1,2,...,N,d=1,2,...,D, where, The rigorous forward response of subarray 1 at depth d for the j-th model calculated in step s3 is: The subarray 1 approximation response at depth point d for the j-th model calculated in step s4.

[0061] s52. Construct input data and rearrange the pixelated grid resistivity of the N models generated in step s2 into a three-dimensional tensor , where K is the total number of pixelated grids in the window corresponding to a single depth point (determined by step s41) range calculation), each sample Represents the conductivity sequence of the j-th model in the window at depth point d, and the output label is the residual tensor , obtained from step s51 constitute.

[0062] s53. A two-level stacked ensemble learning architecture is used to enhance residual prediction accuracy: the first-level base model layer consists of a bagging ensemble tree (containing 500 random forest decision trees, self-sampling 70% of the data and randomly selecting 60% of the features), a gradient boosting regression tree (LightGBM framework, learning rate 0.05, 1000 iterations), and a Gaussian process regression (Matern5 / 2 kernel function, Nelder-Mead optimization hyperparameters), each trained independently; the second-level meta-model uses the LSBoost (Least Squares Boosting) framework, based on the least squares loss function, iterated 500 times, with a learning rate of 0.01, and the optimal parameters are determined through 10-fold cross-validation; cross-validation prediction is used to generate the base model output during training, which is then fused with the original features and input into the meta-model. Electromagnetic constraints are also embedded in the meta-model training ( =0.05, =0.02), and optimizes model fusion through step-by-step training and dynamic weight adjustment mechanisms; this architecture effectively improves the accuracy and generalization ability of residual prediction through the synergy of multi-model complementarity and physical law constraints.

[0063] s54. Divide N models into training, validation, and test sets in a ratio of 7:2:1. Use mean absolute error (MAE) or mean square error (MSE) as the loss function. Use an early stopping strategy during training, and stop training when the validation set loss stops decreasing for multiple consecutive rounds.

[0064] After training, the residual prediction accuracy is evaluated on the test set, requiring the mean absolute error (MAE) to be less than 0.1% × And the coefficient of determination μ>0.95, if it does not meet the standard, return to step s5.3 to adjust the network structure and retrain; if it meets the standard, save the trained subarray 1 residual prediction model , for step s7 to call, the process diagram is as follows Figure 8 As shown in the figure, since each depth point is independent, samples from different models at different depth points can be mixed, which will increase the amount of training data and help train models with strong generalization capabilities.

[0065] s6. Repeat steps s3 to s5 for all subarrays to obtain the trained deep neural network models for all subarrays.

[0066] s7. Randomly establish a horizontal layered model and use the deep neural network model in step s6 to quickly calculate the array resistivity logging response of the model.

[0067] Step s7 is described in further detail below:

[0068] s71. Randomly generate a new horizontal layered stratum model. The number of layers is consistent with the number of layers in the training model. The thickness and resistivity of each layer are randomly selected within the range of the stratum parameters set in step s2. The resolution of the pixelated grid is the same as that in step s2. = 0.05 m, the stratum is pixelated and discretized to generate a resistivity sequence vector along the depth direction, as shown in Figure 9A shown.

[0069] s72. For the response of subarray 1, load the pre-stored geometric factor weight sequence G1 of the subarray and determine the calculation window range ; Traverse each depth point of the model, extract the resistivity sequence in the window centered at the current depth point, perform weighted summation of the resistivity sequence and the geometric factor, and obtain the geometric factor approximation response curve of subarray 1 .

[0070] s73. Call the subarray 1 deep neural network model trained in step s6 , the pixelated grid resistivity of the entire model is input into the network, and the network outputs the predicted residual curve , the residual contains high-frequency response components that are not captured by the geometric factor approximation; the geometric factor approximation response and the neural network residual are added at each depth point to obtain the intelligent forward response of subarray 1 .

