Machine Learning Clustering of Reservoir Heterogeneity with Petrophysical and Production Data

Reservoir development decisions strongly depend on our understanding on reservoir heterogeneity, which is often subject to sparse and conflicting data, interpretational bias and constraints imposed by the modelling assumptions. The work tackles a challenging task of accurately and quickly identifying and describing uncertainty in the spatial distribution of reservoir heterogeneity derived from geological well data and with respect to a geological concept. We propose a metric based machine-learning approach to identify and describe spatial trends in reservoir heterogeneity/facies property distribution using wireline and production data.
We demonstrate how the proposed method can help to partition reservoir heterogeneity and discover and verify spatial trends for a real mature producing field in the Western Siberia. The obtained clustering of reservoir facies based on the wireline logs (alpha-SP) demonstrated a good agreement with the reservoir zonation based on manual log interpretation and the geological concept. Clustering based on individual well production profiles has confirmed the reservoir partitioning and matched some of the reservoir features aligned with the prevailing geological concept. The outcome of the proposed method helps to improve the facies distribution model by integrating the discovered spatial trends into a geostatistical model and account for uncertainty in the depositional scenario that is difficult to quantify based on manual interpretation.