Presenter
Jan Tveranger from NORCE
Co-authors
Bjarte Lønøy, NORCE - Norwegian Research Centre AS
Abstract
Providing realistic training data is key to developing machine learning algorithms that can support geological interpretation of LWD logs. Ample training data can be provided by conventional geocellular models aimed at capturing spatial property distributions and expected uncertainty along planned well trajectories. However, for synthetic logs to be comparable to actual LWD data, the generation of synthetic logs from geo-models must emulate the relevant tool configuration and the spatial/geometric constraints of the actual measurement. This study addresses this issue by investigating sampling techniques for generating synthetic logs from geocellular models, which approximate the physical reality of LWD resistivity log measurements in the subsurface. The techniques presented here use industry-standard reservoir modelling software and simple Python scripts. This allows pre-drill automated extraction of synthetic logs for machine learning from case specific 3D geologic training models. The approach bridges the gap between forecasts of subsurface architecture, properties, and associated uncertainties as captured by target geomodels, and expected well log responses.
Biography
Geologist and senior researcher at NORCE. 25 years + experience from academia and industry. Specializing in reservoir geology and geomodelling.