Three-dimensional gravity-magnetic joint inversion method based on deep learning pre-optimization and model fine-tuning

Abstract

This application belongs to the field of geophysical exploration technology, specifically disclosing a three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning. Through this application, after determining the pre-optimized joint inversion network parameters, the pre-optimized three-dimensional gravity and magnetic field joint inversion result is determined based on the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted. The pre-optimized joint inversion network parameters are then adjusted, and the optimal three-dimensional gravity and magnetic field joint inversion result is determined based on the adjusted optimal joint inversion network parameters. By combining network pre-optimization and network parameter fine-tuning, the method retains the generalization ability of the pre-optimized joint inversion network parameters while mitigating the impact of deviations between the observation data to be predicted and the pre-trained features of the dataset. Finally, the optimal three-dimensional gravity and magnetic field joint inversion result is determined based on the adjusted optimal joint inversion network parameters, thereby effectively improving the accuracy and efficiency of the three-dimensional gravity and magnetic field joint inversion.

Description

Technical Field

[0001] This application belongs to the field of geophysical exploration technology, and more specifically, relates to a three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning. Background Technology

[0002] Gravity and magnetic exploration, as important means of geophysical exploration, plays an irreplaceable role in fields such as deep mineral resource prediction, regional geological structure analysis, and engineering geological exploration by measuring surface gravity anomaly observation data and magnetic anomaly observation data to invert the density and magnetization intensity distribution of underground target geological bodies.

[0003] Currently, the common approach for 3D gravity and magnetic field joint inversion is based on model-driven optimization algorithms. However, these algorithms are computationally slow, resulting in low overall inversion efficiency. Furthermore, they are overly reliant on the initial model selection, exhibiting a high degree of ambiguity. To improve the efficiency of 3D gravity and magnetic field joint inversion, data-driven deep learning algorithms have emerged. Research shows that these algorithms use synthetic data during network training, and the features of synthetic data deviate from the feature distribution of measured data, leading to a significant decrease in the accuracy of 3D gravity and magnetic field joint inversion. Therefore, the aforementioned methods offer low accuracy and efficiency for 3D gravity and magnetic field joint inversion. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application aims to provide a three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning, which solves the problem of low accuracy and efficiency of three-dimensional gravity and magnetic field joint inversion due to slow computing speed or bias in network training data features.

[0005] To achieve the above objectives, in a first aspect, this application provides a three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning, comprising:

[0006] The pre-optimized joint inversion network parameters are determined based on the gravity and magnetic dataset and the target multi-branch neural network.

[0007] Based on the pre-optimized joint inversion network parameters, the pre-optimized three-dimensional gravity and magnetic joint inversion results are determined according to the two-dimensional gravity observation data to be predicted and the two-dimensional magnetic observation data to be predicted.

[0008] The multidimensional loss function value is calculated based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results.

[0009] The parameters of the pre-optimized joint inversion network are adjusted based on the multi-dimensional loss function value, and the optimal three-dimensional gravity and magnetic joint inversion result is determined based on the adjusted optimal joint inversion network parameters.

[0010] In one embodiment, the step of determining the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network includes:

[0011] Based on the gravity and magnetic dataset, we obtain the gravity and magnetic training dataset and the gravity and magnetic validation dataset;

[0012] The gravity and magnetic training dataset is divided into batches, and the subsets of gravity and magnetic training data from each batch are input into the target multi-branch neural network for network training.

[0013] The two-dimensional observation data in the gravity and magnetic training dataset is input into the trained target multi-branch neural network, and the three-dimensional physical property model prediction results output by the target multi-branch neural network are obtained.

[0014] Based on the prediction results of the three-dimensional physical property model and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset, the gravity pre-training loss function value and the magnetic pre-training loss function value are calculated respectively.

[0015] The target multi-branch neural network is pre-optimized based on the gravity pre-training loss function value and the magnetic force pre-training loss function value to obtain the pre-optimized joint inversion network parameters.

[0016] In one embodiment, the step of calculating the gravity pre-training loss function value and the magnetic pre-training loss function value based on the prediction results of the three-dimensional physical property model and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset, respectively, includes:

[0017] Extract the three-dimensional density model prediction results of a single batch from the three-dimensional physical property model prediction results, and calculate the first loss function value of the target gravity label corresponding to each of the three-dimensional density model prediction results and the two-dimensional gravity observation data in the gravity and magnetic training dataset.

[0018] The values ​​of each of the first loss functions are summed, and the average of the total number of first loss function values ​​is calculated to obtain the gravity pre-training loss function value.

[0019] Extract the prediction results of the three-dimensional magnetic intensity model for a single batch from the prediction results of the three-dimensional physical property model, and calculate the second loss function value of the target magnetic label corresponding to each prediction result of the three-dimensional magnetic intensity model and the two-dimensional magnetic observation data in the gravity and magnetic training dataset;

[0020] The values ​​of the second loss function are summed, and the average of the total number of second loss function values ​​is calculated to obtain the magnetic pre-training loss function value.

[0021] In one embodiment, the multi-dimensional loss function values ​​include: gravity data fitting loss function values, magnetic data fitting loss function values, structural coupling loss function values, gravity model smoothing loss function values, and magnetic model smoothing loss function values;

[0022] The step of calculating the multidimensional loss function value based on the current two-dimensional gravity observation data, current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results includes:

[0023] Based on the three-dimensional forward modeling strategy of gravity and magnetic geophysics, the current two-dimensional gravity observation data and the current two-dimensional magnetic observation data on the surface of the pre-optimized three-dimensional gravity and magnetic joint inversion results are determined respectively.

[0024] The gravity data fitting loss function value is calculated based on the current two-dimensional gravity observation data and the two-dimensional gravity observation data to be predicted, and the magnetic data fitting loss function value is calculated based on the current two-dimensional magnetic observation data and the two-dimensional magnetic observation data to be predicted.

[0025] Based on the aforementioned pre-optimized three-dimensional gravity and magnetic field joint inversion results, the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results are obtained.

[0026] Calculate the structural coupling loss function values ​​of the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results;

[0027] The gravity model smoothing loss function value is calculated based on the pre-optimized three-dimensional density inversion results, and the magnetic model smoothing loss function value is calculated based on the pre-optimized three-dimensional magnetization inversion results.

[0028] In one embodiment, the step of adjusting the pre-optimized joint inversion network parameters based on the multi-dimensional loss function value, and determining the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters, includes:

[0029] The gravity data fitting loss function value, the structural coupling loss function value, and the gravity model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the gravity model fine-tuning loss function value.

