Transient electromagnetic forward modeling method and system considering induced polarization effect based on deep learning

Abstract

The invention belongs to the crossing field of artificial intelligence and geophysics, and particularly discloses a transient electromagnetic forward modeling method and system considering an induced polarization effect based on deep learning, and the method comprises the steps: inputting Cole-Cole key parameters representing the induced polarization effect into a CNN-LSTM mixed deep learning network model for forward modeling, and obtaining transient electromagnetic response data; wherein Cole-Cole key parameters representing the induced polarization effect serve as multi-dimensional input features of the CNN-LSTM mixed deep learning network model, transient electromagnetic response data of a corresponding time channel serve as output, and the CNN-LSTM mixed deep learning network model is trained by adopting a segmented weighted mean square error loss function. According to the method, the segmented weighted mean square error loss function is adopted in the training stage, higher weights are given to the symbol inversion points and the prediction errors of the adjacent time windows, and the reliability of all-time response prediction is guaranteed.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of artificial intelligence and geophysics, and more specifically, relates to a transient electromagnetic forward modeling method and system based on deep learning that takes into account the induced polarization effect. Background Technology

[0002] Traditional Time-Domain Electromagnetic Methods (TEM) are an important artificial-source geophysical exploration method. Its working principle involves emitting a primary pulse magnetic field into the subsurface using an ungrounded or grounded source. During the intervals between primary field pulses, the secondary eddy current field induced by the subsurface medium is observed through a receiving coil or grounded electrode. Because this method directly observes the pure secondary field and can be conducted in near-field conditions, it offers advantages such as large detection depth, high vertical resolution, and high observation accuracy. In recent years, it has been widely used in fields such as metal mineral exploration, hydrogeological surveys, and environmental engineering exploration.

[0003] In actual exploration, transient electromagnetic responses sometimes exhibit anomalous characteristics, such as sign reversal in decay curves. This phenomenon is closely related to the induced polarization (IP) effect generated by subsurface polarimetric bodies (such as disseminated ore bodies). However, traditional TEM forward modeling methods have significant limitations in simulating responses with IP effects: on the one hand, traditional methods are usually based on pure conductivity models, making it difficult to accurately characterize the complex resistivity dispersion characteristics caused by IP effects, resulting in insufficient explanatory power for anomalous phenomena such as sign reversal; on the other hand, even with the introduction of dispersion models such as Cole-Cole for coupled calculations, problems such as complex digital implementation and high computational costs remain. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application aims to provide a transient electromagnetic forward modeling method and system based on deep learning that considers the induced polarization effect. This method is intended to solve the problem that traditional TEM forward modeling methods are unable to accurately reflect the complex resistivity dispersion characteristics caused by the induced polarization effect when simulating responses containing the IP effect.

[0005] The first aspect of this application relates to a transient electromagnetic forward modeling method based on deep learning that considers induced polarization effects, comprising the following steps: The key Cole-Cole parameters characterizing the induced electrostatic effect are input into a CNN-LSTM hybrid deep learning network model for forward modeling to obtain transient electromagnetic response data. The training method for the CNN-LSTM hybrid deep learning network model includes the following steps: Step 1: Construct the frequency domain electromagnetic response function and transform it into the time domain; Step 2: In the time-domain electromagnetic response function, the Cole-Cole complex resistivity model is used to replace the actual resistivity of the formation to obtain the transient electromagnetic response function containing the induced polarization effect. Step 3: Based on the transient electromagnetic response function containing the induced polarization effect, the Cole-Cole key parameters characterizing the induced polarization effect are used as the multidimensional input features of the CNN-LSTM hybrid deep learning network model, and the transient electromagnetic response data corresponding to the time channel are used as the output. The CNN-LSTM hybrid deep learning network model is trained using a piecewise weighted mean square error loss function.

[0006] In some implementations, the piecewise weighted mean square error loss function is: ; Among them, adaptive weight coefficients for: ; in, This refers to the sensitive time window for sign reversal triggered by the induced polarization effect; To determine the threshold at which the response value approaches 0; Basic weights; For higher-order penalty weights in the sign-inverted region; The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of a time-channel; The output of the CNN-LSTM hybrid deep learning network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel; For the first A timeline.

[0007] In some implementations, key Cole-Cole parameters characterizing the induced polarization effect include the resistivity of each layer, polarizability, time constant, and frequency correlation coefficient.

[0008] In some specific implementations, step three specifically includes the following steps: Based on the transient electromagnetic response function of the induced polarization effect, a forward numerical simulation was performed on an n-layered geological model to obtain the characteristic laws of the transient electromagnetic response containing the induced polarization effect. Based on the transient response characteristics of the induced polarization effect, B-spline interpolation is performed on an n-layered geological model to randomly construct a vertically continuous and relatively smooth m-layered geological model and obtain a training dataset. The training dataset includes Cole-Cole key parameters and electromagnetic response data that characterize the induced polarization effect. The key Cole-Cole parameters characterizing the induced polarization effect are input into a CNN-LSTM hybrid deep learning network model. The transient electromagnetic response data corresponding to the time channel are used as the output, and a piecewise weighted mean square error loss function is used to train the CNN-LSTM hybrid deep learning network model. Among them, the convolutional neural network model is used to extract the spatial distribution characteristics of the m-layered geological model, and the long short-term memory network is used to learn the time series dependence of the transient electromagnetic response. m>n, and m and n are both integers greater than 2.

