def setup_solver(self):
""" Sometimes it is necessary to set up the solver again after breaking
important references, e.g. after re-meshing.
"""
self._governing_form = self.governing_form()
self._boundary_conditions = self.boundary_conditions()
self._problem = fenics.NonlinearVariationalProblem(
F=self._governing_form,
u=self.solution,
bcs=self._boundary_conditions,
J=fenics.derivative(form=self._governing_form, u=self.solution))
save_parameters = False
if hasattr(self, "solver"):
save_parameters = True
if save_parameters:
solver_parameters = self.solver.parameters.copy()
self.solver = fenics.NonlinearVariationalSolver(problem=self._problem)
if save_parameters:
self.solver.parameters = solver_parameters.copy()
self._adaptive_goal = self.adaptive_goal()
if self._adaptive_goal is not None:
save_parameters = False
if self.adaptive_solver is not None:
save_parameters = True
if save_parameters:
adaptive_solver_parameters = self.adaptive_solver.parameters.copy(
)
self.adaptive_solver = fenics.AdaptiveNonlinearVariationalSolver(
problem=self._problem, goal=self._adaptive_goal)
if save_parameters:
self.adaptive_solver.parameters = adaptive_solver_parameters.copy(
)
self.solver_needs_setup = False
def setup_solver(self):
""" Set up the solver, which is a `fenics.AdaptiveNonlinearVariationalSolver`. """
self.solver = fenics.AdaptiveNonlinearVariationalSolver(
problem=self.problem, goal=self.adaptive_goal_form)
self.solver.parameters["nonlinear_variational_solver"]["newton_solver"]\
["maximum_iterations"] = self.nonlinear_solver_max_iterations
self.solver.parameters["nonlinear_variational_solver"]["newton_solver"]\
["absolute_tolerance"] = self.nonlinear_solver_absolute_tolerance
self.solver.parameters["nonlinear_variational_solver"]["newton_solver"]\
["relative_tolerance"] = self.nonlinear_solver_relative_tolerance
self.solver.parameters["nonlinear_variational_solver"]["newton_solver"]\
["relaxation_parameter"] = self.nonlinear_solver_relaxation
mass = -psi_p * div(u)
F = (momentum + mass) * fenics.dx
JF = fenics.derivative(F, w, fenics.TrialFunction(W))
lid_velocity = (1., 0.)
lid_location = "near(x[1], 1.)"
fixed_wall_velocity = (0., 0.)
fixed_wall_locations = "near(x[0], 0.) | near(x[0], 1.) | near(x[1], 0.)"
V = W.sub(0)
boundary_conditions = [
fenics.DirichletBC(V, lid_velocity, lid_location),
fenics.DirichletBC(V, fixed_wall_velocity, fixed_wall_locations)
]
problem = fenics.NonlinearVariationalProblem(F, w, boundary_conditions, JF)
M = u[0]**2 * fenics.dx
epsilon_M = 1.e-4
solver = fenics.AdaptiveNonlinearVariationalSolver(problem, M)
solver.solve(epsilon_M)
def run(output_dir="output/wang2010_natural_convection_air",
rayleigh_number=1.e6,
prandtl_number=0.71,
stefan_number=0.045,
heat_capacity=1.,
thermal_conductivity=1.,
liquid_viscosity=1.,
solid_viscosity=1.e8,
gravity=(0., -1.),
m_B=None,
ddT_m_B=None,
penalty_parameter=1.e-7,
temperature_of_fusion=-1.e12,
regularization_smoothing_factor=0.005,
mesh=fenics.UnitSquareMesh(fenics.dolfin.mpi_comm_world(), 20, 20,
"crossed"),
initial_values_expression=("0.", "0.", "0.",
"0.5*near(x[0], 0.) -0.5*near(x[0], 1.)"),
boundary_conditions=[{
"subspace": 0,
"value_expression": ("0.", "0."),
"degree": 3,
"location_expression":
"near(x[0], 0.) | near(x[0], 1.) | near(x[1], 0.) | near(x[1], 1.)",
"method": "topological"
}, {
"subspace": 2,
"value_expression": "0.5",
"degree": 2,
"location_expression": "near(x[0], 0.)",
"method": "topological"
}, {
"subspace": 2,
"value_expression": "-0.5",
"degree": 2,
"location_expression": "near(x[0], 1.)",
"method": "topological"
}],
start_time=0.,
end_time=10.,
time_step_size=1.e-3,
stop_when_steady=True,
steady_relative_tolerance=1.e-4,
adaptive=False,
adaptive_metric="all",
adaptive_solver_tolerance=1.e-4,
nlp_absolute_tolerance=1.e-8,
nlp_relative_tolerance=1.e-8,
nlp_max_iterations=50,
restart=False,
restart_filepath=""):
"""Run Phaseflow.