[0030] The magnetic data fitting loss function value, the structural coupling loss function value, and the magnetic model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the magnetic model fine-tuning loss function value.

[0031] Based on the target parameter isolation strategy, the parameters of the pre-optimized joint inversion network are adjusted according to the fine-tuning loss function values ​​of the gravity model and the magnetic model, respectively.

[0032] The two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted are respectively input into the adjusted optimal joint inversion network parameters, and the optimal three-dimensional gravity and magnetic joint inversion results output by the adjusted optimal joint inversion network parameters are obtained.

[0033] In one embodiment, the step of adjusting the parameters of the pre-optimized joint inversion network based on the target parameter isolation strategy, according to the fine-tuning loss function values ​​of the gravity model and the magnetic model, respectively, includes:

[0034] Based on the pre-optimized joint inversion network parameters, the gravity branch network parameters, magnetic branch network parameters, and feature fusion network parameters are obtained;

[0035] Based on the target parameter isolation strategy, the parameters of the gravity branch network are adjusted by backpropagation update according to the fine-tuning loss function value of the gravity model.

[0036] Based on the target parameter isolation strategy, the parameters of the magnetic branch network are adjusted by backpropagation update according to the loss function value of the magnetic model.

[0037] The mean values ​​of the fine-tuning loss function for the gravity model and the fine-tuning loss function for the magnetic model are calculated to obtain the average value of the fine-tuning loss function.

[0038] Based on the target parameter isolation strategy, the parameters of the feature fusion network are adjusted using backpropagation update based on the average value of the fine-tuning loss function.

[0039] Secondly, this application provides a three-dimensional gravity and magnetic field joint inversion device based on deep learning pre-optimization and model fine-tuning, comprising:

[0040] The determination module is used to determine the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network;

[0041] The determining module is further configured to determine the pre-optimized three-dimensional gravity and magnetic joint inversion result based on the pre-optimized joint inversion network parameters and the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted.

[0042] The calculation module is used to calculate the multidimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results.

[0043] The determining module is further configured to adjust the parameters of the pre-optimized joint inversion network based on the multi-dimensional loss function value, and determine the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters.

[0044] 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.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art:

[0049] (1) In the pre-optimization stage, this application uses a synthetic gravity and magnetic data set to train a target multi-branch neural network through deep learning end-to-end to determine the parameters of the pre-optimized joint inversion network. At this time, the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted can be input into the parameters of the pre-optimized joint inversion network to determine the pre-optimized three-dimensional gravity and magnetic joint inversion results. This process, through data-driven deep learning, enables the network to learn a certain generalization ability on the synthetic training data and can efficiently generate joint inversion results based on the newly input gravity and magnetic observation data to be predicted, providing high-quality physical property parameters for the subsequent model fine-tuning of network parameters;

[0050] (2) In the network parameter fine-tuning stage of this application, the pre-optimized three-dimensional gravity and magnetic joint inversion results are used as the initial model, and the pre-optimized joint inversion network parameters are adjusted through multi-dimensional loss function values. By introducing forward modeling physical constraints through model fine-tuning, the optimal joint inversion network parameters are obtained after a certain number of training rounds. At this time, the optimal three-dimensional gravity and magnetic joint inversion results are output by the optimal joint inversion network parameters. The network parameter fine-tuning process does not rely on the three-dimensional physical property model label. It only optimizes the network through three constraints: data fitting, structural coupling, and model smoothing. Based on the pre-optimized inversion results, targeted training is carried out to finally obtain the gravity and magnetic three-dimensional joint inversion results that simultaneously satisfy the constraints of data fitting, structural coupling, and model smoothing. This process, while preserving the generalization performance of the pre-optimized network, alleviates the problem that the inversion results cannot fit the observed data due to the feature distribution shift between the new input data and the training data. The introduction of structural coupling loss function values ​​and model smoothing loss function values ​​reduces the ambiguity of the inversion and constrains the inversion results to avoid spurious shapes, thereby effectively improving the accuracy and efficiency of three-dimensional gravity and magnetic joint inversion.

[0051] (3) The core idea of ​​this application for three-dimensional gravity and magnetic field joint inversion is based on the deep learning technology framework, which unfolds around the same target multi-branch neural network. The pre-optimized joint inversion network parameters can be used repeatedly as pre-optimized inversion parameters for any new input gravity and magnetic field observation data to be predicted. No retraining is required. Only the network parameters need to be fine-tuned for the new input data to obtain the three-dimensional joint inversion results that meet the characteristics of the new data. It can be applied to multiple scenarios and has strong applicability.

[0052] In summary, this application determines the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network; based on the pre-optimized joint inversion network parameters, it determines the pre-optimized three-dimensional gravity and magnetic joint inversion result based on the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted; it calculates the multi-dimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion result; it adjusts the pre-optimized joint inversion network parameters based on the multi-dimensional loss function value, and determines the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters. Through this method, combining network pre-optimization and network parameter fine-tuning, while inheriting the generalization ability of the pre-optimized joint inversion network parameters, it reduces the impact caused by the deviation of the observation data to be predicted from the pre-trained features of the dataset. Then, based on the adjusted optimal joint inversion network parameters, it determines the optimal three-dimensional gravity and magnetic joint inversion result, thereby effectively improving the accuracy and efficiency of three-dimensional gravity and magnetic joint inversion. Attached Figure Description

[0053] Figure 1This is one of the flowcharts of the three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning provided in the embodiments of this application;

[0054] Figure 2(a) is a schematic diagram of the two-dimensional gravity observation data to be predicted provided in an embodiment of this application;

[0055] Figure 2(b) is a schematic diagram of the two-dimensional magnetic observation data to be predicted provided in an embodiment of this application;

[0056] Figure 3(a) is a schematic diagram of the pre-optimized three-dimensional density inversion results provided in the embodiments of this application;

[0057] Figure 3(b) is a schematic diagram of the pre-optimized three-dimensional magnetization inversion results provided in the embodiments of this application;

[0058] Figure 4(a) is a schematic diagram of the current two-dimensional gravity observation data provided in the embodiments of this application;

[0059] Figure 4(b) is a schematic diagram of the current two-dimensional magnetic observation data provided in the embodiments of this application;

[0060] Figure 5(a) is a schematic diagram of the optimal three-dimensional density inversion result provided in the embodiment of this application;

[0061] Figure 5(b) is a schematic diagram of the optimal three-dimensional magnetization inversion result provided in the embodiment of this application;

[0062] Figure 6 This is the second flowchart of the three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning provided in the embodiments of this application;

[0063] Figure 7 This is a schematic diagram of the module structure of the three-dimensional gravity and magnetic field joint inversion device based on deep learning pre-optimization and model fine-tuning provided in the embodiments of this application;

[0064] Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0065] 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.

