class SaveVTK():
def __init__(self, mesh, name='unnamed'):
self.name = name
# Initiate a VTK object
self.VTK = VTK(mesh,
directory='{}_vtks'.format(name),
filename=name
)
def save_vtk(self, m1, Ms, step=0, vtkname='m'):
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(Ms, name='Ms')
self.VTK.save_vector(m1, name='spins')
# Now we set the name of the file to start with m_ as default
self.VTK.filename = vtkname
self.VTK.write_file(step=step)
def save_vtk_scalar(self, scalar_data_array, step=0, vtkname='skx'):
# This saves the skyrmion number or any scalar data
# CHECK: that the array skx_num is being passed with the correct shape
# TODO: put more generic names as default names
self.VTK.save_scalar(scalar_data_array, name='skx_num')
# These files will be saved starting with: skx_ or other name
# and saved to a different folder instead of {}_vtks
self.VTK.filename = vtkname
self.VTK.directory = self.name + '_' + vtkname
self.VTK.write_file(step=step)
def __init__(self, mesh, spin, magnetisation, magnetisation_inv, field,
pins, interactions, name, data_saver):
# ---------------------------------------------------------------------
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
# A reference to either mu_s or Ms to use a common lib for this
# minimiser
self._magnetisation = magnetisation
self._magnetisation_inv = magnetisation_inv
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
self.data_saver = data_saver
# ---------------------------------------------------------------------
# Variables defined in this class
self.spin_last = np.ones_like(self.spin)
self.n = self.mesh.n
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
self.scale = 1.
class SaveVTK():
def __init__(self, mesh, name='unnamed'):
self.name = name
# Initiate a VTK object
self.VTK = VTK(mesh, directory='{}_vtks'.format(name), filename=name)
def save_vtk(self, m1, Ms, step=0, vtkname='m'):
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(Ms, name='Ms')
self.VTK.save_vector(m1, name='spins')
# Now we set the name of the file to start with m_ as default
self.VTK.filename = vtkname
self.VTK.write_file(step=step)
def save_vtk_scalar(self, scalar_data_array, step=0, vtkname='skx'):
# This saves the skyrmion number or any scalar data
# CHECK: that the array skx_num is being passed with the correct shape
# TODO: put more generic names as default names
self.VTK.save_scalar(scalar_data_array, name='skx_num')
# These files will be saved starting with: skx_ or other name
# and saved to a different folder instead of {}_vtks
self.VTK.filename = vtkname
self.VTK.directory = self.name + '_' + vtkname
self.VTK.write_file(step=step)
def __init__(self, mesh, name='unnamed', directory=None):
self.name = name
if directory is None:
self.directory = '{}_vtks'.format(name)
else:
self.directory = directory
# Initiate a VTK object
self.VTK = VTK(mesh, directory=self.directory, filename=name)
def __init__(self, mesh, name='unnamed'):
self.name = name
# Initiate a VTK object
self.VTK = VTK(mesh,
directory='{}_vtks'.format(name),
filename=name
)
def __init__(self, mesh, spin, mu_s, mu_s_inv, field, pins, interactions, name, data_saver, use_jac):
super(AtomisticDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._mu_s = mu_s
self._mu_s_inv = mu_s_inv
# Only for LLG STT: (??)
self.mu_s_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
if use_jac is not True:
self.integrator = cvode.CvodeSolver(self.spin, self.sundials_rhs)
else:
self.integrator = cvode.CvodeSolver(self.spin, self.sundials_rhs, self.sundials_jtn)
self.set_tols()
# When initialising the integrator in the self.integrator call, the
# CVOde class calls the set_initial_value function (with flag_m=0),
# which initialises a new integrator and allocates memory in this
# process. Now, when we set the magnetisation, we will use the same
# memory setting this flag_m to 1, so instead of calling CVodeInit we
# call CVodeReInit. If don't, memory is allocated in every call of
# set_m
self.flag_m = 1
# Factor for the dmdt magnitude in the relaxation function
self._dmdt_factor = 1.0
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh, directory="{}_vtks".format(self.name), filename="m")
self.data_saver = data_saver
def __init__(self, mesh, name='unnamed', directory=None):
self.name = name
if directory is None:
self.directory = '{}_vtks'.format(name)
else:
self.directory = directory
# Initiate a VTK object
self.VTK = VTK(mesh,
directory=self.directory,
filename=name
)
def __init__(self, sim,
initial_images, interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False
):
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# Spring constant (we could use an array in the future)
self.k = spring_constant
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='m'
)
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the NEBM band
self.n_band = self.n_images * self.n_dofs_image
# Climbing Image ------------------------------------------------------
if climbing_image is None:
self.climbing_image = -1
elif climbing_image in range(self.n_images):
self.climbing_image = climbing_image
else:
raise ValueError('The climbing image must be in the band. '
'Specify a valid integer.')
# NEBM Arrays ---------------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
self.last_Y = np.zeros_like(self.band)
class AtomisticDriver(DriverBase):
"""
A class with shared methods and properties for different drivers to solve
the Landau-Lifshitz-Gilbert equation
Variables that are proper of the driver class:
* alpha (damping)
* mu_s_const
* t
* spin_last
* dm_dt
* integrator_tolerances_set
* step
* n (from mesh)
* n_nonzero (from mesh and set_Ms method in Simulation)
* gamma
* do_precession
* default_c (correction factor to keep the magnetisation normalised
during the LLG equation integration. See the
fidimag/atomistic/lib/llg.c file for more details)
From DriverBase: t, spin_last, dm_dt, integrator_tolerances_set, step
"""
def __init__(self, mesh, spin, mu_s, mu_s_inv, field, pins, interactions, name, data_saver, use_jac):
super(AtomisticDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._mu_s = mu_s
self._mu_s_inv = mu_s_inv
# Only for LLG STT: (??)
self.mu_s_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
if use_jac is not True:
self.integrator = cvode.CvodeSolver(self.spin, self.sundials_rhs)
else:
self.integrator = cvode.CvodeSolver(self.spin, self.sundials_rhs, self.sundials_jtn)
self.set_tols()
# When initialising the integrator in the self.integrator call, the
# CVOde class calls the set_initial_value function (with flag_m=0),
# which initialises a new integrator and allocates memory in this
# process. Now, when we set the magnetisation, we will use the same
# memory setting this flag_m to 1, so instead of calling CVodeInit we
# call CVodeReInit. If don't, memory is allocated in every call of
# set_m
self.flag_m = 1
# Factor for the dmdt magnitude in the relaxation function
self._dmdt_factor = 1.0
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh, directory="{}_vtks".format(self.name), filename="m")
self.data_saver = data_saver
# Initialise the table for the data file with the simulation
# information:
# self.saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
# self.saver.update_entity_order()
# ---------------------------------------------------------------------
def set_default_options(self, gamma=1, mu_s=1, alpha=0.1):
"""
Default option for the integrator
* The value of gamma (gyromagnetic ratio) for a free electron
is 1.76e11
"""
self.default_c = -1
self._alpha[:] = alpha
# When we create the simulation, mu_s is set to the default value. This
# is overriden when calling the set_mu_s method from the Simulation
# class or when setting mu_s directly (property)
self._mu_s[:] = mu_s
self.mu_s_const = mu_s
self.gamma = gamma
self.do_precession = True
# Don't know the uselfulness of this function:
# def set_options(self, rtol=1e-8, atol=1e-10):
# self.set_tols(rtol, atol)
def sundials_rhs(self, t, y, ydot):
"""
Defined in the corresponding driver class
"""
pass
def sundials_jtn(self, mp, Jmp, t, m, fy):
# we can not copy mp to self.spin since m and self.spin is one object.
# self.spin[:] = mp[:]
print("NO jac...........")
self.compute_effective_field_jac(t, mp)
clib.compute_llg_jtimes(
Jmp, m, fy, mp, self.field, self.alpha, self._pins, self.gamma, self.n, self.do_precession, self.default_c
)
return 0
# -------------------------------------------------------------------------
# Save functions ----------------------------------------------------------
# -------------------------------------------------------------------------
def save_vtk(self):
"""
Save a VTK file with the magnetisation vector field and magnetic
moments as cell data. Magnetic moments are saved in units of
Bohr magnetons
NOTE: It is recommended to use a *cell to point data* filter in
Paraview or Mayavi to plot the vector field
"""
self.VTK.reset_data()
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(self._mu_s / const.mu_B, name="mu_s")
self.VTK.save_vector(self.spin.reshape(-1, 3), name="spins")
self.VTK.write_file(step=self.step)
def test_save_vector_field(tmpdir):
mesh = CuboidMesh(4, 3, 2, 4, 3, 2)
s = vector_field(mesh, lambda r: (r[0], r[1], r[2]))
vtk = VTK(mesh, directory=str(tmpdir), filename="save_vector")
vtk.save_vector(s, name="s")
assert same_as_ref(vtk.write_file(), REF_DIR)
def test_save_scalar_field():
mesh = CuboidMesh(4, 3, 2, 4, 3, 2)
s = scalar_field(mesh, lambda r: r[0] + r[1] + r[2])
vtk = VTK(mesh, directory=OUTPUT_DIR, filename="save_scalar")
vtk.save_scalar(s, name="s")
class AtomisticDriver(DriverBase):
"""
A class with shared methods and properties for different drivers to solve
the Landau-Lifshitz-Gilbert equation
Variables that are proper of the driver class:
* alpha (damping)
* mu_s_const
* t
* spin_last
* dm_dt
* integrator_tolerances_set
* step
* n (from mesh)
* n_nonzero (from mesh and set_Ms method in Simulation)
* gamma
* do_precession
* default_c (correction factor to keep the magnetisation normalised
during the LLG equation integration. See the
fidimag/atomistic/lib/llg.c file for more details)
From DriverBase: t, spin_last, dm_dt, integrator_tolerances_set, step
"""
def __init__(self,
mesh,
spin,
mu_s,
mu_s_inv,
field,
pins,
interactions,
name,
data_saver,
use_jac,
integrator='sundials'):
super(AtomisticDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._mu_s = mu_s
self._mu_s_inv = mu_s_inv
# Only for LLG STT: (??)
self.mu_s_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
# In the old code, it seemed that self.sundials_jtn was not defined
# anywhere else. Here, we will use the same integrators than in the
# micromagnetic code, where we have self.sundials_jtimes instead of jtn
self.set_integrator(integrator, use_jac)
self.set_tols()
# When initialising the integrator in the self.integrator call, the
# CVOde class calls the set_initial_value function (with flag_m=0),
# which initialises a new integrator and allocates memory in this
# process. Now, when we set the magnetisation, we will use the same
# memory setting this flag_m to 1, so instead of calling CVodeInit we
# call CVodeReInit. If don't, memory is allocated in every call of
# set_m
self.flag_m = 1
# Factor for the dmdt magnitude in the relaxation function
self._dmdt_factor = 1.
