nonlinear/newtonsolver.hh

Go to the documentation of this file.

// -*- mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*-
// vi: set et ts=4 sw=4 sts=4:
//
// SPDX-FileCopyrightText: Copyright © DuMux Project contributors, see AUTHORS.md in root folder
// SPDX-License-Identifier: GPL-3.0-or-later
//
#ifndef DUMUX_NEWTON_SOLVER_HH
#define DUMUX_NEWTON_SOLVER_HH
 
#include <cmath>
#include <memory>
#include <iostream>
#include <type_traits>
#include <algorithm>
#include <numeric>
 
#include <dune/common/timer.hh>
#include <dune/common/exceptions.hh>
#include <dune/common/parallel/mpicommunication.hh>
#include <dune/common/parallel/mpihelper.hh>
#include <dune/common/std/type_traits.hh>
#include <dune/common/indices.hh>
#include <dune/common/hybridutilities.hh>
 
#include <dune/istl/bvector.hh>
#include <dune/istl/multitypeblockvector.hh>
 
#include <dumux/common/parameters.hh>
#include <dumux/common/exceptions.hh>
#include <dumux/common/typetraits/vector.hh>
#include <dumux/common/typetraits/isvalid.hh>
#include <dumux/common/timeloop.hh>
#include <dumux/common/pdesolver.hh>
#include <dumux/common/variablesbackend.hh>
 
#include <dumux/io/format.hh>
 
#include <dumux/linear/matrixconverter.hh>
#include <dumux/assembly/partialreassembler.hh>
 
#include "newtonconvergencewriter.hh"
#include "primaryvariableswitchadapter.hh"
 
namespace Dumux::Detail::Newton {
 
// Helper boolean that states if the assembler exports grid variables
template<class Assembler> using AssemblerGridVariablesType = typename Assembler::GridVariables;
template<class Assembler>
inline constexpr bool assemblerExportsGridVariables
    = Dune::Std::is_detected_v<AssemblerGridVariablesType, Assembler>;
 
// helper struct to define the variables on which the privarswitch should operate
template<class Assembler, bool exportsGridVars = assemblerExportsGridVariables<Assembler>>
struct PriVarSwitchVariablesType { using Type = typename Assembler::GridVariables; };
 
// if assembler does not export them, use an empty class. These situations either mean
// that there is no privarswitch, or, it is handled by a derived implementation.
template<class Assembler>
struct PriVarSwitchVariablesType<Assembler, false>
{ using Type = struct EmptyGridVariables {}; };
 
// Helper alias to deduce the variables types used in the privarswitch adapter
template<class Assembler>
using PriVarSwitchVariables
    = typename PriVarSwitchVariablesType<Assembler, assemblerExportsGridVariables<Assembler>>::Type;
 
struct supportsPartialReassembly
{
    template<class Assembler>
    auto operator()(Assembler&& a)
    -> decltype(a.assembleJacobianAndResidual(std::declval<const typename Assembler::SolutionVector&>(),
                                              std::declval<const PartialReassembler<Assembler>*>()))
    {}
};
 
// helpers to implement max relative shift
template<class C> using dynamicIndexAccess = decltype(std::declval<C>()[0]);
template<class C> using staticIndexAccess = decltype(std::declval<C>()[Dune::Indices::_0]);
template<class C> static constexpr auto hasDynamicIndexAccess = Dune::Std::is_detected<dynamicIndexAccess, C>{};
template<class C> static constexpr auto hasStaticIndexAccess = Dune::Std::is_detected<staticIndexAccess, C>{};
 
template<class V, class Scalar, class Reduce, class Transform>
auto hybridInnerProduct(const V& v1, const V& v2, Scalar init, Reduce&& r, Transform&& t)
-> std::enable_if_t<hasDynamicIndexAccess<V>(), Scalar>
{
    return std::inner_product(v1.begin(), v1.end(), v2.begin(), init, std::forward<Reduce>(r), std::forward<Transform>(t));
}
 
template<class V, class Scalar, class Reduce, class Transform>
auto hybridInnerProduct(const V& v1, const V& v2, Scalar init, Reduce&& r, Transform&& t)
-> std::enable_if_t<hasStaticIndexAccess<V>() && !hasDynamicIndexAccess<V>(), Scalar>
{
    using namespace Dune::Hybrid;
    forEach(std::make_index_sequence<V::N()>{}, [&](auto i){
        init = r(init, hybridInnerProduct(v1[Dune::index_constant<i>{}], v2[Dune::index_constant<i>{}], init, std::forward<Reduce>(r), std::forward<Transform>(t)));
    });
    return init;
}
 
// Maximum relative shift at a degree of freedom.
// For (primary variables) values below 1.0 we use
// an absolute shift.
template<class Scalar, class V>
auto maxRelativeShift(const V& v1, const V& v2)
-> std::enable_if_t<Dune::IsNumber<V>::value, Scalar>
{
    using std::abs; using std::max;
    return abs(v1 - v2)/max<Scalar>(1.0, abs(v1 + v2)*0.5);
}
 