[0066] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0067] The terms "first" and "second," etc., used in the specification and claims herein are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0068] 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.

[0069] Based on this, embodiments of this application provide a three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning, referring to... Figure 1 , Figure 1 This is one of the flowcharts illustrating the three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning provided in this application embodiment. In this embodiment, the three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning includes steps S10 to S40:

[0070] Step S10: Determine the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network.

[0071] It should be noted that the gravity and magnetic dataset refers to the dataset synthesized during the pre-optimization stage of the network parameters. This dataset includes, but is not limited to, a complex three-dimensional density property model synthesized based on a random walk algorithm and its two-dimensional gravity observation data, and a complex three-dimensional magnetic intensity property model synthesized based on a random walk algorithm and its two-dimensional magnetic observation data. The three-dimensional property model can characterize the labels of the two-dimensional observation data. A morphologically continuous and stochastic three-dimensional density model and a three-dimensional magnetization model can be constructed in a three-dimensional grid space using a random walk algorithm. The two-dimensional gravity observation data and the two-dimensional magnetic observation data are obtained from the three-dimensional density model and the three-dimensional magnetization model through geophysical gravity forward modeling and geophysical magnetic forward modeling, respectively. The pre-optimized joint inversion network parameters refer to the network parameters used to determine and output the pre-optimized three-dimensional gravity and magnetic joint inversion results. These pre-optimized joint inversion network parameters can be repeatedly used as pre-optimized inversion parameters for any newly input gravity and magnetic observation data to be predicted, without the need for repeated training.

[0072] Understandably, the structure of the target multi-branch neural network includes: a gravity branch input module, a magnetic branch input module, a gravity branch initial convolutional unit, a magnetic branch initial convolutional unit, a gravity branch multi-level downsampling and convolutional unit, a magnetic branch multi-level downsampling and convolutional unit, a feature fusion module, a gravity branch multi-level upsampling and convolutional unit, a magnetic branch multi-level upsampling and convolutional unit, a gravity branch output module, and a magnetic branch output module. The gravity branch input module and the magnetic branch input module receive gravity and magnetic data of the same dimension as network input. The initial convolutional unit of the gravity branch is connected to the gravity branch input module and includes a sequentially connected 2D convolutional layer, a first 3D convolutional layer, and a second 3D convolutional layer. This unit performs convolution processing on the input from the gravity branch, adding a depth dimension and transforming the original 2D gravity input features into 3D features. The initial convolutional unit of the magnetic branch is also connected to the magnetic branch input module and has the same structure as the initial convolutional unit of the gravity branch. It performs convolution processing on the input from the magnetic branch, adding a depth dimension and transforming the original 2D magnetic input features into 3D features. The multi-level downsampling and convolutional unit of the gravity branch is connected to the initial convolutional unit of the gravity branch and includes multiple sequentially connected downsampling sub-units. Each downsampling and convolutional sub-unit contains a downsampling layer, several subsequently connected 3D convolutional layers, and a feature addition layer. This unit performs downsampling, convolution, and feature fusion processing on the gravity data. The multi-level downsampling and convolutional unit of the magnetic branch... The first module connects to the initial convolutional unit of the magnetic branch, with the same structure as the multi-level downsampling and convolutional unit of the gravity branch. It is used for downsampling, convolution, and feature fusion of magnetic data. The second module connects to the multi-level downsampling and convolutional units of the gravity and magnetic branches. It includes a combination of sequentially connected feature splicing layers, downsampling layers, multiple consecutive convolutional layers, feature addition layers, and feature decoupling layers. It is used for fusing gravity and magnetic data, collaborative feature extraction, and feature decoupling. The third module connects to the multi-level upsampling and convolutional unit of the gravity branch and the feature fusion module. It includes multiple sequentially connected upsampling sub-units. Each upsampling and convolutional sub-unit contains an upsampling layer, several subsequently connected 3D convolutional layers, and feature addition layers. It is used for upsampling and cross-level fusion of gravity and magnetic data to restore the feature map dimension. The fourth module connects to the multi-level upsampling and convolutional unit of the magnetic branch. It has the same structure as the multi-level upsampling and convolutional unit of the gravity branch. It is used for upsampling and cross-level fusion of gravity and magnetic data to restore the feature map dimension.

[0073] It should be noted that, in addition to the above, it may also include: Task Encoder 1, Task Encoder 2, Task Decoder 1, and Task Decoder 2; wherein, the internal structure of Task Encoder 1 is interconnected in the order of data passage, namely: input layer, 1 two-dimensional convolutional layer, 1 three-dimensional convolutional layer, 1 feature addition layer, 1 downsampling layer, 2 consecutive three-dimensional convolutional layers, 1 feature addition layer, 1 downsampling layer, 3 consecutive three-dimensional convolutional layers, 1 feature addition layer, and 1 downsampling layer; the internal structure of the feature fusion module is interconnected in the order of data passage, namely: 1 feature concatenation layer, 1 downsampling layer, 3 consecutive three-dimensional convolutional layers, 1 feature addition layer, and 1 downsampling layer; the internal structure of Task Decoder 1 is interconnected in the order of data passage, namely: 1 upsampling layer, 3 consecutive three-dimensional convolutional layers, 1 residual module, 1 upsampling layer, 2 consecutive three-dimensional convolutional layers, 1 downsampling layer ... The system consists of a convolutional layer, a feature summing layer, an upsampling layer, a 3D convolutional layer, a feature summing layer, a 1×1×1 3D convolutional layer, and an output layer. Task encoder 2 and task encoder 1 have the same structure, as do task decoder 2 and task decoder 1. Between task encoder 1 and task decoder 1, there is a fine-grained feature skip connection mechanism after each feature summing layer in the encoder, where fine-grained features are concatenated to the feature data processed by each upsampling layer in task decoder 1. Task encoder 2 and task decoder 2 have the same skip connection mechanism as task encoder 1 and task decoder 1. Task encoder 1 and task decoder 1 are used to receive and process gravity feature data independently, while task encoder 2 and task decoder 2 are used to receive and process magnetic feature data independently. The feature fusion module is used for high-dimensional global feature fusion and processing.

