Semi-smooth Newton methods for nonlinear complementarity formulation of compositional two-phase flow in porous media

TL;DR

This paper introduces semi-smooth Newton methods to efficiently solve nonlinear complementarity problems in compositional two-phase flow modeling, improving robustness and allowing larger time steps compared to traditional primary variable switching methods.

Contribution

It develops and compares two semi-smooth Newton approaches, including a new inexact Newton method based on Jacobian smoothing, for better handling phase transitions in porous media flow simulations.

Findings

The new inexact Newton method is robust and efficient.

Methods perform well on benchmark and realistic heterogeneous media.

Approach allows larger time steps than PVS methods.

Abstract

Simulating compositional multiphase flow in porous media is a challenging task, especially when phase transition is taken into account. The main problem with phase transition stems from the inconsistency of the primary variables such as phase pressure and phase saturation, i.e. they become ill-defined when a phase appears or disappears. Recently, a new approach for handling phase transition has been developed, whereby the system is formulated as a nonlinear complementarity problem (NCP). Unlike the widely used primary variable switching (PVS) method which requires a drastic reduction of the time step size when a phase appears or disappears, this approach is more robust and allows for larger time steps. One way to solve an NCP system is to reformulate the inequality constraints as a non-smooth equation using a complementary function (C-function). Because of the non-smoothness of the constraint equations, a semi-smooth Newton method needs to be developed. In this work, we consider two methods for solving NCP systems used to model multiphase flow: (1) a semi-smooth Newton method for two C-functions: the minimum and the Fischer-Burmeister functions, and (2) a new inexact Newton method based on the Jacobian smoothing method for a smooth version of the Fischer-Burmeister function. We show that the new method is robust and efficient for standard benchmark problems as well as for realistic examples with highly heterogeneous media such as the SPE10 benchmark.