CN-121983179-A - Groundwater reactive solute transport prediction method and model based on physical-graphic neural network

CN121983179ACN 121983179 ACN121983179 ACN 121983179ACN-121983179-A

Abstract

The invention discloses a prediction method and model for groundwater reactive solute transport based on a physical-graphic neural network. The method comprises the steps of inputting directed graph structure data of a target underground water chemical system into a parameterized graph neural network for prediction to obtain a reaction rate parameter, pH value and pe residual error result of microbial reaction of the target underground water chemical system in a t time step, inputting the reaction rate parameter into a dynamic physical network for predicting the concentration of reactive solute in the t+1th time step, and inputting the residual error result of the concentration, pH value and pe into a residual error correction network for correction to obtain the concentration of the reactive solute in the target underground water chemical system in the t+1th time step, and the pH value and pe value. The method and the model realize high-efficiency and high-precision prediction results, and can be suitable for complex biological geochemical systems.

Inventors

WANG JINBO
CHEN SHUAI
WANG ZIQI
WU JIAHUI
ZHANG JIA
LIU MINGZHU

Assignees

中国地质大学（北京）

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (10)

1. The groundwater reactive solute transport prediction method based on the physical-graphic neural network is characterized by comprising the following steps of: Acquiring groundwater geochemical state data of a target groundwater chemical system at a t-th time step, wherein the groundwater geochemical state data at least comprises the concentration of a reactive solute, the pH value of a groundwater hydrological geochemical environment parameter and the pe value of an oxidation-reduction potential at the t-th time step; constructing a directed graph of the target groundwater chemistry system based on the groundwater geochemical state data; Inputting the directed graph to a parameterized graph neural network based on a graph attention mechanism to infer, so as to obtain a reaction rate parameter of microbial reaction of the target groundwater chemical system at the t-th time step and a residual result of corresponding pH and pe; inputting the reactive solute concentration of the t time step and the reaction rate parameter of the microbial reaction into a kinetic physical network, and predicting the preliminary prediction result of the reactive solute concentration of the target groundwater chemical system at the t+1th time step; And inputting the preliminary prediction result of the reactive solute concentration at the t+1th time step and the residual results of the pH and the pe into a residual correction network for correction to obtain the prediction result of the reactive solute concentration, the pH and the pe of the target groundwater chemical system at the t+1th time step.
2. The method of claim 1, wherein the directed graph comprises a set of nodes comprising species nodes representing chemical species and reaction nodes representing microbial reaction processes, wherein the initial characteristics of the species nodes consist of logarithmic transformation values of reactive solute concentrations at time step t, groundwater geochemical environmental parameters pH and pe values, and a set of directed edges comprising input edges representing relationships of chemical species participating as reactants in respective microbial reaction processes, output edges representing relationships of microbial reaction processes generating respective chemical species, and inhibition edges representing relationships of chemical species or microbial reaction processes producing inhibition or competition effects on another microbial reaction process.
3. The method of claim 2, wherein inputting the directed graph into a graph attention mechanism-based parameterized graph neural network to infer comprises: The parameterized graph neural network based on the graph attention mechanism comprises a plurality of graph attention network layers, in each graph attention network layer, node characteristics in the directed graph are subjected to linear transformation, attention weights between nodes and neighbor nodes of the nodes are calculated based on the self-attention mechanism, and the neighbor node characteristics are subjected to weighted aggregation to obtain embedded representations of the nodes of each species; carrying out layer-by-layer message transmission and feature updating through the multi-layer graph attention network layer to obtain embedded representation of each reaction node; and polymerizing the embedded representation of the reaction node and the embedded representation of the species node, and outputting the reaction rate parameter of the microbial reaction of the target groundwater chemical system at the t-th time step and the residual results of the corresponding pH and pe.
4. A method according to claim 3, wherein the parametric graph neural network based on graph attention mechanisms comprises three graph attention network layers comprising 4 attention headers, 2 attention headers and 1 attention header, respectively.
5. The method of claim 4, wherein the kinetic physical network is configured to perform kinetic calculation on the microbial reaction based on the concentration of the reactive solute at the t-th time step and the reaction rate parameter of the microbial reaction, and obtain a preliminary prediction result of the concentration of the reactive solute at the t+1th time step.
6. The method of claim 5, wherein the residual correction network comprises three residual correction sublayers for performing multi-layer residual correction on the preliminary prediction result of the concentration of the reactive solute at time step t+1 and the corresponding residual results of pH and pe to obtain the prediction result of the concentration of the reactive solute, pH and pe of the target groundwater chemistry system at time step t+1.
7. The method of claim 1, further comprising periodically invoking a groundwater reactive solute transport simulation model coupled by a hydrodynamic simulator to PHREEQC at predetermined time step intervals to correct predictions of reactive solute concentration, pH and pe for the t+1th time step output by the residual correction network.
8. A physical-graph neural network-based groundwater reactive solute transport prediction model, characterized in that it is applied to the method of claims 1-7, the model comprising: The parameterized graph neural network is used for dynamically predicting a target underground water chemical system based on a graph attention mechanism to obtain a reaction rate parameter of microbial reaction of the target underground water chemical system at the t-th time step and residual results of pH and pe of the underground water hydrologic geochemical environment parameter; The input of the parameterized graph neural network comprises a directed graph constructed according to groundwater geochemical state data of the target groundwater chemical system, and the output of the parameterized graph neural network comprises reaction rate parameters of microbial reactions of the target groundwater chemical system at a t-th time step and residual results of pH and pe; A kinetic physical network for predicting a preliminary prediction result of the reactive solute concentration of the target groundwater chemical system at the t+1th time step according to the reactive solute concentration of the t time step and a reaction rate parameter of the microbial reaction; and the residual error correction network is used for carrying out data-driven residual error correction on the preliminary prediction result of the reactive solute concentration of the target groundwater chemical system in the t+1th time step and the residual error results of the pH and pe in the t time step, and outputting the prediction result of the corrected reactive solute concentration, pH and pe of the target groundwater chemical system in the t+1th time step.
9. The model of claim 8, wherein the directed graph comprises a set of nodes comprising species nodes representing chemical species and reaction nodes representing microbial reaction processes, wherein the initial characteristics of the species nodes consist of logarithmic transformation values of reactive solute concentrations at time step t, groundwater geochemical environment parameters pH values and pe values, and a set of directed edges comprising input edges representing relationships in which chemical species participate as reactants in respective microbial reaction processes, output edges representing relationships in which microbial reaction processes produce respective chemical species, and inhibition edges representing relationships in which chemical species or microbial reaction processes produce inhibition or competition effects on another microbial reaction process.
10. The model of claim 9, further comprising a periodic correction network for periodically correcting the predicted results of the concentration of reactive solutes, pH and pe of the corrected target groundwater chemistry at time step t+1, output by the residual correction network.

