Source code for sfepy.terms.terms_dg

r"""
Discontinous Galekrin method specific terms

Note
----

In einsum calls the following convention is used:

 `i` represents iterating over all cells of a region;

 `n` represents iterating over selected cells of a region, for example
 over cells on boundary;

 `b` represents iterating over basis functions of state variable;

 `d` represents iterating over basis functions of test variable;

 `k`, `l` , `m` represent iterating over geometric dimensions, for example
 coordinates of velocity or facet normal vector or rows and columns of diffusion
 tensor;

 `q` represents iterating over quadrature points;

 `f` represents iterating over facets of cell;

"""
import numpy as nm

# sfepy imports
from sfepy.terms.terms import Term, terms
from sfepy.base.base import output


[docs] class DGTerm(Term): r""" Abstract base class for DG terms, provides alternative call_function and eval_real methods to accommodate returning iels and vals. """ poly_space_base = "legendre"
[docs] def call_function(self, out, fargs): try: out, status = self.function(out, *fargs) except (RuntimeError, ValueError): terms.errclear() raise if status: terms.errclear() raise ValueError('term evaluation failed! (%s)' % self.name) return out, status
[docs] def eval_real(self, shape, fargs, mode='eval', term_mode=None, diff_var=None, **kwargs): out = nm.empty(shape, dtype=nm.float64) if mode == 'weak': out, status = self.call_function(out, fargs) else: status = self.call_function(out, fargs) return out, status
@staticmethod def _get_nbrhd_dof_indexes(cells, nrbhs, field): """Get indexes of DOFs for active and of active neighbouring cells Parameters ---------- cells : array_like cells indicies nrbhs : array_like cell - nbrhd indicies field : DGField Returns ------- iels : ndarray inner and outer DOF indicies, i.e. diagonal indicies and then their corresponding neighbour indicies """ inner_iels = field.bubble_dofs inner_iels = nm.stack((nm.repeat(inner_iels, field.n_el_nod), nm.tile(inner_iels, field.n_el_nod).flatten()), axis=-1) outer_iels = nm.stack( (nm.repeat(field.bubble_dofs[cells], field.n_el_nod), nm.tile(field.bubble_dofs[nrbhs], field.n_el_nod).flatten()), axis=-1) iels = nm.vstack((inner_iels, outer_iels)) return iels
[docs] class AdvectionDGFluxTerm(DGTerm): r""" Lax-Friedrichs flux term for advection of scalar quantity :math:`p` with the advection velocity :math:`\ul{a}` given as a material parameter (a known function of space and time). :Definition: .. math:: \int_{\partial{T_K}} \ul{n} \cdot \ul{f}^{*} (p_{in}, p_{out})q where .. math:: \ul{f}^{*}(p_{in}, p_{out}) = \ul{a} \frac{p_{in} + p_{out}}{2} + (1 - \alpha) \ul{n} C \frac{ p_{in} - p_{out}}{2}, :math:`\alpha \in [0, 1]`; :math:`\alpha = 0` for upwind scheme, :math:`\alpha = 1` for central scheme, and .. math:: C = \max_{p \in [?, ?]}\left\lvert n_x a_1 + n_y a_2 \right\rvert = \max_{p \in [?, ?]} \left\lvert \ul{n} \cdot \ul{a} \right\rvert the :math:`p_{in}` resp. :math:`p_{out}` is solution on the boundary of the element provided by element itself resp. its neighbor and :math:`\ul{a}` is advection velocity. :Arguments 1: - material : :math:`\ul{a}` - virtual : :math:`q` - state : :math:`p` :Arguments 3: - material : :math:`\ul{a}` - virtual : :math:`q` - state : :math:`p` - opt_material : :math:`\alpha` """ alpha = 0 name = "dw_dg_advect_laxfrie_flux" modes = ("weak",) arg_types = ('opt_material', 'material_advelo', 'virtual', 'state') arg_shapes = [{'opt_material' : '.: 1', 'material_advelo': 'D, 1', 'virtual' : (1, 'state'), 'state' : 1 }, {'opt_material': None}] integration = 'cell' symbolic = {'expression': 'div(a*p)*w', 'map' : {'p': 'state', 'a': 'material', 'v': 'virtual'} }