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
self.data_saver = data_saver
# Initialise the table for the data file with the simulation
# information:
# self.saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
# self.saver.update_entity_order()
# ---------------------------------------------------------------------
def set_default_options(self, gamma=1, mu_s=1, alpha=0.1):
"""
Default option for the integrator
* The value of gamma (gyromagnetic ratio) for a free electron
is 1.76e11
"""
self.default_c = -1
self._alpha[:] = alpha
# When we create the simulation, mu_s is set to the default value. This
# is overriden when calling the set_mu_s method from the Simulation
# class or when setting mu_s directly (property)
# David Tue 19 Jun 2018: Not a very clear thing to do, we must set a
# WARNING
self._mu_s[:] = mu_s
self._mu_s_inv[:] = 1. / mu_s
self.mu_s_const = mu_s
self.gamma = gamma
self.do_precession = True
# Don't know the uselfulness of this function:
# def set_options(self, rtol=1e-8, atol=1e-10):
# self.set_tols(rtol, atol)
def sundials_rhs(self, t, y, ydot):
"""
Defined in the corresponding driver class
"""
pass
def sundials_jtn(self, mp, Jmp, t, m, fy):
# we can not copy mp to self.spin since m and self.spin is one object.
#self.spin[:] = mp[:]
print('NO jac...........')
self.compute_effective_field_jac(t, mp)
clib.compute_llg_jtimes(Jmp, m, fy, mp, self.field, self.alpha,
self._pins, self.gamma, self.n,
self.do_precession, self.default_c)
return 0
# -------------------------------------------------------------------------
# Save functions ----------------------------------------------------------
# -------------------------------------------------------------------------
def save_vtk(self):
"""
Save a VTK file with the magnetisation vector field and magnetic
moments as cell data. Magnetic moments are saved in units of
Bohr magnetons
NOTE: It is recommended to use a *cell to point data* filter in
Paraview or Mayavi to plot the vector field
"""
self.VTK.reset_data()
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(self._mu_s / const.mu_B, name='mu_s')
self.VTK.save_vector(self.spin.reshape(-1, 3), name='spins')
self.VTK.write_file(step=self.step)
class MicroDriver(DriverBase):
"""
A class with shared methods and properties for different drivers to solve
the Landau-Lifshitz-Gilbert equation
Variables that are proper of the driver class:
* alpha (damping)
* Ms_const
* t
* spin_last
* dm_dt
* integrator_tolerances_set
* step
* n (from mesh)
* n_nonzero (from mesh and set_Ms method in Simulation)
* gamma
* do_precession
* default_c (correction factor to keep the magnetisation normalised
during the LLG equation integration. See the
fidimag/atomistic/lib/llg.c file for more details)
TODO: Check default_c units in the micromagnetic
context
From DriverBase: t, spin_last, dm_dt, integrator_tolerances_set, step
"""
def __init__(self, mesh, spin, Ms, field, pins,
interactions,
name,
data_saver,
integrator='sundials',
use_jac=False
):
super(MicroDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._Ms = Ms
# Only for LLG STT: (??)
self.Ms_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
# Here we set up the CVODE integrator from Sundials to evolve a
# specific micromagnetic equation. The equations are specified in the
# sundials_rhs function from any of the micromagnetic drivers in the
# micromagnetic folder (LLG, LLG_STT, etc.)
if integrator == "sundials" and use_jac:
self.integrator = CvodeSolver(self.spin, self.sundials_rhs,
self.sundials_jtimes)
elif integrator == "sundials_diag":
self.integrator = CvodeSolver(self.spin, self.sundials_rhs,
linear_solver="diag")
elif integrator == "sundials":
self.integrator = CvodeSolver(self.spin, self.sundials_rhs)
elif integrator == "euler" or integrator == "rk4":
self.integrator = CvodeSolver(self.spin, self.step_rhs,
integrator)
elif integrator == "sundials_openmp" and use_jac:
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs,
self.sundials_jtimes)
elif integrator == "sundials_diag_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs,
linear_solver="diag")
elif integrator == "sundials_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs)
elif integrator == "euler_openmp" or integrator == "rk4_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.step_rhs,
integrator)
else:
raise NotImplemented("integrator must be sundials, euler or rk4")
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m'
)
# Initialise the table for the data file with the simulation
# information:
self.data_saver = data_saver
# This should not be necessary:
# self.data_saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
self.data_saver.entities['rhs_evals'] = {
'unit': '<>',
'get': lambda sim: self.integrator.rhs_evals(),
'header': 'rhs_evals'}
self.data_saver.entities['real_time'] = {
'unit': '<s>',
'get': lambda _: time.time(), # seconds since epoch
'header': 'real_time'}
self.data_saver.update_entity_order()
# ---------------------------------------------------------------------
# OOMMF convention is to check if the spins have moved by ~1 degree in
# a nanosecond in order to stop a simulation, so we set this scale for
# dm/dt
# ONE_DEGREE_PER_NANOSECOND:
self._dmdt_factor = (2 * np.pi / 360) / 1e-9
def set_default_options(self, gamma=2.21e5, Ms=8.0e5, alpha=0.1):
"""
Default option for the integrator
Default gamma is for a free electron
"""
self.default_c = 1e11
self._alpha[:] = alpha
# When we create the simulation, Ms is set to the default value. This
# is overriden when calling the set_Ms method from the Siulation class
# or when setting Ms directly (property)
self._Ms[:] = Ms
self.gamma = gamma
self.do_precession = True
def sundials_rhs(self, t, y, ydot):
"""
Defined in the corresponding driver class
"""
pass
def set_tols(self, rtol=1e-8, atol=1e-10, max_ord=None, reset=True):
"""
Set the relative and absolute tolerances for the CVODE integrator
"""
if max_ord is not None:
self.integrator.set_options(rtol=rtol, atol=atol, max_ord=max_ord)
else:
# not all integrators have max_ord (only VODE does)
# and we don't want to encode a default value here either
self.integrator.set_options(rtol=rtol, atol=atol)
if reset:
self.reset_integrator(self.t)
# -------------------------------------------------------------------------
# Save functions ----------------------------------------------------------
# -------------------------------------------------------------------------
def save_vtk(self):
"""
Save a VTK file with the magnetisation vector field (vector data)
and the saturation magnetisation values (scalar data) as
cell data
"""
self.VTK.reset_data()
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(self._Ms, name='mu_s')
self.VTK.save_vector(self.spin.reshape(-1, 3), name='spins')
self.VTK.write_file(step=self.step)
class MicroDriver(DriverBase):
"""
A class with shared methods and properties for different drivers to solve
the Landau-Lifshitz-Gilbert equation
Variables that are proper of the driver class:
* alpha (damping)
* Ms_const
* t
* spin_last
* dm_dt
* integrator_tolerances_set
* step
* n (from mesh)
* n_nonzero (from mesh and set_Ms method in Simulation)
* gamma
* do_precession
* default_c (correction factor to keep the magnetisation normalised
during the LLG equation integration. See the
fidimag/atomistic/lib/llg.c file for more details)
TODO: Check default_c units in the micromagnetic
context
From DriverBase: t, spin_last, dm_dt, integrator_tolerances_set, step
"""
def __init__(self,
mesh,
spin,
Ms,
field,
pins,
interactions,
name,
data_saver,
integrator='sundials',
use_jac=False):
super(MicroDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._Ms = Ms
# Only for LLG STT: (??)
self.Ms_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
self.set_integrator(integrator, use_jac)
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
# Initialise the table for the data file with the simulation
# information:
self.data_saver = data_saver
# This should not be necessary:
# self.data_saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
self.data_saver.entities['rhs_evals'] = {
'unit': '<>',
'get': lambda sim: self.integrator.rhs_evals(),
'header': 'rhs_evals'
}
self.data_saver.entities['real_time'] = {
'unit': '<s>',
'get': lambda _: time.time(), # seconds since epoch
'header': 'real_time'
}
self.data_saver.update_entity_order()
# ---------------------------------------------------------------------
# OOMMF convention is to check if the spins have moved by ~1 degree in
# a nanosecond in order to stop a simulation, so we set this scale for
# dm/dt
# ONE_DEGREE_PER_NANOSECOND:
self._dmdt_factor = (2 * np.pi / 360) / 1e-9
def set_default_options(self, gamma=2.21e5, Ms=8.0e5, alpha=0.1):
"""
Default option for the integrator
Default gamma is for a free electron
"""
self.default_c = 1e11
self._alpha[:] = alpha
# When we create the simulation, Ms is set to the default value. This
# is overriden when calling the set_Ms method from the Siulation class
# or when setting Ms directly (property)
self._Ms[:] = Ms
self.gamma = gamma
self.do_precession = True
def sundials_rhs(self, t, y, ydot):
"""
Defined in the corresponding driver class
"""
pass
def set_tols(self, rtol=1e-8, atol=1e-10, max_ord=None, reset=True):
"""
Set the relative and absolute tolerances for the CVODE integrator
"""
if max_ord is not None:
self.integrator.set_options(rtol=rtol, atol=atol, max_ord=max_ord)
else:
# not all integrators have max_ord (only VODE does)
# and we don't want to encode a default value here either
self.integrator.set_options(rtol=rtol, atol=atol)
if reset:
self.reset_integrator(self.t)
# -------------------------------------------------------------------------
# Save functions ----------------------------------------------------------
# -------------------------------------------------------------------------
def save_vtk(self):
"""
Save a VTK file with the magnetisation vector field (vector data)
and the saturation magnetisation values (scalar data) as
cell data
"""
self.VTK.reset_data()
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(self._Ms, name='M_s')
self.VTK.save_vector(self.spin.reshape(-1, 3), name='spins')
self.VTK.write_file(step=self.step)
class MinimiserBase(object):
"""
Base class for minimiser class. No dependency on CVODE
"""
def __init__(self, mesh, spin, magnetisation, magnetisation_inv, field,
pins, interactions, name, data_saver):
# ---------------------------------------------------------------------
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
# A reference to either mu_s or Ms to use a common lib for this
# minimiser
self._magnetisation = magnetisation
self._magnetisation_inv = magnetisation_inv
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
self.data_saver = data_saver
# ---------------------------------------------------------------------
# Variables defined in this class
self.spin_last = np.ones_like(self.spin)
self.n = self.mesh.n
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
self.scale = 1.
def normalise_field(self, a):
norm = np.sqrt(np.sum(a.reshape(-1, 3)**2, axis=1))
norm_a = a.reshape(-1, 3) / norm[:, np.newaxis]
norm_a.shape = (-1, )
return norm_a
def field_cross_product(self, a, b):
aXb = np.cross(a.reshape(-1, 3), b.reshape(-1, 3))
return aXb.reshape(-1, )
def run_step(self):
"""
Python implementation of the step
"""
pass
def run_step_CLIB(self):
"""
Cython implementation of the step. Normally you would define
functions called in this method, in the lib/ folder
"""
pass
def minimise(self,
stopping_dm=1e-2,
max_steps=2000,
save_data_steps=10,
save_m_steps=None,
save_vtk_steps=None,
log_steps=1000):
pass
def relax(self):
print('Not implemented for the SD minimiser')
# -------------------------------------------------------------------------
def compute_effective_field(self, t=0):
"""
Compute the effective field from the simulation interactions,
calling the method from the corresponding Energy class
"""
self.field[:] = 0
for obj in self.interactions:
self.field += obj.compute_field(t=0, spin=self.spin)
# -------------------------------------------------------------------------
# SAVERS
def save_vtk(self):
"""
Save a VTK file with the magnetisation vector field and magnetic
moments as cell data. Magnetic moments are saved in units of
Bohr magnetons
NOTE: It is recommended to use a *cell to point data* filter in
Paraview or Mayavi to plot the vector field
"""
self.VTK.reset_data()
# Here we save both Ms and spins as cell data
self.VTK.save_scalar(self._magnetisation / const.mu_B,
name='magnetisation')
self.VTK.save_vector(self.spin.reshape(-1, 3), name='spins')
self.VTK.write_file(step=self.step)
def save_m(self, ZIP=False):
"""
Save the magnetisation/spin vector field as a numpy array in
a NPY file. The files are saved in the `{name}_npys` folder, where
`{name}` is the simulation name, with the file name `m_{step}.npy`
where `{step}` is the simulation step (from the integrator)
"""
if not os.path.exists('%s_npys' % self.name):
os.makedirs('%s_npys' % self.name)
name = '%s_npys/m_%g.npy' % (self.name, self.step)
np.save(name, self.spin)
if ZIP:
with zipfile.ZipFile('%s_m.zip' % self.name, 'a') as myzip:
myzip.write(name)
try:
os.remove(name)
except OSError:
pass
def save_skx(self):
"""
Save the skyrmion number density (sk number per mesh site)
as a numpy array in a NPY file.
