Lithosphere thermal structure parameter rapid inversion method and system based on physical constraint and dynamic sample generation

Abstract

The invention discloses a lithosphere thermal structure parameter rapid inversion method and system based on physical constraint and dynamic sample generation, and belongs to the technical field of geophysical exploration. Aiming at the problems of low calculation efficiency, high multiplicity and insufficient physical constraint of a traditional inversion method, the method comprises the following steps: firstly, constructing a physical constraint model based on a transient heat conduction equation; then dynamically generating a large number of synthetic training samples under the driving of the model; training a multi-output regression machine learning model by using the sample data to establish a mapping relation from the multi-source geophysical observation data to the thermal structure parameters; and finally, realizing minute-level parameter rapid inversion and uncertainty quantization through the trained model. The method is mainly used for efficient and reliable inversion of the lithosphere thermal structure in oil-gas exploration and geothermal resource evaluation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geophysical exploration, and particularly relates to a lithospheric thermal structure parameter rapid inversion method and system based on physical constraints and dynamic sample generation. BACKGROUND

[0002] Lithospheric thermal structure parameter inversion is a core technology in basin analysis, oil and gas accumulation research and geothermal resource assessment, and its accuracy is directly related to the understanding of basin formation mechanism, the prediction of source rock maturity, the determination of oil and gas accumulation period and the assessment of geothermal resource potential. Traditional inversion methods mainly rely on optimization algorithms such as differential evolution algorithm, Monte Carlo method and simulated annealing algorithm. These methods adjust model parameters repeatedly and perform a large number of forward simulations to fit the simulation results with the observed data. However, in practical applications, such methods have significant limitations. First, low computational efficiency is a major bottleneck. Since each forward simulation requires solving a complex transient heat conduction partial differential equation, a single calculation takes several minutes to several tens of minutes on an ordinary workstation, and completing a complete inversion analysis often takes several hours to several days, which seriously restricts the demand for rapid evaluation and decision-making in production practice.

[0003] Secondly, the inherent multi-solution problem of the inversion process is difficult to overcome. Geophysical inversion itself is ill-posed, and different underground parameter combinations can produce very similar surface observation responses. In the absence of sufficient prior information constraints, traditional optimization methods can only provide a single "optimal solution", and cannot effectively represent the uncertainty of the entire solution space, which brings great difficulty and risk to subsequent geological interpretation. Furthermore, the lack of physical constraints makes the geological reasonableness of the inversion results questionable. Pure data-driven inversion methods may produce parameter solutions that are mathematically well-fitted but not geophysically reasonable, reducing the practical guiding value of the inversion results.

[0004] In recent years, machine learning techniques have been introduced to attempt to improve inversion efficiency. However, early purely data-driven methods heavily relied on the quantity and quality of training data, which not only risked producing physically unreasonable results but also performed poorly when extrapolated to situations outside the training set. While subsequent developments in physical information machine learning attempted to incorporate physical laws as constraints, they remained constrained by the scarcity of real observational data, making it difficult to obtain sufficient high-quality training samples. Furthermore, in actual exploration, engineers typically obtain heterogeneous data from multiple sources, such as vitrinite reflectivity, borehole temperature, and subsidence history. Effectively integrating this data is crucial for reducing inversion uncertainty, but the differences in spatiotemporal resolution and error characteristics among different types of data pose challenges to establishing a unified data assimilation framework, further increasing the complexity and computational burden of forward modeling. Therefore, there is an urgent need for a technical solution that can efficiently and reliably integrate physical laws and multi-source data, and quickly provide reasonable inversion results in practical applications. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a rapid inversion method and system for lithospheric thermal structure parameters based on physical constraints and dynamic sample generation. This method achieves efficient and reliable inversion of lithospheric thermal structure parameters, providing an important technical means for energy resource exploration.

[0006] To achieve the above objectives, this invention provides a rapid inversion method for lithospheric thermal structure parameters based on physical constraints and dynamically generated samples, comprising: Construct a physical model based on the transient heat conduction equation; Using the physical model, a sample dataset containing multiple sets of lithospheric thermal structure parameters and their corresponding simulated responses is generated within a preset parameter prior space. A multi-output machine learning model was trained using the sample dataset to establish a mapping relationship from geophysical observation data to lithospheric thermal structure parameters; The actual acquired geophysical observation data is input into the trained multi-output machine learning model, which outputs the inversion results of lithospheric thermal structure parameters.