[0071] Repeat steps s72 and s73 for all subarrays to obtain the array resistivity logging instrument response in the model of step s71. Calculate the reference response for the same model using the rigorous forward modeling method of step s3. Calculate the relative error metrics for each subarray: mean absolute percentage error and maximum local deviation. Verify that the relative error in the main reservoir segment is less than 3%. If the error exceeds this limit, trigger the model retraining mechanism. Figure 9BThe comparison between the intelligent forward response and the strict response of the randomly generated model in step s71 is given for the vertical well case (well inclination = 0°). Figure 9C is the specific relative error of the intelligent forward modeling method under the current model. To verify the applicability of the method, Figure 10A The comparison between the intelligent forward response and the strict response of the randomly generated model in step s71 is given for the case of a deviated well (well inclination = 60°). Figure 10B The specific relative error of the intelligent forward modeling method under the current model is shown in Figure 2. It can be found that the average relative error of the intelligent forward modeling method can be controlled within 3%. The error mainly exists at the interface, and the error in the main reservoir section is relatively small. Figure 11 A comparison of single-point calculation time consumption and speed improvement effects of the strict forward modeling method and the intelligent forward modeling method under different well inclination angles is given. It can be seen that compared with the strict forward modeling algorithm, the intelligent forward modeling method can improve the calculation speed by more than three orders of magnitude, and the speed improvement effect will become more obvious with the increase in the number of layers.

[0072] This method enables rapid calculation of array resistivity logging responses in horizontal layered models, laying the foundation for forward modeling inversion of deviated array resistivity logging data. It effectively utilizes the propagation patterns of low-frequency electromagnetic fields and reduces the difficulty of intelligent forward modeling. Furthermore, compared to traditional methods that directly train logging responses, the residual response range is narrower, ensuring higher training accuracy with the same number of training sets.

[0073] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. An intelligent forward modeling method for induction logging response in layered media, characterized by: The steps include: s1. Input the array resistivity logging instrument parameters, including the number of subarrays, the source distance between each subarray, and the transmitting coil frequency; s2. Fixed the number of formation layers and well inclination, set the formation thickness and resistivity range, randomly generate N one-dimensional horizontal layered formation models, and generate a pixelated grid along the vertical depth direction for each model; s3. Using a rigorous forward modeling method, obtain the logging response of subarray 1 of the N models generated in step s2; s4. Set the calculation window range based on the detection performance of subarray 1, convolve the resistivity of each pixelated grid within the window with the low-order geometric factors at different vertical depths, and approximate the logging response of subarray 1 for the N models calculated in step s2; s5. Establish a deep neural network model for subarray 1 for training, with the input set to the pixelated grid resistivity of N models and the output set to the residual between the rigorous forward modeling response in step s3 and the approximate forward modeling response in step s4 corresponding to the N models; s6. Repeat steps s3 to s5 for all subarrays to obtain the trained deep neural network models for all subarrays; s7. Randomly establish a horizontal layered model and use the deep neural network model in step s6 to quickly calculate the array resistivity logging response of the model.

2. The intelligent forward modeling method for induction logging response in layered media according to claim 1, characterized in that: Step s4 specifically includes: s41. Using the subarray information input in step s1, based on Doll geometry factor theory, obtain the longitudinal differential geometry factor function of the composite coil system corresponding to subarray 1 , for the longitudinal differential geometric factor function Integrate to obtain the longitudinal integral geometric factor ,use Equivalence is performed on the strata on both sides of the current calculation window to obtain the equivalent conductivity of the strata on both sides of the calculation window; s42. Function The weight sequence is discretized into a pixelated grid, and discrete weight values ​​are generated within the calculation window of subarray 1. The formula is: ,i=-k,…,0,…,+k, , is the longitudinal position of the receiving coil, is the resolution of the pixelated grid; s43. Select the j-th model and its pixelated grid generated in step s2. Assuming that the model has a total of D depth points, select the d-th depth point of the current model, d = 1, ..., D, and use this depth point as the origin to draw a calculation window range upward and downward. Extract the conductivity sequence within the calculation window from the pixelated grid of the current model. , load pre-generated geometry factor weights , weighted summation of the conductivity series and geometric factor weights within the calculation window , get the sub-array 1 approximation response at the d-th depth point; repeat the above operation for all depth points to get the sub-array 1 approximation response of the j-th model ; s44. Execute step s43 cyclically for the N models to obtain the approximate responses of subarray 1 corresponding to the N models.

