(sec:benchmark:operator-splitting)=
# Benchmarks for operator splitting
*This section was contributed by Juliane Dannberg.*
Models of mantle convection and lithosphere dynamics often also contain
reactions between materials with different chemical compositions, or processes
that can be described as reactions. The most common example is mantle melting:
When mantle temperatures exceed the solidus, rocks start to melt. As this is
only partial melting, and rocks are a mixture of different minerals, which all
contain different chemical components, melting is not only a phase transition,
but also leads to reactions between solid and molten rock. Some components are
more compatible with the mineral structure, and preferentially stay in the
solid rock, other components will mainly move into the mantle melt. This means
that the composition of both solid and melt change over time depending on the
melt fraction.
Usually, it is assumed that these reactions are much faster than convection in
the mantle. In other words, these reactions are so fast that melt is assumed
to be always in equilibrium with the surrounding solid rock. In some cases,
the formation of new oceanic crust, which is also caused by partial melting,
is approximated by a conversion from an average, peridotitic mantle
composition to mid-ocean ridge basalt, forming the crust, and harzburgitic
lithosphere, once material reaches a given depth. This process can also be
considered as a reaction between different compositional fields.
This can cause accuracy problems in geodynamic simulations: The way the
equations are formulated (see equations {math:numref}`eq:stokes-1`--{math:numref}`eq:compositional`), ideally
we would need to know reaction rates (the $q_i$) between the different
components instead of the equilibrium value (which would then have to be
compared with some sort of "old solution" of the compositional
fields). Sometimes we also may not know the equilibrium, and would only be
able to find it iteratively, starting from the current composition. In
addition, the reaction rate for a given compositional field usually depends on
the value of the field itself, but can also depend on other compositional
fields or the temperature and pressure, and the dependence can be nonlinear.
Hence, has the option to decouple the advection from reactions between
compositional fields, using operator splitting.
Instead of solving the coupled equation
```{math}
\begin{aligned}
\frac{\partial \mathfrak{c}(t)}{\partial t} + \mathbf u\cdot\nabla \mathfrak{c}(t)
&=
q(\mathfrak{c}(t)),
\end{aligned}
```
and directly obtaining the composition
value $\mathfrak{c}(t^{n+1})$ for the time step $n+1$ from the value
$\mathfrak{c}(t^{n})$ from the previous time step $n$, we do a first-order
operator split, first solving the advection problem
```{math}
\begin{aligned}
\frac{\partial \mathfrak{c}(t)}{\partial t} + \mathbf u\cdot\nabla \mathfrak{c}(t)
&=
0,
&
\text{obtaining } \Delta \mathfrak{c}_A(t^{n+1}) \text{ from } \mathfrak{c}(t^{n}),
\end{aligned}
```
using the advection time step $\Delta t_A = t^{n+1} - t^{n}$. Then we solve
the reactions as a series of coupled ordinary differential equations
```{math}
\begin{aligned}
\frac{\partial \Delta \mathfrak{c}_R(t)}{\partial t}
&=
q(\mathfrak{c}(t^n))+\Delta \mathfrak{c}_A(t^{n+1})+\Delta \mathfrak{c}_R(t),
&
\text{obtaining } \Delta \mathfrak{c}_R(t^{n+1}) \text{ from } \mathfrak{c}(t^n)+\Delta \mathfrak{c}_A(t^{n+1}).
\end{aligned}
```
This can be done in several iterations, choosing a different, smaller time
step size $\Delta t_R \leq \Delta t_A$ for the time discretization. The
updated value of the compositional field after the overall (advection +
reaction) time step is then obtained as
```{math}
\begin{aligned}
\mathfrak{c}(t^{n+1})
&=
\mathfrak{c}(t^{n}) + \Delta \mathfrak{c}_A(t^{n+1})+\Delta \mathfrak{c}_R(t^{n+1}).
\end{aligned}
```
This is very useful if the time scales of reactions are different from the
time scales of convection. The same scheme can also be used for the
temperature: If we want to model latent heat of melting, the temperature
evolution is controlled by the melting rate, and hence the temperature changes
on the same time scale as the reactions.
