src/layered/implicit.h

    Time-implicit barotropic integration

    This implements a semi-implicit scheme for the evolution of the free-surface elevation \eta of the multilayer solver. The scheme can be summarised as \displaystyle \begin{aligned} \frac{\eta^{n + 1} - \eta^n}{\Delta t} & = - \sum_k \nabla \cdot [\theta_H (hu)_k^{n + 1} + (1 - \theta_H) (hu)^n_k] \\ \frac{(hu)^{n + 1}_k - (hu)_k^n}{\Delta t} & = - \Delta tgh^{n + 1 / 2}_k (\theta_H \nabla \eta^{n + 1} + (1 - \theta_H) \nabla \eta^n) \end{aligned} where \theta_H is the “implicitness parameter” typically set to 1/2.

    The resulting Poisson–Helmholtz equation for \eta^{n+1} is solved using the multigrid Poisson solver. The convergence statistics are stored in mgH.

    #include "poisson.h"
    
    mgstats mgH;
    double theta_H = 0.5;
    
    #define IMPLICIT_H 1

    The scheme is unconditionally stable for gravity waves, so the gravity wave CFL is set to \infty, if it has not already been set (typically by the user).

    event defaults0 (i = 0)
    {
      if (CFL_H == 1e40)
        CFL_H = HUGE;
      mgH.nrelax = 4;
    }

    The relaxation and residual functions of the multigrid solver are derived from the Poisson–Helmholtz equation for \eta^{n+1} derived from the equations above \displaystyle \begin{aligned} \eta^{n + 1} + \nabla \cdot (\alpha \nabla \eta^{n + 1}) & = \eta^n - \Delta t \sum_k \nabla \cdot (hu)_k^{\star}\\ \alpha & \equiv - g (\theta \Delta t)^2 \sum_k h^{n + 1 / 2}_k\\ (hu)_k^{\star} & \equiv (hu)_k^n - \Delta tgh^{n + 1 / 2}_k \theta (1 - \theta) \nabla \eta^n \end{aligned}

    trace
    static void relax_hydro (scalar * ql, scalar * rhsl, int lev, void * data)
    {
      scalar eta = ql[0], rhs_eta = rhsl[0];
      face vector alpha = *((vector *)data);
      foreach_level_or_leaf (lev) {
        double d = - cm[]*Delta;
        double n = d*rhs_eta[];
        eta[] = 0.;
        foreach_dimension() {
          n += alpha.x[0]*a_baro (eta, 0) - alpha.x[1]*a_baro (eta, 1);
          diagonalize (eta) {
    	d -= alpha.x[0]*a_baro (eta, 0) - alpha.x[1]*a_baro (eta, 1);
          }
        }
        eta[] = n/d;
      }
    }
    
    trace
    static double residual_hydro (scalar * ql, scalar * rhsl,
    			      scalar * resl, void * data)
    {
      scalar eta = ql[0], rhs_eta = rhsl[0], res_eta = resl[0];
      face vector alpha = *((vector *)data);
      double maxres = 0.;
      face vector g[];
      foreach_face()
        g.x[] = alpha.x[]*a_baro (eta, 0);
      
      foreach (reduction(max:maxres)) {
        res_eta[] = rhs_eta[] - eta[];
        foreach_dimension()
          res_eta[] += (g.x[1] - g.x[])/(Delta*cm[]);
        if (fabs(res_eta[]) > maxres)
          maxres = fabs(res_eta[]);
      }
    
      return maxres;
    }

    This can be used to optionally store the residual (for debugging).

    scalar res_eta = {-1};
    
    scalar rhs_eta;
    face vector alpha_eta;

    The semi-implicit update of the layer heights is done in two steps. The first step is the explicit advection to time t + (1 - \theta_H)\Delta t of all tracers (including layer heights) i.e. \displaystyle \begin{aligned} h_k^{n + \theta} & = h_k^n - (1 - \theta_H) \Delta t \nabla \cdot (hu)^n_k \end{aligned}

    event half_advection (i++) {
      if (theta_H < 1.)
        advect (tracers, hu, hf, (1. - theta_H)*dt);
    }

