K-fold Cross-validation of Multiple-point Statistical Simulations
P. Juda, P. Renard and J. Straubhaar
Event name: Petroleum Geostatistics 2019
Session: Multi-point Geostatistics
Publication date: 02 September 2019
Info: Extended abstract, PDF ( 636.32Kb )
Price: € 20
In reservoir models, the choice of spatial interpolation or stochastic simulation methods for subsurface properties is crucial when dealing with heterogeneous media. Multiple-point statistics (MPS) algorithms allow to simulate complex structures but they are controlled by hyper-parameters whose identification can be tedious. Furthermore, many different geostatistical methods and models are available. In this work, we present an application of K-fold cross-validation for the selection of a spatial simulation method. The proposed technique allows to rank models based on their predictive accuracy and is completely generic: it can handle categorical and continuous variables, as well as compare MPS algorithms to variogram-based, or object based models. It can be used for the selection of any type of parameters, including the choice of the training image. We demonstrate the performance of the method on a synthetic test case used previously for benchmarking training image selection techniques and on a real field application including non-stationarity.