time_dependent.cc

Go to the documentation of this file.

// ------------------------------------------------------------------------
//
// SPDX-License-Identifier: LGPL-2.1-or-later
// Copyright (C) 1999 - 2025 by the deal.II authors
//
// This file is part of the deal.II library.
//
// Part of the source code is dual licensed under Apache-2.0 WITH
// LLVM-exception OR LGPL-2.1-or-later. Detailed license information
// governing the source code and code contributions can be found in
// LICENSE.md and CONTRIBUTING.md at the top level directory of deal.II.
//
// ------------------------------------------------------------------------
 
#include <deal.II/base/memory_consumption.h>
#include <deal.II/base/parallel.h>
#include <deal.II/base/utilities.h>
 
#include <deal.II/grid/grid_refinement.h>
#include <deal.II/grid/tria.h>
#include <deal.II/grid/tria_accessor.h>
#include <deal.II/grid/tria_iterator.h>
 
#include <deal.II/lac/vector.h>
 
#include <deal.II/numerics/time_dependent.h>
 
#include <algorithm>
#include <functional>
#include <numeric>
 
DEAL_II_NAMESPACE_OPEN
 
TimeDependent::TimeSteppingData::TimeSteppingData(const unsigned int look_ahead,
                                                  const unsigned int look_back)
  : look_ahead(look_ahead)
  , look_back(look_back)
{}
 
 
TimeDependent::TimeDependent(const TimeSteppingData &data_primal,
                             const TimeSteppingData &data_dual,
                             const TimeSteppingData &data_postprocess)
  : sweep_no(numbers::invalid_unsigned_int)
  , timestepping_data_primal(data_primal)
  , timestepping_data_dual(data_dual)
  , timestepping_data_postprocess(data_postprocess)
{}
 
 
TimeDependent::~TimeDependent()
{
  try
    {
      while (timesteps.size() != 0)
        delete_timestep(0);
    }
  catch (...)
    {}
}
 
 
void
TimeDependent::insert_timestep(const TimeStepBase *position,
                               TimeStepBase       *new_timestep)
{
  Assert((std::find(timesteps.begin(), timesteps.end(), position) !=
          timesteps.end()) ||
           (position == nullptr),
         ExcInvalidPosition());
  // first insert the new time step
  // into the doubly linked list
  // of timesteps
  if (position == nullptr)
    {
      // at the end
      new_timestep->set_next_timestep(nullptr);
      if (timesteps.size() > 0)
        {
          timesteps.back()->set_next_timestep(new_timestep);
          new_timestep->set_previous_timestep(timesteps.back());
        }
      else
        new_timestep->set_previous_timestep(nullptr);
    }
  else if (position == timesteps[0])
    {
      // at the beginning
      new_timestep->set_previous_timestep(nullptr);
      if (timesteps.size() > 0)
        {
          timesteps[0]->set_previous_timestep(new_timestep);
          new_timestep->set_next_timestep(timesteps[0]);
        }
      else
        new_timestep->set_next_timestep(nullptr);
    }
  else
    {
      // inner time step
      const std::vector<ObserverPointer<TimeStepBase, TimeDependent>>::iterator
        insert_position =
          std::find(timesteps.begin(), timesteps.end(), position);
      // check iterators again to satisfy coverity: both insert_position and
      // insert_position - 1 must be valid iterators
      Assert(insert_position != timesteps.begin() &&
               insert_position != timesteps.end(),
             ExcInternalError());
 
      (*(insert_position - 1))->set_next_timestep(new_timestep);
      new_timestep->set_previous_timestep(*(insert_position - 1));
      new_timestep->set_next_timestep(*insert_position);
      (*insert_position)->set_previous_timestep(new_timestep);
    }
 
  // finally enter it into the
  // array
  timesteps.insert((position == nullptr ?
                      timesteps.end() :
                      std::find(timesteps.begin(), timesteps.end(), position)),
                   new_timestep);
}
 
 
void
TimeDependent::add_timestep(TimeStepBase *new_timestep)
{
  insert_timestep(nullptr, new_timestep);
}
 
 
void
TimeDependent::delete_timestep(const unsigned int position)
{
  Assert(position < timesteps.size(), ExcInvalidPosition());
 
  // Remember time step object for
  // later deletion and unlock
  // ObserverPointer
  TimeStepBase *t     = timesteps[position];
  timesteps[position] = nullptr;
  // Now delete unsubscribed object
  delete t;
 
  timesteps.erase(timesteps.begin() + position);
 