[0009] More preferably, the CNN-LSTM hybrid deep learning network model includes four convolutional layers for feature extraction, one long short-term memory network layer, and two fully connected layers.

[0010] In some implementations, the Cole-Cole complex resistivity model is as follows: ; in, Zero-frequency resistivity; Polarizability; It is a time constant; The frequency correlation coefficient; It is angular frequency.

[0011] The second aspect of this application relates to a transient electromagnetic forward modeling system based on deep learning and considering induced polarization effects, comprising: The forward modeling module is used to input the key Cole-Cole parameters characterizing the induced electrostatic effect into the CNN-LSTM hybrid deep learning network model for forward modeling to obtain transient electromagnetic response data. The model training module includes: The first function processing unit is used to construct the frequency domain electromagnetic response function and transform it into the time domain. The second function processing unit is used to obtain the transient electromagnetic response function containing the induced polarization effect by replacing the actual resistivity of the formation with the Cole-Cole complex resistivity model in the time domain electromagnetic response function. The training unit is used to train the CNN-LSTM hybrid deep learning network model based on the transient electromagnetic response function containing the induced polarization effect. The Cole-Cole key parameters characterizing the induced polarization effect are used as multidimensional input features of the CNN-LSTM hybrid deep learning network model, and the transient electromagnetic response data corresponding to the time channel are used as output. The piecewise weighted mean square error loss function is used to train the CNN-LSTM hybrid deep learning network model.

[0012] In some implementations, the piecewise weighted mean square error loss function in the training unit is: ; Among them, adaptive weight coefficients for: ; in, This refers to the sensitive time window for sign reversal triggered by the induced polarization effect; To determine the threshold at which the response value approaches 0; Basic weights; For higher-order penalty weights in the sign-inverted region; The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of a time-channel; The output of the CNN-LSTM hybrid deep learning network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel; For the first A timeline.

[0013] In some implementations, the training unit includes: The response characteristic analysis subunit is used to generate n-layered geological models based on the induced polarization effect transient electromagnetic response function and perform forward numerical simulation to obtain the characteristic laws of transient electromagnetic response containing the induced polarization effect. The training data acquisition subunit is used to perform B-spline interpolation on the transient electromagnetic response function of the induced polarization effect, randomly construct a vertically continuous and relatively smooth m-layered geological model, and acquire the training dataset. The training dataset includes Cole-Cole key parameters and electromagnetic response data characterizing the induced polarization effect; m>n, where m and n are both integers greater than 2. The training subunit is used to input the key Cole-Cole parameters characterizing the induced polarization effect into the CNN-LSTM hybrid deep learning network model, and uses the transient electromagnetic response data of the corresponding time channel as the output. The CNN-LSTM hybrid deep learning network model is trained using a piecewise weighted mean square error loss function. Among them, the convolutional neural network model is used to extract the spatial distribution characteristics of the m-layered geological model, and the long short-term memory network is used to learn the time series dependence of the transient electromagnetic response.

[0014] In some implementations, the Cole-Cole complex resistivity model is as follows: ; in, Zero-frequency resistivity; Polarizability; It is a time constant; The frequency correlation coefficient; It is angular frequency.

[0015] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0016] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0017] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0018] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0019] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: Significantly improving computational efficiency and providing a fast forward modeling operator for real-time inversion: This application combines a CNN-LSTM hybrid deep learning network to construct an end-to-end nonlinear mapping model from subsurface medium parameters to transient electromagnetic responses. After model training, the computation time for a single forward modeling operation is only in the milliseconds, compared to the several minutes required by traditional numerical methods, representing a computational efficiency improvement of more than three orders of magnitude. This advantage provides a reliable fast forward modeling operator for subsequent real-time, large-scale inversion interpretation.

[0020] Deep integration of induced polarization effect enhances model physical assurance and generalization ability: This application innovatively integrates the induced polarization effect into a data-driven forward modeling framework. Unlike traditional pure resistivity forward modeling, this application explicitly introduces key Cole-Cole parameters (polarizability η, time parameter τ, and frequency correlation coefficient c) to characterize the induced polarization effect in the construction of training data and network input. This design enables the model to be effectively applied to complex geoelectric structures with polarization layers, significantly expanding the application boundaries of deep learning forward modeling methods.

[0021] To address the sign inversion challenge and ensure reliable response prediction across all time periods, this application proposes an effective network learning and optimization method for linear sign inversion in transient electromagnetic responses. By employing a piecewise weighted mean square error loss function during training, higher weights are given to the prediction errors of the sign inversion point and its adjacent time windows, forcing the network to focus on learning the evolution of this key feature. This method effectively overcomes the problem of decreased prediction accuracy caused by the lack of dedicated handling of sign inversion points in conventional methods, significantly improving the model's prediction performance on sign inversion phenomena, thereby ensuring the prediction accuracy and physical consistency of the entire transient electromagnetic response curve from early to late stages.

[0022] This application forms a complete technological closed loop with significant engineering application value: From theory (induced polarization effect modeling), data (multi-parameter sample library), algorithm (CNN-LSTM hybrid network and customized loss function) to application (fast forward modeling operator), this closed loop allows the trained model to achieve instantaneous prediction of unknown models without repeated mesh partitioning and complex numerical calculations, facilitating seamless integration into the inversion process. This method provides a powerful tool for rapidly and accurately interpreting measured data containing induced polarization effects, lowers the computational threshold, and has significant engineering application value. Attached Figure Description

[0023] Figure 1 This is a flowchart of the transient electromagnetic forward modeling method based on deep learning and considering the induced polarization effect provided in the embodiments of this application.

