Leveraging Machine Learning for Enhanced Geostatistical Modelling of Reservoir Properties

Summary This presentation looks at a new method leveraging both classical geostatistical modelling and a specially modified machine learning algorithm to provide reservoir models. The method differs significantly from classical geostatistical methods in that it starts with direct estimates of the conditional distributions of the variable to be studied. As well as being far simpler for the user than the classical route, it avoids some of the potentially damaging effects of the 'stationary hypothesis' that is needed for traditional property models. Since it provides the conditional distribution at each location, it immediately provides the user with an estimate of the reservoir value, a realistic uncertainty measurement, quantiles and truncations (e.g probability than porosity is greater than 20%). The method also provides simulations of the spatial distribution of the target variable(s) of interest, starting from the set of conditional distributions. Since there was no hypothesis of stationarity made during estimation, this allows the simulations to rapidly adapt to local variations in the reservoir (e.g seismic quality or porosity variability). We apply this method to a Fluvio-Deltaic Triassic Gas Field with two differing hydrocarbon-bearing geological formations.