  // reset "next" pointer of previous
  // time step if possible
  //
  // note that if now position==size,
  // then we deleted the last time step
  if (position != 0)
    timesteps[position - 1]->set_next_timestep(
      (position < timesteps.size()) ?
        timesteps[position] :
        /*null*/ ObserverPointer<TimeStepBase, TimeDependent>());
 
  // same for "previous" pointer of next
  // time step
  if (position < timesteps.size())
    timesteps[position]->set_previous_timestep(
      (position != 0) ?
        timesteps[position - 1] :
        /*null*/ ObserverPointer<TimeStepBase, TimeDependent>());
}
 
 
void
TimeDependent::solve_primal_problem()
{
  do_loop(
    [](TimeStepBase *const time_step) { time_step->init_for_primal_problem(); },
    [](TimeStepBase *const time_step) { time_step->solve_primal_problem(); },
    timestepping_data_primal,
    forward);
}
 
 
void
TimeDependent::solve_dual_problem()
{
  do_loop(
    [](TimeStepBase *const time_step) { time_step->init_for_dual_problem(); },
    [](TimeStepBase *const time_step) { time_step->init_for_dual_problem(); },
    timestepping_data_dual,
    backward);
}
 
 
void
TimeDependent::postprocess()
{
  do_loop(
    [](TimeStepBase *const time_step) { time_step->init_for_postprocessing(); },
    [](TimeStepBase *const time_step) { time_step->postprocess_timestep(); },
    timestepping_data_postprocess,
    forward);
}
 
 
 
void
TimeDependent::start_sweep(const unsigned int s)
{
  sweep_no = s;
 
  // reset the number each
  // time step has, since some time
  // steps might have been added since
  // the last time we visited them
  //
  // also set the sweep we will
  // process in the sequel
  for (unsigned int step = 0; step < timesteps.size(); ++step)
    {
      timesteps[step]->set_timestep_no(step);
      timesteps[step]->set_sweep_no(sweep_no);
    }
 
  for (const auto &timestep : timesteps)
    timestep->start_sweep();
}
 
 
 
void
TimeDependent::end_sweep()
{
  parallel::apply_to_subranges(
    0U,
    timesteps.size(),
    [this](const unsigned int begin, const unsigned int end) {
      this->end_sweep(begin, end);
    },
    1);
}
 
 
 
void
TimeDependent::end_sweep(const unsigned int begin, const unsigned int end)
{
  for (unsigned int step = begin; step < end; ++step)
    timesteps[step]->end_sweep();
}
 
 
 
std::size_t
TimeDependent::memory_consumption() const
{
  std::size_t mem =
    (MemoryConsumption::memory_consumption(timesteps) +
     MemoryConsumption::memory_consumption(sweep_no) +
     sizeof(timestepping_data_primal) + sizeof(timestepping_data_dual) +
     sizeof(timestepping_data_postprocess));
  for (const auto &timestep : timesteps)
    mem += MemoryConsumption::memory_consumption(*timestep);
 
  return mem;
}
 
 
 
/* --------------------------------------------------------------------- */
 
 
TimeStepBase::TimeStepBase(const double time)
  : previous_timestep(nullptr)
  , next_timestep(nullptr)
  , sweep_no(numbers::invalid_unsigned_int)
  , timestep_no(numbers::invalid_unsigned_int)
  , time(time)
  , next_action(numbers::invalid_unsigned_int)
{}
 
 
 
void
TimeStepBase::wake_up(const unsigned int)
{}
 
 
 
void
TimeStepBase::sleep(const unsigned)
{}
 
 
 
void
TimeStepBase::start_sweep()
{}
 
 
 
void
TimeStepBase::end_sweep()
{}
 
 
 
void
TimeStepBase::init_for_primal_problem()
{
  next_action = primal_problem;
}
 
 
 
void
TimeStepBase::init_for_dual_problem()
{
  next_action = dual_problem;
}
 
 
 
void
TimeStepBase::init_for_postprocessing()
{
  next_action = postprocess;
}
 
 
 
void
TimeStepBase::solve_dual_problem()
{
  Assert(false, ExcPureFunctionCalled());
}
 
 
 
void
TimeStepBase::postprocess_timestep()
{
  Assert(false, ExcPureFunctionCalled());
}
 
 
 
double
TimeStepBase::get_time() const
{
  return time;
}
 
 
 
unsigned int
TimeStepBase::get_timestep_no() const
{
  return timestep_no;
}
 
 
 
double
TimeStepBase::get_backward_timestep() const
{
  Assert(previous_timestep != nullptr,
         ExcMessage("The backward time step cannot be computed because "
                    "there is no previous time step."));
  return time - previous_timestep->time;
}
 
