Geological feature selection in reservoir modelling and history matching with Multiple Kernel Learning

There is a continuous challenge in identifying and propagating geologically realistic features into reservoir models. Many of the contemporary geostatistical algorithms are limited by various modelling assumptions, like stationarity or Gaussianity. Another related challenge is to ensure the realistic geological features introduced into a geomodel are preserved during the model update in history matching studies, when the model properties are tuned to fit the flow response to production data. The above challenges motivate exploration and application of other statistical approaches to build and calibrate reservoir models, in particular, methods based on statistical learning.The paper proposes a novel data driven approach – Multiple Kernel Learning (MKL) – for modelling porous property distributions in sub-surface reservoirs. Multiple Kernel Learning aims to extract relevant spatial features from spatial patterns and to combine them in a non-linear way. This ability allows to handle multiple geological scenarios, which represent different spatial scales and a range of modelling concepts/assumptions. Multiple Kernel Learning is not restricted by deterministic or statistical modelling assumptions and, therefore, is more flexible for modelling heterogeneity at different scales and integrating data and knowledge.We demonstrate an MKL application to a problem of history matching based on a diverse prior information embedded into a range of possible geological scenarios. MKL was able to select the most influential prior geological scenarios and fuse the selected spatial features into a multi-scale property model. The MKL was applied to Brugge history matching benchmark example by calibrating the parameters of the MKL reservoir model parameters to production data. The history matching results were compared to the ones obtained from other contemporary approaches – EnKF and kernel PCA with stochastic optimisation.