[0007] Optionally, constructing a physical model based on the transient heat conduction equation includes: Establish transient heat conduction control equations that take into account the lithospheric layered structure, the spatial distribution of radioactive heat generation rate, and the heat advection effect; The governing equations are solved numerically using an implicit difference scheme, and the settlement is calculated based on the principle of local equilibrium.

[0008] Optionally, the physical model may also include the following during the modeling process: When simulating lithospheric stretching events, the lithospheric thickness is dynamically adjusted based on the stretching factor. The temperature field was corrected by using a material resampling method based on the ratio of lithosphere thickness before and after stretching, in order to characterize the thermal advection effect.

[0009] Optionally, the generated sample dataset includes: The Latin hypercube sampling method is used to randomly generate the multiple sets of lithospheric thermal structure parameter combinations in the parameter prior space; The parameter types in the combination of lithospheric thermal structure parameters include crustal stretching coefficient, mantle stretching coefficient, stretching event start time, and event duration.

[0010] Optionally, the multi-output machine learning model is one of a random forest model, a gradient boosting tree model, or a neural network model; The multi-output machine learning model is configured to simultaneously output parameters related to multiple tension events.

[0011] Optionally, the process of outputting the inversion results of lithospheric thermal structure parameters also includes: Based on the prediction results of each base learner in the multi-output machine learning model, the prediction variance of each parameter in the inversion result is calculated to quantify the uncertainty of the inversion result.

[0012] On the other hand, to achieve the above objectives, the present invention also provides a rapid inversion system for lithospheric thermal structure parameters based on physical constraints and dynamically generated samples, comprising: The physical constraint modeling module is used to build and solve physical models based on transient heat conduction equations; The dynamic sample generation module is used to call the physical constraint modeling module to generate a sample dataset in the parameter prior space; A machine learning training module is used to train a multi-output machine learning model using the sample dataset. The parameter fast inversion module is used to input actual observation data into the trained multi-output machine learning model and output the inversion results.

[0013] Optionally, the system may also include a graphical user interface module; The graphical user interface module provides a multi-source observation data import interface, an interactive parameter configuration panel, and a visualization display function for inversion results.

[0014] Technical Advantages of this Invention: This invention discloses a rapid inversion method and system for lithospheric thermal structure parameters based on physical constraints and dynamic sample generation. It combines a physical constraint model with machine learning, generating a large number of synthetic training samples through the physical model, effectively overcoming the limitation of insufficient actual observational data on machine learning model training. Using a trained multi-output regression model, a rapid mapping relationship from multi-source geophysical data to lithospheric thermal structure parameters is established, thereby significantly reducing the inversion time from orders of magnitude to minutes compared to traditional optimization methods, significantly improving computational efficiency. This method ensures the geological rationality of the inversion results through embedded physical laws and provides uncertainty quantification of the inversion results based on the machine learning model, helping to reduce the ambiguity of the inversion and support reliable geological interpretation. This invention can effectively integrate multi-source observational data such as vitrinite reflectivity, borehole temperature, and subsidence history to form a unified rapid inversion framework. Its systematic implementation also facilitates its application in practical oil and gas exploration and geothermal resource assessment. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a rapid inversion method for lithospheric thermal structure parameters based on physical constraints and dynamic sample generation, according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] like Figure 1 As shown, this embodiment provides a fast inversion method for lithospheric thermal structure parameters based on physical constraints and dynamically generated samples, including: Construct a physical model based on the transient heat conduction equation; Using the physical model, a sample dataset containing multiple sets of lithospheric thermal structure parameters and their corresponding simulated responses is generated within a preset parameter prior space. A multi-output machine learning model was trained using the sample dataset to establish a mapping relationship from geophysical observation data to lithospheric thermal structure parameters; The actual acquired geophysical observation data is input into the trained multi-output machine learning model, which outputs the inversion results of lithospheric thermal structure parameters.

