FEProblemSolve.C

Go to the documentation of this file.

//* This file is part of the MOOSE framework
//* https://mooseframework.inl.gov
//*
//* All rights reserved, see COPYRIGHT for full restrictions
//* https://github.com/idaholab/moose/blob/master/COPYRIGHT
//*
//* Licensed under LGPL 2.1, please see LICENSE for details
//* https://www.gnu.org/licenses/lgpl-2.1.html

#include "FEProblemSolve.h"

#include "FEProblem.h"
#include "NonlinearSystemBase.h"
#include "LinearSystem.h"
#include "Convergence.h"
#include "Executioner.h"
#include "ConvergenceIterationTypes.h"
#include "MooseUtils.h"

std::set<std::string> const FEProblemSolve::_moose_line_searches = {"contact", "project"};

const std::set<std::string> &
FEProblemSolve::mooseLineSearches()
{
  return _moose_line_searches;
}

InputParameters
FEProblemSolve::feProblemDefaultConvergenceParams()
{
  InputParameters params = emptyInputParameters();

  params.addParam<unsigned int>("nl_max_its", 50, "Max Nonlinear Iterations");
  params.addParam<unsigned int>("nl_forced_its", 0, "The Number of Forced Nonlinear Iterations");
  params.addParam<unsigned int>("nl_max_funcs", 10000, "Max Nonlinear solver function evaluations");
  params.addParam<Real>("nl_abs_tol", 1.0e-50, "Nonlinear Absolute Tolerance");
  params.addParam<Real>("nl_rel_tol", 1.0e-8, "Nonlinear Relative Tolerance");
  params.addParam<Real>(
      "nl_div_tol",
      1.0e10,
      "Nonlinear Relative Divergence Tolerance. A negative value disables this check.");
  params.addParam<Real>(
      "nl_abs_div_tol",
      1.0e50,
      "Nonlinear Absolute Divergence Tolerance. A negative value disables this check.");
  params.addParam<Real>("nl_abs_step_tol", 0., "Nonlinear Absolute step Tolerance");
  params.addParam<Real>("nl_rel_step_tol", 0., "Nonlinear Relative step Tolerance");
  params.addParam<unsigned int>("n_max_nonlinear_pingpong",
                                100,
                                "The maximum number of times the nonlinear residual can ping pong "
                                "before requesting halting the current evaluation and requesting "
                                "timestep cut for transient simulations");

  params.addParamNamesToGroup(
      "nl_max_its nl_forced_its nl_max_funcs nl_abs_tol nl_rel_tol "
      "nl_rel_step_tol nl_abs_step_tol nl_div_tol nl_abs_div_tol n_max_nonlinear_pingpong",
      "Nonlinear Solver");

  return params;
}

InputParameters
FEProblemSolve::validParams()
{
  InputParameters params = MultiSystemSolveObject::validParams();
  params += FEProblemSolve::feProblemDefaultConvergenceParams();

  std::set<std::string> line_searches = mooseLineSearches();

  std::set<std::string> alias_line_searches = {"default", "none", "basic"};
  line_searches.insert(alias_line_searches.begin(), alias_line_searches.end());
  std::set<std::string> petsc_line_searches = Moose::PetscSupport::getPetscValidLineSearches();
  line_searches.insert(petsc_line_searches.begin(), petsc_line_searches.end());
  std::string line_search_string = Moose::stringify(line_searches, " ");
  MooseEnum line_search(line_search_string, "default");
  std::string addtl_doc_str(" (Note: none = basic)");
  params.addParam<MooseEnum>(
      "line_search", line_search, "Specifies the line search type" + addtl_doc_str);
  MooseEnum line_search_package("petsc moose", "petsc");
  params.addParam<MooseEnum>("line_search_package",
                             line_search_package,
                             "The solver package to use to conduct the line-search");