3. The intelligent forward modeling method for induction logging response in layered media according to claim 1, characterized in that: Step s5 specifically includes: s51. Define the residual response of subarray 1 of the j-th model at the d-th depth point as: ,j=1,2,...,N,d=1,2,...,D, where, The rigorous forward response of subarray 1 at depth d for the j-th model calculated in step s3 is: The approximate response of subarray 1 at depth d for the j-th model calculated in step s4; s52. Construct input data and rearrange the pixelated grids of the N models generated in step s2 into a three-dimensional tensor , where K is the total number of pixelated grids in the window corresponding to a single depth point, and each sample Represents the conductivity sequence of the j-th model in the window at depth point d, and the output label is the residual tensor , obtained from step s51 constitute; s53. A two-level stacked ensemble learning architecture is used to enhance residual prediction accuracy. The first-level base model layer consists of a bagging ensemble tree, a gradient boosted regression tree, and a Gaussian process regression model, each trained independently. The second-level meta-model uses the LSBoost framework. Cross-validation predictions are used during training to generate base model outputs, which are then fused with the original features and fed into the meta-model. Electromagnetic constraints are embedded in the meta-model training, and model fusion is optimized through step-by-step training and a dynamic weight adjustment mechanism. s54. Split the N models into training, validation, and test sets in a ratio of 7:2:

1. Use mean absolute error (MAE) or mean square error (MSE) as the loss function. Use an early stopping strategy during training, stopping training when the validation set loss stops decreasing for several consecutive rounds. After training, the residual prediction accuracy is evaluated on the test set, requiring the mean absolute error (MAE) to be less than 0.1% × And the coefficient of determination μ>0.95, if it does not meet the standard, return to step s5.3 to adjust the network structure and retrain; if it meets the standard, save the trained subarray 1 residual prediction model , for step s7 to call.

4. The intelligent forward modeling method for induction logging response in layered media according to claim 1, characterized in that: Step s7 specifically includes: s71. Randomly generate a new horizontal layered stratum model. The number of layers is consistent with the number of layers in the training model. The thickness and resistivity of each layer are randomly selected within the range of the stratum parameters set in step s2. The resolution of the pixelated grid is the same as that in step s2. Discretize the stratum into pixels and generate resistivity sequence vectors along the depth direction; s72. For the response of subarray 1, load the pre-stored geometric factor weight sequence G1 of the subarray and determine the calculation window range ; Traverse each depth point of the model, extract the resistivity sequence in the window centered at the current depth point, perform weighted summation of the resistivity sequence and the geometric factor, and obtain the geometric factor approximation response curve of subarray 1 ; s73. Call the subarray 1 deep neural network model trained in step s6 , the pixelated resistivity grid of the entire model is input into the network, and the network outputs the predicted residual curve , the residual contains high-frequency response components that are not captured by the geometric factor approximation; the geometric factor approximation response and the neural network residual are added at each depth point to obtain the intelligent forward response of subarray 1 ; Repeat steps s72 and s73 for all subarrays to obtain the array resistivity logging instrument response in the model of step s71. Calculate the reference response for the same model using the rigorous forward modeling method of step s3. Calculate the relative error metrics for each subarray: mean absolute percentage error and maximum local deviation. Verify that the relative error in the main reservoir interval is <3%. If the error exceeds this limit, trigger the model retraining mechanism.

Images