Non-stationary autoregressive filters for prediction of subsurface geological structure

Accurate decision-making in the petroleum industry is highly contingent on building a reliable model of the subsurface geological structure. Building a model of the subsurface typically involves solving an inverse problem with acquired data for various model parameters of interest like P-wave velocity, rock porosity etc. However, issues of poor data quality necessitate regularizing the inverse problem, where prior geological information is incorporated. A big issue is that geology is typically highly heterogeneous (non-stationary) and conventional regularization operators fail to capture this non-stationarity. This project uses autoregressive filters that can learn on training images (TIs) of the geological structure, for use as regularization operators in inverse problems. A nonstationary approach is employed, in which multiple filters are trained for a single grid of the subsurface structure. After training, the filters are expected to provide a way of incorporating non-stationary prior information about geological structure into inverse problems.