The files are saved in the `{name}_skx_npys` folder, where
`{name}` is the simulation name, with the file name `skx_{step}.npy`
where `{step}` is the simulation step (from the integrator)
"""
if not os.path.exists('%s_skx_npys' % self.name):
os.makedirs('%s_skx_npys' % self.name)
name = '%s_skx_npys/m_%g.npy' % (self.name, self.step)
# The _skx_number array is defined in the SimBase class in Common
np.save(name, self._skx_number)
def test_save_vector_field():
mesh = CuboidMesh(4, 3, 2, 4, 3, 2)
s = vector_field(mesh, lambda r: (r[0], r[1], r[2]))
vtk = VTK(mesh, directory=OUTPUT_DIR, filename="save_vector")
vtk.save_vector(s, name="s")
class NEBMBase(object):
"""
Base Class for the NEBM codes.
Abstract Methods: ---------------------------------------------------------
compute_distances :: Function to compute the distances between
corresponding images of two bands. So the
inputs are two arrays with at least one
full image. The output is a 1D array with
the distances. So if the inputs have x
images we must return an array with x
entries.
compute_effective_field_and_energy :: Calculate effective field and
the energies of the images for
the energy band, according to
the number of degrees of
freedom (total number of spin
components). Effective field is
stored in the self.gradient and
the enrgies in self.energies
initialise_energies :: Populate the self.energies array with the
energies of every image for the 0th step
of the algorithm
Sundials_RHS :: Right hand side of the NEB equation for
Sundials
Methods -------------------------------------------------------------------
compute_spring_force :: Compute the spring force for every degree
of freedom of the band (we usually do
this using the C library)
compute_tangents :: Compute the tangents of the system (use
the function defined in
neb_method/nebm_clib.c). This tangents
need to be normalised / projected
according to the variation of the NEB
method.
compute_maximum_dYdt :: Compute the maximum distance between the
images (not counting the extremes) of the
last computed band in the evolver
(self.integrator.y) with the band of the
previous step (stored in self.last_Y)
divided by the last time step of the
evolver.
create_tablewriter :: Start the data frameworks to output the
energies ( *_energy.ndt) and distances
(*_dYs.ndt) for every step of the
integrator
generate_initial_band :: Use the initial states and inteprolations
lists to generate the initial band. The
interpolations are linearly made in
spherical coordinates, using the
nebm_tools.linear_interpolation_spherical
function. We may change this in the
future to generate an inteprolation using
a geodesic path
initialise_integrator :: Start the CVODE integrator and define it
in self.integrator
relax :: Relax the band for a specific number of
steps. Arguments of this function
includes saving VTK and NPY files every
certain number of steps and some criteria
for stopping the integrator
run_until :: This function is called from the relax
function (maybe is not very useful, we
may check this in the future)
save_VTKs :: Function to save a VTK of every image for
the current step of the integrator, which
is defined in the self.iterations
variable
save_npys :: Same for NPY files
ARGUMENTS -----------------------------------------------------------------
sim :: An instance of a micromagnetic or an atomistic
simulation
initial_images :: A sequence of arrays or functions to set up the
magnetisation field profile
inteprolations ::
dof :: Degrees of freedom per spin. Spherical coordinates
have dof=2 and Cartesian have dof=3
VARIABLES -----------------------------------------------------------------
self.dof :: Degree of freedom for the coordinates used in
the band (e.g. Spherical has self.dof=2)
self.sim :: Fidimag atomistic or micromagnetic simulation
object
self.mesh :: Fidimag simulation mesh object
self.name :: Name of the NEBM simulation
self.n_spins :: Number of spins per image
self.k :: Spring constant
self.VTK :: Fidimag VTK object to save VTK files
self.files_convert_f :: Function to convert the coordinates from the
band to Cartesian coordinates
self.initial_images :: List with the initial images for the band
(Numpy arrays or space functions with the
magnetisation/spin field)
self.interpolations :: Optional List with integers indicating
interpolations between images
self.n_images :: Number of images in the band calculated from
the number of initial images and
interpolations between them
self.n_images_inner_band :: Number of images without considering the
images at the extremes
self.n_dofs_image :: Number of degrees of freedom per image, which
is the total number of spins multiplied by the
self.dof
self.n_band :: Total number of degrees of freedom in the
whole band
self.band :: The array containing all the degrees of
freedom (spin directions). It is ordered in
the XYZ format
self.gradientE :: Array with the components of the energy
gradient
self.G :: Array with the effective force from the NEB
method definition
self.tangents :: Array with the NEBM tangents
self.energies :: Array with the energies of every image in the
band (length = self.n_images)
self.spring_force :: Array with the spring force components
self.distances :: Array with the distances between adjacent
images: [0-1 1-2 2-3 .. etc ]
self.last_Y :: Array with the last computed band (like
self.band) from the integrator
self._material :: Array of Booleans of size (self.dof*n_spins),
i.e. the number of dofs in an image.
It is True where Ms or mu_s are larger than
zero. This is necessary to filter spins that
not need to be counted in, for example, a
scaled norm.
self.n_dofs_image_material :: Number of dofs where mu_s or Ms > 0
(in a single image)
"""
def __init__(self,
sim,
initial_images,
interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False):
self.openmp = openmp
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# Spring constant (we could use an array in the future)
self.k = spring_constant
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
# self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='image')
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the NEBM band
self.n_band = self.n_images * self.n_dofs_image
# Climbing Image ------------------------------------------------------
if climbing_image is None:
self.climbing_image = -1
elif climbing_image in range(self.n_images):
self.climbing_image = climbing_image
else:
raise ValueError('The climbing image must be in the band. '
'Specify a valid integer.')
# NEBM Arrays ---------------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
self.last_Y = np.zeros_like(self.band)
# ---------------------------------------------------------------------
def initialise_energies(self):
pass
def save_VTKs(self, coordinates_function=None):
"""
Save VTK files in different folders, according to the simulation name
and step. Files are saved as vtks/simname_simstep_vtk/image_00000x.vtk
coordinates_function :: A function to transform the coordinates of
the band to Cartesian coordinates. For
example, in spherical coordinates we need
the spherical2cartesian function from
nebm_tools
"""
# Create the directory
directory = 'vtks/%s_%d' % (self.name, self.iterations)
self.VTK.directory = directory
self.band.shape = (self.n_images, -1)
# We use Ms from the simulation assuming that all the images are the
# same
for i in range(self.n_images):
self.VTK.reset_data()
# We will try to save for the micromagnetic simulation (Ms) or an
# atomistic simulation (mu_s) TODO: maybe this can be done with an:
# isinstance
if self.sim._micromagnetic:
self.VTK.save_scalar(self.sim.Ms, name='M_s')
else:
self.VTK.save_scalar(self.sim.mu_s, name='mu_s')
if coordinates_function:
self.VTK.save_vector(coordinates_function(
self.band[i]).reshape(-1, 3),
name='spins')
else:
self.VTK.save_vector(self.band[i].reshape(-1, 3), name='spins')
self.VTK.write_file(step=i)
self.band.shape = (-1, )
def save_npys(self, coordinates_function=None):
"""
Save npy files in different folders according to
the simulation name and step
Files are saved as: npys/simname_simstep/image_x.npy
"""
# Create directory as simname_simstep
directory = 'npys/%s_%d' % (self.name, self.iterations)
if not os.path.exists(directory):
os.makedirs(directory)
# Save the images with the format: 'image_{}.npy'
# where {} is the image number, starting from 0
self.band.shape = (self.n_images, -1)
for i in range(self.n_images):
name = os.path.join(directory, 'image_{:06}.npy'.format(i))
if coordinates_function:
np.save(name, coordinates_function(self.band[i, :]))
else:
np.save(name, self.band[i])
self.band.shape = (-1)
def initialise_integrator(self, rtol=1e-6, atol=1e-6):
self.t = 0
self.iterations = 0
self.ode_count = 1
if not self.openmp:
self.integrator = cvode.CvodeSolver(self.band, self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
else:
self.integrator = cvode.CvodeSolver_OpenMP(self.band,
self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
def create_tablewriter(self):
entities_energy = {
'step': {
'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'
},
'energy': {
'unit': '<J>',
'get': lambda sim: sim.energies,
'header': ['image_%d' % i for i in range(self.n_images)]
}
}
self.tablewriter = DataSaver(self,
'%s_energy.ndt' % (self.name),
entities=entities_energy)
entities_dm = {
'step': {
'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'
},
'dYs': {
'unit':
'<1>',
'get':
lambda sim: sim.distances,
'header':
['image_%d_%d' % (i, i + 1) for i in range(self.n_images - 1)]
}
}
self.tablewriter_dm = DataSaver(self,
'%s_dYs.ndt' % (self.name),
entities=entities_dm)
# ---------------------------------------------------------------------
def generate_initial_band(self):
pass
def compute_effective_field_and_energy(self, y):
pass
def compute_tangents(self, y):
pass
def compute_spring_force(self, y):
pass
def compute_distances(self):
pass
# -------------------------------------------------------------------------
# CVODE solver ------------------------------------------------------------
# -------------------------------------------------------------------------
def compute_norms(self, A, B):
"""
Compute the norms of the difference between corresponding images of the
bands A and B. Every norm is scaled by the number of degrees of freedom
The norm is computed only using mesh/lattice sites with material, i.e.
mu_s or Ms > 0
"""
A_minus_B = A - B
A_minus_B.shape = (-1, self.n_dofs_image)
A_minus_B = np.apply_along_axis(
lambda y: compute_norm(y[self._material], scale=True),
axis=1,
arr=A_minus_B)
return A_minus_B.reshape(-1)
def Sundials_RHS(self, t, y, ydot):
"""
This function is called on every iteration of the integrator (CVODE
solver). ydot refers to the Right Hand Side of the equation, since
we are solving dy/dt = 0
"""
self.ode_count += 1
# Update the effective field, energies, spring forces and tangents
# using the *y* array
self.nebm_step(y)
# Now set the RHS of the equation as the effective force on the energy
# band, which is stored on the self.G array
ydot[:] = self.G[:]
# The effective force at the extreme images should already be zero, but
# we will manually remove any value
ydot[:self.n_dofs_image] = 0
ydot[-self.n_dofs_image:] = 0
return 0
def compute_maximum_dYdt(self, A, B, dt):
"""
Compute the maximum difference from the images of the *A* array with
the images of the *B* array, divided by dt
The differences are not computed for the images at the extremes, since
these images are fixed and do not change with the iterator
For instance, in spherical coordinates, if we have a band of (N + 1)
images, labeled from 0 to N, we start by
dY = [A1_theta0 A1_phi0 A1_theta1 ... A(N-1)_theta0 A(N-1)_phi0 A(N-1)_theta1 ... ]
- [B1_theta0 B1_phi0 B1_theta1 ... B(N-1)_theta0 B(N-1)_phi0 B(N-1)_theta1 ... ]
where A(i)_theta(j) is the theta componenet of the j-th spin of the
i-th image in the band
Then we calculate the norm of every difference (the norm is computed
only using mesh/lattice sites with material, i.e. mu_s or Ms > 0):
||dY|| = [ || dY1_theta0 dY1_phi0 dY1_theta1 ... ||
...
