src/conservation.h

    A generic solver for systems of conservation laws

    Using the ideas of Kurganov and Tadmor, 2000 it is possible to write a generic solver for systems of conservation laws of the form \displaystyle \partial_t\left(\begin{array}{c} s_i\\ \mathbf{v}_j\\ \end{array}\right) + \nabla\cdot\left(\begin{array}{c} \mathbf{F}_i\\ \mathbf{T}_j\\ \end{array}\right) = 0 where s_i is a list of scalar fields, \mathbf{v}_j a list of vector fields and \mathbf{F}_i, \mathbf{T}_j are the corresponding vector (resp. tensor) fluxes.

    Note that the Saint-Venant solver is a particular case of this generic algorithm.

    The user must provide the lists of conserved scalar and vector fields

    extern scalar * scalars;
    extern vector * vectors;

    as well as a function which, given the state of each quantity, returns the fluxes and the minimum/maximum eigenvalues (i.e. eigenvalue[0]/eigenvalue[1]) of the corresponding Riemann problem.

    void flux (const double * state, double * flux, double * eigenvalue);

    Time-integration

    Setup

    Time integration will be done with a generic predictor-corrector scheme.

    The generic time-integration scheme in predictor-corrector.h needs to know which fields are updated i.e. all the scalars + the components of all the vector fields. It also needs a function to compute the time-derivatives of the evolving variables.

    scalar * evolving;
    double update_conservation (scalar * conserved, scalar * updates, double dtmax);
    
    event defaults (i = 0)
    {
      evolving = list_concat (scalars, (scalar *) vectors);
      update = update_conservation;

    We switch to a pure minmod limiter by default for increased robustness.

      theta = 1.;

    With the MUSCL scheme we use the CFL depends on the dimension of the problem.

      if (CFL > 1./dimension)
        CFL = 1./dimension;

    On trees we need to replace the default bilinear refinement/prolongation with linear so that reconstructed values also use slope limiting.

      #if TREE
      for (scalar s in evolving) {
        s.refine = s.prolongation = refine_linear;
        s.restriction = restriction_volume_average;
        s.dirty = true; // boundary conditions need to be updated
      }
      #endif
    }

    At the end of the run we need to free the list (to avoid a memory leak).

    event cleanup (i = end) free (evolving);

    User initialisation happens here.

    event init (i = 0);

    Computing fluxes

    The core of the central-upwind scheme (see e.g. section 3.1 of Kurganov & Levy, 2002) is the approximate solution of the Riemann problem given by the left and right states to get the fluxes f.

    static double riemann (const double * right, const double * left,
    		       double Delta, double * f, int len, 
    		       double dtmax)
    {
      double fr[len], fl[len], er[2], el[2];
      flux (right, fr, er);
      flux (left,  fl, el);
      double ap = max(er[1], el[1]); ap = max(ap, 0.);
      double am = min(er[0], el[0]); am = min(am, 0.);
      double a = max(ap, -am); 
    
      if (a > 0.) {
        for (int i = 0; i < len; i++)
          f[i] = (ap*fl[i] - am*fr[i] + ap*am*(right[i] - left[i]))/(ap - am);
        double dt = CFL*Delta/a;
        if (dt < dtmax)
          dtmax = dt;
      }
      else
        for (int i = 0; i < len; i++)
          f[i] = 0.;
      return dtmax;
    }
    
    double update_conservation (scalar * conserved, scalar * updates, double dtmax)
    {

    The gradients of each quantity are stored in a list of dynamically-allocated fields. First-order reconstruction is used for the gradient fields.

      vector * slopes = NULL;
      for (scalar s in conserved) {
        vector slope = new vector;
        foreach_dimension() {
          slope.x.gradient = zero;
          #if TREE
          slope.x.prolongation = refine_linear;
          #endif
        }
        slopes = vectors_append (slopes, slope);
      }
      gradients (conserved, slopes);

    We allocated fields for storing fluxes for each scalar and vector quantity.

      vector * lflux = NULL;
      int len = list_len (conserved);
      for (scalar s in conserved) {
        vector f1 = new face vector;
        lflux = vectors_append (lflux, f1);
      }

    The predictor-corrector scheme treats all fields as scalars (stored in the conserved list). We need to recover vector and tensor quantities from these lists. To do so, knowing the number of scalar fields, we split the scalar list into a list of scalars and a list of vectors.

      int scalars_len = list_len (scalars);
    
      scalar * scalars = list_copy (conserved);
      if (scalars) scalars[scalars_len].i = -1;
      vector * vectors = vectors_from_scalars (&conserved[scalars_len]);

    We then do the same for the gradients i.e. split the list of vectors into a list of vectors and a list of tensors.

      vector * scalar_slopes = vectors_copy (slopes);
      if (scalar_slopes) scalar_slopes[scalars_len] = (vector){{-1}};
      tensor * vector_slopes = tensors_from_vectors (&slopes[scalars_len]);

    And again for the fluxes.

      vector * scalar_fluxes = vectors_copy (lflux);
      if (scalar_fluxes) scalar_fluxes[scalars_len] = (vector){{-1}};
      tensor * vector_fluxes = tensors_from_vectors (&lflux[scalars_len]);

    We are ready to compute the fluxes through each face of the domain.

      foreach_face (reduction (min:dtmax)) {

    Left/right state reconstruction

    We use the central values of each scalar/vector quantity and the pre-computed gradients to compute the left and right states.

        double r[len], l[len]; // right/left Riemann states
        double f[len];         // fluxes for each conserved quantity
        double dx = Delta/2.;
        int i = 0;
        scalar s;
        vector g;
        for (s,g in scalars,scalar_slopes) {
          r[i] = s[] - dx*g.x[];
          l[i++] = s[-1] + dx*g.x[-1];
        }
        vector v;
        tensor t;
        for (v,t in vectors,vector_slopes) {      
          r[i] = v.x[] - dx*t.x.x[];
          l[i++] = v.x[-1] + dx*t.x.x[-1];
          #if dimension > 1
            r[i] = v.y[] - dx*t.y.x[];
    	l[i++] = v.y[-1] + dx*t.y.x[-1];
          #endif
          #if dimension > 2
            r[i] = v.z[] - dx*t.z.x[];
    	l[i++] = v.z[-1] + dx*t.z.x[-1];
          #endif
        }

    Riemann problem

    We then call the generic Riemann solver and store the resulting fluxes in the pre-allocated fields.

        dtmax = riemann (r, l, Delta*cm[]/fm.x[], f, len, dtmax);
        i = 0;
        for (vector fs in scalar_fluxes)
          fs.x[] = fm.x[]*f[i++];
        for (tensor fv in vector_fluxes) {
          fv.x.x[] = fm.x[]*f[i++];
          #if dimension > 1
            fv.y.x[] = fm.x[]*f[i++];
          #endif
          #if dimension > 2
            fv.z.x[] = fm.x[]*f[i++];
          #endif
        }
      }

    Update

    The update for each scalar quantity is the divergence of the fluxes.

      foreach() {
        scalar ds;
        vector f;
        for (ds,f in updates,lflux) {
          ds[] = 0.;
          foreach_dimension()
    	ds[] += (f.x[] - f.x[1])/(cm[]*Delta);
        }
      }

    Cleanup

    We finally deallocate the memory used to store lists and gradient fields.

      free (scalars);
      free (vectors);
      free (scalar_slopes);
      free (vector_slopes);
      free (scalar_fluxes);
      free (vector_fluxes);
      delete ((scalar *) slopes);
      free (slopes);
      delete ((scalar *) lflux);
      free (lflux);
      
      return dtmax;
    }

    Usage

    Tests