def fit_model(guess, data, dbf, comps, phases, **kwargs):
"""
Fit model parameters to input data based on an initial guess for parameters.
Parameters
----------
guess : dict
Parameter names to fit with initial guesses.
data : list of DataFrames
Input data to fit.
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : list
Names of components to consider in the calculation.
phases : list
Names of phases to consider in the calculation.
Returns
-------
(Dictionary of fit key:values), (lmfit minimize result)
Examples
--------
None yet.
"""
if 'maxfev' not in kwargs:
kwargs['maxfev'] = 100
fit_params = lmfit.Parameters()
for guess_name, guess_value in guess.items():
fit_params.add(guess_name, value=guess_value)
param_names = fit_params.valuesdict().keys()
fit_models = {name: Model(dbf, comps, name) for name in phases}
fit_variables = dict()
for name, mod in fit_models.items():
fit_variables[name] = sorted(mod.energy.atoms(v.StateVariable).union({v.T, v.P}), key=str)
# Extra factor '1e-100...' is to work around an annoying broadcasting bug for zero gradient entries
fit_models[name].models['_broadcaster'] = 1e-100 * sympy.Mul(*fit_variables[name]) ** 3
callables = {name: make_callable(mod.ast,
itertools.chain(param_names, fit_variables[name]))
for name, mod in fit_models.items()}
grad_callables = {name: make_callable(sympy.Matrix([mod.ast]).jacobian(fit_variables[name]),
itertools.chain(param_names, fit_variables[name]))
for name, mod in fit_models.items()}
#out = leastsq(residual_equilibrium, param_values,
# args=(data, dbf, comps, fit_models, callables),
# full_output=True, **kwargs)
out = lmfit.minimize(residual_thermochemical, fit_params,
args=(data, dbf, comps, fit_models, callables, grad_callables))
#fit = residual_equilibrium(data, dbf, comps, fit_models, callables)
#ssq = np.linalg.norm(residual_equilibrium(out[0], data, dbf, comps,
# fit_models, callables))
#return dict(zip(param_names, out[0])), out[1:]
return fit_params.valuesdict(), out
def calculate_output(model, variables, output, mode='sympy'):
"""
Calculate the value of the energy at a point.
Parameters
----------
model, Model
Energy model for a phase.
variables, dict
Dictionary of all input variables.
output : str
String of the property to calculate, e.g. 'ast'
mode, ['numpy', 'sympy'], optional
Optimization method for the abstract syntax tree.
"""
# Generate a callable function
# Normally we would use model.subs(variables) here, but we want to ensure
# our optimization functions are working.
prop = make_callable(getattr(model, output),
list(variables.keys()),
mode=mode)
# Unpack all the values in the dict and use them to call the function
return prop(*(list(variables.values())))
def setup_dataset(file_obj, dbf, params, mode=None):
# params should be a list of pymc variables corresponding to parameters
data = json.load(file_obj)
if data['solver']['mode'] != 'manual':
raise NotImplemented
fit_models = {name: Model(dbf, data['components'], name) for name in data['phases']}
param_vars = []
for key, mod in fit_models.items():
param_vars.extend(sorted(set(mod.ast.atoms(Symbol)) - set(mod.variables), key=str))
param_vars = sorted(param_vars, key=str)
if len(params) != len(param_vars):
raise ValueError('Input parameter vector length doesn\'t match the free parameters'
' in the phase models: {0} != {1}'.format(len(params), len(param_vars)))
indep_vars = [v.P, v.T]
site_fracs = {key: sorted(mod.ast.atoms(v.SiteFraction), key=str) for key, mod in fit_models.items()}
# Call this from a routine that pulls in all datasets and generates the variance vars + Potentials
callables = {name: make_callable(getattr(mod, data['output']),
itertools.chain(param_vars, indep_vars, site_fracs[name]), mode=mode)
for name, mod in fit_models.items()}
extra_conds = OrderedDict({key: np.atleast_1d(value) for key, value in data['conditions'].items()})
exp_values = xarray.DataArray(np.array(data['values'], dtype=np.float),
dims=list(extra_conds.keys())+['points'], coords=extra_conds)
def compute_error(*args):
prefill_callables = {key: functools.partial(*itertools.chain([func], args[:len(params)]))
for key, func in callables.items()}
result = calculate(dbf, data['components'], data['phases'], output=data['output'],
points=np.atleast_2d(data['solver']['sublattice_configuration']).astype(np.float),
callables=prefill_callables, model=fit_models, **extra_conds)
# Eliminate data below 300 K for now
error = (result[data['output']] - exp_values).sel(T=slice(300, None)).values.flatten()
return error
def compute_values(*args):
prefill_callables = {key: functools.partial(*itertools.chain([func], args[:len(params)]))
for key, func in callables.items()}
result = calculate(dbf, data['components'], data['phases'], output=data['output'],
points=np.atleast_2d(data['solver']['sublattice_configuration']).astype(np.float),
callables=prefill_callables, model=fit_models, **extra_conds)
return result
return compute_error, compute_values, exp_values, data
def residual_thermochemical(fit_params, input_data, dbf, comps, mods, callables, grad_callables):
"Return an array with the residuals for thermochemical data in 'input_data'."
global_comps = [x.upper() for x in sorted(comps) if x != 'VA']
param_names = fit_params.valuesdict().keys()
parvalues = fit_params.valuesdict().values()
# Prefill parameter values before passing to energy calculator
iter_callables = {name: functools.partial(func, *parvalues)
for name, func in callables.items()}
iter_grad_callables = {name: functools.partial(func, *parvalues)
for name, func in grad_callables.items()}
res = np.zeros(len(input_data))
# TODO: This should definitely be vectorized
# It will probably require an update to equilibrium()
for idx, row in input_data.iterrows():
conditions = dict()
if 'T' in row:
conditions[v.T] = row['T']
if 'P' in row:
conditions[v.P] = row['P']
for comp in global_comps[:-1]:
conditions[v.X(comp)] = row['X('+comp+')']
statevars = dict((str(key), value) for key, value in conditions.items() if key in [v.T, v.P])
eq = equilibrium(dbf, comps, row['Phase'], conditions,
model=mods, callables=iter_callables,
grad_callables=iter_grad_callables, verbose=False)
# TODO: Support for miscibility gaps, i.e., FCC_A1#1 specification
eq_values = eq['Y'].sel(vertex=0).values
#print(eq_values)
# TODO: All the needed 'Types' should be precalculated and looked up
variables = sorted(mods[row['Phase']].energy.atoms(v.StateVariable).union({v.T, v.P}), key=str)
output_callables = {row['Phase']: functools.partial(make_callable(getattr(mods[row['Phase']],
row['Type']),
itertools.chain(param_names, variables)),
*parvalues)}
calculated_value = calculate(dbf, comps, row['Phase'], output=row['Type'],
model=mods, callables=output_callables,
points=eq_values, **statevars)
res[idx] = float(row['Value']) - float(calculated_value[row['Type']].values)
#print('res', idx, res[idx])
return res
def calculate_energy(model, variables, mode='numpy'):
"""
Calculate the value of the energy at a point.
