Block Preconditioners with Algebraic Multigrid Block Solve in Stratigraphic Modeling for oil exploration

Masson Roland

Institut Francais du Petrole, 1 et 4 Avenue Bois Preau, 92000 Rueil Malmaison, France, roland.masson@ifp.fr

Gervais Veronique


Abstract

Stratigraphic models simulate the erosion and sedimentation of sedimentary basins at geological time scales given the sea level variations, the tectonics displacements of the basement, and the sediments fluxes at the boundary of the basin. We consider in this talk a sediments transport model coupling three main processes: a gravity driven transport of the sediments for which the fluxes are proportional to the gradient of the topography, a weather limited transport taking into account the disymmetry between erosion and sedimentation, and a fluvial transport model for which the sediments fluxes are proportional to the water discharge. The main variables of the problem are the sediment thickness, a flux limitor, and the L sediments concentrations in basic lithologies such as sand or shale or carbonates. Such model is applied in oil exploration for a better prediction of potential reservoirs location. The model is derived writing the mass conservation of each lithology leading to a system of mixed parabolic hyperbolic type. It is discretized by a finite volume scheme in space and a fully implicit time integration, leading to the solution at each time step of a non linear systems of L+2 variables on the 2D mesh. After Newton type linearization, we are left with the solution of an ill conditioned linear system with sharp jumps in the diffusion coefficients and coupling L+2 variables of mixed types. These linear systems are solved using an iterative solver and a block approach for the preconditioner in order to separate the different variables. In a first step, a mixture equation is obtained by linear combinations of the rows that should concentrate the ellipticity of the system and as much as possible decouple the first two variables (sediment thickness and flux limitor) from the L concentrations variables. Then the overall system is solved using a block Gauss Seidel preconditioner. The First block (sediment thickness and flux limitor variables) is preconditioned either by a direct sparse solver or by an ILU0 incomplete factorization, or by a vcycle of an algebraic multigrid solver (AMG1R5 from Ruge and Stuben) on the sediment thickness sub-block only in order to avoid non diagonal dominance. The remaining blocks are solve using a gauss seidel sweep in topographical order leading to an exact inversion of these blocks. These block preconditioners are compared with a sparse direct solver and an ILU0 incomplete factorization on the overall system. The comparison is made in terms of CPU time, and scalability with respect to the mesh size and the jump of the diffusion coefficients on two real test cases: the Paris basin and a Rift test case with water discharge. It shows that the block approach combined with a multigrid preconditioning of the the sediment thickness sub-block is a very efficient method, nearly scalable for this problem.