[0024] Figure 2 This is a transient electromagnetic response curve considering the induced polarization effect provided in the embodiments of this application.

[0025] Figure 3(a) is a resistivity diagram in a one-dimensional layered geological model provided in an embodiment of this application.

[0026] Figure 3(b) is a polarizability diagram in a one-dimensional layered geological model provided in the embodiments of this application.

[0027] Figure 3(c) is a relaxation parameter diagram in a one-dimensional layered geological model provided in the embodiments of this application.

[0028] Figure 3(d) is a time constant diagram in the one-dimensional layered geological model provided in the embodiments of this application.

[0029] Figure 4 This is a transient electromagnetic response curve with altered polarization provided in an embodiment of this application.

[0030] Figure 5 This is a transient electromagnetic response curve with varying frequency correlation coefficients provided in an embodiment of this application.

[0031] Figure 6This is a transient electromagnetic response curve with a changed time constant provided in an embodiment of this application.

[0032] Figure 7 This is a transient electromagnetic response curve of changing zero-frequency resistivity provided in an embodiment of this application.

[0033] Figure 8 This is a transient electromagnetic response curve showing the change in polarization layer thickness provided in an embodiment of this application.

[0034] Figure 9 This is a transient electromagnetic response curve showing the change in polarization layer burial depth provided in an embodiment of this application.

[0035] Figure 10 This is a network structure diagram provided in the embodiments of this application.

[0036] Figure 11 This is a scatter plot (60 time channels) of transient electromagnetic response provided in an embodiment of this application.

[0037] Figure 12 This is a graph showing the loss function of the training set and validation set provided in the embodiments of this application.

[0038] Figure 13(a) is a schematic diagram of the first set of forward modeling results provided in the embodiments of this application.

[0039] Figure 13(b) is a schematic diagram of the relative error of the first set of data provided in the embodiments of this application.

[0040] Figure 13(c) is a schematic diagram of the second set of forward modeling results provided in the embodiments of this application.

[0041] Figure 13(d) is a schematic diagram of the relative error of the second set of data provided in the embodiments of this application.

[0042] Figure 14(a) is a schematic diagram of the forward modeling results of the third set of data provided in the embodiments of this application.

[0043] Figure 14(b) is a schematic diagram of the relative error of the third set of data provided in the embodiments of this application.

[0044] Figure 14(c) is a schematic diagram of the forward modeling results of the fourth set of data provided in the embodiments of this application.

[0045] Figure 14(d) is a schematic diagram of the relative error of the fourth set of data provided in the embodiments of this application.

[0046] Figure 15(a) is a schematic diagram of the forward modeling results of the fifth set of data provided in the embodiments of this application.

[0047] Figure 15(b) is a schematic diagram of the relative error of the fifth set of data provided in the embodiments of this application.

[0048] Figure 15(c) is a schematic diagram of the forward modeling results of the sixth set of data provided in the embodiments of this application.

[0049] Figure 15(d) is a schematic diagram of the relative error of the sixth set of data provided in the embodiments of this application.

[0050] Figure 16(a) is a schematic diagram of the forward modeling results of the seventh set of data provided in the embodiments of this application.

[0051] Figure 16(b) is a schematic diagram of the relative error of the seventh set of data provided in the embodiments of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0054] In this application, the terms “first” and “second” are used to distinguish different objects, rather than to describe a specific order of objects.

[0055] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0056] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more.

[0057] The embodiments of this application are described below with reference to the accompanying drawings.

[0058] This application provides a transient electromagnetic forward modeling method based on deep learning and considering induced polarization effects. First, a convolutional neural network (CNN) is used to extract multi-level spatial features from multivariate input data. Second, a long short-term memory network (LSTM) is introduced to model the temporal dependencies of the extracted feature sequences. Finally, a fully connected layer achieves high-precision mapping from features to target parameters. This application realizes the collaborative learning of spatial features and time series, improving the automation level, feature representation ability, and prediction accuracy of forward modeling, thus overcoming the limitations of traditional forward modeling methods or single-architecture networks when processing data with strong spatiotemporal coupling characteristics.

[0059] Deep Learning (DL) is one of the most popular data-driven methods, widely applied in fields such as image processing, natural language processing, speech recognition, and autonomous driving. DL is a family of algorithms containing multiple processing layers used to learn data representations (the most useful information contained in the original data), with multiple levels of abstraction. DL uses the backpropagation algorithm to discover complex structures in large datasets, instructing the machine how it should change its internal parameters to compute the representation of each layer from the representations of previous layers. This approach has the potential to solve highly nonlinear inversion problems with non-uniqueness, such as recovering resistivity from TEM data.

[0060] Numerous studies have applied deep learning (DL), particularly convolutional neural networks (CNNs) and long short-term memory (LSTM), to resistivity inversion of TEM data. Resistivity recovery demonstrates that DL methods can avoid gradient calculations in traditional methods, thus improving efficiency. Therefore, exploring and developing a deep learning-based method for transient electromagnetic forward modeling with induced polarization effects that balances high computational efficiency and high simulation accuracy is not only an important supplement and innovation to traditional numerical computation paradigms but also possesses significant theoretical value and urgent practical needs.