Parameters
----------
model, Model
Energy model for a phase.
variables, dict
Dictionary of all input variables.
mode, ['numpy', 'sympy'], optional
Optimization method for the abstract syntax tree.
"""
# Generate a callable energy function
# Normally we would use model.subs(variables) here, but we want to ensure
# our optimization functions are working.
energy = make_callable(model.ast, list(variables.keys()), mode=mode)
# Unpack all the values in the dict and use them to call the function
return energy(*(list(variables.values())))
def calculate_energy(model, variables, mode="numpy"):
"""
Calculate the value of the energy at a point.
Parameters
----------
model, Model
Energy model for a phase.
variables, dict
Dictionary of all input variables.
mode, ['numpy', 'sympy'], optional
Optimization method for the abstract syntax tree.
"""
# Generate a callable energy function
# Normally we would use model.subs(variables) here, but we want to ensure
# our optimization functions are working.
energy = make_callable(model.ast, list(variables.keys()), mode=mode)
# Unpack all the values in the dict and use them to call the function
return energy(*(list(variables.values())))
def calculate_output(model, variables, output, mode='sympy'):
"""
Calculate the value of the energy at a point.
Parameters
----------
model, Model
Energy model for a phase.
variables, dict
Dictionary of all input variables.
output : str
String of the property to calculate, e.g. 'ast'
mode, ['numpy', 'sympy'], optional
Optimization method for the abstract syntax tree.
"""
# Generate a callable function
# Normally we would use model.subs(variables) here, but we want to ensure
# our optimization functions are working.
prop = make_callable(getattr(model, output), list(variables.keys()), mode=mode)
# Unpack all the values in the dict and use them to call the function
return prop(*(list(variables.values())))
def calculate(dbf,
comps,
phases,
mode=None,
output='GM',
fake_points=False,
**kwargs):
"""
Sample the property surface of 'output' containing the specified
components and phases. Model parameters are taken from 'dbf' and any
state variables (T, P, etc.) can be specified as keyword arguments.
Parameters
----------
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : str or sequence
Names of components to consider in the calculation.
phases : str or sequence
Names of phases to consider in the calculation.
mode : string, optional
See 'make_callable' docstring for details.
output : string, optional
Model attribute to sample.
fake_points : bool, optional (Default: False)
If True, the first few points of the output surface will be fictitious
points used to define an equilibrium hyperplane guaranteed to be above
all the other points. This is used for convex hull computations.
points : ndarray or a dict of phase names to ndarray, optional
Columns of ndarrays must be internal degrees of freedom (site fractions), sorted.
If this is not specified, points will be generated automatically.
pdens : int, a dict of phase names to int, or a seq of both, optional
Number of points to sample per degree of freedom.
model : Model, a dict of phase names to Model, or a seq of both, optional
Model class to use for each phase.
Returns
-------
xray.Dataset of the sampled attribute as a function of state variables
Examples
--------
None yet.
"""
# Here we check for any keyword arguments that are special, i.e.,
# there may be keyword arguments that aren't state variables
pdens_dict = unpack_kwarg(kwargs.pop('pdens', 2000), default_arg=2000)
points_dict = unpack_kwarg(kwargs.pop('points', None), default_arg=None)
model_dict = unpack_kwarg(kwargs.pop('model', Model), default_arg=Model)
callable_dict = unpack_kwarg(kwargs.pop('callables', None),
default_arg=None)
if isinstance(phases, str):
phases = [phases]
if isinstance(comps, str):
comps = [comps]
components = [x for x in sorted(comps) if not x.startswith('VA')]
# Convert keyword strings to proper state variable objects
# If we don't do this, sympy will get confused during substitution
statevar_dict = collections.OrderedDict((v.StateVariable(key), unpack_condition(value)) \
for (key, value) in sorted(kwargs.items()))
str_statevar_dict = collections.OrderedDict((str(key), unpack_condition(value)) \
for (key, value) in statevar_dict.items())
all_phase_data = []
comp_sets = {}
largest_energy = -np.inf
maximum_internal_dof = 0
# Consider only the active phases
active_phases = dict((name.upper(), dbf.phases[name.upper()]) \
for name in unpack_phases(phases))
for phase_name, phase_obj in sorted(active_phases.items()):
# Build the symbolic representation of the energy
mod = model_dict[phase_name]
# if this is an object type, we need to construct it
if isinstance(mod, type):
try:
model_dict[phase_name] = mod = mod(dbf, comps, phase_name)
except DofError:
# we can't build the specified phase because the
# specified components aren't found in every sublattice
# we'll just skip it
logger.warning(
"""Suspending specified phase %s due to
some sublattices containing only unspecified components""",
phase_name)
continue
if points_dict[phase_name] is None:
try:
out = getattr(mod, output)
maximum_internal_dof = max(maximum_internal_dof,
len(out.atoms(v.SiteFraction)))
except AttributeError:
raise AttributeError(
'Missing Model attribute {0} specified for {1}'.format(
output, mod.__class__))
else:
maximum_internal_dof = max(
maximum_internal_dof,
np.asarray(points_dict[phase_name]).shape[-1])
for phase_name, phase_obj in sorted(active_phases.items()):
try:
mod = model_dict[phase_name]
except KeyError:
continue
# Construct an ordered list of the variables
variables, sublattice_dof = generate_dof(phase_obj, mod.components)
# Build the "fast" representation of that model
if callable_dict[phase_name] is None:
out = getattr(mod, output)