 
 
double
TimeStepBase::get_forward_timestep() const
{
  Assert(next_timestep != nullptr,
         ExcMessage("The forward time step cannot be computed because "
                    "there is no next time step."));
  return next_timestep->time - time;
}
 
 
 
void
TimeStepBase::set_previous_timestep(const TimeStepBase *previous)
{
  previous_timestep = previous;
}
 
 
 
void
TimeStepBase::set_next_timestep(const TimeStepBase *next)
{
  next_timestep = next;
}
 
 
 
void
TimeStepBase::set_timestep_no(const unsigned int step_no)
{
  timestep_no = step_no;
}
 
 
 
void
TimeStepBase::set_sweep_no(const unsigned int sweep)
{
  sweep_no = sweep;
}
 
 
 
std::size_t
TimeStepBase::memory_consumption() const
{
  // only simple data types
  return sizeof(*this);
}
 
 
 
template <int dim>
TimeStepBase_Tria<dim>::TimeStepBase_Tria()
  : TimeStepBase(0)
  , tria(nullptr, typeid(*this).name())
  , coarse_grid(nullptr, typeid(*this).name())
  , flags()
  , refinement_flags(0)
{
  Assert(false, ExcPureFunctionCalled());
}
 
 
 
#ifndef DOXYGEN
template <int dim>
TimeStepBase_Tria<dim>::TimeStepBase_Tria(
  const double              time,
  const Triangulation<dim> &coarse_grid,
  const Flags              &flags,
  const RefinementFlags    &refinement_flags)
  : TimeStepBase(time)
  , tria(nullptr, typeid(*this).name())
  , coarse_grid(&coarse_grid, typeid(*this).name())
  , flags(flags)
  , refinement_flags(refinement_flags)
{}
#endif
 
 
 
template <int dim>
TimeStepBase_Tria<dim>::~TimeStepBase_Tria()
{
  if (!flags.delete_and_rebuild_tria)
    {
      Triangulation<dim> *t = tria;
      tria                  = nullptr;
      delete t;
    }
  else
    AssertNothrow(tria == nullptr, ExcInternalError());
 
  coarse_grid = nullptr;
}
 
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::wake_up(const unsigned int wakeup_level)
{
  TimeStepBase::wake_up(wakeup_level);
 
  if (wakeup_level == flags.wakeup_level_to_build_grid)
    if (flags.delete_and_rebuild_tria || !tria)
      restore_grid();
}
 
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::sleep(const unsigned int sleep_level)
{
  if (sleep_level == flags.sleep_level_to_delete_grid)
    {
      Assert(tria != nullptr, ExcInternalError());
 
      if (flags.delete_and_rebuild_tria)
        {
          Triangulation<dim> *t = tria;
          tria                  = nullptr;
          delete t;
        }
    }
 
  TimeStepBase::sleep(sleep_level);
}
 
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::save_refine_flags()
{
  // for any of the non-initial grids
  // store the refinement flags
  refine_flags.emplace_back();
  coarsen_flags.emplace_back();
  tria->save_refine_flags(refine_flags.back());
  tria->save_coarsen_flags(coarsen_flags.back());
}
 
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::restore_grid()
{
  Assert(tria == nullptr, ExcGridNotDeleted());
  Assert(refine_flags.size() == coarsen_flags.size(), ExcInternalError());
 
  // create an empty triangulation and
  // set it to a copy of the coarse grid
  tria = new Triangulation<dim>();
  tria->copy_triangulation(*coarse_grid);
 
  // for each of the previous refinement
  // sweeps
  for (unsigned int previous_sweep = 0; previous_sweep < refine_flags.size();
       ++previous_sweep)
    {
      // get flags
      tria->load_refine_flags(refine_flags[previous_sweep]);
      tria->load_coarsen_flags(coarsen_flags[previous_sweep]);
 
      // limit refinement depth if the user
      // desired so
      //    if (flags.max_refinement_level != 0)
      //      {
      //        typename Triangulation<dim>::active_cell_iterator cell, endc;
      //        for (const auto &cell : tria->active_cell_iterators())
      //          if (static_cast<unsigned int>(cell->level()) >=
      //              flags.max_refinement_level)
      //            cell->clear_refine_flag();
      //      }
 
      tria->execute_coarsening_and_refinement();
    }
}
 
 
 