[docs] def get_fargs(self, alpha, advelo, test, state, mode=None, term_mode=None, diff_var=None, **kwargs): if alpha is not None: self.alpha = alpha field = state.field region = field.region if "DG" not in field.family_name: raise ValueError("Used DG term with non DG field {} of family {}" .format(field.name, field.family_name)) fargs = (state, diff_var, field, region, advelo[:, 0, :, 0]) return fargs
[docs] def function(self, out, state, diff_var, field, region, advelo): if diff_var is not None: fc_n = field.get_cell_normals_per_facet(region) C = nm.abs(nm.einsum("ifk,ik->if", fc_n, advelo)) nbrhd_idx = field.get_facet_neighbor_idx(region, state.eq_map) active_cells, active_facets = nm.where(nbrhd_idx[:, :, 0] >= 0) active_nrbhs = nbrhd_idx[active_cells, active_facets, 0] in_fc_b, out_fc_b, whs = field.get_both_facet_base_vals(state, region) # TODO broadcast advelo to facets # - maybe somehow get values of advelo at them? # compute values inner_diff = nm.einsum("nfk, nfk->nf", fc_n, advelo[:, None, :] + nm.einsum("nfk, nf->nfk", (1 - self.alpha) * fc_n, C)) / 2. outer_diff = nm.einsum("nfk, nfk->nf", fc_n, advelo[:, None, :] - nm.einsum("nfk, nf->nfk", (1 - self.alpha) * fc_n, C)) / 2. inner_vals = nm.einsum("nf, ndfq, nbfq, nfq -> ndb", inner_diff, in_fc_b, in_fc_b, whs) outer_vals = nm.einsum("i, idq, ibq, iq -> idb", outer_diff[active_cells, active_facets], in_fc_b[active_cells, :, active_facets], out_fc_b[active_cells, :, active_facets], whs[active_cells, active_facets]) vals = nm.vstack((inner_vals, outer_vals)) vals = vals.flatten() # compute positions within matrix iels = self._get_nbrhd_dof_indexes(active_cells, active_nrbhs, field) out = (vals, iels[:, 0], iels[:, 1], state, state) else: fc_n = field.get_cell_normals_per_facet(region) # get maximal wave speeds at facets C = nm.abs(nm.einsum("ifk,ik->if", fc_n, advelo)) facet_base_vals = field.get_facet_base(base_only=True) in_fc_v, out_fc_v, weights = field.get_both_facet_state_vals(state, region) # get sane facet base shape fc_b = facet_base_vals[:, 0, :, 0, :].T # (n_el_nod, n_el_facet, n_qp) fc_v_avg = (in_fc_v + out_fc_v)/2. fc_v_jmp = in_fc_v - out_fc_v central = nm.einsum("ik,ifq->ifkq", advelo, fc_v_avg) upwind = (1 - self.alpha) / 2. * nm.einsum("if,ifk,ifq->ifkq", C, fc_n, fc_v_jmp) cell_fluxes = nm.einsum("ifk,ifkq,dfq,ifq->id", fc_n, central + upwind, fc_b, weights) out[:] = 0.0 n_el_nod = field.n_el_nod for i in range(n_el_nod): out[:, :, i, 0] = cell_fluxes[:, i, None] status = None return out, status