[0074] Further, step S10 includes: obtaining a gravity and magnetic training dataset and a gravity and magnetic verification dataset based on the gravity and magnetic dataset; dividing the gravity and magnetic training dataset into batches and inputting the subsets of gravity and magnetic training data from each batch into a target multi-branch neural network for network training; inputting the two-dimensional observation data in the gravity and magnetic training dataset into the trained target multi-branch neural network and obtaining the three-dimensional physical property model prediction results output by the target multi-branch neural network; calculating the gravity pre-training loss function value and the magnetic pre-training loss function value based on the three-dimensional physical property model prediction results and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset; and pre-optimizing the target multi-branch neural network based on the gravity pre-training loss function value and the magnetic pre-training loss function value to obtain the pre-optimized joint inversion network parameters.

[0075] It should be understood that after obtaining the gravity and magnetic dataset, it can be divided into two parts, with a ratio of 7:3. For example, using a random walk algorithm, 512 different morphological continuous and random 3D property models were constructed in a 3D mesh space. Each morphological random model generated 150 sub-models at different spatial locations. The number of mesh subdivisions in the 3D mesh space was 56 × 64 × 24 = 86016. The ground observation point mesh size was 56 × 64, and the mesh spacing was 300 m. The residual density value of the 3D density model was set to 0.6 g / cm³. 3 The magnetization value of the three-dimensional magnetic intensity model is 27 A / m. Based on the geophysical gravity forward modeling and magnetic forward modeling formulas, the three-dimensional physical property model is simulated by forward modeling, generating two-dimensional gravity observation data and two-dimensional magnetic observation data at the ground. The training set and validation set are divided into training set and validation set with a ratio of 7:3, resulting in a gravity training dataset with a sample size of 53,760, a magnetic training dataset with a sample size of 53,760, a gravity validation dataset with a sample size of 23,040, and a magnetic validation dataset with a sample size of 23,040.

[0076] It should be noted that, since the gravity and magnetic training dataset needs to be input into the target multi-branch neural network in batches for network training, the dataset needs to be batch-divided. The network training parameters can be: batch size N=16, upper limit of training epochs set to 60, using the Adam optimizer with an initial learning rate of 1×10⁻², and setting the learning rate to decay to 0.1 times its original size every 20 training epochs. Furthermore, to avoid the impact of the difference in magnitude between gravity and magnetic observation data on network training, mean-variance standardization needs to be performed on the gravity and magnetic training dataset input to the network. This standardization should be performed on the gravity and magnetic training dataset, and the mean and variance of the two-dimensional gravity and magnetic observation data in the dataset should be recorded separately. After the mean-variance standardization is completed, the N gravity two-dimensional observation data and N magnetic two-dimensional observation data are simultaneously input into the two encoders of the target multi-branch neural network. The two decoders then output the corresponding N three-dimensional density model and N three-dimensional magnetization model prediction results, which are the three-dimensional property model prediction results. After obtaining the gravity pre-training loss function value and the magnetic pre-training loss function value in each round of training, the target parameter isolation strategy can be used to pre-optimize the parameters of different modules of the target multi-branch neural network to obtain the pre-optimized joint inversion network parameters.

[0077] Further, the step of calculating the gravity pre-training loss function value and the magnetic pre-training loss function value based on the prediction results of the three-dimensional physical property model and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset includes: extracting the prediction results of a single batch of three-dimensional density models from the prediction results of the three-dimensional physical property model, and calculating the first loss function value of the target gravity label corresponding to each prediction result of the three-dimensional density model and the two-dimensional gravity observation data in the gravity and magnetic training dataset; summing the first loss function values ​​and averaging the total number of first loss function values ​​to obtain the gravity pre-training loss function value; extracting the prediction results of a single batch of three-dimensional magnetic intensity models from the prediction results of the three-dimensional physical property model, and calculating the second loss function value of the target magnetic label corresponding to each prediction result of the three-dimensional magnetic intensity model and the two-dimensional magnetic observation data in the gravity and magnetic training dataset; summing the second loss function values ​​and averaging the total number of second loss function values ​​to obtain the magnetic pre-training loss function value.

[0078] Understandably, after extracting the prediction results of a single batch of three-dimensional density models from the prediction results of the three-dimensional physical property models, the first loss function value of the target gravity label corresponding to each three-dimensional density model prediction result and the two-dimensional gravity observation data in the gravity and magnetic training dataset is calculated, and the second loss function value of the target magnetic label corresponding to each three-dimensional magnetic intensity model prediction result and the two-dimensional magnetic observation data in the gravity and magnetic training dataset is calculated. The deep learning pre-optimized loss function used to calculate the above loss function values ​​can be:

[0079]

[0080] in, This indicates the batch size, which is the number of samples used in a single training session. This represents the total number of elements after flattening a single sample. To represent an extremely small quantity, used to avoid the numerator or denominator being 0. This represents a batch dimension index, with values ​​ranging from 1 to... Used to iterate through each training sample. For flattening index of a single sample space dimension, the value ranges from 1 to... This is used to iterate through each element of a single sample after it has been flattened.

[0081] It should be noted that, to facilitate the calculation of the aforementioned loss function values, a mask of the prediction results from the 3D physical property model can be used in the calculation. This mask can be obtained by processing the Sigmoid function. For example, when calculating the first loss function value, A mask representing the prediction results of a three-dimensional density model. This represents a mask indicating the target gravity label corresponding to the two-dimensional gravity observation data in the gravity and magnetic training dataset; when calculating the second loss function, A mask representing the prediction results of a three-dimensional magnetic field strength model. The mask representing the target magnetic label corresponding to the two-dimensional magnetic observation data in the gravity and magnetic training dataset.

[0082] It should be understood that after calculating the first loss function value, the individual first loss function values ​​are summed, and the average of the total first loss function values ​​is calculated to obtain the gravity pre-training loss function value. This gravity pre-training loss function value is then used as the gravity pre-optimization loss function value for a batch. Each training round requires inputting all data from the gravity training dataset into the network through 53760 / N batches, with the gravity pre-optimization loss function value for each round being the average of the 53760 / N batches. Similarly, after calculating the second loss function value, the individual second loss function values ​​are summed, and the average of the total second loss function values ​​is calculated to obtain the magnetic pre-training loss function value. This magnetic pre-training loss function value is then used as the magnetic pre-optimization loss function value for a batch. Each training round requires inputting all data from the magnetic training dataset into the network through 53760 / N batches, with the magnetic pre-optimization loss function value for each round being the average of the 53760 / N batches.