// Maximum relative shift for generic vector types.
// Recursively calls maxRelativeShift until Dune::IsNumber is true.
template<class Scalar, class V>
auto maxRelativeShift(const V& v1, const V& v2)
-> std::enable_if_t<!Dune::IsNumber<V>::value, Scalar>
{
    return hybridInnerProduct(v1, v2, Scalar(0.0),
        [](const auto& a, const auto& b){ using std::max; return max(a, b); },
        [](const auto& a, const auto& b){ return maxRelativeShift<Scalar>(a, b); }
    );
}
 
template<class To, class From>
void assign(To& to, const From& from)
{
    if constexpr (hasStaticIndexAccess<To>() && hasStaticIndexAccess<To>() && !hasDynamicIndexAccess<From>() && !hasDynamicIndexAccess<From>())
    {
        using namespace Dune::Hybrid;
        forEach(std::make_index_sequence<To::N()>{}, [&](auto i){
            assign(to[Dune::index_constant<i>{}], from[Dune::index_constant<i>{}]);
        });
    }
 
    else if constexpr (std::is_assignable<To&, From>::value)
        to = from;
 
    else if constexpr (hasDynamicIndexAccess<To>() && hasDynamicIndexAccess<From>())
        for (decltype(to.size()) i = 0; i < to.size(); ++i)
            assign(to[i], from[i]);
 
    else if constexpr (hasDynamicIndexAccess<To>() && Dune::IsNumber<From>::value)
    {
        assert(to.size() == 1);
        assign(to[0], from);
    }
 
    else if constexpr (Dune::IsNumber<To>::value && hasDynamicIndexAccess<From>())
    {
        assert(from.size() == 1);
        assign(to, from[0]);
    }
 
    else
        DUNE_THROW(Dune::Exception, "Values are not assignable to each other!");
}
 
} // end namespace Dumux::Detail::Newton
 
namespace Dumux {
 
template <class Assembler, class LinearSolver,
          class Reassembler = PartialReassembler<Assembler>,
          class Comm = Dune::Communication<Dune::MPIHelper::MPICommunicator> >
class NewtonSolver : public PDESolver<Assembler, LinearSolver>
{
    using ParentType = PDESolver<Assembler, LinearSolver>;
 
protected:
    using Backend = VariablesBackend<typename ParentType::Variables>;
    using SolutionVector = typename Backend::DofVector;
    using ResidualVector = typename Assembler::ResidualType;
    using LinearAlgebraNativeBackend = VariablesBackend<ResidualVector>;
private:
    using Scalar = typename Assembler::Scalar;
    using JacobianMatrix = typename Assembler::JacobianMatrix;
    using ConvergenceWriter = ConvergenceWriterInterface<SolutionVector, ResidualVector>;
    using TimeLoop = TimeLoopBase<Scalar>;
 
    // enable models with primary variable switch
    // TODO: Always use ParentType::Variables once we require assemblers to export variables
    static constexpr bool assemblerExportsVariables = Detail::PDESolver::assemblerExportsVariables<Assembler>;
    using PriVarSwitchVariables
        = std::conditional_t<assemblerExportsVariables,
                             typename ParentType::Variables,
                             Detail::Newton::PriVarSwitchVariables<Assembler>>;
    using PrimaryVariableSwitchAdapter = Dumux::PrimaryVariableSwitchAdapter<PriVarSwitchVariables>;
 
public:
    using typename ParentType::Variables;
    using Communication = Comm;
 
    NewtonSolver(std::shared_ptr<Assembler> assembler,
                 std::shared_ptr<LinearSolver> linearSolver,
                 const Communication& comm = Dune::MPIHelper::getCommunication(),
                 const std::string& paramGroup = "",
                 const std::string& paramGroupName = "Newton",
                 int verbosity = 2)
    : ParentType(assembler, linearSolver)
    , endIterMsgStream_(std::ostringstream::out)
    , comm_(comm)
    , paramGroup_(paramGroup)
    , solverName_(paramGroupName)
    , priVarSwitchAdapter_(std::make_unique<PrimaryVariableSwitchAdapter>(paramGroup))
    {
        verbosity_ = comm_.rank() == 0 ? getParamFromGroup<int>(paramGroup, solverName_ + ".Verbosity", verbosity) : 0;
 
        initParams_(paramGroup);
 
        // set the linear system (matrix & residual) in the assembler
        this->assembler().setLinearSystem();
 
        // set a different default for the linear solver residual reduction
        // within the Newton the linear solver doesn't need to solve too exact
        this->linearSolver().setResidualReduction(getParamFromGroup<Scalar>(paramGroup, "LinearSolver.ResidualReduction", 1e-6));
 
        // initialize the partial reassembler
        if (enablePartialReassembly_)
            partialReassembler_ = std::make_unique<Reassembler>(this->assembler());
    }
 
    const Communication& comm() const
    { return comm_; }
 
    void setMaxRelativeShift(Scalar tolerance)
    { shiftTolerance_ = tolerance; }
 
    void setMaxAbsoluteResidual(Scalar tolerance)
    { residualTolerance_ = tolerance; }
 
    void setResidualReduction(Scalar tolerance)
    { reductionTolerance_ = tolerance; }
 
    void setTargetSteps(int targetSteps)
    { targetSteps_ = targetSteps; }
 
    void setMinSteps(int minSteps)
    { minSteps_ = minSteps; }
 
    void setMaxSteps(int maxSteps)
    { maxSteps_ = maxSteps; }
 
    void solve(Variables& vars, TimeLoop& timeLoop) override
    {
        if constexpr (!assemblerExportsVariables)
        {
            if (this->assembler().isStationaryProblem())
                DUNE_THROW(Dune::InvalidStateException, "Using time step control with stationary problem makes no sense!");
        }
 
