Research On Inverse Problem Of Aquifer Parameters Based On Deep Learning

In the modeling process of groundwater flow model,a basic problem is to estimate the aquifer parameters through numerical models or observations obtained from field measurements(e.g.,head value,solute concentration value)and a priori information.Accurate aquifer parameters are the key to whether the forward model can reflect the hydrogeology of the region.At the same time,in the problem of groundwater pollution,accurate aquifer parameters are crucial for predicting the impact of pollutants and subsequent scientific remediation decisions..However,in practical problems,especially when the parameter dimension is high,it is challenging to invert aquifer parameters by forward modeling the state variables of the model.When dealing with the problem of aquifer parameters with strong heterogeneity,even if the parameters are reduced by some dimensionality reduction methods(e.g.,Karhunen-Loève method),usually a large number of random variables are needed to express the heterogeneity of aquifer parameters,So we need to design a method to solve such problems.In recent years,with the improvement of computer computing power,deep learning has opened a new door for knowledge representation and complex pattern recognition.In particular,the development of Convoluted Neural Network(CNN),which is based on the design principle of receptive field,extracts local features of the image,and then up-samples by transposing the convolution layer,which is very suitable for capturing subtle Feature,to achieve translation between pictures,this idea is very suitable to be extended to solve the problem of groundwater aquifer parameter inversion.Based on CNN network,this paper designs a deep learning model for solving the inverse problem of aquifer parameters.On this basis,by using the characteristics of the generative adversarial network(GAN)in learning cross-domain mapping for image translation,this paper proposes an adversarial generative network model HP-GAN for aquifer parameter identification for simultaneous learning.The bidirectional mapping between the groundwater high-dimensional parameter space and the corresponding state space modeled by the forward model,and using HP-GAN to demonstrate the representative problem of underground flow modeling,compared with the results of inversion directly through the CNN model.The experimental results show that HP-GAN has achieved satisfactory accuracy in identifying the mapping of head values and aquifer parameters,and provides a new method based on deep learning for solving complex hydrogeological inverse problems.