// have a few helper functions
namespace
{
  template <int dim>
  void
  mirror_refinement_flags(
    const typename Triangulation<dim>::cell_iterator &new_cell,
    const typename Triangulation<dim>::cell_iterator &old_cell)
  {
    // mirror the refinement
    // flags from the present time level to
    // the previous if the dual problem was
    // used for the refinement, since the
    // error is computed on a time-space cell
    //
    // we don't mirror the coarsening flags
    // since we want stronger refinement. if
    // this was the wrong decision, the error
    // on the child cells of the previous
    // time slab will indicate coarsening
    // in the next iteration, so this is not
    // so dangerous here.
    //
    // also, we only have to check whether
    // the present cell flagged for
    // refinement and the previous one is on
    // the same level and also active. If it
    // already has children, then there is
    // no problem at all, if it is on a lower
    // level than the present one, then it
    // will be refined below anyway.
    if (new_cell->is_active())
      {
        if (new_cell->refine_flag_set() && old_cell->is_active())
          {
            if (old_cell->coarsen_flag_set())
              old_cell->clear_coarsen_flag();
 
            old_cell->set_refine_flag();
          }
 
        return;
      }
 
    if (old_cell->has_children() && new_cell->has_children())
      {
        Assert(old_cell->n_children() == new_cell->n_children(),
               ExcNotImplemented());
        for (unsigned int c = 0; c < new_cell->n_children(); ++c)
          ::mirror_refinement_flags<dim>(new_cell->child(c),
                                               old_cell->child(c));
      }
  }
 
 
 
  template <int dim>
  bool
  adapt_grid_cells(const typename Triangulation<dim>::cell_iterator &cell1,
                   const typename Triangulation<dim>::cell_iterator &cell2)
  {
    if (cell2->has_children() && cell1->has_children())
      {
        bool grids_changed = false;
 
        Assert(cell2->n_children() == cell1->n_children(), ExcNotImplemented());
        for (unsigned int c = 0; c < cell1->n_children(); ++c)
          grids_changed |=
            ::adapt_grid_cells<dim>(cell1->child(c), cell2->child(c));
        return grids_changed;
      }
 
 
    if (!cell1->has_children() && !cell2->has_children())
      // none of the two have children, so
      // make sure that not one is flagged
      // for refinement and the other for
      // coarsening
      {
        if (cell1->refine_flag_set() && cell2->coarsen_flag_set())
          {
            cell2->clear_coarsen_flag();
            return true;
          }
        else if (cell1->coarsen_flag_set() && cell2->refine_flag_set())
          {
            cell1->clear_coarsen_flag();
            return true;
          }
 
        return false;
      }
 
 
    if (cell1->has_children() && !cell2->has_children())
      // cell1 has children, cell2 has not
      // -> cell2 needs to be refined if any
      // of cell1's children is flagged
      // for refinement. None of them should
      // be refined further, since then in the
      // last round something must have gone
      // wrong
      //
      // if cell2 was flagged for coarsening,
      // we need to clear that flag in any
      // case. The only exception would be
      // if all children of cell1 were
      // flagged for coarsening, but rules
      // for coarsening are so complicated
      // that we will not attempt to cover
      // them. Rather accept one cell which
      // is not coarsened...
      {
        bool changed_grid = false;
        if (cell2->coarsen_flag_set())
          {
            cell2->clear_coarsen_flag();
            changed_grid = true;
          }
 
        if (!cell2->refine_flag_set())
          for (unsigned int c = 0; c < cell1->n_children(); ++c)
            if (cell1->child(c)->refine_flag_set() ||
                cell1->child(c)->has_children())
              {
                cell2->set_refine_flag();
                changed_grid = true;
                break;
              }
        return changed_grid;
      }
 
    if (!cell1->has_children() && cell2->has_children())
      // same thing, other way round...
      {
        bool changed_grid = false;
        if (cell1->coarsen_flag_set())
          {
            cell1->clear_coarsen_flag();
            changed_grid = true;
          }
 
        if (!cell1->refine_flag_set())
          for (unsigned int c = 0; c < cell2->n_children(); ++c)
            if (cell2->child(c)->refine_flag_set() ||
                cell2->child(c)->has_children())
              {
                cell1->set_refine_flag();
                changed_grid = true;
                break;
              }
        return changed_grid;
      }
 
    DEAL_II_ASSERT_UNREACHABLE();
    return false;
  }
 
 
 
  template <int dim>
  bool
  adapt_grids(Triangulation<dim> &tria1, Triangulation<dim> &tria2)
  {
    bool grids_changed = false;
 
    typename Triangulation<dim>::cell_iterator cell1 = tria1.begin(),
                                               cell2 = tria2.begin();
    typename Triangulation<dim>::cell_iterator endc;
    endc = (tria1.n_levels() == 1 ?
              typename Triangulation<dim>::cell_iterator(tria1.end()) :
              tria1.begin(1));
    for (; cell1 != endc; ++cell1, ++cell2)
      grids_changed |= ::adapt_grid_cells<dim>(cell1, cell2);
 
    return grids_changed;
  }
} // namespace
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::refine_grid(const RefinementData refinement_data)
{
  Vector<float> criteria;
  get_tria_refinement_criteria(criteria);
 