  params.addParam<unsigned>("contact_line_search_allowed_lambda_cuts",
                            2,
                            "The number of times lambda is allowed to be cut in half in the "
                            "contact line search. We recommend this number be roughly bounded by 0 "
                            "<= allowed_lambda_cuts <= 3");
  params.addParam<Real>("contact_line_search_ltol",
                        "The linear relative tolerance to be used while the contact state is "
                        "changing between non-linear iterations. We recommend that this tolerance "
                        "be looser than the standard linear tolerance");

  params += Moose::PetscSupport::getPetscValidParams();
  params.addParam<Real>("l_tol", 1.0e-5, "Linear Relative Tolerance");
  params.addParam<Real>("l_abs_tol", 1.0e-50, "Linear Absolute Tolerance");
  params.addParam<unsigned int>("l_max_its", 10000, "Max Linear Iterations");
  params.addParam<std::vector<ConvergenceName>>(
      "nonlinear_convergence",
      "Name of the Convergence object(s) to use to assess convergence of the "
      "nonlinear system(s) solve. If not provided, the default Convergence "
      "associated with the Problem will be constructed internally.");
  params.addParam<std::vector<ConvergenceName>>(
      "linear_convergence",
      "Name of the Convergence object(s) to use to assess convergence of the "
      "linear system(s) solve. If not provided, the linear solver tolerance parameters are used");
  params.addParam<bool>(
      "snesmf_reuse_base",
      true,
      "Specifies whether or not to reuse the base vector for matrix-free calculation");
  params.addParam<bool>(
      "skip_exception_check", false, "Specifies whether or not to skip exception check");
  params.addParam<bool>(
      "use_pre_SMO_residual",
      false,
      "Compute the pre-SMO residual norm and use it in the relative convergence check. The "
      "pre-SMO residual is computed at the begining of the time step before solution-modifying "
      "objects are executed. Solution-modifying objects include preset BCs, constraints, "
      "predictors, etc.");
  params.addParam<bool>("automatic_scaling", "Whether to use automatic scaling for the variables.");
  params.addParam<std::vector<bool>>(
      "compute_scaling_once",
      {true},
      "Whether the scaling factors should only be computed once at the beginning of the simulation "
      "through an extra Jacobian evaluation. If this is set to false, then the scaling factors "
      "will be computed during an extra Jacobian evaluation at the beginning of every time step. "
      "Vector entries correspond to each nonlinear system.");
  params.addParam<std::vector<bool>>(
      "off_diagonals_in_auto_scaling",
      {false},
      "Whether to consider off-diagonals when determining automatic scaling factors. Vector "
      "entries correspond to each nonlinear system.");
  params.addRangeCheckedParam<std::vector<Real>>(
      "resid_vs_jac_scaling_param",
      {0},
      "0<=resid_vs_jac_scaling_param<=1",
      "A parameter that indicates the weighting of the residual vs the Jacobian in determining "
      "variable scaling parameters. A value of 1 indicates pure residual-based scaling. A value of "
      "0 indicates pure Jacobian-based scaling. Vector entries correspond to each nonlinear "
      "system.");
  params.addParam<std::vector<std::vector<std::vector<std::string>>>>(
      "scaling_group_variables",
      "Name of variables that are grouped together for determining scale factors. (Multiple "
      "groups can be provided, separated by semicolon). Vector entries correspond to each "
      "nonlinear system.");
  params.addParam<std::vector<std::vector<std::string>>>(
      "ignore_variables_for_autoscaling",
      "List of variables that do not participate in autoscaling. Vector entries correspond to each "
      "nonlinear system.");
  params.addRangeCheckedParam<unsigned int>(
      "num_grids",
      1,
      "num_grids>0",
      "The number of grids to use for a grid sequencing algorithm. This includes the final grid, "
      "so num_grids = 1 indicates just one solve in a time-step");
  params.addParam<std::vector<bool>>("residual_and_jacobian_together",
                                     {false},
                                     "Whether to compute the residual and Jacobian together. "
                                     "Vector entries correspond to each nonlinear system.");

