Identification Of Aquifer Facies Structures Integrating Data Fusion And Deep Learning

Groundwater is present within various facies types,such as sand,clay,and gravel.Different facies types directly determine the range of parameters for groundwater flow and solute transport,including permeability,porosity,and dispersivity.These facies types form a complex spatial distribution known as the aquifer facies structure.Accurate identification of the aquifer facies structure is crucial for obtaining reliable numerical simulation results for groundwater,as inaccurate representation of the aquifer facies structure often leads to larger simulation errors compared to inappropriate model parameters.Previous research has primarily relied on costly borehole data or geophysical data to infer or estimate the spatial distribution of aquifer facies.In contrast,readily available indirect observation data,such as hydraulic head,concentration,and temperature,can provide valuable information about the interaction between groundwater and facies.Consequently,these indirect observation data can also be utilized for identifying the facies structure of the aquifer.By integrating and analyzing these data,along with data fusion algorithms,the identification accuracy of the aquifer facies structure can be improved while reducing uncertainties.There are currently two commonly used data fusion algorithms,one based on stochastic models and the other on deep learning models,but both have certain limitations in practical applications.The aquifer facies structure identified by stochastic models usually still has some uncertainty.On the other hand,deep learning models rely on a substantial amount of reliable samples of aquifer facies structures,which are difficult to obtain in practical applications.To overcome these limitations,this study couples these two models and develops a stage-wise stochastic deep learning inversion framework for reliable identification of aquifer facies structures.Firstly,this paper derives the analytical solution for the transition probability,which enables the precise calculation of the transition probability model through the facies volume proportion and the mean length.Then,the data fusion algorithm is used to identify these two parameters,and the transition probability-based indicator kriging simulation is developed to generate a set of initial aquifer facies structures.Finally,the developed octave convolution adversarial autoencoder is used to learn the initial facies structure,and the data fusion algorithm is combined to optimize and adjust the facies structures,thereby improving the accuracy of the identified aquifer facies structures.Additionally,this paper also develops a deep octave convolution residual network as the forward simulation surrogate model to improve the efficiency of data fusion.In addition to developing appropriate algorithms,the rational selection of observations for data fusion is also crucial for accurately identifying aquifer facies structures.Therefore,this study utilizes the developed algorithm to explore the impact of three common observation types,including hydraulic head,temperature,and concentration,under different observation frequencies and noise levels on the identification results.Finally,the developed algorithm is validated through numerical experiments and applied to the identification of the unconfined aquifer facies structure in Tsitsihar.The main conclusions obtained from this study are presented below:(1)The developed stage-wise stochastic deep learning inversion algorithm can achieve reliable identification of aquifer facies structures without requiring additional training samples,using borehole data and dynamic response observations(e.g.,hydraulic head,concentration).(2)The instability problem in the high-dimensional identification process of aquifer facies structure can be effectively alleviated by the transition probability-based indicator kriging simulation algorithm and the octave convolution adversarial autoencoder.Since these two algorithms can reduce the number of parameters required for the identification process by more than two orders of magnitude.(3)The developed deep octave convolution residual network can achieve fast and accurate prediction of dynamic response distribution fields.In the test scenario of this paper,the deep octave convolution residual network improved the efficiency of dynamic response prediction by more than 50 times,and the overall identification efficiency of aquifer facies structure increased by more than 3 times.(4)Fusing one or multiple types of observations can significantly improve the accuracy of identifying aquifer facies structures.However,fusing more types of observations is not always beneficial for obtaining more accurate dynamic response predictions.High observation noise can decrease the accuracy of identifying aquifer facies structures,while high observation frequency can mitigate the negative impact of high observation noise.(5)The developed algorithm can effectively and reliably identify the aquifer facies structures in the study area.The identified structures can well reflect the facies distribution characteristics of the aquifer in the region.The water level prediction values based on the identified structures are basically consistent with the observations.This indicates that reliable groundwater flow simulation can be achieved based on the identified aquifer facies structures,which is helpful for the subsequent ecological water replenishment,pollution monitoring,and remediation projects in the study area.The aforementioned study enriches and extends the theoretical basis and technical connotation of aquifer facies structures identification,providing a practical and feasible method for applications such as aquifer pollution prediction and early warning,groundwater resource management,geological carbon dioxide storage,and geological disposal of nuclear waste.