        // try solving the non-linear system
        for (std::size_t i = 0; i <= maxTimeStepDivisions_; ++i)
        {
            // linearize & solve
            const bool converged = solve_(vars);
 
            if (converged)
                return;
 
            else if (!converged && i < maxTimeStepDivisions_)
            {
                if constexpr (assemblerExportsVariables)
                    DUNE_THROW(Dune::NotImplemented, "Time step reset for new assembly methods");
                else
                {
                    // set solution to previous solution & reset time step
                    Backend::update(vars, this->assembler().prevSol());
                    this->assembler().resetTimeStep(Backend::dofs(vars));
 
                    if (verbosity_ >= 1)
                    {
                        const auto dt = timeLoop.timeStepSize();
                        std::cout << Fmt::format("{} solver did not converge with dt = {} seconds. ", solverName_, dt)
                                  << Fmt::format("Retrying with time step of dt = {} seconds.\n", dt*retryTimeStepReductionFactor_);
                    }
 
                    // try again with dt = dt * retryTimeStepReductionFactor_
                    timeLoop.setTimeStepSize(timeLoop.timeStepSize() * retryTimeStepReductionFactor_);
                }
            }
 
            else
            {
                DUNE_THROW(NumericalProblem,
                    Fmt::format("{} solver didn't converge after {} time-step divisions; dt = {}.\n",
                                solverName_, maxTimeStepDivisions_, timeLoop.timeStepSize()));
            }
        }
    }
 
    void solve(Variables& vars) override
    {
        const bool converged = solve_(vars);
        if (!converged)
            DUNE_THROW(NumericalProblem,
                Fmt::format("{} solver didn't converge after {} iterations.\n", solverName_, numSteps_));
    }
 
    bool apply(Variables& vars) override
    {
        return solve_(vars);
    }
 
    virtual void newtonBegin(Variables& initVars)
    {
        numSteps_ = 0;
 
        if constexpr (hasPriVarsSwitch<PriVarSwitchVariables>)
        {
            if constexpr (assemblerExportsVariables)
                priVarSwitchAdapter_->initialize(Backend::dofs(initVars), initVars);
            else // this assumes assembly with solution (i.e. Variables=SolutionVector)
                priVarSwitchAdapter_->initialize(initVars, this->assembler().gridVariables());
        }
 
 
        const auto& initSol = Backend::dofs(initVars);
 
        // write the initial residual if a convergence writer was set
        if (convergenceWriter_)
        {
            this->assembler().assembleResidual(initVars);
 
            // dummy vector, there is no delta before solving the linear system
            ResidualVector delta = LinearAlgebraNativeBackend::zeros(Backend::size(initSol));
            convergenceWriter_->write(initSol, delta, this->assembler().residual());
        }
 
        if (enablePartialReassembly_)
        {
            partialReassembler_->resetColors();
            resizeDistanceFromLastLinearization_(initSol, distanceFromLastLinearization_);
        }
    }
 
    virtual bool newtonProceed(const Variables &varsCurrentIter, bool converged)
    {
        if (numSteps_ < minSteps_)
            return true;
        else if (converged)
            return false; // we are below the desired tolerance
        else if (numSteps_ >= maxSteps_)
        {
            // We have exceeded the allowed number of steps. If the
            // maximum relative shift was reduced by a factor of at least 4,
            // we proceed even if we are above the maximum number of steps.
            if (enableShiftCriterion_)
                return shift_*4.0 < lastShift_;
            else
                return reduction_*4.0 < lastReduction_;
        }
 
        return true;
    }
 
    virtual void newtonBeginStep(const Variables& vars)
    {
        lastShift_ = shift_;
        if (numSteps_ == 0)
        {
            lastReduction_ = 1.0;
        }
        else
        {
            lastReduction_ = reduction_;
        }
    }
 
    virtual void assembleLinearSystem(const Variables& vars)
    {
        assembleLinearSystem_(this->assembler(), vars);
 
        if (enablePartialReassembly_)
            partialReassembler_->report(comm_, endIterMsgStream_);
    }
 
    void solveLinearSystem(ResidualVector& deltaU)
    {
        bool converged = false;
 
        try
        {
            if (numSteps_ == 0)
                initialResidual_ = this->linearSolver().norm(this->assembler().residual());
 
            // solve by calling the appropriate implementation depending on whether the linear solver
            // is capable of handling MultiType matrices or not
            converged = solveLinearSystem_(deltaU);
        }
        catch (const Dune::Exception &e)
        {
            if (verbosity_ >= 1)
                std::cout << solverName_ << " caught exception in  the linear solver: \"" << e.what() << "\"\n";
 
            converged = false;
        }
 
        // make sure all processes converged
        int convergedRemote = converged;
        if (comm_.size() > 1)
            convergedRemote = comm_.min(converged);
 
        if (!converged)
        {
            DUNE_THROW(NumericalProblem, "Linear solver did not converge");
            ++numLinearSolverBreakdowns_;
        }
        else if (!convergedRemote)
        {
            DUNE_THROW(NumericalProblem, "Linear solver did not converge on a remote process");
            ++numLinearSolverBreakdowns_; // we keep correct count for process 0
        }
    }
 
    void newtonUpdate(Variables& vars,
                      const SolutionVector& uLastIter,
                      const ResidualVector& deltaU)
    {
        callAndCheck_("newtonUpdate", [&]{
            if (useLineSearch_)
                lineSearchUpdate_(vars, uLastIter, deltaU);
 