[0061] This application provides a transient electromagnetic forward modeling method based on deep learning that considers the induced polarization effect, including the following steps: Step S1: Calculate the frequency domain electromagnetic response without induced polarization and convert it to the time domain; More specifically, it includes the following steps: Step S1.1: Construct a horizontally uniform layered model and calculate the electromagnetic response of the central loop device in the frequency domain: (1); in, The magnetic field strength in the z-direction ( ); The input current (A) of the transmit return line; The radius of the emission loop (m); The wave impedance of the first layer of medium ( ); Wave impedance of the air layer ( ); The integral variable is (1 / m); It is a first-order Bessel function; Step S1.2: Differentiate the magnetic field strength and transform it to the time domain using a cosine transform to obtain the time-domain transient electromagnetic response without induced polarization: (2); in, , , , , , , , and These are all functions related to layer parameters and frequency.

[0062] Step S2: Introduce the Cole-Cole complex resistivity model to calculate the transient electromagnetic response with induced polarization effect; The Cole-Cole complex resistivity model is used to replace the actual resistivity of the formation; (3); in, Zero-frequency resistivity; Polarizability; It is a time constant; Here is the frequency correlation coefficient; the distribution range of each parameter is: , c = 0.1~0.6; the above complex resistivity is obtained by wavenumber analysis. Integrating layered medium wave impedance In the calculation, the electromagnetic response integral equation is substituted, and then a cosine transform is performed to obtain the time-domain transient electromagnetic response containing the induced polarization effect: (4).

[0063] Step S3: Construct a layered geological model containing induced polarization parameters and perform response characteristic analysis; Based on the Cole-Cole model in step S2, a multi-layered geological model containing resistivity, polarizability, time constant and frequency correlation coefficient of each layer is randomly generated. Forward numerical simulation is performed to analyze the influence of induced polarization on transient electromagnetic response. Step S4: Construct a CNN-LSTM hybrid deep learning network. More specifically, construct a hybrid deep learning model of convolutional neural network and long short-term memory network. The convolutional neural network is used to extract the spatial distribution characteristics of geological model parameters, and the long short-term memory network is used to model the time series dependence of transient electromagnetic response.

[0064] Step S5: Generate a training sample set and train the network model; A smooth and continuous layered geological model is constructed based on the B-spline interpolation method. The multidimensional model parameters, including resistivity and induced polarization parameters, are used as network inputs, and the transient electromagnetic response of the corresponding time channel is used as network outputs to train the hybrid network model. Step S6: Optimize and train the network using a piecewise weighted mean square error loss function; The network is optimized and trained using a piecewise weighted mean square error loss function: ; Among them, the weighting coefficient The method for determining it is as follows: ; in, This is a sensitive time window for the excitation effect to trigger sign reversal. To determine the tiny threshold at which the response value approaches 0, Based on the weights, For higher-order penalty weights in the sign-inverted region; For the first A timeline.

[0065] Step S7: Model Validation and Application; The convergent training model is used as the final forward modeling prediction model to achieve fast and high-precision transient electromagnetic forward modeling calculations for any given geoelectric model containing induced polarization.

[0066] Example 1 like Figure 1 As shown, this application provides a transient electromagnetic forward modeling method based on deep learning that considers the induced polarization effect, specifically including the following steps: Step S1: Construct a horizontal uniform layered model, calculate the electromagnetic field distribution of the central loop device in the uniform half-space model in the frequency domain, and obtain the electromagnetic response of the layered medium in the frequency domain. A one-dimensional forward modeling calculation is performed to determine the electromagnetic field distribution of the central loop device in a uniform half-space model in the frequency domain. Then, a horizontally uniform layered model is introduced for solution to obtain the electromagnetic response of the layered medium in the frequency domain. ; in, The magnetic field strength in the z-direction ( ); The input current (A) of the transmit return line; The radius of the emission loop (m); The wave impedance of the first layer of medium ( ); Wave impedance of the air layer ( ); The integral variable is (1 / m); It is a first-order Bessel function; After differentiation, we get: ; in, It is the angular frequency (rad / s); The vacuum permeability is typically taken as... By transforming the signal to the time domain using a cosine transform, the transient electromagnetic response in the time domain can be obtained. ; Taking the derivative, we get: ; in, , , , , , , , and These are all functions related to layer parameters and frequency.

[0067] Step S2: Replace the formation real resistivity with the Cole-Cole complex resistivity model. and through wavenumber Integrating layered medium wave impedance The calculation, substituted into the electromagnetic response integral equation, yields the time-domain transient electromagnetic response containing the induced polarization effect; the mathematical expression of the Cole-Cole complex resistivity model is: ; in, Zero-frequency resistivity; Polarizability; It is a time constant; Here is the frequency correlation coefficient; the distribution range of each parameter is: , c = 0.1~0.6.

