CN-122021176-A - Basin construction parameter inversion method, system and medium based on proxy model optimization

CN122021176ACN 122021176 ACN122021176 ACN 122021176ACN-122021176-A

Abstract

The invention provides a basin construction parameter inversion method, system and medium based on agent model optimization, which comprises the steps of obtaining construction-stratum profile observation data of a target basin, establishing a flexible cantilever forward model of geometric-balance-rheological coupling, carrying out initial sampling in a parameter space, constructing an initial training set, establishing a mapping relation between parameters and objective function values by means of Gaussian process regression, selecting a parameter combination to be evaluated based on a Bayesian optimization framework through expected improved acquisition functions, calling the forward model to calculate and update the training set, iterating until convergence, and finally carrying out importance resampling based on the trained agent model to realize uncertainty quantization of inversion parameters and confidence interval output. The method solves the problems of high calculation cost, slow convergence, unquantifiable multi-solution property and insufficient physical coupling of the forward model in the traditional method, and remarkably improves the efficiency and the result reliability of the inversion of the extended basin construction parameters.

Inventors

TANG XIAOYIN
GAO WANLI
PENG BO

Assignees

中国地质科学院地质力学研究所

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. The basin construction parameter inversion method based on proxy model optimization is characterized by comprising the following steps of: S1, constructing a geometric-balance-rheological coupling deflection cantilever forward model based on acquired structure-stratum profile observation data of a target basin, wherein the method comprises the steps of adopting a finite difference format to discretely solve a fourth-order differential equation for controlling deflection balance of a rock ring, calculating a total expansion factor for a multi-fault system through linear superposition of fault crust thinning distribution, processing effective elastic thickness into a function based on time attenuation of a viscoelastic rheological model so as to dynamically update bending stiffness; s2, initial sampling is carried out in the parameter space of the forward model of the flexible cantilever beam, a corresponding objective function value is obtained through calculation of the forward model of the flexible cantilever beam, an initial training set is built and set as a current training set, and the objective function value is a negative mean square error between observed data and model prediction; S3, training a Gaussian process regression proxy model based on a current training set to establish a mapping relation from geological parameters to objective function values and obtain a prediction mean value and a prediction variance; S4, calculating an acquisition function based on the Gaussian process regression proxy model according to the prediction mean value and the prediction variance, and determining a next group of parameter combinations to be evaluated through the maximized acquisition function; S5, calling the forward model of the deflection cantilever beam to calculate the real objective function value of the group of parameter combinations, and adding the real objective function value into the current training set to update data; s6, repeating the steps S3 to S5 for iteration until a preset convergence condition is met; s7, carrying out uncertainty quantification based on the finally trained agent model, generating an approximate posterior sample set through resampling based on the importance of the prediction distribution, and outputting the statistical characteristics and the confidence interval of the inversion parameters.
2. The method of claim 1, wherein S2, performing initial sampling in the flexible cantilever forward model parameter space, calculating a corresponding objective function value by the flexible cantilever forward model, and constructing an initial training set includes: in a parameter space formed by fault expansion amount, inclination angle and rock ring initial effective elastic thickness, generating by Latin hypercube sampling method Sample set of initial parameters ,; Invoking a forward model of the deflection cantilever beam to calculate objective function values corresponding to each group of parameters Form an initial training set ; The objective function is a fitting measure between observed data and model prediction, and the method further comprises the steps of constructing the objective function and setting physical constraints, namely constructing the objective function with robustness: defining an optimization target as a log-likelihood function for maximizing the fitting degree of model prediction and observed data, wherein a body of the likelihood function is modeled by using Student's t distribution instead of standard Gaussian distribution: ; Wherein, the And Respectively the first Individual observed data points and their parameters The model predictive value of the lower model is calculated, For the corresponding data error estimate, A degree of freedom parameter for the t distribution; and applying at least one geological physical constraint including fault cutting depth constraint, effective elastic thickness range constraint and fault dip angle priori distribution constraint.
3. The method of claim 1, wherein S1, constructing a geometry-equalization-rheological coupled forward model of the flexural cantilever based on the acquired formation-formation profile observations of the target basin comprises: Step S11, numerical solution of geometric-equilibrium coupling And adopting a five-point finite difference format to discretely solve a fourth-order differential equation for controlling the flexural balance of the rock ring: ; ; step S12, linear superposition of multiple fault system loads for a system having Complex system of strip faults, total crust thinning distribution The linear superposition of the contributions of all faults is formed by: ; Wherein, the Is the first The crust thinning amount generated by the strip fault activity; Calculating the total stretch factor of the whole section Wherein For an initial crust thickness, the Will be used to drive the thermal settlement calculation and balance the load at the same time Is updated according to the update of (a); Step S13, time Domain correction of viscoelastic rheology To simulate the long-term stress relaxation of a rock ring, an effective elastic thickness is applied Is processed as time Function of attenuation Description using Maxwell or Standard Linear Solids (SLS) rheology model: Maxwell model: ; SLS model: ; Wherein, the To be in viscosity with the mantle An associated relaxation time; In forward modeling, the time of evolution after an event is constructed Dynamic update The values and recalculate the flexural rigidity 。