# As a last resort, treat undefined symbols as zero
# But warn the user when we do this
# This is consistent with TC's behavior
undefs = list(out.atoms(Symbol) - out.atoms(v.StateVariable))
for undef in undefs:
out = out.xreplace({undef: float(0)})
logger.warning(
'Setting undefined symbol %s for phase %s to zero', undef,
phase_name)
comp_sets[phase_name] = make_callable(out, \
list(statevar_dict.keys()) + variables, mode=mode)
else:
comp_sets[phase_name] = callable_dict[phase_name]
points = points_dict[phase_name]
if points is None:
# Eliminate pure vacancy endmembers from the calculation
vacancy_indices = list()
for idx, sublattice in enumerate(phase_obj.constituents):
active_in_subl = sorted(
set(phase_obj.constituents[idx]).intersection(comps))
if 'VA' in active_in_subl and 'VA' in sorted(comps):
vacancy_indices.append(active_in_subl.index('VA'))
if len(vacancy_indices) != len(phase_obj.constituents):
vacancy_indices = None
logger.debug('vacancy_indices: %s', vacancy_indices)
# Add all endmembers to guarantee their presence
points = endmember_matrix(sublattice_dof,
vacancy_indices=vacancy_indices)
# Sample composition space for more points
if sum(sublattice_dof) > len(sublattice_dof):
points = np.concatenate(
(points,
point_sample(sublattice_dof,
pdof=pdens_dict[phase_name])))
# If there are nontrivial sublattices with vacancies in them,
# generate a set of points where their fraction is zero and renormalize
for idx, sublattice in enumerate(phase_obj.constituents):
if 'VA' in set(sublattice) and len(sublattice) > 1:
var_idx = variables.index(
v.SiteFraction(phase_name, idx, 'VA'))
addtl_pts = np.copy(points)
# set vacancy fraction to log-spaced between 1e-10 and 1e-6
addtl_pts[:, var_idx] = np.power(
10.0, -10.0 * (1.0 - addtl_pts[:, var_idx]))
# renormalize site fractions
cur_idx = 0
for ctx in sublattice_dof:
end_idx = cur_idx + ctx
addtl_pts[:, cur_idx:end_idx] /= \
addtl_pts[:, cur_idx:end_idx].sum(axis=1)[:, None]
cur_idx = end_idx
# add to points matrix
points = np.concatenate((points, addtl_pts), axis=0)
# Filter out nan's that may have slipped in if we sampled too high a vacancy concentration
# Issues with this appear to be platform-dependent
points = points[~np.isnan(points).any(axis=-1)]
# Ensure that points has the correct dimensions and dtype
points = np.atleast_2d(np.asarray(points, dtype=np.float))
phase_ds = _compute_phase_values(phase_obj, components, variables,
str_statevar_dict, points,
comp_sets[phase_name], output,
maximum_internal_dof)
# largest_energy is really only relevant if fake_points is set
if fake_points:
largest_energy = max(phase_ds[output].max(), largest_energy)
all_phase_data.append(phase_ds)
if fake_points:
if output != 'GM':
raise ValueError(
'fake_points=True should only be used with output=\'GM\'')
phase_ds = _generate_fake_points(components, statevar_dict,
largest_energy, output,
maximum_internal_dof)
final_ds = xray.concat(itertools.chain([phase_ds], all_phase_data),
dim='points')
else:
# speedup for single-phase case (found by profiling)
if len(all_phase_data) > 1:
final_ds = xray.concat(all_phase_data, dim='points')
else:
final_ds = all_phase_data[0]
if (not fake_points) and (len(all_phase_data) == 1):
pass
else:
# Reset the points dimension to use a single global index
final_ds['points'] = np.arange(len(final_ds.points))
return final_ds
def equilibrium(dbf, comps, phases, conditions, **kwargs):
"""
Calculate the equilibrium state of a system containing the specified
components and phases, under the specified conditions.
Model parameters are taken from 'dbf'.
Parameters
----------
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : list
Names of components to consider in the calculation.
phases : list or dict
Names of phases to consider in the calculation.
conditions : dict or (list of dict)
StateVariables and their corresponding value.
verbose : bool, optional (Default: True)
Show progress of calculations.
grid_opts : dict, optional
Keyword arguments to pass to the initial grid routine.
Returns
-------
Structured equilibrium calculation.
Examples
--------
None yet.
"""
active_phases = unpack_phases(phases) or sorted(dbf.phases.keys())
comps = sorted(comps)
indep_vars = ['T', 'P']
grid_opts = kwargs.pop('grid_opts', dict())
verbose = kwargs.pop('verbose', True)
phase_records = dict()
callable_dict = kwargs.pop('callables', dict())
grad_callable_dict = kwargs.pop('grad_callables', dict())
points_dict = dict()
maximum_internal_dof = 0
# Construct models for each phase; prioritize user models
models = unpack_kwarg(kwargs.pop('model', Model), default_arg=Model)
if verbose:
print('Components:', ' '.join(comps))
print('Phases:', end=' ')
for name in active_phases:
mod = models[name]
if isinstance(mod, type):
models[name] = mod = mod(dbf, comps, name)
variables = sorted(mod.energy.atoms(v.StateVariable).union(
{key
for key in conditions.keys() if key in [v.T, v.P]}),
key=str)
site_fracs = sorted(mod.energy.atoms(v.SiteFraction), key=str)
maximum_internal_dof = max(maximum_internal_dof, len(site_fracs))
# Extra factor '1e-100...' is to work around an annoying broadcasting bug for zero gradient entries
models[name].models['_broadcaster'] = 1e-100 * Mul(*variables)**3
out = models[name].energy
if name not in callable_dict:
undefs = list(out.atoms(Symbol) - out.atoms(v.StateVariable))
for undef in undefs:
out = out.xreplace({undef: float(0)})
# callable_dict takes variables in a different order due to calculate() pecularities
callable_dict[name] = make_callable(
out,
sorted((key for key in conditions.keys() if key in [v.T, v.P]),
key=str) + site_fracs)
if name not in grad_callable_dict:
grad_func = make_callable(
Matrix([out]).jacobian(variables), variables)
else:
grad_func = grad_callable_dict[name]
# Adjust gradient by the approximate chemical potentials
plane_vars = sorted(models[name].energy.atoms(v.SiteFraction), key=str)
hyperplane = Add(*[
v.MU(i) * mole_fraction(dbf.phases[name], comps, i) for i in comps