[0068] Step S3: Based on the Cole-Cole resistivity model, a five-layered geological model was randomly generated for forward numerical simulation to analyze its response characteristics and patterns. The parameters of the generated model were set as follows: the resistivity of the first layer was 190... No induced polarization effect (polarizability, frequency correlation coefficient, and time constant are all 0); the resistivity of the second layer is 167. The first layer exhibits a significant induced polarization effect, with a polarization rate of 0.25, a frequency correlation coefficient of 0.2, and a time constant of 0.5 s. The resistivities of the third to fifth layers are 170 Ω·m, 206 Ω·m, and 239 Ω·m, respectively, and none exhibit induced polarization effects. Through simulation calculations, the corresponding transient electromagnetic response curves were obtained. Figure 2 Based on this, schematic diagrams 3(a) to 3(d) reflecting the spatial distribution of parameters in the layered geological model were drawn. Figure 2 The curve shape shows that, compared with the transient electromagnetic response curve considering the induced polarization effect, in the early time (… ~ s) The two curves basically overlap; as time progresses to the middle and late stages ( After s), the curve without considering the induced polarization effect (blue solid line) continues its monotonically decreasing trend; while the curve considering the induced polarization effect (red dashed line) shows a decreasing trend after approximately s. s~ The response between s exhibits a "decay first, then rise" characteristic, which is fundamentally different from the classic "slow decay" description and accurately reflects the characteristic response of the induced polarization effect in transient electromagnetic methods.

[0069] Step S4: Change the parameters of the polarization layer and analyze the changing characteristics and influence laws of the transient electromagnetic response curve containing the induced polarization effect, such as... Figures 4-7 As shown; Keeping other parameters constant, this application sets the polarizability to 0.25, 0.45, and 0.65 respectively, plots the corresponding transient electromagnetic response curves, and compares them with curves that do not consider induced polarization effects (e.g., ...). Figure 4 As shown); the results indicate that the greater the polarizability, the earlier the curve exhibits the "first decay, then rise" response characteristic in the middle and late stages; using the same control variable method, the frequency correlation coefficient c was set to 0.1, 0.2, and 0.3 respectively, and compared with the curve without considering the induced polarization effect (as shown); Figure 5 As shown in the figure, it was found that the larger the c value, the earlier the aforementioned "first decay, then rise" response characteristic appeared in the middle and late stages. Further experiments were conducted by adjusting the time constant to 0.5, 1.5, and 10.5 respectively. Figure 6 The results showed that the larger the time constant, the smaller the time difference between the appearance of the "first decay, then rise" characteristic; finally, the value of the zero-frequency resistivity was changed to 267, 167, and 67 respectively. Figure 7The comparison also revealed that the larger the zero-frequency resistivity value, the earlier the "attenuation followed by rise" response characteristic appeared in the mid-to-late stages. Comprehensive analysis shows that regardless of the parameter adjustments, the "attenuation followed by rise" characteristic induced by polarization effect always appeared in the mid-to-late stages. s~ Within the time interval s; at the same time, the larger the value of each parameter, the larger the corresponding transient electromagnetic response value.

[0070] Continue to change the polarization layer thickness and polarization layer burial depth (see...) Figures 8-9 It was found that the thicker and deeper the polarization layer, the earlier the polarization effect appears. Therefore, the Cole-Cole model can be used to equivalently simulate the transient electromagnetic response considering the induced polarization effect. By changing the polarization layer parameters, it was found that the strength of the induced polarization effect is directly proportional to the polarizability, frequency correlation coefficient, zero-frequency resistivity, and thickness of the polarization layer, and inversely proportional to the time constant and burial depth. Each parameter has a different degree of influence on the strength of the induced polarization effect.

[0071] Step S5: Construct a CNN-LSTM hybrid deep learning network; Traditional transient electromagnetic numerical forward modeling has extremely high computational costs (up to minutes per calculation) after introducing the Cole-Cole model. Meanwhile, conventional deep learning forward modeling usually only targets pure conductivity models (pure resistivity), which cannot characterize the complex dispersion characteristics and late reversal phenomenon brought about by induced polarization. Therefore, this application reconstructs the network input, instead of inputting zero-frequency resistivity, it explicitly introduces key Cole-Cole parameters (polarizability m, time constant) characterizing the induced polarization effect. c) Frequency correlation coefficient is used as a multidimensional input feature of the network. In terms of network architecture, a deep fusion architecture of CNN spatial feature extraction and LSTM time series modeling is designed: like Figure 10 As shown, a network structure of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) is constructed. The network structure includes four convolutional layers for feature extraction, one LSTM layer, and two fully connected layers. The kernel size of the convolutional layers is [size missing]. The first three convolutional layers have a stride of 1, the fourth convolutional layer has a stride of 2, and there are 128 convolutional kernels. Each convolutional layer is connected by the LeakyReLU activation function. A batch normalization layer is added to normalize the output of the convolutional layers to improve the stability of model training. After extracting the complex spatial distribution features of multi-layered and multi-dimensional geological parameters using the aforementioned one-dimensional convolutional neural network, the extracted data features are serialized and input into an LSTM layer to accurately capture the dynamic decay law and sign inversion dependency of the transient electromagnetic response curve over time. The output of the LSTM layer is further used as the input to a fully connected layer, which is equipped with a Tanh activation function. In addition, the random deactivation ratio of the Dropout layer is set to 0.2, randomly setting the output of selected neurons to 0, effectively preventing model overfitting. This spatiotemporal collaborative network architecture not only achieves efficient end-to-end forward modeling of geological structures containing polarization layers, but also shows a significant efficiency advantage compared with traditional methods, thereby expanding the application potential of deep learning in induced polarization effect exploration.