Phaseflow is configured entirely through the arguments in this run() function.
See the tests and examples for demonstrations of how to use this.
"""
# Handle default function definitions.
if m_B is None:
def m_B(T, Ra, Pr, Re):
return T * Ra / (Pr * Re**2)
if ddT_m_B is None:
def ddT_m_B(T, Ra, Pr, Re):
return Ra / (Pr * Re**2)
# Report arguments.
phaseflow.helpers.print_once(
"Running Phaseflow with the following arguments:")
phaseflow.helpers.print_once(phaseflow.helpers.arguments())
phaseflow.helpers.mkdir_p(output_dir)
if fenics.MPI.rank(fenics.mpi_comm_world()) is 0:
arguments_file = open(output_dir + "/arguments.txt", "w")
arguments_file.write(str(phaseflow.helpers.arguments()))
arguments_file.close()
# Check if 1D/2D/3D.
dimensionality = mesh.type().dim()
phaseflow.helpers.print_once("Running " + str(dimensionality) +
"D problem")
# Initialize time.
if restart:
with h5py.File(restart_filepath, "r") as h5:
time = h5["t"].value
assert (abs(time - start_time) < TIME_EPS)
else:
time = start_time
# Define the mixed finite element and the solution function space.
W_ele = make_mixed_fe(mesh.ufl_cell())
W = fenics.FunctionSpace(mesh, W_ele)
# Set the initial values.
if restart:
mesh = fenics.Mesh()
with fenics.HDF5File(mesh.mpi_comm(), restart_filepath, "r") as h5:
h5.read(mesh, "mesh", True)
W_ele = make_mixed_fe(mesh.ufl_cell())
W = fenics.FunctionSpace(mesh, W_ele)
w_n = fenics.Function(W)
with fenics.HDF5File(mesh.mpi_comm(), restart_filepath, "r") as h5:
h5.read(w_n, "w")
else:
w_n = fenics.interpolate(
fenics.Expression(initial_values_expression, element=W_ele), W)
# Organize the boundary conditions.
bcs = []
for item in boundary_conditions:
bcs.append(
fenics.DirichletBC(W.sub(item["subspace"]),
item["value_expression"],
item["location_expression"],
method=item["method"]))
# Set the variational form.
"""Set local names for math operators to improve readability."""
inner, dot, grad, div, sym = fenics.inner, fenics.dot, fenics.grad, fenics.div, fenics.sym
"""The linear, bilinear, and trilinear forms b, a, and c, follow the common notation
for applying the finite element method to the incompressible Navier-Stokes equations,
e.g. from danaila2014newton and huerta2003fefluids.
"""
def b(u, q):
return -div(u) * q # Divergence
def D(u):
return sym(grad(u)) # Symmetric part of velocity gradient
def a(mu, u, v):
return 2. * mu * inner(D(u), D(v)) # Stokes stress-strain
def c(w, z, v):
return dot(dot(grad(z), w), v) # Convection of the velocity field
dt = fenics.Constant(time_step_size)
Re = fenics.Constant(reynolds_number)
Ra = fenics.Constant(rayleigh_number)
Pr = fenics.Constant(prandtl_number)
Ste = fenics.Constant(stefan_number)
C = fenics.Constant(heat_capacity)
K = fenics.Constant(thermal_conductivity)
g = fenics.Constant(gravity)
def f_B(T):
return m_B(T=T, Ra=Ra, Pr=Pr, Re=Re) * g # Buoyancy force, $f = ma$
gamma = fenics.Constant(penalty_parameter)
T_f = fenics.Constant(temperature_of_fusion)
r = fenics.Constant(regularization_smoothing_factor)
def P(T):
return 0.5 * (1. - fenics.tanh(
(T_f - T) / r)) # Regularized phase field.
mu_l = fenics.Constant(liquid_viscosity)
mu_s = fenics.Constant(solid_viscosity)
def mu(T):
return mu_s + (mu_l - mu_s) * P(T) # Variable viscosity.
L = C / Ste # Latent heat
u_n, p_n, T_n = fenics.split(w_n)
w_w = fenics.TrialFunction(W)
u_w, p_w, T_w = fenics.split(w_w)
v, q, phi = fenics.TestFunctions(W)
w_k = fenics.Function(W)
u_k, p_k, T_k = fenics.split(w_k)
F = (b(u_k, q) - gamma * p_k * q + dot(u_k - u_n, v) / dt + c(u_k, u_k, v)
+ b(v, p_k) + a(mu(T_k), u_k, v) + dot(f_B(T_k), v) + C / dt *
(T_k - T_n) * phi - dot(C * T_k * u_k, grad(phi)) +
K / Pr * dot(grad(T_k), grad(phi)) + 1. / dt * L *
(P(T_k) - P(T_n)) * phi) * fenics.dx
def ddT_f_B(T):
return ddT_m_B(T=T, Ra=Ra, Pr=Pr, Re=Re) * g
def sech(theta):
return 1. / fenics.cosh(theta)
def dP(T):
return sech((T_f - T) / r)**2 / (2. * r)
def dmu(T):
return (mu_l - mu_s) * dP(T)