Description

Groundwater reactive solute transport prediction method and model based on physical-graphic neural network Technical Field The invention relates to the field of underground water geochemistry, in particular to an underground water reactive solute migration prediction method and model based on a physical-graphic neural network. Background Reactive solute transport simulation requires the simultaneous coupling of hydraulic transport and geochemical reaction processes. The prior art generally solves the fluid transport process based on finite element/finite difference methods and combines high fidelity chemical solvers (e.g., PHREEQC) to handle complex chemical equilibrium and kinetic reactions. Although the method has advantages in accuracy of chemical process description, the method still has significant bottlenecks in practical engineering application, namely firstly, chemical reaction solving involves a rigid equation set, calculation overhead increases in a super-linear manner along with grid refinement and component number increase, large-scale or long-time simulation is difficult to support, secondly, stability problems such as numerical value non-convergence, oscillation and the like are easy to occur under the conditions of high reaction rate, sharp chemical gradient or thin concentration, and in addition, engineering complexity exists in the aspects of coupling integration, parallelization and scene adaptability in the traditional method, so that the application of the traditional method in parameter sensitivity analysis, uncertainty quantification and real-time decision is severely restricted. In order to improve the calculation efficiency, a method of using machine learning as a proxy model to replace part of chemical modules has appeared in recent years, for example, fitting reaction rates by using a neural network or accelerating balance calculation by a table look-up method. However, the existing machine learning method cannot fully embed structural information and chemical constraints of a reaction network, has limited generalization capability in a data sparseness or extrapolation scene, and often lacks physical consistency and engineering integration friendliness. Therefore, a method and model that maintains critical chemical mechanisms, while providing high computational efficiency and good extrapolation capability is needed to facilitate the wide application of reactive solute transport simulation in engineering practice. Disclosure of Invention This disclosure is provided in part to introduce concepts in a simplified form that are further described below in the detailed description. This disclosure is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. In a first aspect, an embodiment of the present application provides a method for predicting groundwater reactive solute transport based on a physical-map neural network, including the steps of: Acquiring groundwater geochemical state data of a target groundwater chemical system at a t-th time step, wherein the groundwater geochemical state data at least comprises the concentration of a reactive solute, the pH value of a groundwater hydrological geochemical environment parameter and the pe value of an oxidation-reduction potential at the t-th time step; constructing a directed graph of the target groundwater chemistry system based on the groundwater geochemical state data; Inputting the directed graph to a parameterized graph neural network based on a graph attention mechanism to infer, so as to obtain a reaction rate parameter of microbial reaction of the target groundwater chemical system at the t-th time step and a residual result of corresponding pH and pe; inputting the reactive solute concentration of the t time step and the reaction rate parameter of the microbial reaction into a kinetic physical network, and predicting the preliminary prediction result of the reactive solute concentration of the target groundwater chemical system at the t+1th time step; And inputting the preliminary prediction result of the reactive solute concentration at the t+1th time step and the residual results of the pH and the pe into a residual correction network for correction to obtain the prediction result of the reactive solute concentration, the pH and the pe of the target groundwater chemical system at the t+1th time step. Further, the directed graph includes a set of nodes including species nodes representing chemical species and reaction nodes representing microbial reaction processes, wherein an initial characteristic of the species nodes is composed of a logarithmic transformation value of reactive solute concentration at a t-th time step, a pH value and a pe value of a groundwater geochemical environment parameter, and a set of directed edges including an input edge for representing a relationship in which the chemical species par