4. The method according to claim 1, characterized in that in step S1 or in the optimization iteration process, a multi-dimensional geological physical constraint is imposed to ensure geological rationality of the inversion result, said geological physical constraint comprising at least one of the following: Geometric kinematics constraint that the maximum cutting depth of the main control fault must not exceed the preset zone crust thickness I.e. This is a hard boundary constraint; rheological constraint of rock ring, effective elastic thickness of rock ring Must be non-negative and its value range should be in conformity with the background range given by regional geophysical studies, i.e ; Fault structure priori constraint, namely, to suppress multiple solutions caused by parameter trade-off relation, fault inclination angle Applying a reasonable prior probability distribution to limit the value range thereof to To the point of In between, this a priori information will be incorporated into the objective function in the form of additional terms, leading the optimization process towards a more geologically trusted solution.
5. The method of claim 1, wherein S3 training a gaussian process regression proxy model based on the current training set comprises: Based on the current training set Optimizing Matern/2 kernel function super-parameters by adopting a maximum likelihood estimation method, wherein the kernel function is defined as: ; Wherein, the For the distance of the parameter space, the super-parameters to be optimized include the signal variance And characteristic length scale ; After training, based on the Gaussian process regression proxy model, any parameter point is calculated Giving the predicted mean value And standard deviation of 。
6. The method of claim 1, wherein S4, calculating an acquisition function based on the gaussian process regression proxy model from the predicted mean and predicted variance, determining the next set of parameter combinations to be evaluated by maximizing the acquisition function comprises: Employing desired improvement criteria as acquisition function Balance exploration and development: ; Wherein, the For the current optimal observation value, In order to adjust the parameters of the device, , And Respectively a cumulative distribution function and a probability density function of standard normal distribution; Global maximization using L-BFGS-B optimization algorithm Determining a next set of parameter combinations to be evaluated 。
7. The method of claim 1, wherein S5, invoking the flexural cantilever forward model to calculate a true objective function value for the set of parameter combinations and adding it to a current training set to update data comprises: Invoking forward model computation Is the true objective function value of (2) Will new data pair Adding training set 。
8. The method of claim 1, wherein S7, performing uncertainty quantization based on the final trained surrogate model, generating an approximate posterior sample set by resampling based on the importance of the prediction distribution, outputting the statistical features and confidence intervals of the inversion parameters comprises: generating a large number of candidate samples in a parameter space Fast prediction of non-normalized log-likelihood approximations for each sample using a gaussian process model ; Calculating importance weights for each candidate sample And normalizing: ; Based on the weight distribution, importance resampling is performed on the candidate sample set to generate a sample set similar to the true posterior distribution ; Based on posterior sample set And calculating the edge probability distribution, the median, the mean value and the confidence interval of the designated quantile of each inversion parameter, and finishing the quantitative description of the uncertainty of the inversion result.
9. A basin construction parameter inversion system based on proxy model optimization, comprising: The data acquisition and modeling module is used for acquiring the structure-stratum profile observation data of a target basin and constructing a geometric-balance-rheological coupling deflection cantilever forward model, and comprises the steps of adopting a finite difference format to discretely solve a fourth-order differential equation for controlling the deflection balance of a rock ring, calculating a total expansion factor through linear superposition of the thinning distribution of each fault crust for a multi-fault system, processing the effective elastic thickness into a function based on the attenuation of a viscoelastic rheological model with time so as to dynamically update bending rigidity, and defining parameters to be inverted and prior constraints thereof; the sampling and initializing module is used for carrying out initial sampling in a parameter space of the forward model of the flexural cantilever beam, calculating a corresponding objective function value through the forward model, constructing an initial training set and setting the initial training set as a current training set, wherein the objective function value is a negative mean square error between observed data and model prediction; The agent model construction module is used for training a Gaussian process regression agent model based on the current training set so as to establish a mapping relation from geological parameters to objective function values and obtain a prediction mean value and a prediction variance; The decision module calculates an acquisition function based on the Gaussian process regression proxy model according to the prediction mean value and the prediction variance, and determines the next group of parameter combinations to be evaluated by maximizing the acquisition function; The forward modeling and updating module is used for calling the forward modeling of the flexural cantilever beam to calculate the real objective function value of the group of parameter combinations and adding the real objective function value into the current training set to update data; The iteration control module is used for controlling the repeated execution of the functions of the agent model construction module, the decision module and the forward modeling and updating module until the preset convergence condition is met; And the uncertainty quantization and output module is used for carrying out uncertainty quantization based on the finally trained proxy model, generating an approximate posterior sample set through resampling based on the importance of the prediction distribution, and outputting the statistical characteristics and the confidence interval of the inversion parameters.
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the proxy model optimization based basin construction parameter inversion method of any one of claims 1 to 8.

