Time-lapse reservoir property change estimation from seismic using machine learning

Abstract

Time-lapse seismic analysis is an important tool in reservoir management, well planning, and reservoir model updating. Existing 4D close-the-loop methods utilize rock-physics models and simulation models, and typically work well in simple 4D cases such as water flood movement (where the primary change is the saturations), but they face challenges in reliably and quantitatively estimating 4D reservoir property changes when the reservoir dynamics are complex. We demonstrate a data-driven quantitative method that uses machine learning to leverage the inherent physics between time-lapse reservoir property changes and seismic attributes. The method can simultaneously utilize multiple seismic attributes, including attributes derived from prestack seismic, from multiple 4D seismic surveys. It significantly improves the efficiency and reduces the cycle time for quantitative 4D interpretation. We evaluate the method through a study based on synthetic data modeled after a complex North Sea reservoir with long production and injection history. For this reservoir, separation and estimation of the time-lapse pressure and saturation change are the critical objectives. Results show high prediction accuracy for all the reservoir properties including pressure, saturations, and compaction changes for the blind validation test. Estimated maps of all the reservoir property changes match very well with the simulation model in the synthetic study. This work also demonstrates that prestack seismic attributes significantly improved the estimation of 4D pressure and saturation changes.