On the Application of Supercomputers for History Matching Multiphase Oil Reservoir Models

This paper discusses the use of algorithm, combined with the multilevel grid method and its parallel implementation for history matching multiphase oil reservoir models on supercomputers. Many existing algorithms for history matching suffer the same difficulty that the quality of the identified parameter distribution is very sensitive to the initial guess. The situation has been improved by the use of the multilevel grid in our algorithm. The block SOR method with red and black ordering is used to solve the algebraic system in parallel for the forward problem, and a special QR factorization method is developed to solve the regularized least square inverse problem. In addition to those algebraic equation solvers, the rest of the algorithm has also been parallelized and vectorized as much as possible. Timing results show that the vectorization and parallelization of that part, which is often neglected in many application programs, can also improve the overall algorithm performance significantly.