|| dY(N-1)_theta0 dY(N-1)_phi0 dY(N-1)_theta1 ... ||
]
We divide by dt:
||dY|| -- > || dY || / dt = dYdt
and finally take the maximum value of dYdt
"""
# We will not consider the images at the extremes to compute dY
# Since we removed the extremes, we only have *n_images_inner_band*
# images
band_no_extremes = slice(self.n_dofs_image, -self.n_dofs_image)
dYdt = self.compute_norms(A[band_no_extremes],
B[band_no_extremes]).reshape(
self.n_images_inner_band, -1)
dYdt /= dt
if np.max(dYdt) > 0:
return np.max(dYdt)
else:
return 0
def run_until(self, t):
if (t) <= self.t:
return
self.integrator.run_until(t)
# Copy the updated energy band to our local array
self.band[:] = self.integrator.y[:]
# Compute the maximum change in the integrator step
max_dYdt = self.compute_maximum_dYdt(self.integrator.y, self.last_Y,
t - self.t)
self.last_Y[:] = self.band[:]
# Update the current step
self.t = t
return max_dYdt
def relax(self,
dt=1e-8,
stopping_dYdt=1,
max_iterations=1000,
save_npys_every=100,
save_vtks_every=100):
"""
"""
log.debug("Relaxation parameters: "
"stopping_dYdt={}, "
"time_step={} s, "
" max_iterations={}.".format(stopping_dYdt, dt,
max_iterations))
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# Save the initial state i=0
self.distances = self.compute_distances(self.band[self.n_dofs_image:],
self.band[:-self.n_dofs_image])
self.tablewriter.save()
self.tablewriter_dm.save()
INNER_DOFS = slice(self.n_dofs_image, -self.n_dofs_image)
for i in range(max_iterations):
# Update the iterations number counter
self.iterations += 1
# Get the current size of the time discretisation from the
# integrator (variable step size)
cvode_dt = self.integrator.get_current_step()
# If the step size of the integrator is larger than the specified
# discretisation, use the current integrator step size to
# compute the next iteration. Otherwise, just stick to the
# specified step
if cvode_dt > dt:
increment_dt = cvode_dt
else:
increment_dt = dt
# max_dYdt = self.run_until(self.t + increment_dt)
if self.t + increment_dt <= self.t:
break
self.integrator.run_until(self.t + increment_dt)
# Copy the updated energy band to our local array
self.band[:] = self.integrator.y[:]
# Compute the maximum change in the integrator step
max_dYdt = self.compute_maximum_dYdt(self.band, self.last_Y,
increment_dt)
self.last_Y[:] = self.band[:]
# Update the current step
self.t = self.t + increment_dt
# Save data -------------------------------------------------------
if self.iterations % save_vtks_every == 0:
self.save_VTKs(coordinates_function=self.files_convert_f)
if self.iterations % save_npys_every == 0:
self.save_npys(coordinates_function=self.files_convert_f)
self.distances = self.compute_distances(
self.band[self.n_dofs_image:], self.band[:-self.n_dofs_image])
self.tablewriter.save()
self.tablewriter_dm.save()
# Print information about the simulation and the NEBM forces.
# The last two terms are the largest gradient and spring
# force norms from the spins (not counting the extrema)
log.debug(
time.strftime("%Y-%m-%d %H:%M:%S ", time.localtime()) +
"step: {:.3g}, step_size: {:.3g},"
"max_dYdt: {:.3g} "
"|max_dE|: {:.3g} "
"and |max_F_k|: {:.3g}".format(
self.iterations, increment_dt, max_dYdt,
np.max(
np.linalg.norm(self.gradientE[INNER_DOFS].reshape(
-1, 3),
axis=1)),
np.max(
np.linalg.norm(self.spring_force[INNER_DOFS].reshape(
-1, 3),
axis=1))))
# -----------------------------------------------------------------
# Stop criteria:
if max_dYdt < stopping_dYdt:
break
log.info("Relaxation finished at time step = {:.4g}, "
"t = {:.2g}, call rhs = {:.4g} "
"and max_dYdt = {:.3g}".format(self.iterations, self.t,
self.ode_count, max_dYdt))
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# -------------------------------------------------------------------------
def compute_polynomial_approximation(self, n_points):
"""
Compute a smooth approximation for the band, using a third order
polynomial approximation. This approximation uses the tangents and
derivatives from each image of the band as information to estimate
the curvatures. The formula can be found in
- Bessarab et al., Computer Physics Communications 196 (2015) 335-347
This function returns a tuple with two elements:
0. An array with the distance of every data point from the 0th
image.
1. A n_points long array, with the cubic interpolated energy band
ARGUMENTS
n_points :: Te number of points for the interpolation
"""
# To be sure, update the effective field and tangents when calling
# this function (this is necesary if we call the function without
# relaxing the band before)
self.compute_effective_field_and_energy(self.band)
self.compute_tangents(self.band)
self.distances = self.compute_distances(self.band[self.n_dofs_image:],
self.band[:-self.n_dofs_image])
deltas = np.zeros(self.n_images)
# Somehow we need to rescale the gradient by the right units. In the
# case of micromag, we use mu0 * Ms, and for the atomistic case we
# simply use mu_s. This must be related to the way we derive the
# effective field to calculate the negative energy gradient, which is
# the functional derivative of the energy
if self.sim._micromagnetic:
scale = np.repeat(
self.mesh.dx * self.mesh.dy * self.mesh.dz *
(self.mesh.unit_length**3.) * const.mu_0 * self.sim.Ms, 3)
else:
scale = np.repeat(self.sim.mu_s, 3)
for i in range(self.n_images):
deltas[i] = np.dot(
scale * self.gradientE.reshape(self.n_images, -1)[i],
self.tangents.reshape(self.n_images, -1)[i])
l = [0]
for i in range(len(self.distances)):
l.append(np.sum(self.distances[:i + 1]))
l = np.array(l)
E = self.energies
# The coefficients for the polynomial approximation
a = np.zeros(self.n_images)
b = np.zeros(self.n_images)
c = deltas
d = E
# Populate the a and b coefficients for every image
for i in range(self.n_images - 1):
a[i] = (deltas[i + 1] + deltas[i]) / (l[i + 1] - l[i])**2.
a[i] -= 2 * (E[i + 1] - E[i]) / (l[i + 1] - l[i])**3.
b[i] = -(deltas[i + 1] + 2 * deltas[i]) / (l[i + 1] - l[i])
b[i] += 3 * (E[i + 1] - E[i]) / (l[i + 1] - l[i])**2.
# The arrays with the data points and the interpolated energy values
x = np.linspace(0, l[-1], n_points)
E_interp = np.zeros(n_points)
i_img = 0
for i, pos in enumerate(x):
if pos > l[i_img + 1]:
i_img += 1
E_interp[i] = (a[i_img] * ((pos - l[i_img])**3.) + b[i_img] *
((pos - l[i_img])**2.) + c[i_img] *
(pos - l[i_img]) + d[i_img])
return x, E_interp
def __init__(self,
sim,
initial_images,
interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False):
self.openmp = openmp
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# Spring constant (we could use an array in the future)
self.k = spring_constant
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
# self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='image')
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the NEBM band
self.n_band = self.n_images * self.n_dofs_image
# Climbing Image ------------------------------------------------------
if climbing_image is None:
self.climbing_image = -1
elif climbing_image in range(self.n_images):
self.climbing_image = climbing_image
else:
raise ValueError('The climbing image must be in the band. '
'Specify a valid integer.')
# NEBM Arrays ---------------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
self.last_Y = np.zeros_like(self.band)
def test_save_scalar_field_hexagonal_mesh(tmpdir):
mesh = HexagonalMesh(1, 3, 3)
s = scalar_field(mesh, lambda r: r[0] + r[1])
vtk = VTK(mesh, directory=str(tmpdir), filename="scalar_hexagonal")
vtk.save_scalar(s, name="s")
assert same_as_ref(vtk.write_file(), REF_DIR)
class NEBMBase(object):
"""
Base Class for the NEBM codes.
Abstract Methods: ---------------------------------------------------------
compute_distances :: Function to compute the distances between
corresponding images of two bands. So the
inputs are two arrays with at least one
full image. The output is a 1D array with
the distances. So if the inputs have x
images we must return an array with x
entries.
compute_effective_field_and_energy :: Calculate effective field and
the energies of the images for
the energy band, according to
the number of degrees of
freedom (total number of spin
components). Effective field is
stored in the self.gradient and
the enrgies in self.energies
initialise_energies :: Populate the self.energies array with the
energies of every image for the 0th step
of the algorithm
Sundials_RHS :: Right hand side of the NEB equation for
Sundials
Methods -------------------------------------------------------------------
compute_spring_force :: Compute the spring force for every degree
of freedom of the band (we usually do
this using the C library)
compute_tangents :: Compute the tangents of the system (use
the function defined in
neb_method/nebm_clib.c). This tangents
need to be normalised / projected
according to the variation of the NEB
method.
compute_maximum_dYdt :: Compute the maximum distance between the
images (not counting the extremes) of the
last computed band in the evolver
(self.integrator.y) with the band of the
previous step (stored in self.last_Y)
divided by the last time step of the
evolver.
create_tablewriter :: Start the data frameworks to output the
energies ( *_energy.ndt) and distances
(*_dYs.ndt) for every step of the
integrator
generate_initial_band :: Use the initial states and inteprolations
lists to generate the initial band. The
interpolations are linearly made in
spherical coordinates, using the
nebm_tools.linear_interpolation_spherical
function. We may change this in the
future to generate an inteprolation using
a geodesic path
initialise_integrator :: Start the CVODE integrator and define it
in self.integrator
relax :: Relax the band for a specific number of
steps. Arguments of this function
includes saving VTK and NPY files every
certain number of steps and some criteria
for stopping the integrator
run_until :: This function is called from the relax
function (maybe is not very useful, we
may check this in the future)
save_VTKs :: Function to save a VTK of every image for
the current step of the integrator, which
is defined in the self.iterations
variable
save_npys :: Same for NPY files
ARGUMENTS -----------------------------------------------------------------
sim :: An instance of a micromagnetic or an atomistic
simulation
initial_images :: A sequence of arrays or functions to set up the
magnetisation field profile
inteprolations ::
dof :: Degrees of freedom per spin. Spherical coordinates
have dof=2 and Cartesian have dof=3
VARIABLES -----------------------------------------------------------------
self.dof :: Degree of freedom for the coordinates used in
the band (e.g. Spherical has self.dof=2)
self.sim :: Fidimag atomistic or micromagnetic simulation
object
self.mesh :: Fidimag simulation mesh object
self.name :: Name of the NEBM simulation
self.n_spins :: Number of spins per image
self.k :: Spring constant
self.VTK :: Fidimag VTK object to save VTK files
self.files_convert_f :: Function to convert the coordinates from the
band to Cartesian coordinates
self.initial_images :: List with the initial images for the band
(Numpy arrays or space functions with the
magnetisation/spin field)
self.interpolations :: Optional List with integers indicating
interpolations between images
self.n_images :: Number of images in the band calculated from
the number of initial images and
interpolations between them
self.n_images_inner_band :: Number of images without considering the
images at the extremes
self.n_dofs_image :: Number of degrees of freedom per image, which
is the total number of spins multiplied by the
self.dof
self.n_band :: Total number of degrees of freedom in the
whole band
self.band :: The array containing all the degrees of
freedom (spin directions). It is ordered in
the XYZ format
self.gradientE :: Array with the components of the energy
gradient
self.G :: Array with the effective force from the NEB
method definition
self.tangents :: Array with the NEBM tangents
self.energies :: Array with the energies of every image in the
band (length = self.n_images)
self.spring_force :: Array with the spring force components
self.distances :: Array with the distances between adjacent
images: [0-1 1-2 2-3 .. etc ]
self.last_Y :: Array with the last computed band (like
self.band) from the integrator
"""
def __init__(self, sim,
initial_images, interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False
):
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# Spring constant (we could use an array in the future)
self.k = spring_constant
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='m'
)
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the NEBM band
self.n_band = self.n_images * self.n_dofs_image
# Climbing Image ------------------------------------------------------
if climbing_image is None:
self.climbing_image = -1
elif climbing_image in range(self.n_images):
self.climbing_image = climbing_image
else:
raise ValueError('The climbing image must be in the band. '
'Specify a valid integer.')