            else if (useChop_)
                choppedUpdate_(vars, uLastIter, deltaU);
 
            else
            {
                auto uCurrentIter = uLastIter;
                Backend::axpy(-1.0, deltaU, uCurrentIter);
                solutionChanged_(vars, uCurrentIter);
 
                if (enableResidualCriterion_)
                    computeResidualReduction_(vars);
            }
        });
 
        if (enableShiftCriterion_ || enablePartialReassembly_)
            newtonComputeShift_(Backend::dofs(vars), uLastIter);
 
        if (enablePartialReassembly_) {
            // Determine the threshold 'eps' that is used for the partial reassembly.
            // Every entity where the primary variables exhibit a relative shift
            // summed up since the last linearization above 'eps' will be colored
            // red yielding a reassembly.
            // The user can provide three parameters to influence the threshold:
            // 'minEps' by 'Newton.ReassemblyMinThreshold' (1e-1*shiftTolerance_ default)
            // 'maxEps' by 'Newton.ReassemblyMaxThreshold' (1e2*shiftTolerance_ default)
            // 'omega'  by 'Newton.ReassemblyShiftWeight'  (1e-3 default)
            // The threshold is calculated from the currently achieved maximum
            // relative shift according to the formula
            // eps = max( minEps, min(maxEps, omega*shift) ).
            // Increasing/decreasing 'minEps' leads to less/more reassembly if
            // 'omega*shift' is small, i.e., for the last Newton iterations.
            // Increasing/decreasing 'maxEps' leads to less/more reassembly if
            // 'omega*shift' is large, i.e., for the first Newton iterations.
            // Increasing/decreasing 'omega' leads to more/less first and last
            // iterations in this sense.
            using std::max;
            using std::min;
            auto reassemblyThreshold = max(reassemblyMinThreshold_,
                                           min(reassemblyMaxThreshold_,
                                               shift_*reassemblyShiftWeight_));
 
            auto actualDeltaU = uLastIter;
            actualDeltaU -= Backend::dofs(vars);
            updateDistanceFromLastLinearization_(uLastIter, actualDeltaU);
            partialReassembler_->computeColors(this->assembler(),
                                               distanceFromLastLinearization_,
                                               reassemblyThreshold);
 
            // set the discrepancy of the red entities to zero
            for (unsigned int i = 0; i < distanceFromLastLinearization_.size(); i++)
                if (partialReassembler_->dofColor(i) == EntityColor::red)
                    distanceFromLastLinearization_[i] = 0;
        }
    }
 
    virtual void newtonEndStep(Variables &vars,
                               const SolutionVector &uLastIter)
    {
        if constexpr (hasPriVarsSwitch<PriVarSwitchVariables>)
        {
            if constexpr (assemblerExportsVariables)
                priVarSwitchAdapter_->invoke(Backend::dofs(vars), vars);
            else // this assumes assembly with solution (i.e. Variables=SolutionVector)
                priVarSwitchAdapter_->invoke(vars, this->assembler().gridVariables());
        }
 
        ++numSteps_;
 
        if (verbosity_ >= 1)
        {
            if (enableDynamicOutput_)
                std::cout << '\r'; // move cursor to beginning of line
 
            const auto width = Fmt::formatted_size("{}", maxSteps_);
            std::cout << Fmt::format("{} iteration {:{}} done", solverName_, numSteps_, width);
 
            if (enableShiftCriterion_)
                std::cout << Fmt::format(", maximum relative shift = {:.4e}", shift_);
            if (enableResidualCriterion_ || enableAbsoluteResidualCriterion_)
                std::cout << Fmt::format(", residual = {:.4e}, residual reduction = {:.4e}", residualNorm_, reduction_);
 
            std::cout << endIterMsgStream_.str() << "\n";
        }
        endIterMsgStream_.str("");
 
        // When the Newton iterations are done: ask the model to check whether it makes sense
        // TODO: how do we realize this? -> do this here in the Newton solver
        // model_().checkPlausibility();
    }
 
    virtual void newtonEnd()  {}
 
    virtual bool newtonConverged() const
    {
        // in case the model has a priVar switch and some some primary variables
        // actually switched their state in the last iteration, enforce another iteration
        if (priVarSwitchAdapter_->switched())
            return false;
 
        if (enableShiftCriterion_ && !enableResidualCriterion_)
        {
            return shift_ <= shiftTolerance_;
        }
        else if (!enableShiftCriterion_ && enableResidualCriterion_)
        {
            if(enableAbsoluteResidualCriterion_)
                return residualNorm_ <= residualTolerance_;
            else
                return reduction_ <= reductionTolerance_;
        }
        else if (satisfyResidualAndShiftCriterion_)
        {
            if(enableAbsoluteResidualCriterion_)
                return shift_ <= shiftTolerance_
                        && residualNorm_ <= residualTolerance_;
            else
                return shift_ <= shiftTolerance_
                        && reduction_ <= reductionTolerance_;
        }
        else if(enableShiftCriterion_ && enableResidualCriterion_)
        {
            if(enableAbsoluteResidualCriterion_)
                return shift_ <= shiftTolerance_
                        || residualNorm_ <= residualTolerance_;
            else
                return shift_ <= shiftTolerance_
                        || reduction_ <= reductionTolerance_;
        }
        else
        {
            return shift_ <= shiftTolerance_
                    || reduction_ <= reductionTolerance_
                    || residualNorm_ <= residualTolerance_;
        }
 
