Mapping discrete fracture networks using inversion of hydraulic tomography data with convolutional neural network: SegNet-Fracture

Abstract

In this paper, we propose a new method to map the fracture network structure in a heterogeneous aquifer from inversion hydraulic head data measured during pumping tests in hydraulic tomography mode. This inversion tool is based on the new concept of convolutional neural networks, which provides a direct approximation to the inverse function linking fracture geometry to hydraulic data. In order to handle the highly nonlinear inverse function more effectively, an advanced neural network is developed from SegNet architecture with encoder-decoder structure, which excels in image processing to translate the water level image associated with the pumping tests at the input into a fracture map at the output. The network is trained with a synthetic dataset where the fracture structure and matrix heterogeneity are randomly generated, and the hydraulic head are obtained by solving the groundwater flow equation. The trained network accurately maps different complexity levels of fractures embedded in a matrix with heterogeneous transmissivity.As a data-driven approach, the accuracy of the mapping depends on the quantity, quality, and relevance of the synthetic dataset used in the training phase. While generating data to train the network requires effort, the trained network performs each inversion instantly. The inversion result appears to be stable even in the presence of data noise, reliably interprets the hydraulic data if they share comparable fracture and matrix properties as specified in the training models.