if i != 'VA'
])
# Workaround an annoying bug with zero gradient entries
# This forces numerically zero entries to broadcast correctly
hyperplane += 1e-100 * Mul(*([v.MU(i) for i in comps if i != 'VA'] +
plane_vars + [v.T, v.P]))**3
plane_grad = make_callable(
Matrix([hyperplane]).jacobian(variables),
[v.MU(i) for i in comps if i != 'VA'] + plane_vars + [v.T, v.P])
plane_hess = make_callable(hessian(hyperplane, variables),
[v.MU(i) for i in comps if i != 'VA'] +
plane_vars + [v.T, v.P])
phase_records[name.upper()] = PhaseRecord(variables=variables,
grad=grad_func,
plane_grad=plane_grad,
plane_hess=plane_hess)
if verbose:
print(name, end=' ')
if verbose:
print('[done]', end='\n')
conds = OrderedDict((key, unpack_condition(value))
for key, value in sorted(conditions.items(), key=str))
str_conds = OrderedDict((str(key), value) for key, value in conds.items())
indep_vals = list([float(x) for x in np.atleast_1d(val)]
for key, val in str_conds.items() if key in indep_vars)
components = [x for x in sorted(comps) if not x.startswith('VA')]
# 'calculate' accepts conditions through its keyword arguments
grid_opts.update(
{key: value
for key, value in str_conds.items() if key in indep_vars})
if 'pdens' not in grid_opts:
grid_opts['pdens'] = 10
coord_dict = str_conds.copy()
coord_dict['vertex'] = np.arange(len(components))
grid_shape = np.meshgrid(*coord_dict.values(), indexing='ij',
sparse=False)[0].shape
coord_dict['component'] = components
if verbose:
print('Computing initial grid', end=' ')
grid = calculate(dbf,
comps,
active_phases,
output='GM',
model=models,
callables=callable_dict,
fake_points=True,
**grid_opts)
if verbose:
print('[{0} points, {1}]'.format(len(grid.points),
sizeof_fmt(grid.nbytes)),
end='\n')
properties = xray.Dataset(
{
'NP': (list(str_conds.keys()) + ['vertex'], np.empty(grid_shape)),
'GM': (list(str_conds.keys()), np.empty(grid_shape[:-1])),
'MU':
(list(str_conds.keys()) + ['component'], np.empty(grid_shape)),
'points': (list(str_conds.keys()) + ['vertex'],
np.empty(grid_shape, dtype=np.int))
},
coords=coord_dict,
attrs={'iterations': 1},
)
# Store the potentials from the previous iteration
current_potentials = properties.MU.copy()
for iteration in range(MAX_ITERATIONS):
if verbose:
print('Computing convex hull [iteration {}]'.format(
properties.attrs['iterations']))
# lower_convex_hull will modify properties
lower_convex_hull(grid, properties)
progress = np.abs(current_potentials - properties.MU).max().values
if verbose:
print('progress', progress)
if progress < MIN_PROGRESS:
if verbose:
print('Convergence achieved')
break
current_potentials[...] = properties.MU.values
if verbose:
print('Refining convex hull')
# Insert extra dimensions for non-T,P conditions so GM broadcasts correctly
energy_broadcast_shape = grid.GM.values.shape[:len(indep_vals)] + \
(1,) * (len(str_conds) - len(indep_vals)) + (grid.GM.values.shape[-1],)
driving_forces = np.einsum('...i,...i',
properties.MU.values[..., np.newaxis, :],
grid.X.values[np.index_exp[...] +
(np.newaxis,) * (len(str_conds) - len(indep_vals)) +
np.index_exp[:, :]]) - \
grid.GM.values.view().reshape(energy_broadcast_shape)
for name in active_phases:
dof = len(models[name].energy.atoms(v.SiteFraction))
current_phase_indices = (grid.Phase.values == name
).reshape(energy_broadcast_shape[:-1] +
(-1, ))
# Broadcast to capture all conditions
current_phase_indices = np.broadcast_arrays(
current_phase_indices, np.empty(driving_forces.shape))[0]
# This reshape is safe as long as phases have the same number of points at all indep. conditions
current_phase_driving_forces = driving_forces[
current_phase_indices].reshape(
current_phase_indices.shape[:-1] + (-1, ))
# Note: This works as long as all points are in the same phase order for all T, P
current_site_fractions = grid.Y.values[..., current_phase_indices[
(0, ) * len(str_conds)], :]
if np.sum(
current_site_fractions[(0, ) *
len(indep_vals)][..., :dof]) == dof:
# All site fractions are 1, aka zero internal degrees of freedom
# Impossible to refine these points, so skip this phase
points_dict[name] = current_site_fractions[
(0, ) * len(indep_vals)][..., :dof]
continue
# Find the N points with largest driving force for a given set of conditions
# Remember that driving force has a sign, so we want the "most positive" values
# N is the number of components, in this context
# N points define a 'best simplex' for every set of conditions
# We also need to restrict ourselves to one phase at a time
trial_indices = np.argpartition(current_phase_driving_forces,
-len(components),
axis=-1)[..., -len(components):]
trial_indices = trial_indices.ravel()
statevar_indices = np.unravel_index(
np.arange(
np.multiply.reduce(properties.GM.values.shape +
(len(components), ))),
properties.GM.values.shape +
(len(components), ))[:len(indep_vals)]
points = current_site_fractions[np.index_exp[statevar_indices +
(trial_indices, )]]
points.shape = properties.points.shape[:-1] + (
-1, maximum_internal_dof)
# The Y arrays have been padded, so we should slice off the padding
points = points[..., :dof]
# Workaround for derivative issues at endmembers
points[points == 0.] = MIN_SITE_FRACTION
if len(points) == 0:
if name in points_dict:
del points_dict[name]
# No nearly stable points: skip this phase
continue
num_vars = len(phase_records[name].variables)
plane_grad = phase_records[name].plane_grad
plane_hess = phase_records[name].plane_hess
statevar_grid = np.meshgrid(*itertools.chain(indep_vals),
sparse=True,
indexing='xy')
# TODO: A more sophisticated treatment of constraints
num_constraints = len(indep_vals) + len(
dbf.phases[name].sublattices)
constraint_jac = np.zeros((num_constraints, num_vars))
# Independent variables are always fixed (in this limited implementation)
for idx in range(len(indep_vals)):
constraint_jac[idx, idx] = 1