[0083] Step S20: Based on the pre-optimized joint inversion network parameters, determine the pre-optimized three-dimensional gravity and magnetic joint inversion results according to the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted.

[0084] Understandably, after obtaining the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted, a mean-variance standardization operation will be performed on the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted. The standardized two-dimensional gravity and magnetic observation data will be used as a set of inputs to the pre-optimized joint inversion network parameters and input into the two task encoders of the network to obtain the pre-optimized three-dimensional gravity and magnetic joint inversion results. The two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted can be a set of measured observation data of the work area. For details, please refer to Figure 2(a) and Figure 2(b). Figure 2(a) shows a schematic diagram of the two-dimensional gravity observation data to be predicted, and Figure 2(b) shows a schematic diagram of the two-dimensional magnetic observation data to be predicted. The horizontal axis represents the east direction and the vertical axis represents the north direction. The pre-optimized three-dimensional density inversion result determined based on the pre-optimized joint inversion network parameters can be referred to Figure 3(a). The horizontal axis represents the east direction, the vertical axis represents the north direction, and the vertical axis represents the depth. The determined pre-optimized three-dimensional magnetization inversion result can be referred to Figure 3(b). The horizontal axis represents the east direction, the vertical axis represents the north direction, and the vertical axis represents the magnetization capacity.

[0085] Step S30: Calculate the multidimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results.

[0086] It should be understood that the multi-dimensional loss function values ​​include the loss function values ​​of the gravity dimension, the loss function values ​​of the magnetic dimension, and the loss function values ​​of the structural coupling dimension. Among them, the loss function values ​​of the gravity dimension include, but are not limited to, the loss function values ​​of gravity data fitting and the loss function values ​​of gravity model smoothing, etc., and the loss function values ​​of the magnetic dimension include, but are not limited to, the loss function values ​​of magnetic data fitting and the loss function values ​​of magnetic model smoothing, etc.

[0087] Further, the multi-dimensional loss function values ​​include: gravity data fitting loss function values, magnetic data fitting loss function values, structural coupling loss function values, gravity model smoothing loss function values, and magnetic model smoothing loss function values; step S30 includes: determining the current two-dimensional gravity observation data and the current two-dimensional magnetic observation data on the surface based on the pre-optimized three-dimensional gravity and magnetic joint inversion results according to the gravity and magnetic geophysical three-dimensional forward modeling strategy; calculating the gravity data fitting loss function value based on the current two-dimensional gravity observation data and the two-dimensional gravity observation data to be predicted, and calculating the magnetic data fitting loss function value based on the current two-dimensional magnetic observation data and the two-dimensional magnetic observation data to be predicted; obtaining the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results based on the pre-optimized three-dimensional gravity and magnetic joint inversion results; calculating the structural coupling loss function value of the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results; calculating the gravity model smoothing loss function value based on the pre-optimized three-dimensional density inversion results, and calculating the magnetic model smoothing loss function value based on the pre-optimized three-dimensional magnetization inversion results.

[0088] Understandably, after obtaining the pre-optimized three-dimensional gravity and magnetic joint inversion results, the current two-dimensional gravity observation data and current two-dimensional magnetic observation data on the surface of the pre-optimized three-dimensional gravity and magnetic joint inversion results are determined based on the three-dimensional forward modeling strategy of gravity and magnetic geophysics. The current two-dimensional gravity observation data on the surface of the pre-optimized three-dimensional gravity and magnetic joint inversion results can be referred to Figure 4(a), and the current two-dimensional magnetic observation data on the surface of the pre-optimized three-dimensional gravity and magnetic joint inversion results can be referred to Figure 4(b).

[0089] It should be noted that, upon obtaining the current two-dimensional gravity observation data and the current two-dimensional magnetic observation data, the gravity data fitting loss function value can be calculated based on the current two-dimensional gravity observation data and the two-dimensional gravity observation data to be predicted, and the magnetic data fitting loss function value can be calculated based on the current two-dimensional magnetic observation data and the two-dimensional magnetic observation data to be predicted. The data fitting loss function used to calculate the gravity data fitting loss function value and the magnetic data fitting loss function value can be:

[0090]

[0091]

[0092] in, Indicates the number of rows in the two-dimensional observation data. This indicates the number of columns in the two-dimensional observation data.

[0093] It is important to emphasize that, to facilitate the calculation of the aforementioned loss function values, a mask of the currently observed data can be used in the calculation. For example, when calculating the fitting loss function value for gravity data, This represents the gravity forward modeling sensitivity matrix. This represents the current two-dimensional gravity observation data. This represents the two-dimensional gravity observation data to be predicted; when calculating the fitting loss function value of the magnetic data, This represents the forward modeling sensitivity matrix of the magnetic force. This represents the current two-dimensional magnetic observation data. This represents the two-dimensional magnetic observation data to be predicted.

[0094] It should be understood that, in addition to calculating the weight data fitting loss function and the magnetic data fitting loss function, the structural coupling loss function values ​​for the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results will also be calculated. The structural coupling loss function used to calculate these values ​​can be:

[0095]

[0096] in, The mask representing the pre-optimized 3D density inversion result. The mask representing the pre-optimized three-dimensional magnetization inversion result. To represent an extremely small quantity, used to avoid the numerator or denominator being 0. This indicates the batch size, which is the number of samples used in a single training session. This represents the total number of elements after flattening a single sample.

[0097] It is understandable that, in addition to calculating the weight data fitting loss function, magnetic data fitting loss function, and structural coupling loss function, the weight model smoothing loss function will also be calculated based on the pre-optimized 3D density inversion results, and the magnetic model smoothing loss function will be calculated based on the pre-optimized 3D magnetization inversion results. The model smoothing loss functions used to calculate the weight model smoothing loss function and the magnetic model smoothing loss function can be:

[0098]

[0099] in, The number of grid cells in the depth direction represents the mask used to represent the prediction results of the pre-optimized 3D model. This represents the number of grid cells in the height direction of the mask used to represent the prediction results of the pre-optimized 3D model. The mask representing the prediction results of the pre-optimized 3D model is located in... The mask representing the prediction results of the pre-optimized 3D model is a discrete grid space difference in the depth direction. The mask representing the prediction results of the pre-optimized 3D model is a discrete grid space difference in the height direction. The mask representing the prediction result of the pre-optimized 3D model is a discrete grid space difference in the width direction.