  params.addParam<bool>("reuse_preconditioner",
                        false,
                        "If true reuse the previously calculated "
                        "preconditioner for the linearized "
                        "system across multiple solves "
                        "spanning nonlinear iterations and time steps. "
                        "The preconditioner resets as controlled by "
                        "reuse_preconditioner_max_linear_its");
  params.addParam<unsigned int>("reuse_preconditioner_max_linear_its",
                                25,
                                "Reuse the previously calculated "
                                "preconditioner for the linear system "
                                "until the number of linear iterations "
                                "exceeds this number");

  params.addParamNamesToGroup("l_tol l_abs_tol l_max_its reuse_preconditioner "
                              "reuse_preconditioner_max_linear_its",
                              "Linear Solver");
  params.addParamNamesToGroup(
      "solve_type snesmf_reuse_base use_pre_SMO_residual "
      "num_grids residual_and_jacobian_together nonlinear_convergence linear_convergence",
      "Nonlinear Solver");
  params.addParamNamesToGroup(
      "automatic_scaling compute_scaling_once off_diagonals_in_auto_scaling "
      "scaling_group_variables resid_vs_jac_scaling_param ignore_variables_for_autoscaling",
      "Solver variable scaling");
  params.addParamNamesToGroup("line_search line_search_package contact_line_search_ltol "
                              "contact_line_search_allowed_lambda_cuts",
                              "Solver line search");
  params.addParamNamesToGroup("skip_exception_check", "Advanced");

  return params;
}

FEProblemSolve::FEProblemSolve(Executioner & ex)
  : MultiSystemSolveObject(ex),
    _num_grid_steps(cast_int<unsigned int>(getParam<unsigned int>("num_grids") - 1))
{
  if (_moose_line_searches.find(getParam<MooseEnum>("line_search").operator std::string()) !=
      _moose_line_searches.end())
    _problem.addLineSearch(_pars);

  auto set_solver_params = [this, &ex](const SolverSystem & sys)
  {
    const auto prefix = sys.prefix();
    if (dynamic_cast<const LinearSystem *>(&sys))
      Moose::PetscSupport::dontAddCommonSNESOptions(_problem, prefix);
    Moose::PetscSupport::storePetscOptions(_problem, prefix, ex);
    Moose::PetscSupport::setConvergedReasonFlags(_problem, prefix);

    // Set solver parameter prefix and system number
    auto & solver_params = _problem.solverParams(sys.number());
    solver_params._prefix = prefix;
    solver_params._solver_sys_num = sys.number();
  };

  // Extract and store PETSc related settings on FEProblemBase
  for (const auto * const sys : _systems)
    set_solver_params(*sys);

  // Set linear solve parameters in the equation system
  // Nonlinear solve parameters are added in the DefaultNonlinearConvergence
  EquationSystems & es = _problem.es();
  es.parameters.set<Real>("linear solver tolerance") = getParam<Real>("l_tol");
  es.parameters.set<Real>("linear solver absolute tolerance") = getParam<Real>("l_abs_tol");
  es.parameters.set<unsigned int>("linear solver maximum iterations") =
      getParam<unsigned int>("l_max_its");
  es.parameters.set<bool>("reuse preconditioner") = getParam<bool>("reuse_preconditioner");
  es.parameters.set<unsigned int>("reuse preconditioner maximum linear iterations") =
      getParam<unsigned int>("reuse_preconditioner_max_linear_its");