# This is for site fraction balance constraints
var_idx = len(indep_vals)
for idx in range(len(dbf.phases[name].sublattices)):
active_in_subl = set(
dbf.phases[name].constituents[idx]).intersection(comps)
constraint_jac[len(indep_vals) + idx,
var_idx:var_idx + len(active_in_subl)] = 1
var_idx += len(active_in_subl)
grad = phase_records[name].grad(
*itertools.chain(statevar_grid, points.T))
if grad.dtype == 'object':
# Workaround a bug in zero gradient entries
grad_zeros = np.zeros(points.T.shape[1:], dtype=np.float)
for i in np.arange(grad.shape[0]):
if isinstance(grad[i], int):
grad[i] = grad_zeros
grad = np.array(grad.tolist(), dtype=np.float)
bcasts = np.broadcast_arrays(
*itertools.chain(properties.MU.values.T, points.T))
cast_grad = -plane_grad(*itertools.chain(bcasts, [0], [0]))
cast_grad = cast_grad.T + grad.T
grad = cast_grad
grad.shape = grad.shape[:-1] # Remove extraneous dimension
# This Hessian is an approximation updated using the BFGS method
# See Nocedal and Wright, ch.3, p. 198
# Initialize as identity matrix
hess = broadcast_to(np.eye(num_vars),
grad.shape + (grad.shape[-1], )).copy()
newton_iteration = 0
while newton_iteration < MAX_NEWTON_ITERATIONS:
e_matrix = np.linalg.inv(hess)
dy_unconstrained = -np.einsum('...ij,...j->...i', e_matrix,
grad)
proj_matrix = np.dot(e_matrix, constraint_jac.T)
inv_matrix = np.rollaxis(np.dot(constraint_jac, proj_matrix),
0, -1)
inv_term = np.linalg.inv(inv_matrix)
first_term = np.einsum('...ij,...jk->...ik', proj_matrix,
inv_term)
# Normally a term for the residual here
# We only choose starting points which obey the constraints, so r = 0
cons_summation = np.einsum('...i,...ji->...j',
dy_unconstrained, constraint_jac)
cons_correction = np.einsum('...ij,...j->...i', first_term,
cons_summation)
dy_constrained = dy_unconstrained - cons_correction
# TODO: Support for adaptive changing independent variable steps
new_direction = dy_constrained[..., len(indep_vals):]
# Backtracking line search
new_points = points + INITIAL_STEP_SIZE * new_direction
alpha = np.full(new_points.shape[:-1],
INITIAL_STEP_SIZE,
dtype=np.float)
negative_points = np.any(new_points < 0., axis=-1)
while np.any(negative_points):
alpha[negative_points] *= 0.1
new_points = points + alpha[...,
np.newaxis] * new_direction
negative_points = np.any(new_points < 0., axis=-1)
# If we made "near" zero progress on any points, don't update the Hessian until
# we've rebuilt the convex hull
# Nocedal and Wright recommend against skipping Hessian updates
# They recommend using a damped update approach, pp. 538-539 of their book
# TODO: Check the projected gradient norm, not the step length
if np.any(
np.max(np.abs(alpha[..., np.newaxis] * new_direction),
axis=-1) < MIN_STEP_LENGTH):
break
# Workaround for derivative issues at endmembers
new_points[new_points == 0.] = 1e-16
# BFGS update to Hessian
new_grad = phase_records[name].grad(
*itertools.chain(statevar_grid, new_points.T))
if new_grad.dtype == 'object':
# Workaround a bug in zero gradient entries
grad_zeros = np.zeros(new_points.T.shape[1:],
dtype=np.float)
for i in np.arange(new_grad.shape[0]):
if isinstance(new_grad[i], int):
new_grad[i] = grad_zeros
new_grad = np.array(new_grad.tolist(), dtype=np.float)
bcasts = np.broadcast_arrays(
*itertools.chain(properties.MU.values.T, new_points.T))
cast_grad = -plane_grad(*itertools.chain(bcasts, [0], [0]))
cast_grad = cast_grad.T + new_grad.T
new_grad = cast_grad
new_grad.shape = new_grad.shape[:
-1] # Remove extraneous dimension
# Notation used here consistent with Nocedal and Wright
s_k = np.empty(points.shape[:-1] +
(points.shape[-1] + len(indep_vals), ))
# Zero out independent variable changes for now
s_k[..., :len(indep_vals)] = 0
s_k[..., len(indep_vals):] = new_points - points
y_k = new_grad - grad
s_s_term = np.einsum('...j,...k->...jk', s_k, s_k)
s_b_s_term = np.einsum('...i,...ij,...j', s_k, hess, s_k)
y_y_y_s_term = np.einsum('...j,...k->...jk', y_k, y_k) / \
np.einsum('...i,...i', y_k, s_k)[..., np.newaxis, np.newaxis]
update = np.einsum('...ij,...jk,...kl->...il', hess, s_s_term, hess) / \
s_b_s_term[..., np.newaxis, np.newaxis] + y_y_y_s_term
hess = hess - update
cast_hess = -plane_hess(
*itertools.chain(bcasts, [0], [0])).T + hess
hess = -cast_hess #TODO: Why does this fix things?
# TODO: Verify that the chosen step lengths reduce the energy
points = new_points
grad = new_grad
newton_iteration += 1
new_points = new_points.reshape(
new_points.shape[:len(indep_vals)] +
(-1, new_points.shape[-1]))
new_points = np.concatenate(
(current_site_fractions[..., :dof], new_points), axis=-2)
points_dict[name] = new_points
if verbose:
print('Rebuilding grid', end=' ')
grid = calculate(dbf,
comps,
active_phases,
output='GM',
model=models,
callables=callable_dict,
fake_points=True,
points=points_dict,
**grid_opts)
if verbose:
print('[{0} points, {1}]'.format(len(grid.points),
sizeof_fmt(grid.nbytes)),
end='\n')
properties.attrs['iterations'] += 1
# One last call to ensure 'properties' and 'grid' are consistent with one another
lower_convex_hull(grid, properties)
ravelled_X_view = grid['X'].values.view().reshape(
-1, grid['X'].values.shape[-1])
ravelled_Y_view = grid['Y'].values.view().reshape(
-1, grid['Y'].values.shape[-1])
ravelled_Phase_view = grid['Phase'].values.view().reshape(-1)
# Copy final point values from the grid and drop the index array
# For some reason direct construction doesn't work. We have to create empty and then assign.
properties['X'] = xray.DataArray(
np.empty_like(ravelled_X_view[properties['points'].values]),
dims=properties['points'].dims + ('component', ))
properties['X'].values[...] = ravelled_X_view[properties['points'].values]
properties['Y'] = xray.DataArray(
np.empty_like(ravelled_Y_view[properties['points'].values]),
dims=properties['points'].dims + ('internal_dof', ))
properties['Y'].values[...] = ravelled_Y_view[properties['points'].values]