[0072] Step S6: Generate a training sample set and train the network model; To address the need for transient electromagnetic forward modeling considering induced ionization (IEE) effects, traditional discrete layered models often suffer from abrupt parameter jumps and fail to reflect the gradual changes in real geological conditions. Therefore, this application proposes a method based on B-spline interpolation to randomly construct a vertically continuous and relatively smooth 30-layered geological model. After model construction, forward modeling calculations of transient electromagnetic responses incorporating IEE effects are performed. Leveraging a deep learning framework, this application overcomes the limitations of single-parameter input by combining a constructed smooth 30-layer multidimensional geoelectric parameter set (including resistivity, polarizability induced by electrostatics, and time constant, etc.) as network input, and using the transient electromagnetic responses of 60 time channels obtained from forward modeling as network output (e.g., ...). Figure 11 As shown in the figure, a complex nonlinear mapping relationship between "multidimensional geological parameters and electromagnetic response labels" is established.

[0073] To complete the training of the network model and address the late transient electromagnetic response caused by the induced polarization effect (IPE), s~ To prevent the network from tending to learn only early features with larger values ​​and ignoring the inverted region (between s), this application employs a piecewise weighted mean squared error loss function (Piecewise Weighted MSE) during the training phase. Its calculation formula is as follows: ; in, The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of each time channel (label); The output of the network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel; segmented weights to account for sign inversion characteristics. The calculation formula is as follows: ; in, Sampling time; This is a sensitive time window for the excitation effect to trigger sign reversal. To determine the tiny threshold at which the response value approaches 0, Set the base weight for the regular region (set to 1). This is the higher-order penalty weight for the sign-inverted region; through this formula, the model is forced to prioritize correcting small errors in the sign-inverted region during backpropagation, thereby completing high-precision iterative training of the network model.

[0074] To minimize the loss function, this application divides network training into three stages. Initially, the Adam optimizer is used with a learning rate of 0.01; the learning rate is then changed to 0.001; finally, the AdaGrad optimizer is used. The loss function after the three training stages is as follows: Figure 12 The relatively small and stable loss values ​​(0.01314 on the test set and 0.01433 on the validation set) further confirm the model's training effectiveness. The low and stable loss values ​​indicate a high degree of model fit to the data, and the difference in loss between the test and validation sets is only about 0.001, again demonstrating the model's good generalization ability and its stable adaptation to data not used in training. Combined with the previous curve characteristics, the current model has achieved effective learning while avoiding the risk of overfitting, indicating a relatively ideal training state.

[0075] Step S7: The model trained and converged based on the piecewise weighted mean square error loss function is used as the final network prediction model; a layered geological model is generated for forward modeling testing to verify the effectiveness of the algorithm; for the case where no sign inversion occurs (the first four sets of transient electromagnetic response curves are shown in Figures 13(a)~13(d) and 14(a)~14(d): In s~ Within the given time range, the relative error in forward modeling is relatively small, demonstrating the high accuracy of this neural network. However, the response curves of the fifth and sixth groups (as shown in Figures 15(a)~15(d)) exhibit sign reversal within the same time range, with a significant increase in relative error at the reversal point, indicating lower accuracy at this location. Based on this anomaly, the main reason is believed to be that the values ​​at the reversal points infinitely approach 0 (e.g., the scatter plots of the transient electromagnetic response curve). Figure 11As shown, the large blue rectangle represents the sign change phenomenon, and the small red rectangle encloses the reversal point (i.e., the minimum point). However, the model training process did not specifically handle values ​​approaching 0, which caused the model to focus on learning early features with larger values, resulting in a significant increase in the relative error of the reversal region and a sharp decline in prediction performance. However, this application uses a piecewise weighted mean square error loss function in model training to assign high-order adaptive weights to the aforementioned reversal minimum point region approaching 0, forcing the model to accurately learn the small reversal features, thereby effectively overcoming this problem and significantly reducing the relative error at the reversal point.

[0076] The forward modeling results of the seventh group (Figures 16(a) to 16(b)) are in 10 -6 s~10 - Within a time range of 2s, the sign reverses and the curve trends remain the same, with small errors; however, within 10... - After 2s, the real model begins to show a potential sign reversal trend. Due to the influence of various factors, the location of the sign reversal is uncertain; the clear distinguishing band of the actual IP phenomenon is mostly around 10. -3 After s, traditional methods are prone to missing IP due to weak late-stage signals. Based on the optimized collaborative network structure and weighted loss function, this application can more sensitively capture and accurately predict the weak reversal trend across the map window in the late time period, effectively avoiding the missing IP phenomenon in complex geological exploration. This fully verifies the high accuracy and robustness of this application in all time periods, especially under strong induced polarization effect conditions.

[0077] It should be noted that the kernel size, stride, number of kernels, and dimensions of the convolutional layers described above are set based on the model training effect. Therefore, the above examples are only illustrative. Without departing from the concept of this application, the kernel size, stride, number of kernels, and dimensions of the convolutional layers can be adjusted.

[0078] Example 2 This application provides a transient electromagnetic forward modeling system based on deep learning and considering the induced polarization effect, including: The forward modeling module is used to input the key Cole-Cole parameters characterizing the induced electrostatic effect into the CNN-LSTM hybrid deep learning network model for forward modeling to obtain transient electromagnetic response data. The model training module includes: The first function processing unit is used to construct the frequency domain electromagnetic response function and transform it into the time domain. The second function processing unit is used to obtain the transient electromagnetic response function containing the induced polarization effect by replacing the actual resistivity of the formation with the Cole-Cole complex resistivity model in the time domain electromagnetic response function. The training unit is used to train the CNN-LSTM hybrid deep learning network model based on the transient electromagnetic response function containing the induced polarization effect. The Cole-Cole key parameters characterizing the induced polarization effect are used as multidimensional input features of the CNN-LSTM hybrid deep learning network model, and the transient electromagnetic response data corresponding to the time channel are used as output. The piecewise weighted mean square error loss function is used to train the CNN-LSTM hybrid deep learning network model.

