Machine learning based deep carbonate reservoir characterization with physical constraints

Abstract

Summary Seismic characterization of deep carbonate reservoir is challenging due to the heterogeneous reservoir properties caused by the complex diagenesis and deep buried physical conditions. We propose a variety of physical constraints (including spatial constraint, continuity constraint, gradient constraint and category constraint) to guide the machine learning (Random Forest method) for reservoir quality prediction using multi-seismic attributes. Taking the carbonate reservoirs in the Tarim Basin, Western China as an example, we demonstrate that, various physical constraints are effective in enhancing the prediction performance based on the well test. The combination of the four proposed physical constraints gives the best prediction performance in terms of identifying reservoir and non-reservoir as well as inferring reservoir quality. We also show that a two-step strategy gives higher F1 score for reservoir quality evaluation. Machine learning based seismic prediction of deep carbonate reservoir with physical constraints suggests that this approach can effectively delineate the heterogeneous reservoir distribution, laying the foundation for geological model building and sweet spot detection.