# NEBM Arrays ---------------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
self.last_Y = np.zeros_like(self.band)
# ---------------------------------------------------------------------
def initialise_energies(self):
pass
def save_VTKs(self, coordinates_function=None):
"""
Save VTK files in different folders, according to the simulation name
and step. Files are saved as vtks/simname_simstep_vtk/image_00000x.vtk
coordinates_function :: A function to transform the coordinates of
the band to Cartesian coordinates. For
example, in spherical coordinates we need
the spherical2cartesian function from
nebm_tools
"""
# Create the directory
directory = 'vtks/%s_%d' % (self.name, self.iterations)
self.VTK.directory = directory
self.band.shape = (self.n_images, -1)
# We use Ms from the simulation assuming that all the images are the
# same
for i in range(self.n_images):
self.VTK.reset_data()
# We will try to save for the micromagnetic simulation (Ms) or an
# atomistic simulation (mu_s) TODO: maybe this can be done with an:
# isinstance
if self.sim._micromagnetic:
self.VTK.save_scalar(self.sim.Ms, name='M_s')
else:
self.VTK.save_scalar(self.sim.mu_s, name='mu_s')
if coordinates_function:
self.VTK.save_vector(
coordinates_function(self.band[i]).reshape(-1, 3),
name='spins'
)
else:
self.VTK.save_vector(
self.band[i].reshape(-1, 3),
name='spins'
)
self.VTK.write_file(step=i)
self.band.shape = (-1, )
def save_npys(self, coordinates_function=None):
"""
Save npy files in different folders according to
the simulation name and step
Files are saved as: npys/simname_simstep/image_x.npy
"""
# Create directory as simname_simstep
directory = 'npys/%s_%d' % (self.name, self.iterations)
if not os.path.exists(directory):
os.makedirs(directory)
# Save the images with the format: 'image_{}.npy'
# where {} is the image number, starting from 0
self.band.shape = (self.n_images, -1)
for i in range(self.n_images):
name = os.path.join(directory, 'image_{:06}.npy'.format(i))
if coordinates_function:
np.save(name, coordinates_function(self.band[i, :]))
else:
np.save(name, self.band[i])
self.band.shape = (-1)
def initialise_integrator(self, rtol=1e-6, atol=1e-6, openmp=False):
self.t = 0
self.iterations = 0
self.ode_count = 1
if not openmp:
self.integrator = cvode.CvodeSolver(self.band,
self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
else:
self.integrator = cvode.CvodeSolver_OpenMP(self.band,
self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
def create_tablewriter(self):
entities_energy = {
'step': {'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'},
'energy': {'unit': '<J>',
'get': lambda sim: sim.energies,
'header': ['image_%d' % i
for i in range(self.n_images)]}
}
self.tablewriter = DataSaver(
self, '%s_energy.ndt' % (self.name), entities=entities_energy)
entities_dm = {
'step': {'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'},
'dYs': {'unit': '<1>',
'get': lambda sim: sim.distances,
'header': ['image_%d_%d' % (i, i + 1)
for i in range(self.n_images - 1)]}
}
self.tablewriter_dm = DataSaver(
self, '%s_dYs.ndt' % (self.name), entities=entities_dm)
# ---------------------------------------------------------------------
def generate_initial_band(self):
pass
def compute_effective_field_and_energy(self, y):
pass
def compute_tangents(self, y):
pass
def compute_spring_force(self, y):
pass
def compute_distances(self):
pass
# -------------------------------------------------------------------------
# CVODE solver ------------------------------------------------------------
# -------------------------------------------------------------------------
def compute_norms(self, A, B):
"""
Compute the norms of the difference between corresponding images of the
bands A and B. Every norm is scaled by the number of degrees of freedom
The norm is computed only using mesh/lattice sites with material, i.e.
mu_s or Ms > 0
"""
A_minus_B = A - B
A_minus_B.shape = (-1, self.n_dofs_image)
A_minus_B = np.apply_along_axis(
lambda y: compute_norm(y[self._material], scale=True),
axis=1,
arr=A_minus_B
)
return A_minus_B.reshape(-1)
def Sundials_RHS(self, t, y, ydot):
"""
This function is called on every iteration of the integrator (CVODE
solver). ydot refers to the Right Hand Side of the equation, since
we are solving dy/dt = 0
"""
self.ode_count += 1
# Update the effective field, energies, spring forces and tangents
# using the *y* array
self.nebm_step(y)
# Now set the RHS of the equation as the effective force on the energy
# band, which is stored on the self.G array
ydot[:] = self.G[:]
# The effective force at the extreme images should already be zero, but
# we will manually remove any value
ydot[:self.n_dofs_image] = 0
ydot[-self.n_dofs_image:] = 0
return 0
def compute_maximum_dYdt(self, A, B, dt):
"""
Compute the maximum difference from the images of the *A* array with
the images of the *B* array, divided by dt
The differences are not computed for the images at the extremes, since
these images are fixed and do not change with the iterator
For instance, in spherical coordinates, if we have a band of (N + 1)
images, labeled from 0 to N, we start by
dY = [A1_theta0 A1_phi0 A1_theta1 ... A(N-1)_theta0 A(N-1)_phi0 A(N-1)_theta1 ... ]
- [B1_theta0 B1_phi0 B1_theta1 ... B(N-1)_theta0 B(N-1)_phi0 B(N-1)_theta1 ... ]
where A(i)_theta(j) is the theta componenet of the j-th spin of the
i-th image in the band
Then we calculate the norm of every difference (the norm is computed
only using mesh/lattice sites with material, i.e. mu_s or Ms > 0):
||dY|| = [ || dY1_theta0 dY1_phi0 dY1_theta1 ... ||
...
|| dY(N-1)_theta0 dY(N-1)_phi0 dY(N-1)_theta1 ... ||
]
We divide by dt:
||dY|| -- > || dY || / dt = dYdt
and finally take the maximum value of dYdt
"""
# We will not consider the images at the extremes to compute dY
# Since we removed the extremes, we only have *n_images_inner_band*
# images
band_no_extremes = slice(self.n_dofs_image, -self.n_dofs_image)
dYdt = self.compute_norms(
A[band_no_extremes],
B[band_no_extremes]).reshape(self.n_images_inner_band, -1)
dYdt /= dt
if np.max(dYdt) > 0:
return np.max(dYdt)
else:
return 0
def run_until(self, t):
if (t) <= self.t:
return
self.integrator.run_until(t)
# Copy the updated energy band to our local array
self.band[:] = self.integrator.y[:]
# Compute the maximum change in the integrator step
max_dYdt = self.compute_maximum_dYdt(self.integrator.y, self.last_Y,
t - self.t)
self.last_Y[:] = self.band[:]
# Update the current step
self.t = t
return max_dYdt
def relax(self, dt=1e-8, stopping_dYdt=1, max_iterations=1000,
save_npys_every=100, save_vtks_every=100
):
"""
"""
log.debug("Relaxation parameters: "
"stopping_dYdt={}, "
"time_step={} s, "
" max_iterations={}.".format(stopping_dYdt,
dt,
max_iterations))
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# Save the initial state i=0
self.distances = self.compute_distances(self.band[self.n_dofs_image:],
self.band[:-self.n_dofs_image]
)
self.tablewriter.save()
self.tablewriter_dm.save()
for i in range(max_iterations):
# Update the iterations number counter
self.iterations += 1
# Get the current size of the time discretisation from the
# integrator (variable step size)
cvode_dt = self.integrator.get_current_step()
# If the step size of the integrator is larger than the specified
# discretisation, use the current integrator step size to
# compute the next iteration. Otherwise, just stick to the
# specified step
if cvode_dt > dt:
increment_dt = cvode_dt
else:
increment_dt = dt
# max_dYdt = self.run_until(self.t + increment_dt)
if self.t + increment_dt <= self.t:
break
self.integrator.run_until(self.t + increment_dt)
# Copy the updated energy band to our local array
self.band[:] = self.integrator.y[:]
# Compute the maximum change in the integrator step
max_dYdt = self.compute_maximum_dYdt(self.band,
self.last_Y,
increment_dt)
self.last_Y[:] = self.band[:]
# Update the current step
self.t = self.t + increment_dt
# Save data -------------------------------------------------------
if self.iterations % save_vtks_every == 0:
self.save_VTKs(coordinates_function=self.files_convert_f)
if self.iterations % save_npys_every == 0:
self.save_npys(coordinates_function=self.files_convert_f)
self.distances = self.compute_distances(
self.band[self.n_dofs_image:],
self.band[:-self.n_dofs_image]
)
self.tablewriter.save()
self.tablewriter_dm.save()
log.debug(time.strftime("%Y-%m-%d %H:%M:%S ", time.localtime()) +
"step: {:.3g}, step_size: {:.3g}"
" and max_dYdt: {:.3g}.".format(self.iterations,
increment_dt,
max_dYdt)
)
# -----------------------------------------------------------------
# Stop criteria:
if max_dYdt < stopping_dYdt:
break
log.info("Relaxation finished at time step = {:.4g}, "
"t = {:.2g}, call rhs = {:.4g} "
"and max_dYdt = {:.3g}".format(self.iterations,
self.t,
self.ode_count,
max_dYdt)
)
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# -------------------------------------------------------------------------
def compute_polynomial_approximation(self, n_points):
"""
Compute a smooth approximation for the band, using a third order
polynomial approximation. This approximation uses the tangents and
derivatives from each image of the band as information to estimate
the curvatures. The formula can be found in
- Bessarab et al., Computer Physics Communications 196 (2015) 335-347
This function returns a tuple with two elements:
0. An array with the distance of every data point from the 0th
image.
1. A n_points long array, with the cubic interpolated energy band
ARGUMENTS
n_points :: Te number of points for the interpolation
"""
# To be sure, update the effective field and tangents when calling
# this function (this is necesary if we call the function without
# relaxing the band before)
self.compute_effective_field_and_energy(self.band)
self.compute_tangents(self.band)
self.distances = self.compute_distances(self.band[self.n_dofs_image:],
self.band[:-self.n_dofs_image]
)
deltas = np.zeros(self.n_images)
# Somehow we need to rescale the gradient by the right units. In the
# case of micromag, we use mu0 * Ms, and for the atomistic case we
# simply use mu_s. This must be related to the way we derive the
# effective field to calculate the negative energy gradient, which is
# the functional derivative of the energy
if self.sim._micromagnetic:
scale = np.repeat(self.mesh.dx * self.mesh.dy * self.mesh.dz *
(self.mesh.unit_length ** 3.) *
const.mu_0 * self.sim.Ms, 3)
else:
scale = np.repeat(self.sim.mu_s, 3)
for i in range(self.n_images):
deltas[i] = np.dot(scale * self.gradientE.reshape(self.n_images, -1)[i],
self.tangents.reshape(self.n_images, -1)[i]
)
l = [0]
for i in range(len(self.distances)):
l.append(np.sum(self.distances[:i + 1]))
l = np.array(l)
E = self.energies
# The coefficients for the polynomial approximation
a = np.zeros(self.n_images)
b = np.zeros(self.n_images)
c = deltas
d = E
# Populate the a and b coefficients for every image
for i in range(self.n_images - 1):
a[i] = (deltas[i + 1] + deltas[i]) / (l[i + 1] - l[i]) ** 2.
a[i] -= 2 * (E[i + 1] - E[i]) / (l[i + 1] - l[i]) ** 3.
b[i] = -(deltas[i + 1] + 2 * deltas[i]) / (l[i + 1] - l[i])
b[i] += 3 * (E[i + 1] - E[i]) / (l[i + 1] - l[i]) ** 2.