[0079] In some implementations, the piecewise weighted mean square error loss function in the training unit is: ; Among them, adaptive weight coefficients for: ; in, This refers to the sensitive time window for sign reversal triggered by the induced polarization effect; To determine the threshold at which the response value approaches 0; Basic weights; For higher-order penalty weights in the sign-inverted region; The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of a time-channel; The output of the CNN-LSTM hybrid deep learning network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel.

[0080] In some implementations, the training unit includes: The response characteristic analysis subunit is used to generate n-layered geological models based on the induced polarization effect transient electromagnetic response function and perform forward numerical simulation to obtain the characteristic laws of transient electromagnetic response containing the induced polarization effect. The training data acquisition subunit is used to perform B-spline interpolation on the transient electromagnetic response function of the induced polarization effect, randomly construct a vertically continuous and relatively smooth m-layered geological model, and acquire the training dataset. The training dataset includes Cole-Cole key parameters and electromagnetic response data characterizing the induced polarization effect; m>n, where m and n are both integers greater than 2. The training subunit is used to input the key Cole-Cole parameters characterizing the induced polarization effect into the CNN-LSTM hybrid deep learning network model, and uses the transient electromagnetic response data of the corresponding time channel as the output. The CNN-LSTM hybrid deep learning network model is trained using a piecewise weighted mean square error loss function. Among them, the convolutional neural network model is used to extract the spatial distribution characteristics of the m-layered geological model, and the long short-term memory network is used to learn the time series dependence of the transient electromagnetic response.

[0081] In some implementations, the Cole-Cole complex resistivity model is as follows: ; in, Zero-frequency resistivity; Polarizability; It is a time constant; The frequency correlation coefficient; It is angular frequency.

[0082] In summary, this application has the following advantages compared with the prior art: Significantly improving computational efficiency and providing a fast forward modeling operator for real-time inversion: This application combines a CNN-LSTM hybrid deep learning network to construct an end-to-end nonlinear mapping model from subsurface medium parameters to transient electromagnetic responses. After model training, the computation time for a single forward modeling operation is only in the milliseconds, compared to the several minutes required by traditional numerical methods, representing a computational efficiency improvement of more than three orders of magnitude. This advantage provides a reliable fast forward modeling operator for subsequent real-time, large-scale inversion interpretation.

[0083] Deep integration of induced polarization effect enhances model physical assurance and generalization ability: This application innovatively integrates the induced polarization effect into a data-driven forward modeling framework. Unlike traditional pure resistivity forward modeling, this application explicitly introduces key Cole-Cole parameters (polarizability η, time parameter τ, and frequency correlation coefficient c) to characterize the induced polarization effect in the construction of training data and network input. This design enables the model to be effectively applied to complex geoelectric structures with polarization layers, significantly expanding the application boundaries of deep learning forward modeling methods.

[0084] To address the sign inversion challenge and ensure reliable response prediction across all time periods, this application proposes an effective network learning and optimization method for linear sign inversion in transient electromagnetic responses. By employing a piecewise weighted mean square error loss function during training, higher weights are given to the prediction errors of the sign inversion point and its adjacent time windows, forcing the network to focus on learning the evolution of this key feature. This method effectively overcomes the problem of decreased prediction accuracy caused by the lack of dedicated handling of sign inversion points in conventional methods, significantly improving the model's prediction performance on sign inversion phenomena, thereby ensuring the prediction accuracy and physical consistency of the entire transient electromagnetic response curve from early to late stages.

[0085] This application forms a complete technological closed loop with significant engineering application value: From theory (induced polarization effect modeling), data (multi-parameter sample library), algorithm (CNN-LSTM hybrid network and customized loss function) to application (fast forward modeling operator), this closed loop allows the trained model to achieve instantaneous prediction of unknown models without repeated mesh partitioning and complex numerical calculations, facilitating seamless integration into the inversion process. This method provides a powerful tool for rapidly and accurately interpreting measured data containing induced polarization effects, lowers the computational threshold, and has significant engineering application value.

[0086] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0087] Based on the methods in the above embodiments, this application provides an electronic device / image signal generator / network device / transmitter / terminal / base station / industrial control computer. This electronic device / image signal generator / network device / transmitter / terminal / base station / industrial control computer may include: a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor can call logical instructions in the memory to execute the methods in the above embodiments.

[0088] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0089] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0090] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0091] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0092] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0093] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0094] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A transient electromagnetic forward modeling method based on deep learning considering induced polarization effects, characterized in that, Includes the following steps: The key Cole-Cole parameters characterizing the induced electrostatic effect are input into a CNN-LSTM hybrid deep learning network model for forward modeling to obtain transient electromagnetic response data. The training method for the CNN-LSTM hybrid deep learning network model includes the following steps: Step 1: Construct the frequency domain electromagnetic response function and transform it into the time domain; Step 2: In the time-domain electromagnetic response function, the Cole-Cole complex resistivity model is used to replace the actual resistivity of the formation to obtain the transient electromagnetic response function containing the induced polarization effect. Step 3: Based on the transient electromagnetic response function containing the induced polarization effect, the Cole-Cole key parameters characterizing the induced polarization effect are used as the multidimensional input features of the CNN-LSTM hybrid deep learning network model, and the transient electromagnetic response data corresponding to the time channel are used as the output. The CNN-LSTM hybrid deep learning network model is trained using a piecewise weighted mean square error loss function.