  // Transfer to the Problem misc nonlinear solve optimization parameters
  _problem.setSNESMFReuseBase(getParam<bool>("snesmf_reuse_base"),
                              _pars.isParamSetByUser("snesmf_reuse_base"));
  _problem.skipExceptionCheck(getParam<bool>("skip_exception_check"));

  if (isParamValid("nonlinear_convergence"))
  {
    if (_problem.onlyAllowDefaultNonlinearConvergence())
      mooseError("The selected problem does not allow 'nonlinear_convergence' to be set.");
    _problem.setNonlinearConvergenceNames(
        getParam<std::vector<ConvergenceName>>("nonlinear_convergence"));
  }
  else
    _problem.setNeedToAddDefaultNonlinearConvergence();
  if (isParamValid("linear_convergence"))
  {
    if (_problem.numLinearSystems() == 0)
      paramError(
          "linear_convergence",
          "Setting 'linear_convergence' is currently only possible for solving linear systems");
    _problem.setLinearConvergenceNames(
        getParam<std::vector<ConvergenceName>>("linear_convergence"));
  }

  // Check whether the user has explicitly requested automatic scaling and is using a solve type
  // without a matrix. If so, then we warn them
  if ((_pars.isParamSetByUser("automatic_scaling") && getParam<bool>("automatic_scaling")) &&
      std::all_of(_systems.begin(),
                  _systems.end(),
                  [this](const auto & solver_sys)
                  { return _problem.solverParams(solver_sys->number())._type == Moose::ST_JFNK; }))
  {
    paramWarning("automatic_scaling",
                 "Automatic scaling isn't implemented for the case where you do not have a "
                 "preconditioning matrix. No scaling will be applied");
    _problem.automaticScaling(false);
  }
  else
    // Check to see whether automatic_scaling has been specified anywhere, including at the
    // application level. No matter what: if we don't have a matrix, we don't do scaling
    _problem.automaticScaling(
        isParamValid("automatic_scaling")
            ? getParam<bool>("automatic_scaling")
            : (getMooseApp().defaultAutomaticScaling() &&
               std::any_of(_systems.begin(),
                           _systems.end(),
                           [this](const auto & solver_sys)
                           {
                             return _problem.solverParams(solver_sys->number())._type !=
                                    Moose::ST_JFNK;
                           })));

  if (!_using_multi_sys_fp_iterations && isParamValid("multi_system_fixed_point_convergence"))
    paramError("multi_system_fixed_point_convergence",
               "Cannot set a convergence object for multi-system fixed point iterations if "
               "'multi_system_fixed_point' is set to false");
  if (_using_multi_sys_fp_iterations && !isParamValid("multi_system_fixed_point_convergence"))
    paramError("multi_system_fixed_point_convergence",
               "Must set a convergence object for multi-system fixed point iterations if using "
               "multi-system fixed point iterations");

  // Set the same parameters to every nonlinear system by default
  int i_nl_sys = -1;
  for (const auto i_sys : index_range(_systems))
  {
    auto nl_ptr = dynamic_cast<NonlinearSystemBase *>(_systems[i_sys]);
    // Linear systems have very different parameters at the moment
    if (!nl_ptr)
      continue;
    auto & nl = *nl_ptr;
    i_nl_sys++;

    nl.setPreSMOResidual(getParam<bool>("use_pre_SMO_residual"));

    const auto res_and_jac =
        getParamFromNonlinearSystemVectorParam<bool>("residual_and_jacobian_together", i_nl_sys);
    if (res_and_jac)
      nl.residualAndJacobianTogether();