# The arrays with the data points and the interpolated energy values
x = np.linspace(0, l[-1], n_points)
E_interp = np.zeros(n_points)
i_img = 0
for i, pos in enumerate(x):
if pos > l[i_img + 1]:
i_img += 1
E_interp[i] = (a[i_img] * ((pos - l[i_img]) ** 3.) +
b[i_img] * ((pos - l[i_img]) ** 2.) +
c[i_img] * (pos - l[i_img]) +
d[i_img]
)
return x, E_interp
def __init__(self,
mesh,
spin,
Ms,
field,
pins,
interactions,
name,
data_saver,
integrator='sundials',
use_jac=False):
super(MicroDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._Ms = Ms
# Only for LLG STT: (??)
self.Ms_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
self.set_integrator(integrator, use_jac)
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
# Initialise the table for the data file with the simulation
# information:
self.data_saver = data_saver
# This should not be necessary:
# self.data_saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
self.data_saver.entities['rhs_evals'] = {
'unit': '<>',
'get': lambda sim: self.integrator.rhs_evals(),
'header': 'rhs_evals'
}
self.data_saver.entities['real_time'] = {
'unit': '<s>',
'get': lambda _: time.time(), # seconds since epoch
'header': 'real_time'
}
self.data_saver.update_entity_order()
# ---------------------------------------------------------------------
# OOMMF convention is to check if the spins have moved by ~1 degree in
# a nanosecond in order to stop a simulation, so we set this scale for
# dm/dt
# ONE_DEGREE_PER_NANOSECOND:
self._dmdt_factor = (2 * np.pi / 360) / 1e-9
def __init__(self,
sim,
initial_images,
interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False):
self.openmp = openmp
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
# self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='image')
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the string/band
self.n_band = self.n_images * self.n_dofs_image
# Spring constant -----------------------------------------------------
# Spring constant (we could use an array in the future)
self.k = spring_constant * np.ones(self.n_images)
# Set to True to update spring constant values relative to the energies
# (TESTING)
self.variable_k = False
self.dk = 1
# Climbing Image ------------------------------------------------------
# Set a list with the images where 1 is for climbing image and 0 for
# normal
self._climbing_image = np.zeros(self.n_images, dtype=np.int32)
if climbing_image is not None:
self.climbing_image = climbing_image
# Chain Method Arrays -------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
# Total distance starting from image_0
# (first element shoud always be zero)
self.path_distances = np.zeros(self.n_images)
self.last_Y = np.zeros_like(self.band)
# ---------------------------------------------------------------------
# If the integrator uses an LLG-like equation to relax the energy band
# we need to set this variable
# This variable only affects the StepIntegrators, NOT Sundials
self._llg_evolve = False
# ---------------------------------------------------------------------
# Factors for interpolating the energy band
# For now we only have a 3rd order polynomial interp, thus we set
# 4 factors
self.interp_factors = np.zeros((4, self.n_images))
# Somehow we need to rescale the gradient by the right units. In the
# case of micromag, we use mu0 * Ms, and for the atomistic case we
# simply use mu_s. This must be related to the way we derive the
# effective field to calculate the negative energy gradient, which is
# the functional derivative of the energy
if self.sim._micromagnetic:
self.scale = np.repeat(
self.mesh.dx * self.mesh.dy * self.mesh.dz *
(self.mesh.unit_length**3.) * const.mu_0 * self.sim.Ms, 3)
else:
self.scale = np.repeat(self.sim.mu_s, 3)
# ---------------------------------------------------------------------
self.G_log = []
def __init__(self, mesh, spin, Ms, field, pins,
interactions,
name,
data_saver,
integrator='sundials',
use_jac=False
):
super(MicroDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._Ms = Ms
# Only for LLG STT: (??)
self.Ms_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
# Here we set up the CVODE integrator from Sundials to evolve a
# specific micromagnetic equation. The equations are specified in the
# sundials_rhs function from any of the micromagnetic drivers in the
# micromagnetic folder (LLG, LLG_STT, etc.)
if integrator == "sundials" and use_jac:
self.integrator = CvodeSolver(self.spin, self.sundials_rhs,
self.sundials_jtimes)
elif integrator == "sundials_diag":
self.integrator = CvodeSolver(self.spin, self.sundials_rhs,
linear_solver="diag")
elif integrator == "sundials":
self.integrator = CvodeSolver(self.spin, self.sundials_rhs)
elif integrator == "euler" or integrator == "rk4":
self.integrator = CvodeSolver(self.spin, self.step_rhs,
integrator)
elif integrator == "sundials_openmp" and use_jac:
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs,
self.sundials_jtimes)
elif integrator == "sundials_diag_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs,
linear_solver="diag")
elif integrator == "sundials_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.sundials_rhs)
elif integrator == "euler_openmp" or integrator == "rk4_openmp":
self.integrator = CvodeSolver_OpenMP(self.spin, self.step_rhs,
integrator)
else:
raise NotImplemented("integrator must be sundials, euler or rk4")
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m'
)
# Initialise the table for the data file with the simulation
# information:
self.data_saver = data_saver
# This should not be necessary:
# self.data_saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
self.data_saver.entities['rhs_evals'] = {
'unit': '<>',
'get': lambda sim: self.integrator.rhs_evals(),
'header': 'rhs_evals'}
self.data_saver.entities['real_time'] = {
'unit': '<s>',
'get': lambda _: time.time(), # seconds since epoch
'header': 'real_time'}
self.data_saver.update_entity_order()
# ---------------------------------------------------------------------
# OOMMF convention is to check if the spins have moved by ~1 degree in
# a nanosecond in order to stop a simulation, so we set this scale for
# dm/dt
# ONE_DEGREE_PER_NANOSECOND:
self._dmdt_factor = (2 * np.pi / 360) / 1e-9
def __init__(self, mesh, name='unnamed'):
self.name = name
# Initiate a VTK object
self.VTK = VTK(mesh, directory='{}_vtks'.format(name), filename=name)
def __init__(self,
mesh,
spin,
mu_s,
mu_s_inv,
field,
pins,
interactions,
name,
data_saver,
use_jac,
integrator='sundials'):
super(AtomisticDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._mu_s = mu_s
self._mu_s_inv = mu_s_inv
# Only for LLG STT: (??)
self.mu_s_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
# In the old code, it seemed that self.sundials_jtn was not defined
# anywhere else. Here, we will use the same integrators than in the
# micromagnetic code, where we have self.sundials_jtimes instead of jtn
self.set_integrator(integrator, use_jac)
self.set_tols()
# When initialising the integrator in the self.integrator call, the
# CVOde class calls the set_initial_value function (with flag_m=0),
# which initialises a new integrator and allocates memory in this
# process. Now, when we set the magnetisation, we will use the same
# memory setting this flag_m to 1, so instead of calling CVodeInit we
# call CVodeReInit. If don't, memory is allocated in every call of
# set_m
self.flag_m = 1
# Factor for the dmdt magnitude in the relaxation function
self._dmdt_factor = 1.
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m')
self.data_saver = data_saver
def test_save_scalar_field_hexagonal_mesh(tmpdir):
mesh = HexagonalMesh(1, 3, 3)
s = scalar_field(mesh, lambda r: r[0] + r[1])
vtk = VTK(mesh, directory=str(tmpdir), filename="scalar_hexagonal")
vtk.save_scalar(s, name="s")
assert same_as_ref(vtk.write_file(), REF_DIR)
def __init__(self, mesh, spin, Ms, field, pins,
interactions,
name,
data_saver,
integrator='sundials',
use_jac=False
):
super(MicroDriver, self).__init__()
# These are (ideally) references to arrays taken from the Simulation
# class. Variables with underscore are arrays changed by a property in
# the simulation class
self.mesh = mesh
self.spin = spin
self._Ms = Ms
# Only for LLG STT: (??)
self.Ms_const = 0
self.field = field
self._pins = pins
self.interactions = interactions
# Strings are not referenced, this is a copy:
self.name = name
# The following are proper of the driver class: (see DriverBase) ------
# See also the set_default_options() function
self.n = self.mesh.n
self.n_nonzero = self.mesh.n # number of spins that are not zero
# We check this in the set_Ms function
self.initiate_variables(self.n)
self.set_default_options()
# Integrator options --------------------------------------------------
self.set_integrator(integrator, use_jac)
# Savers --------------------------------------------------------------
# VTK saver for the magnetisation/spin field
self.VTK = VTK(self.mesh,
directory='{}_vtks'.format(self.name),
filename='m'
)
# Initialise the table for the data file with the simulation
# information:
self.data_saver = data_saver
# This should not be necessary:
# self.data_saver.entities['skx_num'] = {
# 'unit': '<>',
# 'get': lambda sim: sim.skyrmion_number(),
# 'header': 'skx_num'}
self.data_saver.entities['rhs_evals'] = {
'unit': '<>',
'get': lambda sim: self.integrator.rhs_evals(),
'header': 'rhs_evals'}
self.data_saver.entities['real_time'] = {
'unit': '<s>',
'get': lambda _: time.time(), # seconds since epoch
'header': 'real_time'}
self.data_saver.update_entity_order()
# ---------------------------------------------------------------------
# OOMMF convention is to check if the spins have moved by ~1 degree in
# a nanosecond in order to stop a simulation, so we set this scale for
# dm/dt
# ONE_DEGREE_PER_NANOSECOND:
self._dmdt_factor = (2 * np.pi / 360) / 1e-9
def test_save_scalar_field_hexagonal_mesh():
mesh = HexagonalMesh(1, 3, 3)
s = scalar_field(mesh, lambda r: r[0] + r[1])
vtk = VTK(mesh, directory=OUTPUT_DIR, filename="scalar_hexagonal")
vtk.save_scalar(s, name="s")
class ChainMethodBase(object):
"""
Base Class for chain methods, such as NEBM or String Method codes.
Abstract Methods: ---------------------------------------------------------
compute_distances :: Function to compute the distances between
corresponding images of two bands. So the
inputs are two arrays with at least one
full image. The output is a 1D array with
the distances. So if the inputs have x
images we must return an array with x
entries.
compute_effective_field_and_energy :: Calculate effective field and
the energies of the images for
the energy band, according to
the number of degrees of
freedom (total number of spin
components). Effective field is
stored in the self.gradient and
the enrgies in self.energies
initialise_energies :: Populate the self.energies array with the
energies of every image for the 0th step
of the algorithm
Sundials_RHS :: Right hand side of the NEB equation for
Sundials
Methods -------------------------------------------------------------------
compute_maximum_dYdt :: Compute the maximum distance between the
images (not counting the extremes) of the
last computed band in the evolver
(self.integrator.y) with the band of the
previous step (stored in self.last_Y)
divided by the last time step of the
evolver.
create_tablewriter :: Start the data frameworks to output the
energies ( *_energy.ndt) and distances
(*_dYs.ndt) for every step of the
integrator
generate_initial_band :: Use the initial states and inteprolations
lists to generate the initial band. The
interpolations are linearly made in
spherical coordinates, using the
chain_method_tools.linear_interpolation_spherical
function. We may change this in the
future to generate an inteprolation using
a geodesic path
initialise_integrator :: Start the CVODE integrator and define it
in self.integrator
relax :: Relax the band for a specific number of
steps. Arguments of this function
includes saving VTK and NPY files every
certain number of steps and some criteria
for stopping the integrator
run_until :: This function is called from the relax
function (maybe is not very useful, we
may check this in the future)
save_VTKs :: Function to save a VTK of every image for
the current step of the integrator, which
is defined in the self.iterations
variable
save_npys :: Same for NPY files
ARGUMENTS -----------------------------------------------------------------
sim :: An instance of a micromagnetic or an atomistic
simulation
initial_images :: A sequence of arrays or functions to set up the
magnetisation field profile
interpolations ::
dof :: Degrees of freedom per spin. Spherical coordinates
have dof=2 and Cartesian have dof=3
VARIABLES -----------------------------------------------------------------
self.dof :: Degree of freedom for the coordinates used in
the band (e.g. Spherical has self.dof=2)
self.sim :: Fidimag atomistic or micromagnetic simulation
object
self.mesh :: Fidimag simulation mesh object
self.name :: Name of the chain method simulation
self.n_spins :: Number of spins per image
self.k :: Spring constant
self.variable_k(TEST) :: Set True to update k values acc to energy.
Need to specify self.dk var.