[docs] class DiffusionDGFluxTerm(DGTerm): r""" Basic DG diffusion flux term for scalar quantity. :Definition: .. math:: \int_{\partial{T_K}} D \langle \nabla p \rangle [q] \mbox{ , } \int_{\partial{T_K}} D \langle \nabla q \rangle [p] where .. math:: \langle \nabla \phi \rangle = \frac{\nabla\phi_{in} + \nabla\phi_{out}}{2} .. math:: [\phi] = \phi_{in} - \phi_{out} :math: The :math:`p_{in}` resp. :math:`p_{out}` is solution on the boundary of the element provided by element itself resp. its neighbour. :Arguments 1: - material : :math:`D` - state : :math:`p` - virtual : :math:`q` :Arguments 2: - material : :math:`D` - virtual : :math:`q` - state : :math:`p` """ name = "dw_dg_diffusion_flux" arg_types = (('material', 'state', 'virtual'), # left ('material', 'virtual', 'state') # right ) arg_shapes = [{'material': '1, 1', 'virtual/avg_state': (1, None), 'state/avg_state' : 1, 'virtual/avg_virtual': (1, None), 'state/avg_virtual' : 1, }] integration = 'cell' modes = ('avg_state', 'avg_virtual')
[docs] def get_fargs(self, diff_tensor, test, state, mode=None, term_mode=None, diff_var=None, **kwargs): field = state.field region = field.region if "DG" not in field.family_name: raise ValueError("Used DG term with non DG field {} of family {}" .format(field.name, field.family_name)) if self.mode == "avg_state": # put state where it is expected by the function state = test fargs = (state, diff_var, field, region, diff_tensor[:, 0, :, :]) return fargs
[docs] def function(self, out, state, diff_var, field, region, D): D = Term.tile_mat(D, out.shape[0]) if diff_var is not None: # matrix mode # outR = out.copy()[..., 0:1] out = self._function_matrix(out, state, diff_var, field, region, D) # vals, ielsi, ielsj = out[:3] # from scipy.sparse import coo_matrix # extra = coo_matrix((vals, (ielsi, ielsj)), # shape=2*(field.n_el_nod * field.n_cell,)) # M = extra.toarray() # u = state.data[0] # Mu = nm.dot(M, u).reshape((field.n_cell, field.n_el_nod)) # # outR = self._function_residual(outR, state, diff_var, field, # region, D).squeeze() # from matplotlib import pyplot as plt # plt.imshow((Mu - outR).T, aspect="auto") # plt.colorbar() # # nbrhd_idx = field.get_facet_neighbor_idx(region, state.eq_map) # bcells = nm.where(nbrhd_idx[:, :, 0] < 0)[0], # plt.vlines(bcells, -.5, 15, alpha=.3) # plt.show() else: # residual mode out = self._function_residual(out, state, diff_var, field, region, D) status = None return out, status
def _function_matrix(self, out, state, diff_var, field, region, D): fc_n = field.get_cell_normals_per_facet(region) nbrhd_idx = field.get_facet_neighbor_idx(region, state.eq_map) active_cells, active_facets = nm.where(nbrhd_idx[:, :, 0] >= 0) active_nrbhs = nbrhd_idx[active_cells, active_facets, 0] inner_facet_base, outer_facet_base, whs = \ field.get_both_facet_base_vals(state, region, derivative=False) inner_facet_base_d, outer_facet_base_d, _ = \ field.get_both_facet_base_vals(state, region, derivative=True) if self.mode == 'avg_state': # content of diagonal inner_vals = nm.einsum("nkl, nbfkq, nfk, ndfq, nfq->ndb", D, inner_facet_base_d / 2, # state fc_n, inner_facet_base, # test whs) outer_vals = nm.einsum( "ikl, ibkq, ik, idq, iq->idb", D[active_cells], outer_facet_base_d[active_cells, :, active_facets] / 2, # state fc_n[active_cells, active_facets], inner_facet_base[active_cells, :, active_facets], # test whs[active_cells, active_facets]) elif self.mode == 'avg_virtual': # content of diagonal inner_vals = nm.einsum("nkl, ndfkq, nfk, nbfq, nfq->ndb", D, inner_facet_base_d / 2, # test fc_n, inner_facet_base, # state