Panzer: Time-averaged data

[delcmo]

Hello,

We are working on implementing large eddy simulation capability and would need to compute time-averaged quantities. Our plan is to define a nodal fields/variables and closure model on the same model as what is done for initial conditions that initialize a field at the nodes. I am not entirely sure how to add this closure model to the tree and how it should be called at run time. Are there any examples in Panzer that I could rely on to help with the implementation?

Thanks,

Marco

[cgcgcg]

@rppawlo

[delcmo]

All,

as a first attempt I wrote a closure model that works on quadrature points to calculate the time average of a given field. The implementation is as below:

//---------------------------------------------------------------------------//
template<class EvalType, class Traits>
void TimeAverage<EvalType, Traits>::evaluateFields(
    typename Traits::EvalData workset)
{
    _dt = workset.step_size;
    if (workset.time > 0.1)
    {
        _atime += _dt;
        _beta = _dt / _atime;
    }
    else
    {
        _atime = 0.0;
        _beta = 0.0;
    }
    auto policy = panzer::HP::inst().teamPolicy<scalar_type, PHX::Device>(
        workset.num_cells);
    Kokkos::parallel_for(this->getName(), policy, *this);
}

//---------------------------------------------------------------------------//
template<class EvalType, class Traits>
void TimeAverage<EvalType, Traits>::operator()(
    const Kokkos::TeamPolicy<PHX::exec_space>::member_type& team) const
{
    const int cell = team.league_rank();
    const int num_point = _field.extent(1);

    Kokkos::parallel_for(
        Kokkos::TeamThreadRange(team, 0, num_point), [&](const int point) {
            if (_atime > 0.0)
            {
                _avg_field(cell, point) = (1.0 - _beta)
                                              * _avg_field(cell, point)
                                          + _beta * _field(cell, point);
            }
            else
            {
                _avg_field(cell, point) = 0.0;
            }
        });
}

I tested the above logic with two approaches.

In a first approach, the closure model is triggered when writing the time-averaged field time_average_velocity_0 to the solution file (Exodus) for every time step. This seems to be working fine on one or multiple cores:

Figure 1: time-averaged velocity and final velocity.

In a second approach, I call the closure model by setting a Response_Functional since ultimately I want to evaluate the time-averaged field after each time step without writing it to the solution file. I face two issues with this approach:

• when the closure model is called by the Response_Functional, the time and the time step are not set and return nan.
• the time-averaged field is updated by the Response_Functional but also when writing the field to the solution file (Exodus). The time-averaged field is updated twice for time steps that write the solution to the Exodus file.

After testing these two approaches, I am not sure that a Response is the proper way to update the time-averaged field after each time step.

Thanks,
Marco