[0019] Furthermore, constructing a physical model based on the transient heat conduction equation includes: Establish transient heat conduction control equations that take into account the lithospheric layered structure, the spatial distribution of radioactive heat generation rate, and the heat advection effect; The governing equations are solved numerically using an implicit difference scheme, and the settlement is calculated based on the principle of local equilibrium.

[0020] Specifically, the implementation process of this embodiment includes: The physical basis is a one-dimensional transient heat conduction equation: ; in, For temperature, For time, For depth, For density, For specific heat capacity, Thermal conductivity, The rate of radioactive heat generation.

[0021] Numerical discretization is performed using an implicit backward Euler scheme: ; in, For the first i Temperature of each grid point For time step, For spatial step size, For the first i Thermal conductivity of each grid point Let be the radioactive heat generation rate of the i-th grid point.

[0022] The interfacial thermal conductivity is calculated using a harmonic average. ; in, For grid points i and i Interfacial thermal conductivity between +1, , represents the thermal conductivity of adjacent grid points.

[0023] The settlement calculation is based on the equilibrium principle to calculate the total settlement: ; in, To construct the settlement (m), This represents the change in mass per unit area (kg / m²). The density of the mantle (kg / m³) The density of water is (kg / m³).

[0024] Settlement due to thermal contraction: ; in, for t Heat deposition at time (m), The coefficient of thermal expansion (1 / °C) The thickness of the lithosphere (m) For the initial temperature field, Let be the temperature field at time t.

[0025] Furthermore, the physical model also includes the following in the modeling process: When simulating lithospheric stretching events, the lithospheric thickness is dynamically adjusted based on the stretching factor. The temperature field was corrected by using a material resampling method based on the ratio of lithosphere thickness before and after stretching, in order to characterize the thermal advection effect.

[0026] Specifically, the implementation process of this embodiment includes: For stretching events, the change in lithosphere thickness is as follows: ; in, for t The thickness of the lithosphere at any given time (in meters). Initial lithosphere thickness (m) for t The stretching factor at time (dimensionless).

[0027] Temperature field correction is performed using a material resampling method: ; in, Let Z be the temperature at depth z after stretching. The temperature field before stretching. This represents the source depth before stretching.

[0028] The source depth is calculated using mass conservation: ; in, The source depth (m) before stretching. The thickness of the lithosphere before stretching (m), The thickness of the lithosphere after stretching (m). The depth after stretching (m).

[0029] Furthermore, the generated sample dataset includes: The Latin hypercube sampling method is used to randomly generate the multiple sets of lithospheric thermal structure parameter combinations in the parameter prior space; The parameter types in the combination of lithospheric thermal structure parameters include crustal stretching coefficient, mantle stretching coefficient, stretching event start time, and event duration.

[0030] Specifically, the implementation process of this embodiment includes: Latin hypercube sampling is used to ensure uniform coverage of the parameter space and generate parameter combinations: ; in, For the i-th sample parameter combination, Let be the crustal stretching coefficient of the i-th sample. Let be the mantle stretching coefficient of the i-th sample. Let Ma be the start time (Ma) of the tensioning event for the i-th sample. Let Ma be the event duration of the i-th sample. The total number of samples.

[0031] Parallel computing technology is used to accelerate sample generation, and each sample contains complete temperature history, sedimentation history and Ro profile data.

[0032] Furthermore, the multi-output machine learning model is one of the following: random forest model, gradient boosting tree model, or neural network model; The multi-output machine learning model is configured to simultaneously output parameters related to multiple tension events.

[0033] Specifically, the implementation process of this embodiment includes: A multi-output regression model is employed to simultaneously predict parameters of multiple stretching events. The feature vector includes time-series features, Ro profile features, and statistical features. The random forest multi-output regression model is as follows: ; in, For the first j The predicted values ​​of each output parameter, For the number of decision trees, For the first j The output parameter of the first k The predicted values ​​of each decision tree. This is the input feature vector.

[0034] Furthermore, the process of outputting the inversion results of lithospheric thermal structure parameters also includes: Based on the prediction results of each base learner in the multi-output machine learning model, the prediction variance of each parameter in the inversion result is calculated to quantify the uncertainty of the inversion result.