whs) outer_vals = nm.einsum("ikl, idkq, ik, ibq, iq->idb", D[active_cells], inner_facet_base_d[active_cells, :, active_facets] / 2, # test fc_n[active_cells, active_facets], - outer_facet_base[active_cells, :, active_facets], # state whs[active_cells, active_facets]) iels = self._get_nbrhd_dof_indexes(active_cells, active_nrbhs, field) vals = nm.vstack((inner_vals, outer_vals)) vals = vals.flatten() # i j out = (vals, iels[:, 0], iels[:, 1], state, state) return out def _function_residual(self, out, state, diff_var, field, region, D): fc_n = field.get_cell_normals_per_facet(region) # get base values inner_facet_base, outer_facet_base, _ = \ field.get_both_facet_base_vals(state, region, derivative=False) inner_facet_base_d, outer_facet_base_d, _ = \ field.get_both_facet_base_vals(state, region, derivative=True) # get state values inner_facet_state_d, outer_facet_state_d, _ = \ field.get_both_facet_state_vals(state, region, derivative=True) inner_facet_state, outer_facet_state, weights = \ field.get_both_facet_state_vals(state, region, derivative=False) if self.mode == 'avg_state': avgDdState = (nm.einsum("ikl,ifkq->ifkq", D, inner_facet_state_d) + nm.einsum("ikl,ifkq ->ifkq", D, outer_facet_state_d)) / 2. # outer_facet_base is in DG zero - hence the jump is inner value jmpBase = inner_facet_base cell_fluxes = nm.einsum("ifkq , ifk, idfq, ifq -> id", avgDdState, fc_n, jmpBase, weights) elif self.mode == 'avg_virtual': avgDdbase = (nm.einsum("ikl,idfkq->idfkq", D, inner_facet_base_d)) / 2. jmpState = inner_facet_state - outer_facet_state cell_fluxes = nm.einsum("idfkq, ifk, ifq , ifq -> id", avgDdbase, fc_n, jmpState, weights) out[:] = 0.0 n_el_nod = field.n_el_nod for i in range(n_el_nod): out[:, :, i, 0] = cell_fluxes[:, i, None] return out
[docs] class DiffusionInteriorPenaltyTerm(DGTerm): r""" Penalty term used to counteract discontinuity arising when modeling diffusion using Discontinuous Galerkin schemes. :Definition: .. math:: \int_{\partial{T_K}} \bar{D} C_w \frac{Ord^2}{d(\partial{T_K})}[p][q] where .. math:: [\phi] = \phi_{in} - \phi_{out} :math: the :math:`p_{in}` resp. :math:`p_{out}` is solution on the boundary of the element provided by element itself resp. its neighbour. :Arguments: - material : :math:`D` - material : :math:`C_w` - state : :math:`p` - virtual : :math:`q` """ name = "dw_dg_interior_penalty" modes = ("weak",) arg_types = ('material', 'material_Cw', 'virtual', 'state') arg_shapes = [{'material': '1, 1', 'material_Cw': '.: 1', 'virtual' : (1, 'state'), 'state' : 1 }]
[docs] def get_fargs(self, diff_tensor, Cw, test, state, mode=None, term_mode=None, diff_var=None, **kwargs): field = state.field region = field.region if "DG" not in field.family_name: raise ValueError("Used DG term with non DG field {} of family {}" .format(field.name, field.family_name)) fargs = (state, diff_var, field, region, Cw, diff_tensor[:, 0, :, :]) return fargs