    // Automatic scaling parameters
    nl.computeScalingOnce(
        getParamFromNonlinearSystemVectorParam<bool>("compute_scaling_once", i_nl_sys));
    nl.autoScalingParam(
        getParamFromNonlinearSystemVectorParam<Real>("resid_vs_jac_scaling_param", i_nl_sys));
    nl.offDiagonalsInAutoScaling(
        getParamFromNonlinearSystemVectorParam<bool>("off_diagonals_in_auto_scaling", i_nl_sys));
    if (isParamValid("scaling_group_variables"))
      nl.scalingGroupVariables(
          getParamFromNonlinearSystemVectorParam<std::vector<std::vector<std::string>>>(
              "scaling_group_variables", i_nl_sys));
    if (isParamValid("ignore_variables_for_autoscaling"))
    {
      // Before setting ignore_variables_for_autoscaling, check that they are not present in
      // scaling_group_variables
      if (isParamValid("scaling_group_variables"))
      {
        const auto & ignore_variables_for_autoscaling =
            getParamFromNonlinearSystemVectorParam<std::vector<std::string>>(
                "ignore_variables_for_autoscaling", i_nl_sys);
        const auto & scaling_group_variables =
            getParamFromNonlinearSystemVectorParam<std::vector<std::vector<std::string>>>(
                "scaling_group_variables", i_nl_sys);
        for (const auto & group : scaling_group_variables)
          for (const auto & var_name : group)
            if (std::find(ignore_variables_for_autoscaling.begin(),
                          ignore_variables_for_autoscaling.end(),
                          var_name) != ignore_variables_for_autoscaling.end())
              paramError("ignore_variables_for_autoscaling",
                         "Variables cannot be in a scaling grouping and also be ignored");
      }
      nl.ignoreVariablesForAutoscaling(
          getParamFromNonlinearSystemVectorParam<std::vector<std::string>>(
              "ignore_variables_for_autoscaling", i_nl_sys));
    }
  }

  // Multi-grid options
  _problem.numGridSteps(_num_grid_steps);
}

template <typename T>
T
FEProblemSolve::getParamFromNonlinearSystemVectorParam(const std::string & param_name,
                                                       unsigned int index) const
{
  const auto & param_vec = getParam<std::vector<T>>(param_name);
  if (index > _num_nl_systems)
    paramError(param_name,
               "Vector parameter is requested at index (" + std::to_string(index) +
                   ") which is larger than number of nonlinear systems (" +
                   std::to_string(_num_nl_systems) + ").");
  if (param_vec.size() == 0)
    paramError(
        param_name,
        "This parameter was passed to a routine which cannot handle empty vector parameters");
  if (param_vec.size() != 1 && param_vec.size() != _num_nl_systems)
    paramError(param_name,
               "Vector parameter size (" + std::to_string(param_vec.size()) +
                   ") is different than the number of nonlinear systems (" +
                   std::to_string(_num_nl_systems) + ").");

  // User passed only one parameter, assume it applies to all nonlinear systems
  if (param_vec.size() == 1)
    return param_vec[0];
  else
    return param_vec[index];
}

void
FEProblemSolve::initialSetup()
{
  MultiSystemSolveObject::initialSetup();
  convergenceSetup();
  // Keep track of the solution warnings from the setup
  // before a count reset at the beginning of the time step
  if (!_app.isRecovering())
  {
    _app.solutionInvalidity().syncIteration();
    _app.solutionInvalidity().accumulateIterationIntoTimeStepOccurences();
    _app.solutionInvalidity().accumulateTimeStepIntoTotalOccurences(0);
  }
}

void
FEProblemSolve::convergenceSetup()
{
  // nonlinear
  const auto conv_names = _problem.getNonlinearConvergenceNames();
  for (const auto & conv_name : conv_names)
  {
    auto & conv = _problem.getConvergence(conv_name);
    conv.checkIterationType(ConvergenceIterationTypes::NONLINEAR);
  }

  // linear
  if (isParamValid("linear_convergence"))
  {
    const auto conv_names = getParam<std::vector<ConvergenceName>>("linear_convergence");
    for (const auto & conv_name : conv_names)
    {
      auto & conv = _problem.getConvergence(conv_name);
      conv.checkIterationType(ConvergenceIterationTypes::LINEAR);
    }
  }

