A Physics-Driven Deep Learning Network for Subsurface Inversion

Abstract

Subsurface inversion is an essential technique for many applications including seismic processing, oilfield well logging an geosteering. Conventional inverse methods based on optimization are time-consuming and sensitive to initial values. The traditional lookup table approach which is limited by the table size could reduce the computational time but only achieves low accuracy. To solve these issues, we propose a physics-driven Deep Neural Network (PhDNN) for solving non-linear inverse problems. In this framework, the physical forward model is utilized to produce a data misfit. Both the model misfit and data misfit are used to train the network. As an example, we use this framework to solve a geosteering problem which enables the drilling direction adjusted by collected resistivity well logging measurements. Numerical tests indicate that the proposed network could improve the quality of the prediction significantly.