Multiple kernel learning approach for reservoir modelling

The paper proposes a novel data driven approach for modelling petrophysical properties in oil reservoir. We aim to improve realism of reservoir models with a more intelligent way of integrating the raw data and geological knowledge. Multiple Kernel Learning (MKL) provides enhanced interpretability of the model by using separate kernels for input variables performs kernel/feature selection to solve a regression problem in a high-dimensional feature space. MKL has an advantage of rigorous control over the model complexity to achieve the right balance between data fit and prediction accuracy. The MKL reservoir model was designed to integrate data and prior knowledge, which describe geological structure at multiple scales. Geological structures can be detected by applying convolution filters on noisy seismic data to capture changes in gradients, orientations, sizes of meandering channels. Such "geo-features" are added as input variables into the MKL model, which optimises the weighted combination of kernels to fit to the available data. MKL application to a synthetic meandering channel reservoir has demonstrated capacity of selecting the relevant input information for detecting the channel structure. Experiments with noisy seismic inputs highlighted feature selection skills of MKL which was able to filter them out.