# TODO: What about invariant reactions? We should perform a final driving force calculation here.
# We can handle that in the same post-processing step where we identify single-phase regions.
properties['Phase'] = xray.DataArray(np.empty_like(
ravelled_Phase_view[properties['points'].values]),
dims=properties['points'].dims)
properties['Phase'].values[...] = ravelled_Phase_view[
properties['points'].values]
del properties['points']
return properties
def setup_dataset(file_obj, dbf, params, mode=None):
# params should be a list of pymc variables corresponding to parameters
data = json.load(file_obj)
if data['solver']['mode'] != 'manual':
raise NotImplemented
fit_models = {
name: Model(dbf, data['components'], name)
for name in data['phases']
}
param_vars = []
for key, mod in fit_models.items():
param_vars.extend(
sorted(set(mod.ast.atoms(Symbol)) - set(mod.variables), key=str))
param_vars = sorted(param_vars, key=str)
if len(params) != len(param_vars):
raise ValueError(
'Input parameter vector length doesn\'t match the free parameters'
' in the phase models: {0} != {1}'.format(len(params),
len(param_vars)))
indep_vars = [v.P, v.T]
site_fracs = {
key: sorted(mod.ast.atoms(v.SiteFraction), key=str)
for key, mod in fit_models.items()
}
# Call this from a routine that pulls in all datasets and generates the variance vars + Potentials
callables = {
name: make_callable(getattr(mod, data['output']),
itertools.chain(param_vars, indep_vars,
site_fracs[name]),
mode=mode)
for name, mod in fit_models.items()
}
extra_conds = OrderedDict({
key: np.atleast_1d(value)
for key, value in data['conditions'].items()
})
exp_values = xray.DataArray(np.array(data['values'], dtype=np.float),
dims=list(extra_conds.keys()) + ['points'],
coords=extra_conds)
def compute_error(*args):
prefill_callables = {
key:
functools.partial(*itertools.chain([func], args[:len(params)]))
for key, func in callables.items()
}
result = calculate(
dbf,
data['components'],
data['phases'],
output=data['output'],
points=np.atleast_2d(
data['solver']['sublattice_configuration']).astype(np.float),
callables=prefill_callables,
model=fit_models,
**extra_conds)
# Eliminate data below 300 K for now
error = (result[data['output']] -
exp_values).sel(T=slice(300, None)).values.flatten()
return error
def compute_values(*args):
prefill_callables = {
key:
functools.partial(*itertools.chain([func], args[:len(params)]))
for key, func in callables.items()
}
result = calculate(
dbf,
data['components'],
data['phases'],
output=data['output'],
points=np.atleast_2d(
data['solver']['sublattice_configuration']).astype(np.float),
callables=prefill_callables,
model=fit_models,
**extra_conds)
return result
return compute_error, compute_values, exp_values, data
def calculate(dbf, comps, phases, mode=None, output="GM", fake_points=False, **kwargs):
"""
Sample the property surface of 'output' containing the specified
components and phases. Model parameters are taken from 'dbf' and any
state variables (T, P, etc.) can be specified as keyword arguments.
Parameters
----------
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : str or sequence
Names of components to consider in the calculation.
phases : str or sequence
Names of phases to consider in the calculation.
mode : string, optional
See 'make_callable' docstring for details.
output : string, optional
Model attribute to sample.
fake_points : bool, optional (Default: False)
If True, the first few points of the output surface will be fictitious
points used to define an equilibrium hyperplane guaranteed to be above
all the other points. This is used for convex hull computations.
points : ndarray or a dict of phase names to ndarray, optional
Columns of ndarrays must be internal degrees of freedom (site fractions), sorted.
If this is not specified, points will be generated automatically.
pdens : int, a dict of phase names to int, or a seq of both, optional
Number of points to sample per degree of freedom.
model : Model, a dict of phase names to Model, or a seq of both, optional
Model class to use for each phase.
Returns
-------
xray.Dataset of the sampled attribute as a function of state variables
Examples
--------
None yet.
"""
# Here we check for any keyword arguments that are special, i.e.,
# there may be keyword arguments that aren't state variables
pdens_dict = unpack_kwarg(kwargs.pop("pdens", 2000), default_arg=2000)
points_dict = unpack_kwarg(kwargs.pop("points", None), default_arg=None)
model_dict = unpack_kwarg(kwargs.pop("model", Model), default_arg=Model)
callable_dict = unpack_kwarg(kwargs.pop("callables", None), default_arg=None)
if isinstance(phases, str):
phases = [phases]
if isinstance(comps, str):
comps = [comps]
components = [x for x in sorted(comps) if not x.startswith("VA")]
# Convert keyword strings to proper state variable objects
# If we don't do this, sympy will get confused during substitution
statevar_dict = collections.OrderedDict(
(v.StateVariable(key), unpack_condition(value)) for (key, value) in sorted(kwargs.items())
)
str_statevar_dict = collections.OrderedDict(
(str(key), unpack_condition(value)) for (key, value) in statevar_dict.items()
)
all_phase_data = []
comp_sets = {}
largest_energy = -np.inf
maximum_internal_dof = 0
# Consider only the active phases
active_phases = dict((name.upper(), dbf.phases[name.upper()]) for name in unpack_phases(phases))
for phase_name, phase_obj in sorted(active_phases.items()):
# Build the symbolic representation of the energy
mod = model_dict[phase_name]
# if this is an object type, we need to construct it
if isinstance(mod, type):
try:
model_dict[phase_name] = mod = mod(dbf, comps, phase_name)
except DofError:
# we can't build the specified phase because the
# specified components aren't found in every sublattice
# we'll just skip it
logger.warning(
"""Suspending specified phase %s due to
some sublattices containing only unspecified components""",
phase_name,
)
continue
if points_dict[phase_name] is None:
try:
out = getattr(mod, output)
maximum_internal_dof = max(maximum_internal_dof, len(out.atoms(v.SiteFraction)))
except AttributeError:
raise AttributeError("Missing Model attribute {0} specified for {1}".format(output, mod.__class__))
else:
maximum_internal_dof = max(maximum_internal_dof, np.asarray(points_dict[phase_name]).shape[-1])
for phase_name, phase_obj in sorted(active_phases.items()):
try:
mod = model_dict[phase_name]
except KeyError:
continue
# Construct an ordered list of the variables
variables, sublattice_dof = generate_dof(phase_obj, mod.components)
# Build the "fast" representation of that model
if callable_dict[phase_name] is None:
out = getattr(mod, output)
# As a last resort, treat undefined symbols as zero
# But warn the user when we do this
# This is consistent with TC's behavior
undefs = list(out.atoms(Symbol) - out.atoms(v.StateVariable))
for undef in undefs:
out = out.xreplace({undef: float(0)})
logger.warning("Setting undefined symbol %s for phase %s to zero", undef, phase_name)
comp_sets[phase_name] = make_callable(out, list(statevar_dict.keys()) + variables, mode=mode)