  // copy the following two values since
  // we may need modified values in the
  // process of this function
  double refinement_threshold = refinement_data.refinement_threshold,
         coarsening_threshold = refinement_data.coarsening_threshold;
 
  // prepare an array where the criteria
  // are stored in a sorted fashion
  // we need this if cell number correction
  // is switched on.
  // the criteria are sorted in ascending
  // order
  // only fill it when needed
  Vector<float> sorted_criteria;
  // two pointers into this array denoting
  // the position where the two thresholds
  // are assumed
  Vector<float>::const_iterator p_refinement_threshold = nullptr,
                                p_coarsening_threshold = nullptr;
 
 
  // if we are to do some cell number
  // correction steps, we have to find out
  // which further cells (beyond
  // refinement_threshold) to refine in case
  // we need more cells, and which cells
  // to not refine in case we need less cells
  // (or even to coarsen, if necessary). to
  // this end, we first define pointers into
  // a sorted array of criteria pointing
  // to the thresholds of refinement or
  // coarsening; moving these pointers amounts
  // to changing the threshold such that the
  // number of cells flagged for refinement
  // or coarsening would be changed by one
  if ((timestep_no != 0) &&
      (sweep_no >= refinement_flags.first_sweep_with_correction) &&
      (refinement_flags.cell_number_correction_steps > 0))
    {
      sorted_criteria = criteria;
      std::sort(sorted_criteria.begin(), sorted_criteria.end());
      p_refinement_threshold =
        Utilities::lower_bound(sorted_criteria.begin(),
                               sorted_criteria.end(),
                               static_cast<float>(refinement_threshold));
      p_coarsening_threshold =
        std::upper_bound(sorted_criteria.begin(),
                         sorted_criteria.end(),
                         static_cast<float>(coarsening_threshold));
    }
 
 
  // actually flag cells the first time
  GridRefinement::refine(*tria, criteria, refinement_threshold);
  GridRefinement::coarsen(*tria, criteria, coarsening_threshold);
 
  // store this number for the following
  // since its computation is rather
  // expensive and since it doesn't change
  const unsigned int n_active_cells = tria->n_active_cells();
 
  // if not on first time level: try to
  // adjust the number of resulting
  // cells to those on the previous
  // time level. Only do the cell number
  // correction for higher sweeps and if
  // there are sufficiently many cells
  // already to avoid "grid stall" i.e.
  // that the grid's evolution is hindered
  // by the correction (this usually
  // happens if there are very few cells,
  // since then the number of cells touched
  // by the correction step may exceed the
  // number of cells which are flagged for
  // refinement; in this case it often
  // happens that the number of cells
  // does not grow between sweeps, which
  // clearly is not the wanted behavior)
  //
  // however, if we do not do anything, we
  // can get into trouble somewhen later.
  // therefore, we also use the correction
  // step for the first sweep or if the
  // number of cells is between 100 and 300
  // (unlike in the first version of the
  // algorithm), but relax the conditions
  // for the correction to allow deviations
  // which are three times as high than
  // allowed (sweep==1 || cell number<200)
  // or twice as high (sweep==2 ||
  // cell number<300). Also, since
  // refinement never does any harm other
  // than increased work, we allow for
  // arbitrary growth of cell number if
  // the estimated cell number is below
  // 200.
  //
  // repeat this loop several times since
  // the first estimate may not be totally
  // correct
  if ((timestep_no != 0) &&
      (sweep_no >= refinement_flags.first_sweep_with_correction))
    for (unsigned int loop = 0;
         loop < refinement_flags.cell_number_correction_steps;
         ++loop)
      {
        Triangulation<dim> *previous_tria =
          dynamic_cast<const TimeStepBase_Tria<dim> *>(previous_timestep)->tria;
 
        // do one adaption step if desired
        // (there are more coming below then
        // also)
        if (refinement_flags.adapt_grids)
          ::adapt_grids<dim>(*previous_tria, *tria);
 
        // perform flagging of cells
        // needed to regularize the
        // triangulation
        tria->prepare_coarsening_and_refinement();
        previous_tria->prepare_coarsening_and_refinement();
 
