New Parallelized Adaptive Implicit Methods for Large Compositional Reservoir Simulations

Simulation of large compositional reservoir models in an acceptable CPU time still remains a key issue for any reservoir simulation software. Parallel computing is thus required to speed up the simulations and robust numerical solutions need to be implemented to ensure accurate results, especially for simulations of complex compositional physics. In that context, classical fully implicit schemes are often not practical or, even unusable, due to memory and CPU restrictions. We propose new time discretization methods which mix the accuracy and stability of the fully implicit methods and CPU performances and low memory consumption of the explicit ones. Within these approaches, a fully implicit scheme is applied in parts of the reservoir with highest physical property variations, and an explicit method is used in the other ones. Explicit schemes are delimited by a stability CFL condition. Explicit cells with large variations on the unknown are switched dynamically to implicit ones to improve global simulation CPU performances. These status computations are based on either threshold or spectral criteria. Implicit grid blocks may also be discretized using fully implicit or quasi-implicit schemes depending on the fluid composition variations.
These different schemes have been parallelized and integrated in a multi-purpose industrial reservoir simulator. The CFL criteria are discussed and tested on academic reference cases and then, comparisons of the available methods are presented for large compositional real case studies run on different parallel hardware platforms.