        return false;
    }
 
    virtual void newtonFail(Variables& u) {}
 
    virtual void newtonSucceed()  {}
 
    void report(std::ostream& sout = std::cout) const
    {
        sout << '\n'
             << solverName_ << " statistics\n"
             << "----------------------------------------------\n"
             << "-- Total iterations:                   " << totalWastedIter_ + totalSucceededIter_ << '\n'
             << "-- Total wasted iterations:            " << totalWastedIter_ << '\n'
             << "-- Total succeeded iterations:         " << totalSucceededIter_ << '\n'
             << "-- Average iterations per solve:       " << std::setprecision(3) << double(totalSucceededIter_) / double(numConverged_) << '\n'
             << "-- Number of linear solver breakdowns: " << numLinearSolverBreakdowns_ << '\n'
             << std::endl;
    }
 
    void resetReport()
    {
        totalWastedIter_ = 0;
        totalSucceededIter_ = 0;
        numConverged_ = 0;
        numLinearSolverBreakdowns_ = 0;
    }
 
    void reportParams(std::ostream& sout = std::cout) const
    {
        sout << "\n" << solverName_ << " solver configured with the following options and parameters:\n";
        // options
        if (useLineSearch_) sout << " -- " << solverName_ << ".UseLineSearch = true\n";
        if (useChop_) sout << " -- " << solverName_ << ".EnableChop = true\n";
        if (enablePartialReassembly_) sout << " -- " << solverName_ << ".EnablePartialReassembly = true\n";
        if (enableAbsoluteResidualCriterion_) sout << " -- " << solverName_ << ".EnableAbsoluteResidualCriterion = true\n";
        if (enableShiftCriterion_) sout << " -- " << solverName_ << ".EnableShiftCriterion = true (relative shift convergence criterion)\n";
        if (enableResidualCriterion_) sout << " -- " << solverName_ << ".EnableResidualCriterion = true\n";
        if (satisfyResidualAndShiftCriterion_) sout << " -- " << solverName_ << ".SatisfyResidualAndShiftCriterion = true\n";
        // parameters
        if (enableShiftCriterion_) sout << " -- " << solverName_ << ".MaxRelativeShift = " << shiftTolerance_ << '\n';
        if (enableAbsoluteResidualCriterion_) sout << " -- " << solverName_ << ".MaxAbsoluteResidual = " << residualTolerance_ << '\n';
        if (enableResidualCriterion_) sout << " -- " << solverName_ << ".ResidualReduction = " << reductionTolerance_ << '\n';
        sout << " -- " << solverName_ << ".MinSteps = " << minSteps_ << '\n';
        sout << " -- " << solverName_ << ".MaxSteps = " << maxSteps_ << '\n';
        sout << " -- " << solverName_ << ".TargetSteps = " << targetSteps_ << '\n';
        if (enablePartialReassembly_)
        {
            sout << " -- " << solverName_ << ".ReassemblyMinThreshold = " << reassemblyMinThreshold_ << '\n';
            sout << " -- " << solverName_ << ".ReassemblyMaxThreshold = " << reassemblyMaxThreshold_ << '\n';
            sout << " -- " << solverName_ << ".ReassemblyShiftWeight = " << reassemblyShiftWeight_ << '\n';
        }
        sout << " -- " << solverName_ << ".RetryTimeStepReductionFactor = " << retryTimeStepReductionFactor_ << '\n';
        sout << " -- " << solverName_ << ".MaxTimeStepDivisions = " << maxTimeStepDivisions_ << '\n';
        sout << std::endl;
    }
 
    Scalar suggestTimeStepSize(Scalar oldTimeStep) const
    {
        // be aggressive reducing the time-step size but
        // conservative when increasing it. the rationale is
        // that we want to avoid failing in the next Newton
        // iteration which would require another linearization
        // of the problem.
        if (numSteps_ > targetSteps_) {
            Scalar percent = Scalar(numSteps_ - targetSteps_)/targetSteps_;
            return oldTimeStep/(1.0 + percent);
        }
 
        Scalar percent = Scalar(targetSteps_ - numSteps_)/targetSteps_;
        return oldTimeStep*(1.0 + percent/1.2);
    }
 
    void setVerbosity(int val)
    { verbosity_ = val; }
 
    int verbosity() const
    { return verbosity_ ; }
 
    void setUseLineSearch(bool val = true)
    { useLineSearch_ = val; }
 
    bool useLineSearch() const
    { return useLineSearch_; }
 
    const std::string& paramGroup() const
    { return paramGroup_; }
 
    void attachConvergenceWriter(std::shared_ptr<ConvergenceWriter> convWriter)
    { convergenceWriter_ = convWriter; }
 
    void detachConvergenceWriter()
    { convergenceWriter_ = nullptr; }
 
    Scalar retryTimeStepReductionFactor() const
    { return retryTimeStepReductionFactor_; }
 
    void setRetryTimeStepReductionFactor(const Scalar factor)
    { retryTimeStepReductionFactor_ = factor; }
 
protected:
 
    virtual void solutionChanged_(Variables& vars, const SolutionVector& uCurrentIter)
    {
        Backend::update(vars, uCurrentIter);
 
        if constexpr (!assemblerExportsVariables)
            this->assembler().updateGridVariables(Backend::dofs(vars));
    }
 
    void computeResidualReduction_(const Variables& vars)
    {
        // we assume that the assembler works on solution vectors
        // if it doesn't export the variables type
        if constexpr (!assemblerExportsVariables)
            this->assembler().assembleResidual(Backend::dofs(vars));
        else
            this->assembler().assembleResidual(vars);
 
        residualNorm_ = this->linearSolver().norm(this->assembler().residual());
 
        reduction_ = residualNorm_/initialResidual_;
    }
 
    bool enableResidualCriterion() const
    { return enableResidualCriterion_; }
 
    int targetSteps_;
    int minSteps_;
    int maxSteps_;
    int numSteps_;
 