  // multisystem fixed point
  if (isParamValid("multi_system_fixed_point_convergence"))
  {
    _multi_sys_fp_convergence =
        &_problem.getConvergence(getParam<ConvergenceName>("multi_system_fixed_point_convergence"));
    _multi_sys_fp_convergence->checkIterationType(
        ConvergenceIterationTypes::MULTISYSTEM_FIXED_POINT);
  }
}

bool
FEProblemSolve::solve()
{
  // Outer loop for multi-grid convergence
  bool converged = false;
  unsigned int fp_iter = 0;

  for (MooseIndex(_num_grid_steps) grid_step = 0; grid_step <= _num_grid_steps; ++grid_step)
  {
    // Multi-system fixed point loop
    fp_iter = 0;
    converged = false;
    while (!converged)
    {
      if (_using_multi_sys_fp_iterations)
        _console << COLOR_MAGENTA << "Multi-system fixed point iteration " << fp_iter << ":"
                 << COLOR_DEFAULT << "\n"
                 << std::endl;

      // Loop over each system
      for (const auto sys_i : index_range(_systems))
      {
        auto * const sys = _systems[sys_i];
        const bool is_nonlinear = (dynamic_cast<NonlinearSystemBase *>(sys) != nullptr);
        const Real fp_relax =
            _using_multi_sys_fp_iterations ? _multi_sys_fp_relax_factors[sys_i] : 1.0;
        const bool apply_fp_relax =
            _using_multi_sys_fp_iterations && !MooseUtils::absoluteFuzzyEqual(fp_relax, 1.0);
        if (apply_fp_relax)
        {
          sys->setFixedPointRelaxationFactor(fp_relax);
          sys->saveOldSolutionForFixedPointRelaxation();
        }

        // Call solve on the problem for that system
        if (is_nonlinear)
          _problem.solve(sys->number());
        else
        {
          const auto linear_sys_number =
              cast_int<unsigned int>(sys->number() - _problem.numNonlinearSystems());
          _problem.solveLinearSystem(linear_sys_number, &_problem.getPetscOptions());
        }

        // Check convergence
        const auto solve_name =
            _systems.size() == 1 ? " Solve" : "System " + sys->name() + ": Solve";
        if (_problem.shouldSolve())
        {
          if (_problem.converged(sys->number()))
          {
            if (apply_fp_relax)
              sys->applyFixedPointRelaxation();
            _console << COLOR_GREEN << solve_name << " Converged!" << COLOR_DEFAULT << "\n"
                     << std::endl;
          }
          else
          {
            _console << COLOR_RED << solve_name << " Did NOT Converge!" << COLOR_DEFAULT << "\n"
                     << std::endl;
            if (apply_fp_relax)
              sys->clearFixedPointRelaxation();
            return false;
          }
        }
        else
          _console << COLOR_GREEN << solve_name << " Skipped!" << COLOR_DEFAULT << "\n"
                   << std::endl;

        if (!is_nonlinear)
        {
          const auto linear_sys_number =
              cast_int<unsigned int>(sys->number() - _problem.numNonlinearSystems());
          auto & linear_sys = _problem.getLinearSystem(linear_sys_number);

          // This is for postprocessing purposes in case none of the objects request the gradients.
          // TODO: Somehow collect information if the postprocessors need gradients and if nothing
          // needs this, just skip it
          linear_sys.computeGradients();
        }

        if (apply_fp_relax)
          sys->clearFixedPointRelaxation();
      }

      // Assess convergence of the multi-system fixed point iteration
      if (!_using_multi_sys_fp_iterations)
        converged = true;
      else
      {
        _problem.execute(EXEC_MULTISYSTEM_FIXED_POINT_CONVERGENCE);

        // checkConvergence expects the number of iterations performed, not the iteration index:
        const auto n_fp_iter = fp_iter + 1;
        const auto convergence_status = _multi_sys_fp_convergence->checkConvergence(n_fp_iter);
        converged = convergence_status == Convergence::MooseConvergenceStatus::CONVERGED;
        if (convergence_status == Convergence::MooseConvergenceStatus::DIVERGED)
          break;
      }
      fp_iter++;
    }

    if (grid_step != _num_grid_steps)
      _problem.uniformRefine();
  }

  return converged;
}