else:
comp_sets[phase_name] = callable_dict[phase_name]
points = points_dict[phase_name]
if points is None:
# Eliminate pure vacancy endmembers from the calculation
vacancy_indices = list()
for idx, sublattice in enumerate(phase_obj.constituents):
active_in_subl = sorted(set(phase_obj.constituents[idx]).intersection(comps))
if "VA" in active_in_subl and "VA" in sorted(comps):
vacancy_indices.append(active_in_subl.index("VA"))
if len(vacancy_indices) != len(phase_obj.constituents):
vacancy_indices = None
logger.debug("vacancy_indices: %s", vacancy_indices)
# Add all endmembers to guarantee their presence
points = endmember_matrix(sublattice_dof, vacancy_indices=vacancy_indices)
# Sample composition space for more points
if sum(sublattice_dof) > len(sublattice_dof):
points = np.concatenate((points, point_sample(sublattice_dof, pdof=pdens_dict[phase_name])))
# If there are nontrivial sublattices with vacancies in them,
# generate a set of points where their fraction is zero and renormalize
for idx, sublattice in enumerate(phase_obj.constituents):
if "VA" in set(sublattice) and len(sublattice) > 1:
var_idx = variables.index(v.SiteFraction(phase_name, idx, "VA"))
addtl_pts = np.copy(points)
# set vacancy fraction to log-spaced between 1e-10 and 1e-6
addtl_pts[:, var_idx] = np.power(10.0, -10.0 * (1.0 - addtl_pts[:, var_idx]))
# renormalize site fractions
cur_idx = 0
for ctx in sublattice_dof:
end_idx = cur_idx + ctx
addtl_pts[:, cur_idx:end_idx] /= addtl_pts[:, cur_idx:end_idx].sum(axis=1)[:, None]
cur_idx = end_idx
# add to points matrix
points = np.concatenate((points, addtl_pts), axis=0)
# Filter out nan's that may have slipped in if we sampled too high a vacancy concentration
# Issues with this appear to be platform-dependent
points = points[~np.isnan(points).any(axis=-1)]
# Ensure that points has the correct dimensions and dtype
points = np.atleast_2d(np.asarray(points, dtype=np.float))
phase_ds = _compute_phase_values(
phase_obj,
components,
variables,
str_statevar_dict,
points,
comp_sets[phase_name],
output,
maximum_internal_dof,
)
# largest_energy is really only relevant if fake_points is set
if fake_points:
largest_energy = max(phase_ds[output].max(), largest_energy)
all_phase_data.append(phase_ds)
if fake_points:
if output != "GM":
raise ValueError("fake_points=True should only be used with output='GM'")
phase_ds = _generate_fake_points(components, statevar_dict, largest_energy, output, maximum_internal_dof)
final_ds = xray.concat(itertools.chain([phase_ds], all_phase_data), dim="points")
else:
# speedup for single-phase case (found by profiling)
if len(all_phase_data) > 1:
final_ds = xray.concat(all_phase_data, dim="points")
else:
final_ds = all_phase_data[0]
if (not fake_points) and (len(all_phase_data) == 1):
pass
else:
# Reset the points dimension to use a single global index
final_ds["points"] = np.arange(len(final_ds.points))
return final_ds
def fit_model(guess, data, dbf, comps, phases, **kwargs):
"""
Fit model parameters to input data based on an initial guess for parameters.
Parameters
----------
guess : dict
Parameter names to fit with initial guesses.
data : list of DataFrames
Input data to fit.
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : list
Names of components to consider in the calculation.
phases : list
Names of phases to consider in the calculation.
Returns
-------
(Dictionary of fit key:values), (lmfit minimize result)
Examples
--------
None yet.
"""
if 'maxfev' not in kwargs:
kwargs['maxfev'] = 100
fit_params = lmfit.Parameters()
for guess_name, guess_value in guess.items():
fit_params.add(guess_name, value=guess_value)
param_names = fit_params.valuesdict().keys()
fit_models = {name: Model(dbf, comps, name) for name in phases}
fit_variables = dict()
for name, mod in fit_models.items():
fit_variables[name] = sorted(mod.energy.atoms(v.StateVariable).union(
{v.T, v.P}),
key=str)
# Extra factor '1e-100...' is to work around an annoying broadcasting bug for zero gradient entries
fit_models[name].models['_broadcaster'] = 1e-100 * sympy.Mul(
*fit_variables[name])**3
callables = {
name: make_callable(mod.ast,
itertools.chain(param_names, fit_variables[name]))
for name, mod in fit_models.items()
}
grad_callables = {
name: make_callable(
sympy.Matrix([mod.ast]).jacobian(fit_variables[name]),
itertools.chain(param_names, fit_variables[name]))
for name, mod in fit_models.items()
}
#out = leastsq(residual_equilibrium, param_values,
# args=(data, dbf, comps, fit_models, callables),
# full_output=True, **kwargs)
out = lmfit.minimize(residual_thermochemical,
fit_params,
args=(data, dbf, comps, fit_models, callables,
grad_callables))
#fit = residual_equilibrium(data, dbf, comps, fit_models, callables)
#ssq = np.linalg.norm(residual_equilibrium(out[0], data, dbf, comps,
# fit_models, callables))
#return dict(zip(param_names, out[0])), out[1:]
return fit_params.valuesdict(), out
def residual_thermochemical(fit_params, input_data, dbf, comps, mods,
callables, grad_callables):
"Return an array with the residuals for thermochemical data in 'input_data'."
global_comps = [x.upper() for x in sorted(comps) if x != 'VA']
param_names = fit_params.valuesdict().keys()
parvalues = fit_params.valuesdict().values()
# Prefill parameter values before passing to energy calculator
iter_callables = {
name: functools.partial(func, *parvalues)
for name, func in callables.items()
}
iter_grad_callables = {
name: functools.partial(func, *parvalues)
for name, func in grad_callables.items()
}
res = np.zeros(len(input_data))
# TODO: This should definitely be vectorized
# It will probably require an update to equilibrium()
for idx, row in input_data.iterrows():
conditions = dict()
if 'T' in row:
conditions[v.T] = row['T']
if 'P' in row:
conditions[v.P] = row['P']
for comp in global_comps[:-1]:
conditions[v.X(comp)] = row['X(' + comp + ')']
statevars = dict((str(key), value)
for key, value in conditions.items()
if key in [v.T, v.P])
eq = equilibrium(dbf,
comps,
row['Phase'],
conditions,
model=mods,
callables=iter_callables,
grad_callables=iter_grad_callables,
verbose=False)
# TODO: Support for miscibility gaps, i.e., FCC_A1#1 specification
eq_values = eq['Y'].sel(vertex=0).values
#print(eq_values)
# TODO: All the needed 'Types' should be precalculated and looked up
variables = sorted(mods[row['Phase']].energy.atoms(
v.StateVariable).union({v.T, v.P}),
key=str)
output_callables = {
row['Phase']:
functools.partial(
make_callable(getattr(mods[row['Phase']], row['Type']),
itertools.chain(param_names, variables)),
*parvalues)
}
calculated_value = calculate(dbf,
comps,
row['Phase'],
output=row['Type'],
model=mods,
callables=output_callables,
points=eq_values,
**statevars)
res[idx] = float(row['Value']) - float(
calculated_value[row['Type']].values)
#print('res', idx, res[idx])
return res
def energy_surf(dbf, comps, phases, mode=None, output='GM', **kwargs):
"""
Sample the property surface of 'output' containing the specified
components and phases. Model parameters are taken from 'dbf' and any
state variables (T, P, etc.) can be specified as keyword arguments.