Default value: False
self.VTK :: Fidimag VTK object to save VTK files
self.files_convert_f :: Function to convert the coordinates from the
band to Cartesian coordinates
self.initial_images :: List with the initial images for the band
(Numpy arrays or space functions with the
magnetisation/spin field)
self.interpolations :: Optional List with integers indicating
interpolations between images
self.n_images :: Number of images in the band calculated from
the number of initial images and
interpolations between them
self.n_images_inner_band :: Number of images without considering the
images at the extremes
self.n_dofs_image :: Number of degrees of freedom per image, which
is the total number of spins multiplied by the
self.dof
self.n_band :: Total number of degrees of freedom in the
whole band
self.band :: The array containing all the degrees of
freedom (spin directions). It is ordered in
the XYZ format
self.gradientE :: Array with the components of the energy
gradient
self.G :: Array with the effective force from the NEB
method definition
self.tangents :: Array with the NEBM tangents
self.energies :: Array with the energies of every image in the
band (length = self.n_images)
self.spring_force :: Array with the spring force components
self.distances :: Array with the distances between adjacent
images: [0-1 1-2 2-3 .. etc ]
self.last_Y :: Array with the last computed band (like
self.band) from the integrator
self._material :: Array of Booleans of size (self.dof*n_spins),
i.e. the number of dofs in an image.
It is True where Ms or mu_s are larger than
zero. This is necessary to filter spins that
not need to be counted in, for example, a
scaled norm.
self.n_dofs_image_material :: Number of dofs where mu_s or Ms > 0
(in a single image)
climbing_image :: Any iterable with the indexes of the climbing
images. Climbing images are stored in the
self._climbing_image array, which the
self.climbing_image decorator populates
"""
def __init__(self,
sim,
initial_images,
interpolations=None,
spring_constant=1e5,
name='unnamed',
climbing_image=None,
dof=2,
openmp=False):
self.openmp = openmp
# Degrees of Freedom per spin
self.dof = dof
self.sim = sim
self.mesh = self.sim.mesh
self.name = name
# Number of spins in the system
self.n_spins = len(self.mesh.coordinates)
# We will use this filter to know which sites of the system has
# material, i.e. M_s or mu_s > 0 and norm(m) = 1
if self.sim._micromagnetic:
self._material = np.repeat(self.sim.Ms, self.dof) > 1e-10
else:
# We will assume, for now, that the magnetic moment in atomistic
# simulations is in units of mu_B
self._material = np.repeat(self.sim.mu_s / const.mu_B,
self.dof) > 1e-10
# self._material = self._material
# For C, we use 1 and 0s
self._material_int = np.copy(self._material).astype(np.int32)
self.n_dofs_image_material = np.sum(self._material)
# VTK saver for the magnetisation/spin field --------------------------
self.VTK = VTK(self.mesh,
directory='vtks'.format(self.name),
filename='image')
# Functions to convert the energy band coordinates to Cartesian
# coordinates when saving VTK and NPY files We assume Cartesian
# coordinates by default, i.e. we do not transform anything
self.files_convert_f = None
# Initial states ------------------------------------------------------
# We assume the extremes are fixed
self.initial_images = initial_images
if interpolations:
self.interpolations = interpolations
else:
self.interpolations = [0 for i in range(len(initial_images) - 1)]
# Number of images with/without the extremes
self.n_images = len(self.initial_images) + np.sum(self.interpolations)
self.n_images_inner_band = self.n_images - 2
# Number of degrees of freedom per image
self.n_dofs_image = (self.dof * self.n_spins)
# Total number of degrees of freedom in the string/band
self.n_band = self.n_images * self.n_dofs_image
# Spring constant -----------------------------------------------------
# Spring constant (we could use an array in the future)
self.k = spring_constant * np.ones(self.n_images)
# Set to True to update spring constant values relative to the energies
# (TESTING)
self.variable_k = False
self.dk = 1
# Climbing Image ------------------------------------------------------
# Set a list with the images where 1 is for climbing image and 0 for
# normal
self._climbing_image = np.zeros(self.n_images, dtype=np.int32)
if climbing_image is not None:
self.climbing_image = climbing_image
# Chain Method Arrays -------------------------------------------------
# We will initialise every array using the total number of images,
# but we must have in mind that the images at the extrema of the band
# are kept fixed, so we do not compute their gradient, tangents, etc.
# This might be not memory efficient but the code is understood better
# when we perform the loops when calculating the effective fields
# and forces
# The array containing every degree of freedom
self.band = np.zeros(self.n_band)
# The gradient with respect to the magnetisation (effective field)
self.gradientE = np.zeros_like(self.band)
# The effective force
self.G = np.zeros_like(self.band)
self.tangents = np.zeros_like(self.band)
self.energies = np.zeros(self.n_images)
self.spring_force = np.zeros_like(self.band)
self.distances = np.zeros(self.n_images - 1)
# Total distance starting from image_0
# (first element shoud always be zero)
self.path_distances = np.zeros(self.n_images)
self.last_Y = np.zeros_like(self.band)
# ---------------------------------------------------------------------
# If the integrator uses an LLG-like equation to relax the energy band
# we need to set this variable
# This variable only affects the StepIntegrators, NOT Sundials
self._llg_evolve = False
# ---------------------------------------------------------------------
# Factors for interpolating the energy band
# For now we only have a 3rd order polynomial interp, thus we set
# 4 factors
self.interp_factors = np.zeros((4, self.n_images))
# Somehow we need to rescale the gradient by the right units. In the
# case of micromag, we use mu0 * Ms, and for the atomistic case we
# simply use mu_s. This must be related to the way we derive the
# effective field to calculate the negative energy gradient, which is
# the functional derivative of the energy
if self.sim._micromagnetic:
self.scale = np.repeat(
self.mesh.dx * self.mesh.dy * self.mesh.dz *
(self.mesh.unit_length**3.) * const.mu_0 * self.sim.Ms, 3)
else:
self.scale = np.repeat(self.sim.mu_s, 3)
# ---------------------------------------------------------------------
self.G_log = []
# TODO: Move this property to the NEBM classes because they are only
# relevant to the NEBM and not the string method
@property
def climbing_image(self):
return self._climbing_image
@climbing_image.setter
def climbing_image(self, climbing_image_list):
"""
Set climbing images passing a list of image indexes
Negative indexes mean a falling image (total force = gradient only)
"""
self._climbing_image[:] = 0
images = range(self.n_images)[1:-1]
for ci in np.array([climbing_image_list]).flatten():
if abs(ci) in images:
if ci > 0:
self._climbing_image[ci] = 1
# Falling images are specified with negative index
elif ci < 0:
self._climbing_image[abs(ci)] = -1
else:
raise Exception(
'Cannot set image={} as climbing image'.format(ci))
@climbing_image.deleter
def climbing_image(self):
self._climbing_image[:] = 0
def initialise_energies(self):
pass
def save_VTKs(self, coordinates_function=None):
"""
Save VTK files in different folders, according to the simulation name
and step. Files are saved as vtks/simname_simstep_vtk/image_00000x.vtk
coordinates_function :: A function to transform the coordinates of
the band to Cartesian coordinates. For
example, in spherical coordinates we need
the spherical2cartesian function from
chain_method_tools
"""
# Create the directory
directory = "vtks/{}_{:05d}".format(self.name, self.iterations)
self.VTK.directory = directory
self.band.shape = (self.n_images, -1)
# We use Ms from the simulation assuming that all the images are the
# same
for i in range(self.n_images):
self.VTK.reset_data()
# We will try to save for the micromagnetic simulation (Ms) or an
# atomistic simulation (mu_s) TODO: maybe this can be done with an:
# isinstance
if self.sim._micromagnetic:
self.VTK.save_scalar(self.sim.Ms, name='M_s')
else:
self.VTK.save_scalar(self.sim.mu_s, name='mu_s')
if coordinates_function:
self.VTK.save_vector(coordinates_function(
self.band[i]).reshape(-1, 3),
name='spins')
else:
self.VTK.save_vector(self.band[i].reshape(-1, 3), name='spins')
self.VTK.write_file(step=i)
self.band.shape = (-1, )
def save_npys(self, coordinates_function=None):
"""
Save npy files in different folders according to
the simulation name and step
Files are saved as: npys/simname_simstep/image_x.npy
"""
# Create directory as simname_simstep
directory = 'npys/%s_%d' % (self.name, self.iterations)
if not os.path.exists(directory):
os.makedirs(directory)
# Save the images with the format: 'image_{}.npy'
# where {} is the image number, starting from 0
self.band.shape = (self.n_images, -1)
for i in range(self.n_images):
name = os.path.join(directory, 'image_{:06}.npy'.format(i))
if coordinates_function:
np.save(name, coordinates_function(self.band[i, :]))
else:
np.save(name, self.band[i])
self.band.shape = (-1)
def initialise_integrator(self,
integrator='sundials',
rtol=1e-6,
atol=1e-6):
self.t = 0
self.iterations = 0
self.ode_count = 1
if integrator == 'sundials':
if not self.openmp:
self.integrator = cvode.CvodeSolver(self.band,
self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
else:
self.integrator = cvode.CvodeSolver_OpenMP(
self.band, self.Sundials_RHS)
self.integrator.set_options(rtol, atol)
# elif integrator == 'scipy':
# self.integrator = ScipyIntegrator(self.band, self.step_RHS)
# self.integrator.set_options()
elif integrator == 'rk4' or integrator == 'euler':
self.integrator = StepIntegrator(self.band,
self.step_RHS,
step=self.integrator,
stepsize=1e-3)
self.integrator.set_options()
self._llg_evolve = True
elif integrator == 'verlet':
self.integrator = VerletIntegrator(
self.band, # y
self.G, # forces
self.step_RHS,
self.n_images,
self.n_dofs_image,
mass=1,
stepsize=1e-4)
self.integrator.set_options()
# In Verlet algorithm we only use the total force G and not YxYxG:
self._llg_evolve = False
else:
raise Exception('No valid integrator specified. Available: '
'"sundials", "scipy"')
def create_tablewriter(self):
entities_energy = {
'step': {
'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'
},
'energy': {
'unit': '<J>',
'get': lambda sim: sim.energies,
'header': ['image_%d' % i for i in range(self.n_images)]
}
}
self.tablewriter = DataSaver(self,
'%s_energy.ndt' % (self.name),
entities=entities_energy)
entities_dm = {
'step': {
'unit': '<1>',
'get': lambda sim: sim.iterations,
'header': 'iterations'
},
'dYs': {
'unit':
'<1>',
'get':
lambda sim: sim.distances,
'header':
['image_%d_%d' % (i, i + 1) for i in range(self.n_images - 1)]
}
}
self.tablewriter_dm = DataSaver(self,
'%s_dYs.ndt' % (self.name),
entities=entities_dm)
# ---------------------------------------------------------------------
def generate_initial_band(self):
pass
def compute_effective_field_and_energy(self, y):
pass
# NEBM only:
# def compute_tangents(self, y):
# pass
# def compute_spring_force(self, y):
# pass
def compute_distances(self):
pass
# -------------------------------------------------------------------------
# CVODE solver ------------------------------------------------------------
# -------------------------------------------------------------------------
def compute_norms(self, A, B):
"""
Compute the norms of the difference between corresponding images of the
bands A and B. Every norm is scaled by the number of degrees of freedom
The norm is computed only using mesh/lattice sites with material, i.e.
mu_s or Ms > 0
"""
A_minus_B = A - B
A_minus_B.shape = (-1, self.n_dofs_image)
A_minus_B = np.apply_along_axis(
lambda y: compute_norm(y[self._material], scale=True),
axis=1,
arr=A_minus_B)
return A_minus_B.reshape(-1)
def step_RHS(self, t, y):
"""
The RHS of the ODE to be solved for the Chain Method
This function is specified for the Scipy integrator
"""
pass
def Sundials_RHS(self, t, y, ydot):
"""
This function is called on every iteration of the integrator (CVODE
solver). ydot refers to the Right Hand Side of the equation, since
we are solving dy/dt = 0
"""
pass
def compute_maximum_dYdt(self, A, B, dt):
"""
Compute the maximum difference from the images of the *A* array with
the images of the *B* array, divided by dt
The differences are not computed for the images at the extremes, since
these images are fixed and do not change with the iterator
For instance, in spherical coordinates, if we have a band of (N + 1)
images, labeled from 0 to N, we start by
dY = [A1_theta0 A1_phi0 A1_theta1 ... A(N-1)_theta0 A(N-1)_phi0 A(N-1)_theta1 ... ]
- [B1_theta0 B1_phi0 B1_theta1 ... B(N-1)_theta0 B(N-1)_phi0 B(N-1)_theta1 ... ]
where A(i)_theta(j) is the theta componenet of the j-th spin of the
i-th image in the band
Then we calculate the norm of every difference (the norm is computed
only using mesh/lattice sites with material, i.e. mu_s or Ms > 0):
||dY|| = [ || dY1_theta0 dY1_phi0 dY1_theta1 ... ||
...