[0035] Specifically, the implementation process of this embodiment includes: Input observation data The parameter estimates are obtained as follows: ; in, The results are from the inversion of lithospheric thermal structure parameters. For a well-trained machine learning model, For observational data.

[0036] Uncertainty in predictions based on random forests: ; in, For the first j The prediction variance of each output parameter.

[0037] This embodiment also provides a fast inversion system for lithospheric thermal structure parameters based on physical constraints and dynamically generated samples, including: The physical constraint modeling module is used to build and solve physical models based on transient heat conduction equations; The dynamic sample generation module is used to call the physical constraint modeling module to generate a sample dataset in the parameter prior space; A machine learning training module is used to train a multi-output machine learning model using the sample dataset. The parameter fast inversion module is used to input actual observation data into the trained multi-output machine learning model and output the inversion results.

[0038] Furthermore, the system also includes a graphical user interface module; The graphical user interface module provides a multi-source observation data import interface, an interactive parameter configuration panel, and a visualization display function for inversion results.

[0039] Example 1: Basic System Configuration The system requires a multi-core processor and ample memory; an eight-core or higher CPU and at least 16GB of RAM are recommended. The software environment requires scientific computing libraries and machine learning frameworks.

[0040] During system initialization, users need to configure basic physical model parameters, including initial crustal thickness, initial mantle thickness, crustal thermal conductivity, and crustal heat generation rate. These parameters can be adjusted appropriately according to the geological characteristics of the specific study area.

[0041] Example 2: Complete Inversion Process Before the inversion process begins, observational data needs to be prepared. The system supports the import of various data formats, including vitrinite reflectance profile data, borehole temperature measurement data, and tectonic subsidence history data. After data import, the system will automatically perform quality checks and preprocessing.

[0042] Next, the system enters the dynamic sample generation phase. Based on the user-defined parameter range, the system generates a large number of training samples using the Latin hypercube sampling method. The parameter range is typically set as follows: crustal stretching coefficient 0.5 to 5.0, mantle stretching coefficient 0.5 to 5.0, extensional event start time 10 to 200 million years, and event duration 1 to 50 million years. The number of samples can be set according to computing resources, typically ranging from ten thousand to one hundred thousand.

[0043] After sample generation is complete, the system enters the machine learning training phase. Users can choose different regression algorithms, such as random forests, gradient boosting trees, or neural networks. During training, the system automatically optimizes the model's hyperparameters and evaluates the model's predictive performance.

[0044] Once training is complete, rapid inversion can be performed. The user inputs observational data, and the system outputs inversion results in a short time, including parameter estimates and corresponding confidence assessments for each period of tension events.

[0045] Example 3: Industrial Application Case In an application in an oil and gas basin, the system of this invention was used to conduct a study on the inversion of thermal structural parameters. First, vitrinite reflectance data from 15 wells in the region, system temperature measurement data from 8 wells, and regional tectonic subsidence history data were collected.

[0046] The system generated 50,000 training samples, covering three possible combinations of tension event parameters. The random forest algorithm was used for training, and the model training time was approximately one hour. After training, the actual inversion calculation only took three minutes.

[0047] The inversion results identified three major extensional events, occurring over periods of 200–180 million years, 150–130 million years, and 100–80 million years, respectively. Compared to traditional differential evolution inversion methods, the inversion results obtained by the method of this invention show better agreement with measured heat flow data, achieving a coefficient of determination of 0.89. The correlation coefficient between the predicted and measured values ​​of vitrinite reflectance reaches 0.94.

[0048] This inversion result provides a reliable thermal constraint for the analysis of hydrocarbon accumulation stages in the region, significantly improving the accuracy of exploration evaluation. Compared with traditional methods, the computation time is reduced by more than 95%, providing timely technical support for exploration decisions.

[0049] This invention has been successfully applied to basin analysis in multiple oil and gas fields, and has significant application value in oil and gas exploration, geothermal resource assessment, basic geological research, and mineral exploration. Through an innovative technical approach, it achieves efficient and reliable inversion of lithospheric thermal structure parameters, providing an important technical means for energy resource exploration.