We here illustrate the way this operator splitting works using the simple
example of exponential decay in a stationary advection field. We will start
with a model that has a constant initial temperature and composition and no
advection. The reactions for exponential decay
```{math}
\begin{aligned}
\mathfrak{c}(t)
&=
\mathfrak{c}_0 e^{\lambda t} \text{ with } \lambda = - \log(2)/t_{1/2},
\end{aligned}
```
where $\mathfrak{c}_0$ is the initial composition and $t_{1/2}$ is the half
life, are implemented in a shared library
([benchmarks/operator_splitting/exponential_decay/exponential_decay.cc](https://www.github.com/geodynamics/aspect/blob/main/benchmarks/operator_splitting/exponential_decay/exponential_decay.cc)). As
we split the time-stepping of advection and reactions, there are now two
different time steps in the model: We control the advection time step using
the 'Maximum time step' parameter (as the velocity is essentially
0, we can not use the CFL number), and we set the reaction time step using the
'Reaction time step' parameter.
```{literalinclude} exponential_decay.part.1.prm
```
To illustrate convergence, we will vary both parameters in different model
runs.
In our example, we choose $\mathfrak{c}_0=1$, and specify this as initial
condition using the `function` plugin for both composition and temperature. We
also set $t_{1/2}=10$, which is implemented as a parameter in the
`exponential decay` material model and the `exponential decay heating` model.
```{literalinclude} exponential_decay.part.2.prm
```
The complete parameter file for this setup can be found in
[benchmarks/operator_splitting/exponential_decay/exponential_decay.base.prm](https://www.github.com/geodynamics/aspect/blob/main/benchmarks/operator_splitting/exponential_decay/exponential_decay.base.prm).
{numref}`fig:exponential-decay` shows the convergence behavior of these models: As there is
no advection, the advection time step does not influence the error (blue data
points). As we use a first-order operator split, the error is expected to
converge linearly with the reaction time step $\Delta t_R$, which is indeed
the case (red data points). Errors are the same for both composition and
temperature, as both fields have identical initial conditions and reactions,
and we use the same methods to solve for these variables.
```{figure-md} fig:exponential-decay
Error for both compositional field and temperature compared to the analytical solution, varying the time steps of advection (blue data points and and top/blue x axis) and reactions (red data points and and bottom/red x axis), while keeping the other one constant, respectively.
```
For the second benchmark case, we want to see the effect of advection on
convergence. In order to do this, we choose an initial temperature and
composition that depends on x (in this case a sine), a decay rate that
linearly depends on z, and we apply a constant velocity in x-direction on all
boundaries. Our new analytical solution for the evolution of composition is
now
```{math}
\begin{aligned}
\mathfrak{c}(t)
&=
\sin (2\pi(x-t v_0)) \, \mathfrak{c}_0 e^{\lambda z t}.
\end{aligned}
```
$v_0$ is the constant velocity, which we set to 0.01 m/s. The parameter file for
this setup can be found in
[benchmarks/operator_splitting/advection_reaction/advection_reaction.base.prm](https://www.github.com/geodynamics/aspect/blob/main/benchmarks/operator_splitting/advection_reaction/advection_reaction.base.prm).
```{figure-md} fig:advection-reaction
Error for both compositional field and temperature compared to the analytical solution, varying the time steps of advection (blue data points and top/blue x axis) and reactions (red data points and bottom/red x axis), while keeping the other one constant, respectively.
```
{numref}`fig:advection-reaction` shows the convergence behavior in this second set of models:
If we choose the same resolution as in the previous example (left panel), for
large advection time steps $\Delta t_A > 0.1$ the error is dominated by
advection, and converges with decreasing advection time step size (blue data
points). However, for smaller advection time steps, the error stagnates. The
data series where the reaction time step varies also shows a stagnating error.
The reason for that is probably that our analytical solution is not in the
finite element space we chose, and so neither decreasing the advection or the
reaction time step will improve the error. If we increase the resolution by a
factor of 4 (right panel), we see that that errors converge both with
decreasing advection and reaction time steps.
The results shown here can be reproduced using the bash scripts `run.sh` in
the corresponding benchmark folders.