    // residual criterion variables
    Scalar reduction_;
    Scalar residualNorm_;
    Scalar lastReduction_;
    Scalar initialResidual_;
 
    // shift criterion variables
    Scalar shift_;
    Scalar lastShift_;
 
    std::ostringstream endIterMsgStream_;
 
 
private:
    bool solve_(Variables& vars)
    {
        bool converged = false;
 
        Dune::Timer assembleTimer(false);
        Dune::Timer solveTimer(false);
        Dune::Timer updateTimer(false);
 
        try {
            converged = solveImpl_(
                vars, assembleTimer, solveTimer, updateTimer
            );
        }
 
        // Dumux::NumericalProblem may be recovered from
        catch (const Dumux::NumericalProblem &e)
        {
            if (verbosity_ >= 1)
                std::cout << solverName_ << " caught exception: \"" << e.what() << "\"\n";
            converged = false;
        }
 
        // make sure all processes were successful
        int allProcessesConverged = static_cast<int>(converged);
        if (comm_.size() > 1)
            allProcessesConverged = comm_.min(static_cast<int>(converged));
 
        if (allProcessesConverged)
        {
            totalSucceededIter_ += numSteps_;
            numConverged_++;
            newtonSucceed();
 
            if (verbosity_ >= 1) {
                const auto elapsedTot = assembleTimer.elapsed() + solveTimer.elapsed() + updateTimer.elapsed();
                std::cout << Fmt::format("Assemble/solve/update time: {:.2g}({:.2f}%)/{:.2g}({:.2f}%)/{:.2g}({:.2f}%)\n",
                                            assembleTimer.elapsed(), 100*assembleTimer.elapsed()/elapsedTot,
                                            solveTimer.elapsed(), 100*solveTimer.elapsed()/elapsedTot,
                                            updateTimer.elapsed(), 100*updateTimer.elapsed()/elapsedTot);
            }
        }
        else
        {
            totalWastedIter_ += numSteps_;
            newtonFail(vars);
        }
 
        return static_cast<bool>(allProcessesConverged);
    }
 
    bool solveImpl_(Variables& vars,
        Dune::Timer& assembleTimer,
        Dune::Timer& solveTimer,
        Dune::Timer& updateTimer)
    {
        // newtonBegin may manipulate the solution
        newtonBegin(vars);
 
        // the given solution is the initial guess
        auto uLastIter = Backend::dofs(vars);
        ResidualVector deltaU = LinearAlgebraNativeBackend::zeros(Backend::size(Backend::dofs(vars)));
        Detail::Newton::assign(deltaU, Backend::dofs(vars));
 
        // execute the method as long as the solver thinks
        // that we should do another iteration
        bool converged = false;
        while (newtonProceed(vars, converged))
        {
            // notify the solver that we're about to start
            // a new iteration
            newtonBeginStep(vars);
 
            // make the current solution to the old one
            if (numSteps_ > 0)
                uLastIter = Backend::dofs(vars);
 
            if (verbosity_ >= 1 && enableDynamicOutput_)
                std::cout << "Assemble: r(x^k) = dS/dt + div F - q;   M = grad r"
                            << std::flush;
 
            // assemble
 
            // linearize the problem at the current solution
            assembleTimer.start();
            callAndCheck_("assemble", [&]{ assembleLinearSystem(vars); });
            assembleTimer.stop();
 
            // linear solve
 
            // Clear the current line using an ansi escape
            // sequence.  for an explanation see
            // http://en.wikipedia.org/wiki/ANSI_escape_code
            const char clearRemainingLine[] = { 0x1b, '[', 'K', 0 };
 
            if (verbosity_ >= 1 && enableDynamicOutput_)
                std::cout << "\rSolve: M deltax^k = r"
                            << clearRemainingLine << std::flush;
 
            // solve the resulting linear equation system
            solveTimer.start();
 
            // set the delta vector to zero before solving the linear system!
            deltaU = 0;
 
            solveLinearSystem(deltaU);
            solveTimer.stop();
 
            // update
            if (verbosity_ >= 1 && enableDynamicOutput_)
                std::cout << "\rUpdate: x^(k+1) = x^k - deltax^k"
                            << clearRemainingLine << std::flush;
 
            updateTimer.start();
            // update the current solution (i.e. uOld) with the delta
            // (i.e. u). The result is stored in u
            newtonUpdate(vars, uLastIter, deltaU);
            updateTimer.stop();
 
            // tell the solver that we're done with this iteration
            newtonEndStep(vars, uLastIter);
 
            // if a convergence writer was specified compute residual and write output
            if (convergenceWriter_)
            {
                this->assembler().assembleResidual(vars);
                convergenceWriter_->write(Backend::dofs(vars), deltaU, this->assembler().residual());
            }
 