[0050] This invention discloses a rapid inversion method and system for lithospheric thermal structure parameters based on physical constraints and dynamic sample generation. It combines a physical constraint model with machine learning, generating a large number of synthetic training samples through the physical model, effectively overcoming the limitation of insufficient actual observational data on machine learning model training. Using a trained multi-output regression model, a rapid mapping relationship from multi-source geophysical data to lithospheric thermal structure parameters is established, thereby significantly reducing the inversion time from orders of magnitude to minutes compared to traditional optimization methods, significantly improving computational efficiency. This method ensures the geological rationality of the inversion results through embedded physical laws and provides uncertainty quantification of the inversion results based on the machine learning model, helping to reduce the ambiguity of the inversion and support reliable geological interpretation. This invention can effectively integrate multi-source observational data such as vitrinite reflectivity, borehole temperature, and subsidence history to form a unified rapid inversion framework. Its systematic implementation also facilitates its application in practical oil and gas exploration and geothermal resource assessment.

[0051] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for fast inversion of lithospheric thermal structure parameters based on physical constraints and dynamic sample generation, characterized in that, The method comprises the following steps: constructing a physical model based on a transient heat conduction equation; generating a sample data set containing multiple sets of lithospheric thermal structure parameters and their corresponding simulation responses in a preset parameter prior space through the physical model; training a multi-output machine learning model using the sample data set to establish a mapping relationship from geophysical observation data to lithospheric thermal structure parameters; inputting actual acquired geophysical observation data into the trained multi-output machine learning model to output an inversion result of the lithospheric thermal structure parameters.

2. The method of claim 1, wherein the method is characterized by, The step of constructing a physical model based on a transient heat conduction equation comprises the following steps: establishing a transient heat conduction control equation considering the layered structure of the lithosphere, the spatial distribution of radioactive heat generation rate, and the thermal advection effect; numerically solving the control equation using an implicit difference format, and calculating and constructing the subsidence based on the principle of local equilibrium.

3. The method of claim 2, wherein the method further comprises: The physical model further comprises the following steps in the modeling process: when simulating a lithospheric extension event, dynamically adjusting the lithospheric thickness according to the extension factor; using a material resampling method to correct the temperature field according to the ratio of the lithospheric thickness before and after the extension to represent the thermal advection effect.

4. The method of claim 1, wherein the method is characterized by, The step of generating a sample data set comprises the following steps: randomly generating the multiple sets of lithospheric thermal structure parameter combinations in the parameter prior space using a Latin hypercube sampling method; wherein the parameter types in the lithospheric thermal structure parameter combinations include the crustal extension coefficient, the mantle extension coefficient, the starting time of the extension event, and the event duration.

5. The method of claim 1, wherein the method is characterized by, The multi-output machine learning model is one of a random forest model, a gradient boosting tree model, or a neural network model; The multi-output machine learning model is configured to simultaneously output parameters related to multiple extension events.

6. The method of claim 1, wherein the method is characterized by, The process of outputting the inversion result of the lithospheric thermal structure parameters further comprises the following steps: based on the prediction results of each base learner in the multi-output machine learning model, calculating the prediction variance of each parameter in the inversion result to quantify the uncertainty of the inversion result.

7. A lithospheric thermal structure parameter rapid inversion system based on physical constraints and dynamic sample generation, characterized in that, The system for implementing the method for quickly inverting lithospheric thermal structure parameters based on physical constraints and dynamic sample generation according to any one of claims 1-6 comprises: a physical constraint modeling module for constructing and solving a physical model based on a transient heat conduction equation; a dynamic sample generation module for calling the physical constraint modeling module to generate a sample data set in a parameter prior space; a machine learning training module for training a multi-output machine learning model using the sample data set; a parameter quick inversion module for inputting actual observation data into the trained multi-output machine learning model and outputting an inversion result.

8. The system for fast inversion of lithospheric thermal structure parameters based on physical constraints and dynamic sample generation of claim 7, wherein, The system further comprises a graphical user interface module; The graphical user interface module provides a multi-source observation data import interface, an interactive parameter configuration panel, and an inversion result visualization display function.

Images