CN-121683548-B - Underground medium multi-scale forward and backward modeling method and system based on decoupling neural network

CN121683548BCN 121683548 BCN121683548 BCN 121683548BCN-121683548-B

Abstract

The invention discloses an underground medium multi-scale forward and backward modeling method and system based on a decoupling neural network, and belongs to the technical field of underground multi-physical field coupling. The underground medium multi-scale forward and backward modeling method based on the decoupling neural network comprises the steps of generating a space-time evolution data set of multiple physical field variables, constructing a physical information neural network architecture of physical field decoupling, wherein the physical information neural network architecture comprises three special sub-networks, each sub-network takes space-time coordinates as a basis for input, dynamically receives the output of other sub-networks as auxiliary input characteristics, realizes strong coupling modeling between physical fields through characteristic sharing and joint loss functions, executes a two-stage training strategy, and executes forward intelligent prediction or key physical parameter inversion based on the trained neural network architecture. The invention constructs a unified neural network framework consisting of a plurality of special sub-networks, realizes physical consistency from the core scale to the site scale, realizes data efficient cross-scale modeling, and synchronously supports forward simulation and parameter inversion.

Inventors

XIONG QINGRONG
LI BO
WANG LIGE
Qu Yongxiao
WANG FENGLONG
KOU LEI
ZHOU DENGWANG

Assignees

山东大学

Dates

Publication Date: 20260508
Application Date: 20260210

Claims (6)