            // detect if the method has converged
            converged = newtonConverged();
        }
 
        // tell solver we are done
        newtonEnd();
 
        // check final convergence status
        converged = newtonConverged();
 
        // return status
        return converged;
    }
 
    template<std::invocable Func>
    void callAndCheck_(const std::string& stepName, Func&& run)
    {
        bool successful = false;
        try {
            run();
            successful = true;
        }
        catch (const Dumux::NumericalProblem& e)
        {
            successful = false;
 
            if (verbosity_ >= 1)
                std::cout << solverName_ << " caught exception: \"" << e.what() << "\"\n";
        }
 
        // make sure all processes converged
        int successfulRemote = static_cast<int>(successful);
        if (comm_.size() > 1)
            successfulRemote = comm_.min(static_cast<int>(successful));
 
        if (!successful)
            DUNE_THROW(NumericalProblem, "" << solverName_ << " caught exception during " << stepName << " on process " << comm_.rank());
 
        else if (!successfulRemote)
            DUNE_THROW(NumericalProblem, "" << solverName_ << " caught exception during " << stepName << " on a remote process");
    }
 
    template<class A>
    auto assembleLinearSystem_(const A& assembler, const Variables& vars)
    -> typename std::enable_if_t<decltype(isValid(Detail::Newton::supportsPartialReassembly())(assembler))::value, void>
    {
        this->assembler().assembleJacobianAndResidual(vars, partialReassembler_.get());
    }
 
    template<class A>
    auto assembleLinearSystem_(const A& assembler, const Variables& vars)
    -> typename std::enable_if_t<!decltype(isValid(Detail::Newton::supportsPartialReassembly())(assembler))::value, void>
    {
        this->assembler().assembleJacobianAndResidual(vars);
    }
 
    [[deprecated("Use computeShift_(u1, u2) instead")]]
    virtual void newtonUpdateShift_(const SolutionVector &uLastIter,
                                    const ResidualVector &deltaU)
    {
        auto uNew = uLastIter;
        Backend::axpy(-1.0, deltaU, uNew);
        newtonComputeShift_(uLastIter, uNew);
    }
 
    virtual void newtonComputeShift_(const SolutionVector &u1,
                                     const SolutionVector &u2)
    {
        shift_ = Detail::Newton::maxRelativeShift<Scalar>(u1, u2);
        if (comm_.size() > 1)
            shift_ = comm_.max(shift_);
    }
 
    virtual void lineSearchUpdate_(Variables &vars,
                                   const SolutionVector &uLastIter,
                                   const ResidualVector &deltaU)
    {
        Scalar lambda = 1.0;
        auto uCurrentIter = uLastIter;
 
        while (true)
        {
            Backend::axpy(-lambda, deltaU, uCurrentIter);
            solutionChanged_(vars, uCurrentIter);
 
            computeResidualReduction_(vars);
 
            if (reduction_ < lastReduction_ || lambda <= lineSearchMinRelaxationFactor_)
            {
                endIterMsgStream_ << Fmt::format(", residual reduction {:.4e}->{:.4e}@lambda={:.4f}", lastReduction_, reduction_, lambda);
                return;
            }
 
            // try with a smaller update and reset solution
            lambda *= 0.5;
            uCurrentIter = uLastIter;
        }
    }
 
    virtual void choppedUpdate_(Variables& vars,
                                const SolutionVector& uLastIter,
                                const ResidualVector& deltaU)
    {
        DUNE_THROW(Dune::NotImplemented,
                   "Chopped " << solverName_ << " solver update strategy not implemented.");
    }
 
    virtual bool solveLinearSystem_(ResidualVector& deltaU)
    {
        assert(this->checkSizesOfSubMatrices(this->assembler().jacobian()) && "Matrix blocks have wrong sizes!");
 
        return this->linearSolver().solve(
            this->assembler().jacobian(),
            deltaU,
            this->assembler().residual()
        );
    }
 
    void initParams_(const std::string& group = "")
    {
        useLineSearch_ = getParamFromGroup<bool>(group, solverName_ + ".UseLineSearch", false);
        lineSearchMinRelaxationFactor_ = getParamFromGroup<Scalar>(group, solverName_ + ".LineSearchMinRelaxationFactor", 0.125);
        useChop_ = getParamFromGroup<bool>(group, solverName_ + ".EnableChop", false);
        if(useLineSearch_ && useChop_)
            DUNE_THROW(Dune::InvalidStateException, "Use either linesearch OR chop!");
 
        enableAbsoluteResidualCriterion_ = getParamFromGroup<bool>(group, solverName_ + ".EnableAbsoluteResidualCriterion", false);
        enableShiftCriterion_ = getParamFromGroup<bool>(group, solverName_ + ".EnableShiftCriterion", true);
        enableResidualCriterion_ = getParamFromGroup<bool>(group, solverName_ + ".EnableResidualCriterion", false) || enableAbsoluteResidualCriterion_;
        satisfyResidualAndShiftCriterion_ = getParamFromGroup<bool>(group, solverName_ + ".SatisfyResidualAndShiftCriterion", false);
        enableDynamicOutput_ = getParamFromGroup<bool>(group, solverName_ + ".EnableDynamicOutput", true);
 
