With
growing physical complexity, the importance of nonlinear and linear
solution processes is underscored. Our aim is to make fundamental
advances in several challenge areas that arise ubiquitously in complex
application scenarios. Leveraging these fundamental advances, we also
aim to
develop lasting, general, and robust solution technologies and methods.
Our ultimate measure of success will be the wide and successful
adoption of our methods in industrial and academic simulators to
perform full-resolution, full-Physics modeling of seriously complex
large-scale emerging processes.
Locality is inherent to
all transient flow and transport phenomena. The superposition of the
two disparate spatiotemporal characteristic scales that underly flow
and transport leads to a problem of dimensionality. While the extent of
an adequate simulation domain is dictated by the nature of flow, the
resolution of its discrete form is restricted by the transport
variables.
As a first response to this issue of dimensionality, Adaptive
Discretization methods have been devised to exploit an a priori
understanding of locality in order to reduce the computational cost of
simulations. While such methods have provided various degrees of
success in applications, it remains that they are inherently restricted
by the fidelity of the discretization under aggressive adaptivity.
Our work in this areas seeks an alternate strategy. We are devising
nonlinear solution processes that internally adapt the level of
computation to precisely match that of the underlying spatiotemporal
changes in state. This will enable the efficient use of high-fidelity,
static, and full-resolution discretizations that do not require global
computations unless absolutely necessary.
Fundamental Question: Over
a Newton iteration, which cells will experience a nonzero update in
state?
Encouraged by an
initial result,
we are answering this question for general flow and transport,
independent
of the spatial discretization that is used. The result is obtained by
the application of the infinite dimensional Newton iteration process to
the semidiscrete governing equations. That is, suppose that the
governing equation has the form,
Then the semidiscrete fully-implicit form is simply,
In reservoir simulation, the generally nonlinear and heterogeneous flow
term is discretized in space using an accurate method. We will not
perform this discretization. rather we will work with the infinite
dimensional (continuous) problem itself.
Over a single time-step, the initial state is,
and the subsequent state at the end of the time-step is,
Newton's method can be applied to any differentiable nonlinear
operator, including an infinite dimensional form. The iteration
produces the sequence of iterations,
Note that this is a sequence of functions defined over the entire
domain. The equation for the updates can be written as,
where the generalized Frechet derivative of the infinite dimensional
residual equation is a linear partial differential equation. The Newton
updates in a reservoir simulator are numerical approximations to the
infinite dimensional update that can be obtained analytically in
certain cases.
We have derived sharp conservative estimates for the analytical
infinite dimensional Newton updates for general sequential flow and
transport!
Using these estimates which are essentially free to evaluate, we can
tell which cells within the domain will experience change over a Newton
step, and which ones will not. A whitepaper describing this process is
available
here.
Our current research in this area is focused on:
- Deriving estimates for the coupled form of reactive
flow and transport in multiple dimensions.
- Extensions to general nonlinear multicomponent
advection, diffusion, and reaction problems including thermal.
- Exploration of applications in fractured and
stimulated media.
Application 1: solvers
that only solve for cells that will experience change.
In any iterative nonlinear solution process that generates iterates
that are the solution of a large sparse linear system, why solve for
the entire system if we know that only a portion of the resulting
vector has nonzero entries? Moreover, at initial iterates, far from the
solution, it may be unnecessary to compute the linear solution very
accurately. Our universal results on locality, provide an a priori
identification of these vector components that is available at
negligible cost.
We are developing application grade solvers that do this. Compared to
detection methods that are designed at the discrete level such as in
this
article, our methods are general to both flow and transport, and
are available by evaluating a single formula!
Our current research in this area is focused on:
- Devising solvers that do not degrade the convergence
rate of the nonlinear iteration.
- Developing strategies that adaptively alter the
target linear convergence tolerance.
Application 2: Adaptive
Mesh Refinement inside the nonlinear solver and not within the
discretization.
Particularly in the case of fast physics such as flow, the nonlinear
iteration updates may have relatively global support. In cases where
these updates decay rapidly around certain points and very slowly near
the boundaries, it is desirable to use adaptively localized meshes.
Keeping the discretization mesh static, we are devising methods that
solve the nonlinear residual using adaptive solver meshes. These ideas
are inspired by nonlinear control optimization methods that adapt the
resolution of control variables to reduce computational costs.
Application 3: Multirate
nonlinear methods
Newton's method is an explicit Euler discretization of the Newton Flow
system of Ordinary Differential Equations using a stepsize of one. We
propose to develop multirate or local timestepping methods to integrate
the Newton Flow system. Using our results on spatiotemporal locality,
we can decompose the spatial domain into several sets based on the
anticipated rate of change in state.
Research
is underway to
develop nonlinear iterations that have an expanded spatiotemporal
support. Our initial point of departure is to seek methods that are
motivated by higher-order ODE integration. While expanded support has
been observed with such methods, the additional computational costs of
each step render them value-neutral. Current research is focused on
retaining the
expanded support properties while reducing the computational cost per
iteration. Our approach to achieving this uses a differential geometry
perspective, where the dynamics of motion can be abstracted with linear
transformation operators. A promising idea is to develop extrapolation
approaches on the Frenet Frame of the solution path.
Application 2: If we have
bounds on the asymptotic convergence rates, we may better control
timestep size selection for convergence.
A
direct application of our results that characterize convergence rate is
to
use them in order to select timestep size in a manner that minimizes
the
computational cost per unit time advancement in implicit simulation.
Suppose that at a specific timestep, time truncation error control
dictates a maximum allowable timestep size,
Then, using our estimate for the number of iterations required for
convergence as function of timestep size,
the objective is to select that target timestep size that satisfies the
minimization problem,