[docs] def function(self, out, state, diff_var, field, region, Cw, diff_tensor): approx_order = field.approx_order inner_facet_base, outer_facet_base, whs = \ field.get_both_facet_base_vals(state, region, derivative=False) facet_vols = nm.sum(whs, axis=-1) # nu characterizes diffusion tensor, so far we use diagonal average nu = nm.trace(diff_tensor, axis1=-2, axis2=-1)[..., None] / \ diff_tensor.shape[1] sigma = nu * Cw * approx_order ** 2 / facet_vols if diff_var is not None: nbrhd_idx = field.get_facet_neighbor_idx(region, state.eq_map) active_cells, active_facets = nm.where(nbrhd_idx[:, :, 0] >= 0) active_nrbhs = nbrhd_idx[active_cells, active_facets, 0] inner = nm.einsum("nf, ndfq, nbfq, nfq -> ndb", sigma, inner_facet_base, # test inner_facet_base, # state whs) outer = nm.einsum("i, idq, ibq, iq -> idb", sigma[active_cells, active_facets], inner_facet_base[active_cells, :, active_facets], # test - outer_facet_base[active_cells, :, active_facets], # state whs[active_cells, active_facets]) vals = nm.vstack((inner, outer)) vals = vals.flatten() iels = self._get_nbrhd_dof_indexes(active_cells, active_nrbhs, field) out = (vals, iels[:, 0], iels[:, 1], state, state) else: inner_facet_state, outer_facet_state, whs = \ field.get_both_facet_state_vals(state, region, derivative=False ) inner_facet_base, outer_facet_base, _ = \ field.get_both_facet_base_vals(state, region, derivative=False ) jmp_state = inner_facet_state - outer_facet_state jmp_base = inner_facet_base # - outer_facet_base n_el_nod = nm.shape(inner_facet_base)[1] cell_penalty = nm.einsum("nf,nfq,ndfq,nfq->nd", sigma, jmp_state, jmp_base, whs) out[:] = 0.0 for i in range(n_el_nod): out[:, :, i, 0] = cell_penalty[:, i, None] status = None return out, status
[docs] class NonlinearHyperbolicDGFluxTerm(DGTerm): r""" Lax-Friedrichs flux term for nonlinear hyperpolic term of scalar quantity :math:`p` with the vector function :math:`\ul{f}` given as a material parameter. :Definition: .. math:: \int_{\partial{T_K}} \ul{n} \cdot f^{*} (p_{in}, p_{out})q where .. math:: \ul{f}^{*}(p_{in}, p_{out}) = \frac{\ul{f}(p_{in}) + \ul{f}(p_{out})}{2} + (1 - \alpha) \ul{n} C \frac{ p_{in} - p_{out}}{2}, :math:`\alpha \in [0, 1]`; :math:`\alpha = 0` for upwind scheme, :math:`\alpha = 1` for central scheme, and .. math:: C = \max_{p \in [?, ?]}\left\lvert n_x \frac{d f_1}{d p} + n_y \frac{d f_2}{d p} + \cdots \right\rvert = \max_{p \in [?, ?]} \left\lvert \vec{n}\cdot\frac{d\ul{f}}{dp}(p) \right\rvert the :math:`p_{in}` resp. :math:`p_{out}` is solution on the boundary of the element provided by element itself resp. its neighbor. :Arguments 1: - material : :math:`\ul{f}` - material : :math:`\frac{d\ul{f}}{d p}` - virtual : :math:`q` - state : :math:`p` :Arguments 3: - material : :math:`\ul{f}` - material : :math:`\frac{d\ul{f}}{d p}` - virtual : :math:`q` - state : :math:`p` - opt_material : :math:`\alpha` """ alf = 0 name = "dw_dg_nonlinear_laxfrie_flux" modes = ("weak",) arg_types = ('opt_material', 'fun', 'fun_d', 'virtual', 'state') arg_shapes = [{'opt_material' : '.: 1', 'material_fun' : '.: 1', 'material_fun_d': '.: 1', 'virtual' : (1, 'state'), 'state' : 1 }, {'opt_material': None}] integration = 'cell' symbolic = {'expression': 'div(f(p))*w', 'map' : {'p': 'state', 'v': 'virtual', 'f': 'function'} }
[docs] def get_fargs(self, alpha, fun, dfun, test, state, mode=None, term_mode=None, diff_var=None, **kwargs): if alpha is not None: self.alf = nm.max(alpha) # extract alpha value regardless of shape self.fun = fun self.dfun = dfun if diff_var is not None: output("Diff var is not None in nonlinear, residual only " + "term"" {} ! Skipping.".format(self.name)) return None, None, None, 0, 0 else: field = state.field region = field.region if "DG" not in field.family_name: raise ValueError( "Used DG term with non DG field {} of family {}!" .format(field.name, field.family_name)) fargs = (state, field, region, fun, dfun) return fargs