|| dY(N-1)_theta0 dY(N-1)_phi0 dY(N-1)_theta1 ... ||
]
We divide by dt:
||dY|| -- > || dY || / dt = dYdt
and finally take the maximum value of dYdt
"""
# We will not consider the images at the extremes to compute dY.
# Since we removed the extremes, we only have *n_images_inner_band*
# images
band_no_extremes = slice(self.n_dofs_image, -self.n_dofs_image)
dYdt = self.compute_norms(A[band_no_extremes],
B[band_no_extremes]).reshape(
self.n_images_inner_band, -1)
dYdt /= dt
if np.max(dYdt) > 0:
return np.max(dYdt)
else:
return 0
def run_until(self, t):
if (t) <= self.t:
return
self.integrator.run_until(t)
# Copy the updated energy band to our local array
self.band[:] = self.integrator.y[:]
# Compute the maximum change in the integrator step
max_dYdt = self.compute_maximum_dYdt(self.integrator.y, self.last_Y,
t - self.t)
self.last_Y[:] = self.band[:]
# Update the current step
self.t = t
return max_dYdt
def relax(self,
dt=1e-8,
stopping_dYdt=1,
max_iterations=1000,
save_npys_every=100,
save_vtks_every=100,
save_initial_state=True):
"""
Relax the energy band according to the specified integrator
Sundials: dt is the initial stepsize which is updated
by CVODE. Number of calls is given by CVODE
StepIntegrators: dt is the relaxation stepsize. These integrators
evolve using an internal `stepsize`, thus the
number of evaluations is dt / stepsize.
You can update the integrator evolve step using:
self.integrator.stepsize = 1e-4
"""
log.debug("Relaxation parameters: "
"stopping_dYdt={}, "
"time_step={} s, "
" max_iterations={}.".format(stopping_dYdt, dt,
max_iterations))
if save_initial_state:
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# Save the initial state i=0 in the data table
# Update self.distances and self.path_distances:
self.compute_distances()
self.tablewriter.save()
self.tablewriter_dm.save()
INNER_DOFS = slice(self.n_dofs_image, -self.n_dofs_image)
for i in range(max_iterations):
# Update the iterations number counter
self.iterations += 1
# Get the current size of the time discretisation from the
# integrator (variable step size)
try:
cvode_dt = self.integrator.get_current_step()
except AttributeError:
cvode_dt = dt
# If the step size of the integrator is larger than the specified
# discretisation, use the current integrator step size to
# compute the next iteration. Otherwise, just stick to the
# specified step
if cvode_dt > dt:
increment_dt = cvode_dt
else:
increment_dt = dt
# Integrator steps: ***********************************************
# This can be redefined according to the chosen chain method
max_dYdt = self.run_until(self.t + increment_dt)
# *****************************************************************
# Update the current step
self.t = self.t + increment_dt
# Save data -------------------------------------------------------
if self.iterations % save_vtks_every == 0:
self.save_VTKs(coordinates_function=self.files_convert_f)
if self.iterations % save_npys_every == 0:
self.save_npys(coordinates_function=self.files_convert_f)
self.compute_distances()
self.tablewriter.save()
self.tablewriter_dm.save()
# Print information about the simulation and the forces.
# The last two terms are the largest gradient and spring
# force norms from the spins (not counting the extrema)
G_norms = np.linalg.norm(self.G[INNER_DOFS].reshape(-1, 3), axis=1)
gradE_norms = np.linalg.norm(self.gradientE[INNER_DOFS].reshape(
-1, 3),
axis=1)
Fk_norms = np.linalg.norm(self.spring_force[INNER_DOFS].reshape(
-1, 3),
axis=1)
# For DEBUGGING purposes: -----------------------------------------
# mean_G_norms_per_image = np.mean(G_norms.reshape(self.n_images - 2, -1), axis=1)
# print(mean_G_norms_per_image)
# gradE_dot_t = np.einsum('ij,ij->i',
# self.gradientE.reshape(self.n_images, -1),
# self.tangents.reshape(self.n_images, -1))
# gradE_perp = (self.gradientE.reshape(self.n_images, -1)
# - np.einsum('i,ij->ij', gradE_dot_t,
# self.tangents.reshape(self.n_images, -1))
# )
# print(np.max(gradE_perp))
# print(np.max(gradE_norms.reshape(self.n_images - 2, -1), axis=1))
# print(np.max(Fk_norms.reshape(self.n_images - 2, -1), axis=1))
# -----------------------------------------------------------------
log.debug(
time.strftime("%Y-%m-%d %H:%M:%S ", time.localtime()) +
"step: {:.6g}, step_size: {:.3g}, "
"max dYdt: {:.3g} "
"max|G|: {:.3g} "
"max|gradE|: {:.3g} "
"and max|F_k|: {:.3g}".format(
self.iterations, increment_dt, max_dYdt, np.max(G_norms),
np.max(gradE_norms), np.max(Fk_norms)))
self.G_log.append(np.max(G_norms))
# -----------------------------------------------------------------
# Stop criteria:
if max_dYdt < stopping_dYdt:
break
np.savetxt(self.name + '_G_log.txt', np.array(self.G_log))
log.info("Relaxation finished at time step = {:.4g}, "
"t = {:.2g}, call rhs = {:.4g} "
"and max_dYdt = {:.3g}".format(self.iterations, self.t,
self.ode_count, max_dYdt))
self.save_VTKs(coordinates_function=self.files_convert_f)
self.save_npys(coordinates_function=self.files_convert_f)
# -------------------------------------------------------------------------
# Interpolations
def compute_polynomial_factors(self, compute_fields=True):
"""
Compute a smooth approximation for the band, using a third order
polynomial approximation. Prefactors are stored in the
self.interp_factors array
This approximation uses the tangents and derivatives from each image of
the band as information to estimate the curvatures. The formula can be
found in
- Bessarab et al., Computer Physics Communications 196 (2015) 335-347
"""
# To be sure, update the effective field and tangents when calling
# this function (this is necesary if we call the function without
# relaxing the band before)
if compute_fields:
self.compute_effective_field_and_energy(self.band)
self.compute_tangents(self.band)
self.compute_distances()
deltas = np.zeros(self.n_images)
for i in range(self.n_images):
deltas[i] = np.dot(
self.scale * (self.gradientE).reshape(self.n_images, -1)[i],
self.tangents.reshape(self.n_images, -1)[i])
ds = self.path_distances
E = self.energies
# The coefficients for the polynomial approximation
self.interp_factors[2][:] = deltas
self.interp_factors[3][:] = E
# Populate the a and b coefficients for every image
for i in range(self.n_images - 1):
# a factor
self.interp_factors[0][i] = (deltas[i + 1] +
deltas[i]) / (ds[i + 1] - ds[i])**2.
self.interp_factors[0][i] -= 2 * (E[i + 1] - E[i]) / (ds[i + 1] -
ds[i])**3.
# b factor
self.interp_factors[1][i] = -(deltas[i + 1] + 2 * deltas[i]) / (
ds[i + 1] - ds[i])
self.interp_factors[1][i] += 3 * (E[i + 1] - E[i]) / (ds[i + 1] -
ds[i])**2.
def compute_polynomial_approximation_energy(self, n_points):
"""
Compute a smooth approximation of the energy band, using a third order
polynomial approximation. The pre-factors in the polynomial are
computed from the self.compute_polynomial_factors function
This function returns a tuple with two elements:
0. An array with the distance of every data point from the 0th
image.
1. A n_points long array, with the cubic interpolated energy band
ARGUMENTS
n_points :: The number of points for the interpolation
"""
ds = self.path_distances
# The arrays with the data points and the interpolated energy values
x = np.linspace(0, ds[-1], n_points)
E_interp = np.array(
[self._compute_polynomial_approximation_energy(i) for i in x])
return x, E_interp
def _compute_polynomial_approximation_energy(self, x):
"""
Return interpolated energy value for a point x
"""
ds = self.path_distances
if x < 0.0 or x > ds[-1]:
raise Exception('x lies outside the valid interpolation range')
# Find index of the ds array for the value that is closest to x
ds_idx = np.abs(x - ds).argmin()
# If x is smaller than the given ds, use the previous ds value so
# that we use ds(i) when x lies in the interval ds(i) < x < ds(i+1)
if x < ds[ds_idx]:
ds_idx -= 1
E_interp = (self.interp_factors[0][ds_idx] * ((x - ds[ds_idx])**3.) +
self.interp_factors[1][ds_idx] * ((x - ds[ds_idx])**2.) +
self.interp_factors[2][ds_idx] * ((x - ds[ds_idx])) +
self.interp_factors[3][ds_idx])
return E_interp
# -------------------------------------------------------------------------
def compute_Bernstein_polynomials(self, compute_fields=True):
"""
Compute Bernstein polynomials to approximate the the energy curve.
The functions are stored in the self.Bernstein_polynomials variable
These polynomials can be used with the
self.compute_Bernstein_approximation_energy method
"""
# To be sure, update the effective field and tangents when calling
# this function (this is necesary if we call the function without
# relaxing the band before)
if compute_fields:
self.compute_effective_field_and_energy(self.band)
self.compute_tangents(self.band)
self.compute_distances()
derivatives = np.zeros(self.n_images)
for i in range(self.n_images):
derivatives[i] = np.dot(
self.scale * (self.gradientE).reshape(self.n_images, -1)[i],
self.tangents.reshape(self.n_images, -1)[i])
E = self.energies
# The coefficients for the polynomial approximation
# self.interp_factors[0][:] = E
# self.interp_factors[1][:] = deltas
# Store the polynomial functions
self.Bernstein_polynomials = []
for i, ds in enumerate(self.distances):
self.Bernstein_polynomials.append(
si.BPoly.from_derivatives(
[self.path_distances[i], self.path_distances[i + 1]],
[[E[i], derivatives[i]], [E[i + 1], derivatives[i + 1]]]))
def _compute_Bernstein_approximation_energy(self, x):
"""
Return interpolated energy value for a point x
"""
ds = self.path_distances
if x < 0.0 or x > ds[-1]:
raise Exception('x lies outside the valid interpolation range')
elif x == 0.0:
return self.distances[0]
elif x == ds[-1]:
return self.distances[-1]
# Find index of the ds array for the value that is closest to x
ds_idx = np.abs(x - ds).argmin()
# If x is smaller than the given ds, use the previous ds value so
# that we use ds(i) when x lies in the interval ds(i) < x < ds(i+1)
if x < ds[ds_idx]:
ds_idx -= 1
return self.Bernstein_polynomials[ds_idx](x)
def compute_Bernstein_approximation_energy(self, n_points):
"""
Return interpolated energy values of the energy band with n_points
resolution
"""
ds = self.path_distances
# The arrays with the data points and the interpolated energy values
x = np.linspace(0, ds[-1], n_points)
E_interp = np.zeros_like(x)
E_interp[0] = self.energies[0]
E_interp[-1] = self.energies[-1]
E_interp[1:-1] = np.array(
[self._compute_Bernstein_approximation_energy(i) for i in x[1:-1]])
return x, E_interp
return E_interp
def test_save_vector_field(tmpdir):
mesh = CuboidMesh(4, 3, 2, 4, 3, 2)
s = vector_field(mesh, lambda r: (r[0], r[1], r[2]))
vtk = VTK(mesh, directory=str(tmpdir), filename="save_vector")
vtk.save_vector(s, name="s")
assert same_as_ref(vtk.write_file(), REF_DIR)