# Set the Jacobian (formally the Gateaux derivative) in variational form.
JF = (b(u_w, q) - gamma * p_w * q + dot(u_w, v) / dt + c(u_k, u_w, v) +
c(u_w, u_k, v) + b(v, p_w) + a(T_w * dmu(T_k), u_k, v) +
a(mu(T_k), u_w, v) + dot(T_w * ddT_f_B(T_k), v) +
C / dt * T_w * phi - dot(C * T_k * u_w, grad(phi)) -
dot(C * T_w * u_k, grad(phi)) + K / Pr * dot(grad(T_w), grad(phi)) +
1. / dt * L * T_w * dP(T_k) * phi) * fenics.dx
# Set the functional metric for the error estimator for adaptive mesh refinement.
"""I haven't found a good way to make this flexible yet.
Ideally the user would be able to write the metric, but this would require giving the user
access to much data that phaseflow is currently hiding.
"""
M = P(T_k) * fenics.dx
if adaptive_metric == "phase_only":
pass
elif adaptive_metric == "all":
M += T_k * fenics.dx
for i in range(dimensionality):
M += u_k[i] * fenics.dx
else:
assert (False)
# Make the problem.
problem = fenics.NonlinearVariationalProblem(F, w_k, bcs, JF)
# Make the solvers.
""" For the purposes of this project, it would be better to just always use the adaptive solver; but
unfortunately the adaptive solver encounters nan's whenever evaluating the error for problems not
involving phase-change. So far my attempts at writing a MWE to reproduce the issue have failed.
"""
adaptive_solver = fenics.AdaptiveNonlinearVariationalSolver(problem, M)
adaptive_solver.parameters["nonlinear_variational_solver"]["newton_solver"]["maximum_iterations"]\
= nlp_max_iterations
adaptive_solver.parameters["nonlinear_variational_solver"]["newton_solver"]["absolute_tolerance"]\
= nlp_absolute_tolerance
adaptive_solver.parameters["nonlinear_variational_solver"]["newton_solver"]["relative_tolerance"]\
= nlp_relative_tolerance
static_solver = fenics.NonlinearVariationalSolver(problem)
static_solver.parameters["newton_solver"][
"maximum_iterations"] = nlp_max_iterations
static_solver.parameters["newton_solver"][
"absolute_tolerance"] = nlp_absolute_tolerance
static_solver.parameters["newton_solver"][
"relative_tolerance"] = nlp_relative_tolerance
# Open a context manager for the output file.
with fenics.XDMFFile(output_dir + "/solution.xdmf") as solution_file:
# Write the initial values.
write_solution(solution_file, w_n, time)
if start_time >= end_time - TIME_EPS:
phaseflow.helpers.print_once(
"Start time is already too close to end time. Only writing initial values."
)
return w_n, mesh
# Solve each time step.
progress = fenics.Progress("Time-stepping")
fenics.set_log_level(fenics.PROGRESS)
for it in range(1, MAX_TIME_STEPS):
if (time > end_time - TIME_EPS):
break
if adaptive:
adaptive_solver.solve(adaptive_solver_tolerance)
else:
static_solver.solve()
time = start_time + it * time_step_size
phaseflow.helpers.print_once("Reached time t = " + str(time))
write_solution(solution_file, w_k, time)
# Write checkpoint/restart files.
restart_filepath = output_dir + "/restart_t" + str(time) + ".h5"
with fenics.HDF5File(fenics.mpi_comm_world(), restart_filepath,
"w") as h5:
h5.write(mesh.leaf_node(), "mesh")
h5.write(w_k.leaf_node(), "w")
if fenics.MPI.rank(fenics.mpi_comm_world()) is 0:
with h5py.File(restart_filepath, "r+") as h5:
h5.create_dataset("t", data=time)
# Check for steady state.
if stop_when_steady and steady(W, w_k, w_n,
steady_relative_tolerance):
phaseflow.helpers.print_once(
"Reached steady state at time t = " + str(time))
break
# Set initial values for next time step.
w_n.leaf_node().vector()[:] = w_k.leaf_node().vector()
# Report progress.
progress.update(time / end_time)
if time >= (end_time - fenics.dolfin.DOLFIN_EPS):
phaseflow.helpers.print_once("Reached end time, t = " +
str(end_time))
break
# Return the interpolant to sample inside of Python.
w_k.rename("w", "state")
return w_k, mesh