 
        // now count the number of elements
        // which will result on the previous
        // grid after it will be refined. The
        // number which will really result
        // should be approximately that that we
        // compute here, since we already
        // performed most of the prepare*
        // steps for the previous grid
        //
        // use a double value since for each
        // four cells (in 2d) that we flagged
        // for coarsening we result in one
        // new. but since we loop over flagged
        // cells, we have to subtract 3/4 of
        // a cell for each flagged cell
        Assert(!tria->get_anisotropic_refinement_flag(), ExcNotImplemented());
        Assert(!previous_tria->get_anisotropic_refinement_flag(),
               ExcNotImplemented());
        double previous_cells = previous_tria->n_active_cells();
        typename Triangulation<dim>::active_cell_iterator cell, endc;
        cell = previous_tria->begin_active();
        endc = previous_tria->end();
        for (; cell != endc; ++cell)
          if (cell->refine_flag_set())
            previous_cells += (GeometryInfo<dim>::max_children_per_cell - 1);
          else if (cell->coarsen_flag_set())
            previous_cells -= static_cast<double>(
                                GeometryInfo<dim>::max_children_per_cell - 1) /
                              GeometryInfo<dim>::max_children_per_cell;
 
        // @p{previous_cells} now gives the
        // number of cells which would result
        // from the flags on the previous grid
        // if we refined it now. However, some
        // more flags will be set when we adapt
        // the previous grid with this one
        // after the flags have been set for
        // this time level; on the other hand,
        // we don't account for this, since the
        // number of cells on this time level
        // will be changed afterwards by the
        // same way, when it is adapted to the
        // next time level
 
        // now estimate the number of cells which
        // will result on this level
        double estimated_cells = n_active_cells;
        cell                   = tria->begin_active();
        endc                   = tria->end();
        for (; cell != endc; ++cell)
          if (cell->refine_flag_set())
            estimated_cells += (GeometryInfo<dim>::max_children_per_cell - 1);
          else if (cell->coarsen_flag_set())
            estimated_cells -= static_cast<double>(
                                 GeometryInfo<dim>::max_children_per_cell - 1) /
                               GeometryInfo<dim>::max_children_per_cell;
 
        // calculate the allowed delta in
        // cell numbers; be more lenient
        // if there are few cells
        double delta_up   = refinement_flags.cell_number_corridor_top,
               delta_down = refinement_flags.cell_number_corridor_bottom;
 
        const std::vector<std::pair<unsigned int, double>> &relaxations =
          (sweep_no >= refinement_flags.correction_relaxations.size() ?
             refinement_flags.correction_relaxations.back() :
             refinement_flags.correction_relaxations[sweep_no]);
        for (const auto &relaxation : relaxations)
          if (n_active_cells < relaxation.first)
            {
              delta_up *= relaxation.second;
              delta_down *= relaxation.second;
              break;
            }
 
        // now, if the number of estimated
        // cells exceeds the number of cells
        // on the old time level by more than
        // delta: cut the top threshold
        //
        // note that for each cell that
        // we unflag we have to diminish the
        // estimated number of cells by
        // @p{children_per_cell}.
        if (estimated_cells > previous_cells * (1. + delta_up))
          {
            // only limit the cell number
            // if there will not be less
            // than some number of cells
            //
            // also note that when using the
            // dual estimator, the initial
            // time level is not refined
            // on its own, so we may not
            // limit the number of the second
            // time level on the basis of
            // the initial one; since for
            // the dual estimator, we
            // mirror the refinement
            // flags, the initial level
            // will be passively refined
            // later on.
            if (estimated_cells > refinement_flags.min_cells_for_correction)
              {
                // number of cells by which the
                // new grid is to be diminished
                double delta_cells =
                  estimated_cells - previous_cells * (1. + delta_up);
 
                // if we need to reduce the
                // number of cells, we need
                // to raise the thresholds,
                // i.e. move ahead in the
                // sorted array, since this
                // is sorted in ascending
                // order. do so by removing
                // cells tagged for refinement
 
                for (unsigned int i = 0; i < delta_cells;
                     i += GeometryInfo<dim>::max_children_per_cell - 1)
                  if (p_refinement_threshold != sorted_criteria.end())
                    ++p_refinement_threshold;
                  else
                    break;
              }
            else
              // too many cells, but we
              // won't do anything about
              // that
              break;
          }
        else
          // likewise: if the estimated number
          // of cells is less than 90 per cent
          // of those at the previous time level:
          // raise threshold by refining
          // additional cells. if we start to
          // run into the area of cells
          // which are to be coarsened, we
          // raise the limit for these too
          if (estimated_cells < previous_cells * (1. - delta_down))
            {
              // number of cells by which the
              // new grid is to be enlarged
              double delta_cells =
                previous_cells * (1. - delta_down) - estimated_cells;
              // heuristics: usually, if we
              // add @p{delta_cells} to the
              // present state, we end up
              // with much more than only
              // (1-delta_down)*prev_cells
              // because of the effect of
              // regularization and because
              // of adaption to the
              // following grid. Therefore,
              // if we are not in the last
              // correction loop, we try not
              // to add as many cells as seem
              // necessary at first and hope
              // to get closer to the limit
              // this way. Only in the last
              // loop do we have to take the
              // full number to guarantee the
              // wanted result.
              //
              // The value 0.9 is taken from
              // practice, as the additional
              // number of cells introduced
              // by regularization is
              // approximately 10 per cent
              // of the flagged cells.
              if (loop != refinement_flags.cell_number_correction_steps - 1)
                delta_cells *= 0.9;
 