1. The underground medium multi-scale forward and backward modeling method based on the decoupling neural network is characterized by comprising the following steps of: Based on a core scale gas-water-rock interaction experiment and corresponding high-fidelity numerical simulation, generating a space-time evolution data set of multiple physical field variables, and carrying out pretreatment and data enhancement; The physical information neural network architecture for decoupling the physical field is constructed, and the physical information neural network architecture comprises a temperature field sub-network, a seepage field sub-network and a stress field sub-network, wherein each sub-network takes space-time coordinates as a basis for input, outputs of other sub-networks are dynamically received as auxiliary input characteristics, and strong coupling modeling between the physical fields is realized through characteristic sharing and joint loss functions; The method comprises the steps of performing a two-stage training strategy on a physical information neural network architecture of physical field decoupling, wherein in the first stage, the physical information neural network of the physical field decoupling is pre-trained by using a rock core scale data set to enable a microcosmic physical rule to be internalized; based on the physical information neural network architecture of the physical field decoupling after the training, executing forward intelligent prediction or key physical property parameter inversion; The joint loss function is specifically a weighted sum of temperature field loss, seepage field loss, stress field loss and cross coupling constraint loss; The cross coupling constraint loss forces each constitutive relation to be strictly established at all sampling points, wherein the constitutive relation comprises Darcy's law, a porosity-strain relation and a state equation; the cross-coupling constraint loss is specifically expressed as: in the formula, Representing cross-coupling constraint loss; Representing a flow velocity vector; represents permeability; representing porosity; Representing permeability function with porosity And changes; Indicating the viscosity of the fluid; Representing the temperature of the output of the temperature field sub-network; representing a fluid viscosity function with temperature output by a temperature field subnetwork And changes; indicating an initial temperature; Represents pore pressure; Representing pore pressure gradients; Representing an initial pore pressure; representing the initial porosity; representing the effective stress coefficient; representing a displacement vector fed back by the stress field sub-network; displacement vector representing output of stress field sub-network Is a degree of divergence of (2); representing fluid density; representing an initial fluid density; representing the physical coefficient; indicating the initial fluid viscosity; representing a temperature dependence coefficient; Representation of Squaring the norm; the first stage is a microscopic pre-training stage, wherein a high-fidelity core scale data set is used as a unique supervision signal, an end-to-end joint optimization temperature field sub-network, a seepage field sub-network and a stress field sub-network are adopted, a minimum joint loss function is used as a target, and a complex physical relationship determined by a microscopic pore structure is internalized; The second stage is a macroscopic fusion fine tuning stage, a sub-network which is pre-trained in the microscopic pre-training stage is used as an initial model with physical priori, a field scale modeling task is accessed, field monitoring data corresponding to a thermal-flow-solid coupling numerical model of the field scale is imported, and end-to-end fine tuning is carried out on the whole network; The total loss function of the macroscopic fusion fine tuning stage is a weighted combination of joint loss and observed loss, expressed as: in the formula, Representing the total loss of the macroscopic fusion fine tuning stage; representing joint loss; representing an actual observed value acquired at the position of the monitoring point; The prediction output of the network at the same time-space position is obtained; to observe the loss weight.
2. The method of multi-scale forward modeling of an underground medium based on a decoupling neural network as claimed in claim 1, wherein the temperature field sub-network, the seepage field sub-network and the stress field sub-network all adopt deep multi-layer perceptron architecture and use Swish functions as activation functions, and the complete input vector of each sub-network comprises the output of basic space-time variables and other sub-networks at the same space-time position.
3. The underground medium multi-scale forward modeling method based on the decoupling neural network according to claim 1, wherein the temperature field loss is based on an energy conservation equation of the coupling effect of solid heat conduction and fluid convection, the seepage field loss is based on a mass conservation equation of the compression property of a pore medium and the influence of skeleton deformation, the stress field loss is based on a quasi-static mechanical balance equation fusing an effective stress principle and a thermoelastic effect, and the cross coupling constraint loss explicitly penalizes inconsistency of physical variables among all sub-networks.
4. The underground medium multi-scale forward modeling method based on the decoupling neural network of claim 1, wherein forward intelligent prediction specifically comprises the step of jointly outputting predicted values of temperature, pore pressure and displacement through single forward propagation according to input target space-time positions, boundary conditions and initial field states after feature sharing closed-loop interaction by a temperature field sub-network, a seepage field sub-network and a stress field sub-network.
5. The method for multi-scale forward modeling of the underground medium based on the decoupling neural network according to claim 1, wherein the inversion of key physical parameters comprises modeling at least one physical parameter of permeability, biot coefficient and thermal conductivity as a learnable field variable associated with a spatial position, and forming a joint optimization vector together with all weights of a temperature field sub-network, a seepage field sub-network and a stress field sub-network, and realizing the inversion of parameters by minimizing the difference between a predicted value and an on-site observed value and synchronously updating network parameters and physical parameter distribution under the constraint of a multi-physical field control equation.
6. Underground medium multiscale inversion system based on decoupling neural network, characterized by comprising: The data acquisition module is configured to generate a space-time evolution data set of multiple physical field variables based on a core scale gas-water-rock interaction experiment and corresponding high-fidelity numerical simulation, and perform preprocessing and data enhancement; the coupling modeling module is configured to construct a physical information neural network architecture of decoupling of a physical field, and comprises a temperature field sub-network, a seepage field sub-network and a stress field sub-network, wherein each sub-network takes space-time coordinates as a basis for input, and dynamically receives the output of other sub-networks as auxiliary input characteristics, and realizes strong coupling modeling between the physical fields through characteristic sharing and joint loss functions; The training module is configured to execute a two-stage training strategy on a physical information neural network architecture of physical field decoupling, wherein in the first stage, the physical information neural network of the physical field decoupling is pre-trained by using a rock core scale data set to enable a microscopic physical rule to be internalized; the execution module is configured to execute forward intelligent prediction or key physical property parameter inversion based on the physical information neural network architecture of the physical field decoupling after training; The joint loss function is specifically a weighted sum of temperature field loss, seepage field loss, stress field loss and cross coupling constraint loss; The cross coupling constraint loss forces each constitutive relation to be strictly established at all sampling points, wherein the constitutive relation comprises Darcy's law, a porosity-strain relation and a state equation; the cross-coupling constraint loss is specifically expressed as: in the formula, Representing cross-coupling constraint loss; Representing a flow velocity vector; represents permeability; representing porosity; Representing permeability function with porosity And changes; Indicating the viscosity of the fluid; Representing the temperature of the output of the temperature field sub-network; representing a fluid viscosity function with temperature output by a temperature field subnetwork And changes; indicating an initial temperature; Represents pore pressure; Representing pore pressure gradients; Representing an initial pore pressure; representing the initial porosity; representing the effective stress coefficient; representing a displacement vector fed back by the stress field sub-network; displacement vector representing output of stress field sub-network Is a degree of divergence of (2); representing fluid density; representing an initial fluid density; representing the physical coefficient; indicating the initial fluid viscosity; representing a temperature dependence coefficient; Representation of Squaring the norm; the first stage is a microscopic pre-training stage, wherein a high-fidelity core scale data set is used as a unique supervision signal, an end-to-end joint optimization temperature field sub-network, a seepage field sub-network and a stress field sub-network are adopted, a minimum joint loss function is used as a target, and a complex physical relationship determined by a microscopic pore structure is internalized; The second stage is a macroscopic fusion fine tuning stage, a sub-network which is pre-trained in the microscopic pre-training stage is used as an initial model with physical priori, a field scale modeling task is accessed, field monitoring data corresponding to a thermal-flow-solid coupling numerical model of the field scale is imported, and end-to-end fine tuning is carried out on the whole network; The total loss function of the macroscopic fusion fine tuning stage is a weighted combination of joint loss and observed loss, expressed as: in the formula, Representing the total loss of the macroscopic fusion fine tuning stage; representing joint loss; representing an actual observed value acquired at the position of the monitoring point; The prediction output of the network at the same time-space position is obtained; to observe the loss weight.