[0100] It is important to emphasize that when calculating the smoothing loss function value of the gravity model, This represents the mask used to represent the pre-optimized 3D density inversion results; it is used when calculating the smoothing loss function value of the magnetic model. The mask representing the pre-optimized three-dimensional magnetization inversion result.

[0101] Step S40: Adjust the parameters of the pre-optimized joint inversion network according to the multi-dimensional loss function value, and determine the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters.

[0102] It should be noted that after obtaining the multi-dimensional loss function value, the pre-optimized joint inversion network parameters can be fine-tuned based on the target parameter isolation strategy. The adjusted joint inversion network parameters are the optimal joint inversion network parameters. At this time, the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted can be input into the adjusted optimal joint inversion network parameters respectively, and the optimal three-dimensional gravity and magnetic joint inversion results output by the adjusted optimal joint inversion network parameters can be obtained. The optimal three-dimensional gravity and magnetic joint inversion results include the optimal three-dimensional density inversion results and the optimal three-dimensional magnetization inversion results. The optimal three-dimensional density inversion results can be referred to in Figure 5(a), and the optimal three-dimensional magnetization inversion results can be referred to in Figure 5(b). By comparison, it can be seen that this embodiment effectively reduces the data fitting difference of the inversion results by fine-tuning the network parameters, making the inversion results more physically meaningful.

[0103] This embodiment determines the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network; based on the pre-optimized joint inversion network parameters, it determines the pre-optimized three-dimensional gravity and magnetic joint inversion result according to the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted; it calculates the multi-dimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion result; it adjusts the pre-optimized joint inversion network parameters according to the multi-dimensional loss function value, and determines the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters. Through the above method, combining network pre-optimization and network parameter fine-tuning, while inheriting the generalization ability of the pre-optimized joint inversion network parameters, it reduces the impact caused by the deviation of the observation data to be predicted from the pre-trained features of the dataset. Then, based on the adjusted optimal joint inversion network parameters, it determines the optimal three-dimensional gravity and magnetic joint inversion result, thereby effectively improving the accuracy and efficiency of three-dimensional gravity and magnetic joint inversion.

[0104] In one specific implementation, this application provides steps for determining the optimal three-dimensional gravity and magnetic field joint inversion results. Please refer to... Figure 6 , Figure 6 This is the second flowchart illustrating the three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning provided in this application embodiment. Step S40 includes steps S401 to S404:

[0105] Step S401: The gravity data fitting loss function value, the structural coupling loss function value, and the gravity model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the gravity model fine-tuning loss function value.

[0106] It should be noted that after obtaining the multi-dimensional loss function values, the gravity model fine-tuning loss function value can be calculated by weighted summation based on the gravity data fitting loss function value, the structural coupling loss function value, and the gravity model smoothing loss function value. The model fine-tuning loss function used to calculate the gravity model fine-tuning loss function value can be:

[0107]

[0108] in, This represents the value of the loss function for fine-tuning the gravity model. This represents the value of the loss function used to fit the gravity data. The weights represent the values ​​of the loss function used to fit the gravity data. This represents the value of the structural coupling loss function. The weights represent the values ​​of the structural coupling loss function. This represents the smoothing loss function value of the gravity model. The weights represent the values ​​of the smoothing loss function for the gravity model.

[0109] Step S402: The magnetic data fitting loss function value, structural coupling loss function value, and magnetic model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the magnetic model fine-tuning loss function value.

[0110] It is understandable that, after obtaining the multi-dimensional loss function values, the magnetic model fine-tuning loss function value can be calculated by weighted summation based on the magnetic data fitting loss function value, the structural coupling loss function value, and the magnetic model smoothing loss function value. The model fine-tuning loss function used to calculate the magnetic model fine-tuning loss function value can be:

[0111]

[0112] in, This represents the value of the loss function for fine-tuning the magnetic model. This represents the value of the loss function used to fit the magnetic data. The weights represent the values ​​of the loss function for fitting magnetic data. This represents the value of the structural coupling loss function. The weights represent the values ​​of the structural coupling loss function. This represents the smoothing loss function value of the magnetic model. The weights represent the values ​​of the smoothing loss function for the magnetic model.

[0113] Step S403: Based on the target parameter isolation strategy, the parameters of the pre-optimized joint inversion network are adjusted according to the fine-tuning loss function values ​​of the gravity model and the magnetic model, respectively.

[0114] It should be understood that the target parameter isolation strategy refers to adjusting the pre-optimized joint inversion network parameters separately in an isolated manner. After obtaining the gravity model fine-tuning loss function value and the magnetic model fine-tuning loss function value in each round of model fine-tuning, the pre-optimized joint inversion network parameters are fine-tuned separately based on the gravity model fine-tuning loss function value and the magnetic model fine-tuning loss function value. The upper limit of the number of training rounds is set to 20000, and the initial learning rate is 1×10. -6 The Adam optimizer is used, and the learning rate is set to decay to 0.5 times its original size every 200 training epochs.

[0115] Further, step S403 includes: obtaining gravity branch network parameters, magnetic branch network parameters, and feature fusion network parameters based on the pre-optimized joint inversion network parameters; adjusting the gravity branch network parameters using backpropagation update based on the gravity model fine-tuning loss function value according to the target parameter isolation strategy; adjusting the magnetic branch network parameters using backpropagation update based on the magnetic model fine-tuning loss function value according to the target parameter isolation strategy; calculating the mean of the gravity model fine-tuning loss function value and the magnetic model fine-tuning loss function value to obtain the average value of the fine-tuning loss function; and adjusting the feature fusion network parameters using backpropagation update based on the average value of the fine-tuning loss function according to the target parameter isolation strategy.

[0116] It is understood that, for this embodiment, the pre-optimized joint inversion network parameters include, but are not limited to, gravity branch network parameters, magnetic branch network parameters, and feature fusion network parameters. The feature fusion network parameters refer to the network parameters of the feature fusion module. This feature fusion module is connected to the gravity branch multi-level downsampling and convolution unit and the magnetic branch multi-level downsampling and convolution unit, and includes a combination of a feature splicing layer, a downsampling layer, multiple consecutive convolutional layers, a feature addition layer, and a feature decoupling layer connected in sequence, used to fuse gravity and magnetic feature maps, perform collaborative feature extraction, and decouple features.