              // if more cells need to be
              // refined, we need to lower
              // the thresholds, i.e. to
              // move to the beginning
              // of sorted_criteria, which is
              // sorted in ascending order
              for (unsigned int i = 0; i < delta_cells;
                   i += (GeometryInfo<dim>::max_children_per_cell - 1))
                if (p_refinement_threshold != p_coarsening_threshold)
                  --refinement_threshold;
                else if (p_coarsening_threshold != sorted_criteria.begin())
                  --p_coarsening_threshold, --p_refinement_threshold;
                else
                  break;
            }
          else
            // estimated cell number is ok,
            // stop correction steps
            break;
 
        if (p_refinement_threshold == sorted_criteria.end())
          {
            Assert(p_coarsening_threshold != p_refinement_threshold,
                   ExcInternalError());
            --p_refinement_threshold;
          }
 
        coarsening_threshold = *p_coarsening_threshold;
        refinement_threshold = *p_refinement_threshold;
 
        if (coarsening_threshold >= refinement_threshold)
          coarsening_threshold = 0.999 * refinement_threshold;
 
        // now that we have re-adjusted
        // thresholds: clear all refine and
        // coarsening flags and do it all
        // over again
        cell = tria->begin_active();
        endc = tria->end();
        for (; cell != endc; ++cell)
          {
            cell->clear_refine_flag();
            cell->clear_coarsen_flag();
          }
 
 
        // flag cells finally
        GridRefinement::refine(*tria, criteria, refinement_threshold);
        GridRefinement::coarsen(*tria, criteria, coarsening_threshold);
      }
 
  // if step number is greater than
  // one: adapt this and the previous
  // grid to each other. Don't do so
  // for the initial grid because
  // it is always taken to be the first
  // grid and needs therefore no
  // treatment of its own.
  if ((timestep_no >= 1) && (refinement_flags.adapt_grids))
    {
      Triangulation<dim> *previous_tria =
        dynamic_cast<const TimeStepBase_Tria<dim> *>(previous_timestep)->tria;
      Assert(previous_tria != nullptr, ExcInternalError());
 
      // if we used the dual estimator, we
      // computed the error information on
      // a time slab, rather than on a level
      // of its own. we then mirror the
      // refinement flags we determined for
      // the present level to the previous
      // one
      //
      // do this mirroring only, if cell number
      // adjustment is on, since otherwise
      // strange things may happen
      if (refinement_flags.mirror_flags_to_previous_grid)
        {
          ::adapt_grids<dim>(*previous_tria, *tria);
 
          typename Triangulation<dim>::cell_iterator old_cell, new_cell, endc;
          old_cell = previous_tria->begin(0);
          new_cell = tria->begin(0);
          endc     = tria->end(0);
          for (; new_cell != endc; ++new_cell, ++old_cell)
            ::mirror_refinement_flags<dim>(new_cell, old_cell);
        }
 
      tria->prepare_coarsening_and_refinement();
      previous_tria->prepare_coarsening_and_refinement();
 
      // adapt present and previous grids
      // to each other: flag additional
      // cells to avoid the previous grid
      // to have cells refined twice more
      // than the present one and vica versa.
      ::adapt_grids<dim>(*previous_tria, *tria);
    }
}
 
 
 
template <int dim>
void
TimeStepBase_Tria<dim>::init_for_refinement()
{
  next_action = grid_refinement;
}
 
 
 
template <int dim>
std::size_t
TimeStepBase_Tria<dim>::memory_consumption() const
{
  return (TimeStepBase::memory_consumption() + sizeof(tria) +
          MemoryConsumption::memory_consumption(coarse_grid) + sizeof(flags) +
          sizeof(refinement_flags) +
          MemoryConsumption::memory_consumption(refine_flags) +
          MemoryConsumption::memory_consumption(coarsen_flags));
}
 