Description

Underground medium multi-scale forward and backward modeling method and system based on decoupling neural network Technical Field The invention belongs to the technical field of underground multi-physical field coupling, and particularly relates to an underground medium multi-scale forward and backward modeling method and system based on a decoupling neural network. Background The statements herein merely provide background information related to the present disclosure and may not necessarily constitute prior art. In deep energy engineering such as oil gas exploitation, high-level waste disposal, carbon dioxide geological storage and the like, the interaction of heat-flow-solid multiple physical fields has an important influence on system evolution. The process is accurately described, and not only a reasonable mathematical model is needed, but also physical parameters and correlations which can reflect the characteristics of a real medium are relied on. Currently, numerical simulations are typically developed on a field scale, describing multi-field coupling behavior using successive medium hypotheses. The constitutive and physical parameters required by such models often result from experimental testing or empirical fitting of the core dimensions. However, core samples represent only localized media features, while actual sites have significant heterogeneity and structural diversity in space. The relationship obtained by the core scale is directly applied to the field scale, so that potential influence of microstructure differences on macroscopic response is easily ignored, and the prediction capability of the model under complex conditions is limited. On the other hand, high-fidelity experiments or numerical simulation of the core scale can provide visual microscopic process information, and a multi-field coupling mechanism is revealed. But there is a lack of effective engagement paths between these microscopic data and the engineering model. In recent years, physical Information Neural Networks (PINN) have demonstrated potential in multi-physical field problems because of their ability to incorporate control equations as constraints into data-driven modeling. However, the existing application mostly adopts a unified network structure to process all physical field variables, and the difference of different fields in evolution characteristics is not fully considered. More importantly, how to organically combine microcosmic physical information of the core scale with monitoring data of the field scale under a unified framework, and realize reasonable transition from local mechanism to regional prediction through a neural network. Disclosure of Invention The invention aims to overcome the defects in the prior art and provide a method and a system for multi-scale forward and backward modeling of an underground medium based on a decoupling neural network, by constructing a unified neural network framework consisting of a plurality of special sub-networks, physical consistency from core scale to site scale, data efficient trans-scale modeling are realized, and forward simulation and parameter inversion are synchronously supported. In order to achieve the above object, the present invention is realized by the following technical scheme: In a first aspect, the present invention provides a method for multi-scale forward modeling of an underground medium based on a decoupled neural network, including: Based on a core scale gas-water-rock interaction experiment and corresponding high-fidelity numerical simulation, generating a space-time evolution data set of multiple physical field variables, and carrying out pretreatment and data enhancement; The physical information neural network architecture for decoupling the physical field is constructed, and the physical information neural network architecture comprises a temperature field sub-network, a seepage field sub-network and a stress field sub-network, wherein each sub-network takes space-time coordinates as a basis for input, outputs of other sub-networks are dynamically received as auxiliary input characteristics, and strong coupling modeling between the physical fields is realized through characteristic sharing and joint loss functions; The method comprises the steps of performing a two-stage training strategy on a physical information neural network architecture of physical field decoupling, wherein in the first stage, the physical information neural network of the physical field decoupling is pre-trained by using a rock core scale data set to enable a microcosmic physical rule to be internalized; based on the physical information neural network architecture of the physical field decoupling after the training, performing forward intelligent prediction or key physical property parameter inversion. In at least one embodiment, the temperature field sub-network, the seepage field sub-network and the stress field sub-network all adopt deep multi-layer perceptron arc