[0117] It should be noted that after obtaining the fine-tuning loss function values ​​of the gravity model and the magnetic model, the average value of the fine-tuning loss function can be obtained by averaging. At this point, a target parameter isolation strategy can be adopted to adjust the parameters of the gravity branch network, the magnetic branch network, and the feature fusion network respectively through backpropagation. That is, the parameters of the gravity branch network are updated through backpropagation using the fine-tuning loss function value of the gravity model, the parameters of the magnetic branch network are updated through backpropagation using the fine-tuning loss function value of the magnetic model, and the parameters of the feature fusion network are updated through backpropagation using the average value of the fine-tuning loss function.

[0118] Step S404: Input the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted into the adjusted optimal joint inversion network parameters, and obtain the optimal three-dimensional gravity and magnetic joint inversion result output by the adjusted optimal joint inversion network parameters.

[0119] Understandably, the adjusted joint inversion network parameters are the optimal joint inversion network parameters. At this point, the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted can be input into the adjusted optimal joint inversion network parameters respectively. The output of the adjusted optimal joint inversion network parameters is the optimal three-dimensional gravity and magnetic joint inversion result. This can inherit the generalization ability of the pre-optimized joint inversion network parameters, reduce the impact caused by the deviation of the observation data to be predicted from the pre-trained features of the dataset, and effectively reduce the multi-solution of the model, thus meeting the actual geophysical exploration requirements of "high precision, high efficiency, and strong applicability".

[0120] In this embodiment, the gravity data fitting loss function value, structural coupling loss function value, and gravity model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the gravity model fine-tuning loss function value; the magnetic data fitting loss function value, structural coupling loss function value, and magnetic model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the magnetic model fine-tuning loss function value; based on the target parameter isolation strategy, the parameters of the pre-optimized joint inversion network are adjusted according to the gravity model fine-tuning loss function value and the magnetic model fine-tuning loss function value; the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted are respectively input into the adjusted optimal joint inversion network parameters, and the optimal three-dimensional gravity and magnetic joint inversion result output by the adjusted optimal joint inversion network parameters is obtained. By using the above method, after obtaining the multi-dimensional loss function, the fine-tuning loss function values ​​of the gravity model and the magnetic model are calculated separately by weighted summation. Then, the optimal joint inversion network parameters are fine-tuned based on the fine-tuning loss function values ​​of the gravity model and the magnetic model. The optimal joint inversion network parameters output the optimal joint inversion results corresponding to the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted. This can effectively improve the accuracy and efficiency of three-dimensional gravity and magnetic joint inversion, effectively reduce the multiple solutions of the model, and meet the actual geophysical exploration requirements of "high precision, high efficiency, and strong applicability".

[0121] The following describes the three-dimensional gravity and magnetic field joint inversion device based on deep learning pre-optimization and model fine-tuning provided in this application. The three-dimensional gravity and magnetic field joint inversion device described below corresponds to the three-dimensional gravity and magnetic field joint inversion method described above. Please refer to... Figure 7 , Figure 7 This is a schematic diagram of the module structure of the three-dimensional gravity and magnetic field joint inversion device based on deep learning pre-optimization and model fine-tuning provided in the embodiments of this application, including:

[0122] The determination module T10 is used to determine the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network.

[0123] The determining module T10 is also used to determine the pre-optimized three-dimensional gravity and magnetic joint inversion result based on the pre-optimized joint inversion network parameters and the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted.

[0124] The calculation module T20 is used to calculate the multidimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results.

[0125] The determining module T10 is also used to adjust the parameters of the pre-optimized joint inversion network according to the multi-dimensional loss function value, and to determine the optimal three-dimensional gravity and magnetic joint inversion result according to the adjusted optimal joint inversion network parameters.

[0126] This embodiment determines the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network; based on the pre-optimized joint inversion network parameters, it determines the pre-optimized three-dimensional gravity and magnetic joint inversion result according to the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted; it calculates the multi-dimensional loss function value based on the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion result; it adjusts the pre-optimized joint inversion network parameters according to the multi-dimensional loss function value, and determines the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters. Through the above method, combining network pre-optimization and network parameter fine-tuning, while inheriting the generalization ability of the pre-optimized joint inversion network parameters, it reduces the impact caused by the deviation of the observation data to be predicted from the pre-trained features of the dataset. Then, based on the adjusted optimal joint inversion network parameters, it determines the optimal three-dimensional gravity and magnetic joint inversion result, thereby effectively improving the accuracy and efficiency of three-dimensional gravity and magnetic joint inversion.

[0127] It is understood that the detailed functional implementation of each of the above modules can be found in the description of the aforementioned method embodiments, and will not be repeated here.

[0128] 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.

[0129] Based on the methods in the above embodiments, this application provides an electronic device, please refer to... Figure 8 , Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application.

[0130] It should be noted that the system may include: a processor 10, a communications interface 20, a memory 30, and a communication bus 40. The processor 10, communications interface 20, and memory 30 communicate with each other via the communication bus 40. The processor 10 can invoke logical instructions stored in the memory 30 to execute the methods described in the above embodiments.

[0131] Furthermore, the logical instructions in the aforementioned memory 30 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0132] 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.

[0133] 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.

[0134] It is understood that the processor in the embodiments of this application can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays, 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.

[0135] 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, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, registers, hard disks, portable hard disks, 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.

[0136] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application. Those skilled in the art will readily understand that the above descriptions are merely preferred embodiments of this application and are 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 protection scope of this application.

Claims

1. A three-dimensional gravity and magnetic field joint inversion method based on deep learning pre-optimization and model fine-tuning, characterized in that, include: The pre-optimized joint inversion network parameters are determined based on the gravity and magnetic dataset and the target multi-branch neural network. Based on the pre-optimized joint inversion network parameters, the pre-optimized three-dimensional gravity and magnetic joint inversion results are determined according to the two-dimensional gravity observation data to be predicted and the two-dimensional magnetic observation data to be predicted. Based on the pre-optimized three-dimensional gravity and magnetic joint inversion results, multi-dimensional loss function values ​​are calculated using the current two-dimensional gravity observation data, the current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface. The multi-dimensional loss function values ​​include: gravity data fitting loss function value, magnetic data fitting loss function value, structural coupling loss function value, gravity model smoothing loss function value, and magnetic model smoothing loss function value. The parameters of the pre-optimized joint inversion network are adjusted based on the multi-dimensional loss function value, and the optimal three-dimensional gravity and magnetic joint inversion result is determined based on the adjusted optimal joint inversion network parameters.