        if (!enableShiftCriterion_ && !enableResidualCriterion_)
        {
            DUNE_THROW(Dune::NotImplemented,
                       "at least one of " << solverName_ << ".EnableShiftCriterion or "
                       << solverName_ << ".EnableResidualCriterion has to be set to true");
        }
 
        setMaxRelativeShift(getParamFromGroup<Scalar>(group, solverName_ + ".MaxRelativeShift", 1e-8));
        setMaxAbsoluteResidual(getParamFromGroup<Scalar>(group, solverName_ + ".MaxAbsoluteResidual", 1e-5));
        setResidualReduction(getParamFromGroup<Scalar>(group, solverName_ + ".ResidualReduction", 1e-5));
        setTargetSteps(getParamFromGroup<int>(group, solverName_ + ".TargetSteps", 10));
        setMinSteps(getParamFromGroup<int>(group, solverName_ + ".MinSteps", 2));
        setMaxSteps(getParamFromGroup<int>(group, solverName_ + ".MaxSteps", 18));
 
        enablePartialReassembly_ = getParamFromGroup<bool>(group, solverName_ + ".EnablePartialReassembly", false);
        reassemblyMinThreshold_ = getParamFromGroup<Scalar>(group, solverName_ + ".ReassemblyMinThreshold", 1e-1*shiftTolerance_);
        reassemblyMaxThreshold_ = getParamFromGroup<Scalar>(group, solverName_ + ".ReassemblyMaxThreshold", 1e2*shiftTolerance_);
        reassemblyShiftWeight_ = getParamFromGroup<Scalar>(group, solverName_ + ".ReassemblyShiftWeight", 1e-3);
 
        maxTimeStepDivisions_ = getParamFromGroup<std::size_t>(group, solverName_ + ".MaxTimeStepDivisions", 10);
        retryTimeStepReductionFactor_ = getParamFromGroup<Scalar>(group, solverName_ + ".RetryTimeStepReductionFactor", 0.5);
 
        numSteps_ = 0;
 
        // output a parameter report
        if (verbosity_ >= 2)
            reportParams();
    }
 
    template<class SolA, class SolB>
    void updateDistanceFromLastLinearization_(const SolA& u, const SolB& uDelta)
    {
        if constexpr (Dune::IsNumber<SolA>::value)
        {
            auto nextPriVars = u;
            nextPriVars -= uDelta;
 
            // add the current relative shift for this degree of freedom
            auto shift = Detail::Newton::maxRelativeShift<Scalar>(u, nextPriVars);
            distanceFromLastLinearization_[0] += shift;
        }
        else
        {
            for (std::size_t i = 0; i < u.size(); ++i)
            {
                const auto& currentPriVars(u[i]);
                auto nextPriVars(currentPriVars);
                nextPriVars -= uDelta[i];
 
                // add the current relative shift for this degree of freedom
                auto shift = Detail::Newton::maxRelativeShift<Scalar>(currentPriVars, nextPriVars);
                distanceFromLastLinearization_[i] += shift;
            }
        }
    }
 
    template<class ...ArgsA, class...ArgsB>
    void updateDistanceFromLastLinearization_(const Dune::MultiTypeBlockVector<ArgsA...>& uLastIter,
                                              const Dune::MultiTypeBlockVector<ArgsB...>& deltaU)
    {
        DUNE_THROW(Dune::NotImplemented, "Reassembly for MultiTypeBlockVector");
    }
 
    template<class Sol>
    void resizeDistanceFromLastLinearization_(const Sol& u, std::vector<Scalar>& dist)
    {
        dist.assign(Backend::size(u), 0.0);
    }
 
    template<class ...Args>
    void resizeDistanceFromLastLinearization_(const Dune::MultiTypeBlockVector<Args...>& u,
                                              std::vector<Scalar>& dist)
    {
        DUNE_THROW(Dune::NotImplemented, "Reassembly for MultiTypeBlockVector");
    }
 
    Communication comm_;
 
    int verbosity_;
 
    Scalar shiftTolerance_;
    Scalar reductionTolerance_;
    Scalar residualTolerance_;
 
    // time step control
    std::size_t maxTimeStepDivisions_;
    Scalar retryTimeStepReductionFactor_;
 
    // further parameters
    bool useLineSearch_;
    Scalar lineSearchMinRelaxationFactor_;
    bool useChop_;
    bool enableAbsoluteResidualCriterion_;
    bool enableShiftCriterion_;
    bool enableResidualCriterion_;
    bool satisfyResidualAndShiftCriterion_;
    bool enableDynamicOutput_;
 
    std::string paramGroup_;
    std::string solverName_;
 
    // infrastructure for partial reassembly
    bool enablePartialReassembly_;
    std::unique_ptr<Reassembler> partialReassembler_;
    std::vector<Scalar> distanceFromLastLinearization_;
    Scalar reassemblyMinThreshold_;
    Scalar reassemblyMaxThreshold_;
    Scalar reassemblyShiftWeight_;
 
    // statistics for the optional report
    std::size_t totalWastedIter_ = 0; 
    std::size_t totalSucceededIter_ = 0; 
    std::size_t numConverged_ = 0; 
    std::size_t numLinearSolverBreakdowns_ = 0; 
 
    std::unique_ptr<PrimaryVariableSwitchAdapter> priVarSwitchAdapter_;
 
    std::shared_ptr<ConvergenceWriter> convergenceWriter_ = nullptr;
};
 
} // end namespace Dumux
 
#endif