    The r.h.s. and \alpha coefficients of the Poisson–Helmholtz equation are computed using the flux values at the “half-timestep”.

    event acceleration (i++)
    {    
      face vector su[];
      alpha_eta = new face vector;
      double C = - sq(theta_H*dt);
      foreach_face() {
        double ax = theta_H*a_baro (eta, 0);
        su.x[] = 0., alpha_eta.x[] = 0.;
        foreach_layer() {
          double hl = h[-1] > dry ? h[-1] : 0.;
          double hr = h[] > dry ? h[] : 0.;
          double uf = hl > 0. || hr > 0. ? (hl*u.x[-1] + hr*u.x[])/(hl + hr) : 0.;
          hu.x[] = (1. - theta_H)*(hu.x[] + dt*hf.x[]*ax) + theta_H*hf.x[]*uf;
          hu.x[] += dt*(theta_H*ha.x[] - hf.x[]*ax);
          ha.x[] -= hf.x[]*ax;
          su.x[] += hu.x[];
          alpha_eta.x[] += hf.x[];
        }
        alpha_eta.x[] *= C;
      }

    The r.h.s. is \displaystyle \text{rhs}_\eta = \eta^n - \Delta t\sum_k\nabla\cdot(hu)^\star_k

      rhs_eta = new scalar;
      foreach() {
        rhs_eta[] = eta[];
        foreach_dimension()
          rhs_eta[] -= dt*(su.x[1] - su.x[])/(Delta*cm[]);
      }

    The fields used by the relaxation function above (and/or by the relaxation function of the non-hydrostatic solver) need to be restricted to all levels.

      restriction ({cm, fm, zb, h, hf, alpha_eta});

    The restriction function for \eta, which has been modified by the multilayer solver, needs to be replaced by the (default) averaging function for the multigrid solver to work properly.

    #if TREE
      eta.restriction = restriction_average;
    #endif
    }

    In the second (implicit) step, the Poisson–Helmholtz equation for \eta^{n+1} is solved and the corresponding values for the fluxes (hu)^{n+1} are obtained by applying the corresponding pressure gradient term.

    event pressure (i++)
    {
      mgH = mg_solve ({eta}, {rhs_eta}, residual_hydro, relax_hydro, &alpha_eta,
    		  res = res_eta.i >= 0 ? (scalar *){res_eta} : NULL,
    		  nrelax = 4, minlevel = 1,
    		  tolerance = TOLERANCE);
      delete ({rhs_eta, alpha_eta});

    The restriction function for \eta is restored.

    #if TREE
      eta.restriction = restriction_eta;
    #endif

    Note that what is stored in hu corresponds to \theta_H(hu)^{n+1} since this is the flux which will be used in the pressure event to perform the “implicit” update of the tracers (including layer heights) i.e. \displaystyle \begin{aligned} h_k^{n + 1} & = h_k^{n + \theta} - \Delta t \nabla \cdot \theta_H (hu)^{n+1}_k \end{aligned}

      foreach_face() {
        double ax = theta_H*a_baro (eta, 0);
        foreach_layer() {
          ha.x[] += hf.x[]*ax;
          double hl = h[-1] > dry ? h[-1] : 0.;
          double hr = h[] > dry ? h[] : 0.;
          double uf = hl > 0. || hr > 0. ? (hl*u.x[-1] + hr*u.x[])/(hl + hr) : 0.;
          hu.x[] = theta_H*(hf.x[]*uf + dt*ha.x[]) - dt*ha.x[];
        }
      }
    }

    References

    [vitousek2013stability]

    Sean Vitousek and Oliver B Fringer. Stability and consistency of nonhydrostatic free-surface models using the semi-implicit θ-method. International Journal for Numerical Methods in Fluids, 72(5):550–582, 2013.

    Usage

    Tests