Use of convolutional neural networks with encoder-decoder structure for predicting the inverse operator in hydraulic tomography

In this manuscript, we discuss the capabilities of a deep learning algorithm implemented with Conventional Neural Network concept to characterize hydraulic properties aquifers. The called CNN-HT is designed predict inverse operator tomography using synthetic training dataset in which head data associated pumping tests are linked transmissivity field. This approach relies on an adaptation SegNet network that was initially developed process image segmentation. composed encoders and decoders networks. encoder, sequential operations multiple filters, as convolution, batch normalization, max-pooling performed identify feature maps input data. decoder, up-sampling, normalization regression used prepare output by recovering loss spatial resolution occurred encoder process. adaptation, least-square iterative formulation at initial iteration Jacobian matrix resize match size (transmissivity field). protocol applied computed numerically solving groundwater flow equation for given field, generated geostatistically Gaussian spherical variograms. A part other test its performance. step confirmed effectiveness tool reconstructing main heterogeneities properties, related nature quantity Moreover, method provided inversion results same quality than those obtained Gauss-Newton finite difference or adjoint state computation matrix. However, computational time longer but can be less order method.