[docs] def function(self, out, state, field, region, f, df): if state is None: out[:] = 0.0 return None fc_n = field.get_cell_normals_per_facet(region) facet_base_vals = field.get_facet_base(base_only=True) in_fc_v, out_fc_v, weights = field.get_both_facet_state_vals(state, region) fc_b = facet_base_vals[:, 0, :, 0, :].T # (n_el_nod, n_el_facet, n_qp) n_el_nod = field.n_el_nod # get maximal wave speeds at facets df_in = df(in_fc_v) df_out = df(out_fc_v) fc_n__dot__df_in = nm.einsum("ifk,ifqk->ifq", fc_n, df_in) fc_n__dot__df_out = nm.einsum("ifk,ifqk->ifq", fc_n, df_out) dfdn = nm.stack((fc_n__dot__df_in, fc_n__dot__df_out), axis=-1) C = nm.amax(nm.abs(dfdn), axis=(-2, -1)) fc_f_avg = (f(in_fc_v) + f(out_fc_v)) / 2. fc_v_jmp = in_fc_v - out_fc_v central = fc_f_avg upwind = (1 - self.alf) / 2. * nm.einsum("if,ifk,ifq->ifqk", C, fc_n, fc_v_jmp) cell_fluxes = nm.einsum("ifk,ifqk,dfq,ifq->id", fc_n, central + upwind, fc_b, weights) out[:] = 0.0 for i in range(n_el_nod): out[:, :, i, 0] = cell_fluxes[:, i, None] status = None return out, status
from sfepy.linalg import dot_sequences
[docs] class NonlinearScalarDotGradTerm(Term): r""" Product of virtual and divergence of vector function of state or volume dot product of vector function of state and gradient of scalar virtual. :Definition: .. math:: \int_{\Omega} q \cdot \nabla \cdot \ul{f}(p) = \int_{\Omega} q \cdot \text{div} \ul{f}(p) \mbox{ , } \int_{\Omega} \ul{f}(p) \cdot \nabla q :Arguments 1: - function : :math:`\ul{f}` - virtual : :math:`q` - state : :math:`p` :Arguments 2: - function : :math:`\ul{f}` - state : :math:`p` - virtual : :math:`q` TODO maybe this term would fit better to terms_dot? """ name = 'dw_ns_dot_grad_s' arg_types = (('fun', 'fun_d', 'virtual', 'state'), ('fun', 'fun_d', 'state', 'virtual')) arg_shapes = [{'material_fun' : '.: 1', 'material_fun_d' : '.: 1', 'virtual/grad_state' : (1, None), 'state/grad_state' : 1, 'virtual/grad_virtual': (1, None), 'state/grad_virtual' : 1}] modes = ('grad_state', 'grad_virtual')
[docs] @staticmethod def function(out, out_qp, geo, fmode): status = geo.integrate(out, out_qp) return status
[docs] def get_fargs(self, fun, dfun, var1, var2, mode=None, term_mode=None, diff_var=None, **kwargs): vg1, _ = self.get_mapping(var1) vg2, _ = self.get_mapping(var2) if diff_var is None: if self.mode == 'grad_state': # TODO rewrite using einsum? geo = vg1 bf_t = vg1.bf.transpose((0, 1, 3, 2)) val_grad_qp = self.get(var2, 'grad') out_qp = dot_sequences(bf_t, val_grad_qp, 'ATB') else: geo = vg2 val_qp = fun(self.get(var1, 'val'))[..., 0, :].swapaxes(-2, -1) out_qp = dot_sequences(vg2.bfg, val_qp, 'ATB') fmode = 0 else: raise ValueError("Matrix mode not supported for {}" .format(self.name)) # however it could be with use of dfun if self.mode == 'grad_state': geo = vg1 bf_t = vg1.bf.transpose((0, 1, 3, 2)) out_qp = dot_sequences(bf_t, vg2.bfg, 'ATB') else: geo = vg2 out_qp = dot_sequences(vg2.bfg, vg1.bf, 'ATB') fmode = 1 return out_qp, geo, fmode