2. The transient electromagnetic forward modeling method according to claim 1, characterized in that, The piecewise weighted mean square error loss function is: ； Among them, adaptive weight coefficients for: ； in, This refers to the sensitive time window for sign reversal triggered by the induced polarization effect; To determine the threshold at which the response value approaches 0; Basic weights; For higher-order penalty weights in the sign-inverted region; The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of a time-channel; The output of the CNN-LSTM hybrid deep learning network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel; For the first A timeline.

3. The transient electromagnetic forward modeling method according to claim 1 or 2, characterized in that, Key Cole-Cole parameters characterizing the induced polarization effect include the resistivity of each layer, polarizability, time parameters, and frequency correlation coefficient.

4. The transient electromagnetic forward modeling method according to claim 1 or 2, characterized in that, Step three specifically includes the following steps: Based on the transient electromagnetic response function of the induced polarization effect, a forward numerical simulation was performed on an n-layered geological model to obtain the characteristic laws of the transient electromagnetic response containing the induced polarization effect. Based on the transient response characteristics of the induced polarization effect, B-spline interpolation is performed on an n-layered geological model to randomly construct a vertically continuous and relatively smooth m-layered geological model and obtain a training dataset. The training dataset includes Cole-Cole key parameters and electromagnetic response data that characterize the induced polarization effect. The key Cole-Cole parameters characterizing the induced polarization effect are input into a CNN-LSTM hybrid deep learning network model. The transient electromagnetic response data corresponding to the time channel are used as the output, and a piecewise weighted mean square error loss function is used to train the CNN-LSTM hybrid deep learning network model. Among them, the convolutional neural network model is used to extract the spatial distribution characteristics of the m-layered geological model, and the long short-term memory network is used to learn the time series dependence of the transient electromagnetic response. m>n, and m and n are both integers greater than 2.

5. The transient electromagnetic forward modeling method according to claim 1, characterized in that, The Cole-Cole complex resistivity model is: ； in, Zero-frequency resistivity; Polarizability; It is a time constant; The frequency correlation coefficient; It is angular frequency.

6. A transient electromagnetic forward modeling system based on deep learning considering induced polarization effects, characterized in that, include: The forward modeling module is used to input the key Cole-Cole parameters characterizing the induced electrostatic effect into the CNN-LSTM hybrid deep learning network model for forward modeling to obtain transient electromagnetic response data. The model training module includes: The first function processing unit is used to construct the frequency domain electromagnetic response function and transform it into the time domain. The second function processing unit is used to obtain the transient electromagnetic response function containing the induced polarization effect by replacing the actual resistivity of the formation with the Cole-Cole complex resistivity model in the time domain electromagnetic response function. The training unit is used to train the CNN-LSTM hybrid deep learning network model based on the transient electromagnetic response function containing the induced polarization effect. The Cole-Cole key parameters characterizing the induced polarization effect are used as multidimensional input features of the CNN-LSTM hybrid deep learning network model, and the transient electromagnetic response data corresponding to the time channel are used as output. The piecewise weighted mean square error loss function is used to train the CNN-LSTM hybrid deep learning network model.

7. The transient electromagnetic forward modeling system according to claim 6, characterized in that, The piecewise weighted mean square error loss function in the training unit is: ； Among them, adaptive weight coefficients for: ； in, This refers to the sensitive time window for sign reversal triggered by the induced polarization effect; To determine the threshold at which the response value approaches 0; Basic weights; For higher-order penalty weights in the sign-inverted region; The total number of time channels in the transient electromagnetic response; For the first The forward modeling true electromagnetic response value of a time-channel; The output of the CNN-LSTM hybrid deep learning network model The predicted response value for each time channel; For the corresponding number Adaptive weighting coefficients for each time channel; For the first A timeline.

8. The transient electromagnetic forward modeling system according to claim 6 or 7, characterized in that, The training unit includes: The response characteristic analysis subunit is used to generate n-layered geological models based on the induced polarization effect transient electromagnetic response function and perform forward numerical simulation to obtain the characteristic laws of transient electromagnetic response containing the induced polarization effect. The training data acquisition subunit is used to perform B-spline interpolation on the transient electromagnetic response function of the induced polarization effect, randomly construct a vertically continuous and relatively smooth m-layered geological model, and acquire the training dataset. The training dataset includes Cole-Cole key parameters and electromagnetic response data characterizing the induced polarization effect; m>n, where m and n are both integers greater than 2. The training subunit is used to input the key Cole-Cole parameters characterizing the induced polarization effect into the CNN-LSTM hybrid deep learning network model, and uses the transient electromagnetic response data of the corresponding time channel as the output. The CNN-LSTM hybrid deep learning network model is trained using a piecewise weighted mean square error loss function. Among them, the convolutional neural network model is used to extract the spatial distribution characteristics of the m-layered geological model, and the long short-term memory network is used to learn the time series dependence of the transient electromagnetic response.

9. The transient electromagnetic forward modeling system according to claim 6 or 7, characterized in that, The Cole-Cole complex resistivity model is: ； in, Zero-frequency resistivity; Polarizability; It is a time constant; The frequency correlation coefficient; It is angular frequency.

10. An electronic device, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1 to 5.

Images