Data-Driven Intelligent Inversion for Remaining Oil Using Nonrepeatable Time-Lapse Seismic: Tarim Clastic Reservoir Study

This study addresses the low prediction accuracy of remaining oil distribution in late-development-stage clastic reservoirs within the Tarim Basin. By integrating multi-phase non-repeatable time-lapse seismic data with dynamic production data, we overcome the limitations of static seismic interpretation and establish a quantitative prediction method for remaining oil distribution in clastic reservoirs. Dynamic production well data are first utilized to calibrate time-variant oil saturation patterns, enabling the extraction of dynamic reservoir response characteristics. Subsequently, a machine learning prediction model is developed to achieve three-dimensional spatial quantitative characterization of remaining oil enrichment zones. Results demonstrate that integrating dynamic production data with time-lapse seismic analysis significantly improves prediction accuracy. Multi-phase non-repeatable seismic data prove critical for reliable predictions, while machine learning algorithms effectively reveal nonlinear relationships between seismic difference volumes and reservoir dynamics. In field tests, the model successfully identified three concealed remaining oil “sweet spots,” enabling optimized well placement. This approach enhances recovery efficiency in mature oilfields by providing spatially precise targets for development adjustments. Key innovations include: (i) a systematic integration of dynamic and seismic data, (ii) a machine learning architecture decoding reservoir heterogeneity patterns, and (iii) a transferable workflow for late-stage reservoir management. The methodology offers both theoretical advancements and practical solutions for sustainable hydrocarbon recovery in aging reservoirs.