 
 
template <int dim>
TimeStepBase_Tria_Flags::Flags<dim>::Flags()
  : delete_and_rebuild_tria(false)
  , wakeup_level_to_build_grid(0)
  , sleep_level_to_delete_grid(0)
{
  DEAL_II_ASSERT_UNREACHABLE();
}
 
 
 
template <int dim>
TimeStepBase_Tria_Flags::Flags<dim>::Flags(
  const bool         delete_and_rebuild_tria,
  const unsigned int wakeup_level_to_build_grid,
  const unsigned int sleep_level_to_delete_grid)
  : delete_and_rebuild_tria(delete_and_rebuild_tria)
  , wakeup_level_to_build_grid(wakeup_level_to_build_grid)
  , sleep_level_to_delete_grid(sleep_level_to_delete_grid)
{
  //   Assert (!delete_and_rebuild_tria || (wakeup_level_to_build_grid>=1),
  //        ExcInvalidParameter(wakeup_level_to_build_grid));
  //   Assert (!delete_and_rebuild_tria || (sleep_level_to_delete_grid>=1),
  //        ExcInvalidParameter(sleep_level_to_delete_grid));
}
 
 
template <int dim>
typename TimeStepBase_Tria_Flags::RefinementFlags<dim>::CorrectionRelaxations
  TimeStepBase_Tria_Flags::RefinementFlags<dim>::default_correction_relaxations(
    1, // one element, denoting the first and all subsequent sweeps
    std::vector<std::pair<unsigned int, double>>(1, // one element, denoting the
                                                    // upper bound for the
                                                    // following relaxation
                                                 std::make_pair(0U, 0.)));
 
 
template <int dim>
TimeStepBase_Tria_Flags::RefinementFlags<dim>::RefinementFlags(
  const unsigned int           max_refinement_level,
  const unsigned int           first_sweep_with_correction,
  const unsigned int           min_cells_for_correction,
  const double                 cell_number_corridor_top,
  const double                 cell_number_corridor_bottom,
  const CorrectionRelaxations &correction_relaxations,
  const unsigned int           cell_number_correction_steps,
  const bool                   mirror_flags_to_previous_grid,
  const bool                   adapt_grids)
  : max_refinement_level(max_refinement_level)
  , first_sweep_with_correction(first_sweep_with_correction)
  , min_cells_for_correction(min_cells_for_correction)
  , cell_number_corridor_top(cell_number_corridor_top)
  , cell_number_corridor_bottom(cell_number_corridor_bottom)
  , correction_relaxations(correction_relaxations.size() != 0 ?
                             correction_relaxations :
                             default_correction_relaxations)
  , cell_number_correction_steps(cell_number_correction_steps)
  , mirror_flags_to_previous_grid(mirror_flags_to_previous_grid)
  , adapt_grids(adapt_grids)
{
  Assert(cell_number_corridor_top >= 0,
         ExcInvalidValue(cell_number_corridor_top));
  Assert(cell_number_corridor_bottom >= 0,
         ExcInvalidValue(cell_number_corridor_bottom));
  Assert(cell_number_corridor_bottom <= 1,
         ExcInvalidValue(cell_number_corridor_bottom));
}
 
 
template <int dim>
TimeStepBase_Tria_Flags::RefinementData<dim>::RefinementData(
  const double _refinement_threshold,
  const double _coarsening_threshold)
  : refinement_threshold(_refinement_threshold)
  ,
  // in some rare cases it may happen that
  // both thresholds are the same (e.g. if
  // there are many cells with the same
  // error indicator). That would mean that
  // all cells will be flagged for
  // refinement or coarsening, but some will
  // be flagged for both, namely those for
  // which the indicator equals the
  // thresholds. This is forbidden, however.
  //
  // In some rare cases with very few cells
  // we also could get integer round off
  // errors and get problems with
  // the top and bottom fractions.
  //
  // In these case we arbitrarily reduce the
  // bottom threshold by one permille below
  // the top threshold
  coarsening_threshold((_coarsening_threshold == _refinement_threshold ?
                          _coarsening_threshold :
                          0.999 * _coarsening_threshold))
{
  Assert(refinement_threshold >= 0, ExcInvalidValue(refinement_threshold));
  Assert(coarsening_threshold >= 0, ExcInvalidValue(coarsening_threshold));
  // allow both thresholds to be zero,
  // since this is needed in case all indicators
  // are zero
  Assert((coarsening_threshold < refinement_threshold) ||
           ((coarsening_threshold == 0) && (refinement_threshold == 0)),
         ExcInvalidValue(coarsening_threshold));
}
 
 
 
/*-------------- Explicit Instantiations -------------------------------*/
#include "numerics/time_dependent.inst"
 
 
DEAL_II_NAMESPACE_CLOSE