Description

Basin construction parameter inversion method, system and medium based on proxy model optimization Technical Field The invention relates to the technical field of computational geophysics and geological exploration, and particularly provides a basin construction parameter inversion method, system and medium based on proxy model optimization. Background The formation of an extended basin is closely related to the tensile thinning of the rock hoop under regional tensile stress. The flexural cantilever model is a classical basin dynamics forward model that simulates the process of stretching basin formation. Unlike the McKenzie model, which assumes uniform stretching, the flexural cantilever Liang Moxing simplifies the rock ring to an elastic plate, the upper crust breaks brittle, and the lower crust and the rock ring flow viscously. By establishing a kinematic relationship between the rigid rotation of the upper crust brittle fracture and the ductile deformation of the lower crust, the structural settlement in the subsidence period and the buckling rising phenomenon of the basin edge shoulder are simulated. The core parameters of the model comprise the horizontal extension of the main control fault, the inclination angle of the section, the conversion depth of the crust brittleness and toughness and the effective elastic thickness of the rock ring. The current common practice is to manually tune parameters or to perform parameter calibration of the model by means of exhaustive grid search. Due to the strong coupling relationship between these parameters. For example, a larger extension with a steeper incline may produce a settling effect that is geometrically very similar to a smaller extension with a softer incline. The traditional algorithm is very easy to be trapped in local optimum, and has stronger multi-solution. In addition, the forward modeling of the cantilever Liang Moxing itself involves the solution of the variable stiffness flexural equation, a single calculation taking several seconds to tens of seconds. When the inversion object involves a multi-fault system, the parameter dimension increases sharply, and the calculation amount of grid search or Monte Carlo simulation increases exponentially, which is difficult to finish in a feasible time. In addition, the fault geometric motion and deflection equilibrium are used as independent modules for series calculation in the existing tool, so that the real-time coupling relation of the fault geometric motion and the deflection equilibrium in the geological process is ignored, and the physical rationality of the model is influenced. In recent years, the Bayesian inversion method has been applied to basin evolution simulation (such as Chandra & Muller, 2020), but the model still has the defects of lack of uncertainty quantization capability, high calculation cost of the MCMC algorithm, imperfect physical coupling mechanism and the like. Therefore, a method for improving the physical reality of the forward model and the inversion calculation efficiency is needed. Disclosure of Invention In order to solve the above problems, the present invention provides, in a first aspect, a basin construction parameter inversion method based on proxy model optimization, including: S1, constructing a geometric-balance-rheological coupling deflection cantilever forward model based on acquired structure-stratum profile observation data of a target basin, wherein the method comprises the steps of adopting a finite difference format to discretely solve a fourth-order differential equation for controlling deflection balance of a rock ring, calculating a total expansion factor for a multi-fault system through linear superposition of fault crust thinning distribution, processing effective elastic thickness into a function based on time attenuation of a viscoelastic rheological model so as to dynamically update bending stiffness; s2, initial sampling is carried out in the parameter space of the forward model of the flexible cantilever beam, a corresponding objective function value is obtained through calculation of the forward model of the flexible cantilever beam, an initial training set is built and set as a current training set, and the objective function value is a negative mean square error between observed data and model prediction; S3, training a Gaussian process regression proxy model based on a current training set to establish a mapping relation from geological parameters to objective function values and obtain a prediction mean value and a prediction variance; S4, calculating an acquisition function based on the Gaussian process regression proxy model according to the prediction mean value and the prediction variance, and determining a next group of parameter combinations to be evaluated through the maximized acquisition function; S5, calling the forward model of the deflection cantilever beam to calculate the real objective function value of the group of parameter co