2. The method as described in claim 1, characterized in that, The step of determining the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network includes: Based on the gravity and magnetic dataset, we obtain the gravity and magnetic training dataset and the gravity and magnetic validation dataset; The gravity and magnetic training dataset is divided into batches, and the subsets of gravity and magnetic training data from each batch are input into the target multi-branch neural network for network training. The two-dimensional observation data in the gravity and magnetic training dataset is input into the trained target multi-branch neural network, and the three-dimensional physical property model prediction results output by the target multi-branch neural network are obtained. Based on the prediction results of the three-dimensional physical property model and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset, the gravity pre-training loss function value and the magnetic pre-training loss function value are calculated respectively. The target multi-branch neural network is pre-optimized based on the gravity pre-training loss function value and the magnetic force pre-training loss function value to obtain the pre-optimized joint inversion network parameters.

3. The method as described in claim 2, characterized in that, The step of calculating the gravity pre-training loss function value and the magnetic pre-training loss function value based on the prediction results of the three-dimensional physical property model and the target labels corresponding to the two-dimensional observation data in the gravity and magnetic training dataset includes: Extract the three-dimensional density model prediction results of a single batch from the three-dimensional physical property model prediction results, and calculate the first loss function value of the target gravity label corresponding to each of the three-dimensional density model prediction results and the two-dimensional gravity observation data in the gravity and magnetic training dataset. The values ​​of each of the first loss functions are summed, and the average of the total number of first loss function values ​​is calculated to obtain the gravity pre-training loss function value. Extract the prediction results of the three-dimensional magnetic intensity model for a single batch from the prediction results of the three-dimensional physical property model, and calculate the second loss function value of the target magnetic label corresponding to each prediction result of the three-dimensional magnetic intensity model and the two-dimensional magnetic observation data in the gravity and magnetic training dataset; The values ​​of the second loss function are summed, and the average of the total number of second loss function values ​​is calculated to obtain the magnetic pre-training loss function value.

4. The method as described in claim 1, characterized in that, The step of calculating the multidimensional loss function value based on the current two-dimensional gravity observation data, current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted on the Earth's surface according to the pre-optimized three-dimensional gravity and magnetic joint inversion results includes: Based on the three-dimensional forward modeling strategy of gravity and magnetic geophysics, the current two-dimensional gravity observation data and the current two-dimensional magnetic observation data on the surface of the pre-optimized three-dimensional gravity and magnetic joint inversion results are determined respectively. The gravity data fitting loss function value is calculated based on the current two-dimensional gravity observation data and the two-dimensional gravity observation data to be predicted, and the magnetic data fitting loss function value is calculated based on the current two-dimensional magnetic observation data and the two-dimensional magnetic observation data to be predicted. Based on the aforementioned pre-optimized three-dimensional gravity and magnetic field joint inversion results, the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results are obtained. Calculate the structural coupling loss function values ​​of the pre-optimized three-dimensional density inversion results and the pre-optimized three-dimensional magnetization inversion results; The gravity model smoothing loss function value is calculated based on the pre-optimized three-dimensional density inversion results, and the magnetic model smoothing loss function value is calculated based on the pre-optimized three-dimensional magnetization inversion results.

5. The method according to any one of claims 1 to 4, characterized in that, The step of adjusting the pre-optimized joint inversion network parameters based on the multi-dimensional loss function value, and determining the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters, includes: The gravity data fitting loss function value, the structural coupling loss function value, and the gravity model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the gravity model fine-tuning loss function value. The magnetic data fitting loss function value, the structural coupling loss function value, and the magnetic model smoothing loss function value in the multi-dimensional loss function value are weighted and summed to obtain the magnetic model fine-tuning loss function value. Based on the target parameter isolation strategy, the parameters of the pre-optimized joint inversion network are adjusted according to the fine-tuning loss function values ​​of the gravity model and the magnetic model, respectively. The two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted are respectively input into the adjusted optimal joint inversion network parameters, and the optimal three-dimensional gravity and magnetic joint inversion results output by the adjusted optimal joint inversion network parameters are obtained.

6. The method as described in claim 5, characterized in that, The step of adjusting the parameters of the pre-optimized joint inversion network based on the target parameter isolation strategy, according to the fine-tuning loss function values ​​of the gravity model and the magnetic model, includes: Based on the pre-optimized joint inversion network parameters, the gravity branch network parameters, magnetic branch network parameters, and feature fusion network parameters are obtained; Based on the target parameter isolation strategy, the parameters of the gravity branch network are adjusted by backpropagation update according to the fine-tuning loss function value of the gravity model. Based on the target parameter isolation strategy, the parameters of the magnetic branch network are adjusted by backpropagation update according to the loss function value of the magnetic model. The mean values ​​of the fine-tuning loss function for the gravity model and the fine-tuning loss function for the magnetic model are calculated to obtain the average value of the fine-tuning loss function. Based on the target parameter isolation strategy, the parameters of the feature fusion network are adjusted using backpropagation update based on the average value of the fine-tuning loss function.

7. A three-dimensional gravity and magnetic field joint inversion device based on deep learning pre-optimization and model fine-tuning, characterized in that, include: The determination module is used to determine the pre-optimized joint inversion network parameters based on the gravity and magnetic dataset and the target multi-branch neural network; The determining module is further configured to determine the pre-optimized three-dimensional gravity and magnetic joint inversion result based on the pre-optimized joint inversion network parameters and the two-dimensional gravity observation data and the two-dimensional magnetic observation data to be predicted. The calculation module is used to calculate multi-dimensional loss function values ​​based on the pre-optimized three-dimensional gravity and magnetic joint inversion results on the surface, including current two-dimensional gravity observation data, current two-dimensional magnetic observation data, the two-dimensional gravity observation data to be predicted, and the two-dimensional magnetic observation data to be predicted. The multi-dimensional loss function values ​​include: gravity data fitting loss function value, magnetic data fitting loss function value, structural coupling loss function value, gravity model smoothing loss function value, and magnetic model smoothing loss function value. The determining module is further configured to adjust the parameters of the pre-optimized joint inversion network based on the multi-dimensional loss function value, and determine the optimal three-dimensional gravity and magnetic joint inversion result based on the adjusted optimal joint inversion network parameters.

8. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, it causes the processor to perform the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, When the computer program product is run on a processor, the processor causes the processor to perform the method as described in any one of claims 1-6.

Images