Parameters
----------
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : list
Names of components to consider in the calculation.
phases : list
Names of phases to consider in the calculation.
mode : string, optional
See 'make_callable' docstring for details.
output : string, optional
Model attribute to sample.
pdens : int, a dict of phase names to int, or a list of both, optional
Number of points to sample per degree of freedom.
model : Model, a dict of phase names to Model, or a list of both, optional
Model class to use for each phase.
Returns
-------
DataFrame of the output as a function of composition, temperature, etc.
Examples
--------
None yet.
"""
warnings.warn('Use pycalphad.calculate() instead', DeprecationWarning, stacklevel=2)
# Here we check for any keyword arguments that are special, i.e.,
# there may be keyword arguments that aren't state variables
pdens_dict = unpack_kwarg(kwargs.pop('pdens', 2000), default_arg=2000)
model_dict = unpack_kwarg(kwargs.pop('model', Model), default_arg=Model)
callable_dict = unpack_kwarg(kwargs.pop('callables', None), default_arg=None)
# Convert keyword strings to proper state variable objects
# If we don't do this, sympy will get confused during substitution
statevar_dict = \
collections.OrderedDict((v.StateVariable(key), value) \
for (key, value) in sorted(kwargs.items()))
# Generate all combinations of state variables for 'map' calculation
# Wrap single values of state variables in lists
# Use 'kwargs' because we want state variable names to be stringified
statevar_values = [_listify(val) for val in statevar_dict.values()]
statevars_to_map = np.array(list(itertools.product(*statevar_values)))
# Consider only the active phases
active_phases = dict((name.upper(), dbf.phases[name.upper()]) \
for name in phases)
comp_sets = {}
# Construct a list to hold all the data
all_phase_data = []
for phase_name, phase_obj in sorted(active_phases.items()):
# Build the symbolic representation of the energy
mod = model_dict[phase_name]
# if this is an object type, we need to construct it
if isinstance(mod, type):
try:
mod = mod(dbf, comps, phase_name)
except DofError:
# we can't build the specified phase because the
# specified components aren't found in every sublattice
# we'll just skip it
logger.warning("""Suspending specified phase %s due to
some sublattices containing only unspecified components""",
phase_name)
continue
try:
out = getattr(mod, output)
except AttributeError:
raise AttributeError('Missing Model attribute {0} specified for {1}'
.format(output, mod.__class__))
# Construct an ordered list of the variables
variables, sublattice_dof = generate_dof(phase_obj, mod.components)
site_ratios = list(phase_obj.sublattices)
# Build the "fast" representation of that model
if callable_dict[phase_name] is None:
# As a last resort, treat undefined symbols as zero
# But warn the user when we do this
# This is consistent with TC's behavior
undefs = list(out.atoms(Symbol) - out.atoms(v.StateVariable))
for undef in undefs:
out = out.xreplace({undef: float(0)})
logger.warning('Setting undefined symbol %s for phase %s to zero',
undef, phase_name)
comp_sets[phase_name] = make_callable(out, \
list(statevar_dict.keys()) + variables, mode=mode)
else:
comp_sets[phase_name] = callable_dict[phase_name]
# Eliminate pure vacancy endmembers from the calculation
vacancy_indices = list()
for idx, sublattice in enumerate(phase_obj.constituents):
if 'VA' in sorted(sublattice) and 'VA' in sorted(comps):
vacancy_indices.append(sorted(sublattice).index('VA'))
if len(vacancy_indices) != len(phase_obj.constituents):
vacancy_indices = None
logger.debug('vacancy_indices: %s', vacancy_indices)
# Add all endmembers to guarantee their presence
points = endmember_matrix(sublattice_dof,
vacancy_indices=vacancy_indices)
# Sample composition space for more points
if sum(sublattice_dof) > len(sublattice_dof):
points = np.concatenate((points,
point_sample(sublattice_dof,
pdof=pdens_dict[phase_name])
))
# If there are nontrivial sublattices with vacancies in them,
# generate a set of points where their fraction is zero and renormalize
for idx, sublattice in enumerate(phase_obj.constituents):
if 'VA' in set(sublattice) and len(sublattice) > 1:
var_idx = variables.index(v.SiteFraction(phase_name, idx, 'VA'))
addtl_pts = np.copy(points)
# set vacancy fraction to log-spaced between 1e-10 and 1e-6
addtl_pts[:, var_idx] = np.power(10.0, -10.0*(1.0 - addtl_pts[:, var_idx]))
# renormalize site fractions
cur_idx = 0
for ctx in sublattice_dof:
end_idx = cur_idx + ctx
addtl_pts[:, cur_idx:end_idx] /= \
addtl_pts[:, cur_idx:end_idx].sum(axis=1)[:, None]
cur_idx = end_idx
# add to points matrix
points = np.concatenate((points, addtl_pts), axis=0)
data_dict = {'Phase': phase_name}
# Broadcast compositions and state variables along orthogonal axes
# This lets us eliminate an expensive Python loop
data_dict[output] = \
comp_sets[phase_name](*itertools.chain(
np.transpose(statevars_to_map[:, :, np.newaxis], (1, 2, 0)),
np.transpose(points[:, :, np.newaxis], (1, 0, 2)))).T.ravel()
# Save state variables, with values indexed appropriately
statevar_vals = np.repeat(statevars_to_map, len(points), axis=0).T
data_dict.update({str(statevar): vals for statevar, vals \
in zip(statevar_dict.keys(), statevar_vals)})
# Map the internal degrees of freedom to global coordinates
# Normalize site ratios by the sum of site ratios times a factor
# related to the site fraction of vacancies
site_ratio_normalization = np.zeros(len(points))
for idx, sublattice in enumerate(phase_obj.constituents):
vacancy_column = np.ones(len(points))
if 'VA' in set(sublattice):
var_idx = variables.index(v.SiteFraction(phase_name, idx, 'VA'))
vacancy_column -= points[:, var_idx]
site_ratio_normalization += site_ratios[idx] * vacancy_column
for comp in sorted(comps):
if comp == 'VA':
continue
avector = [float(vxx.species == comp) * \
site_ratios[vxx.sublattice_index] for vxx in variables]
data_dict['X('+comp+')'] = np.tile(np.divide(np.dot(
points[:, :], avector), site_ratio_normalization),
statevars_to_map.shape[0])
# Copy coordinate information into data_dict
# TODO: Is there a more memory-efficient way to deal with this?
# Perhaps with hierarchical indexing...
var_fmt = 'Y({0},{1},{2})'
data_dict.update({var_fmt.format(vxx.phase_name, vxx.sublattice_index,
vxx.species): \
np.tile(vals, statevars_to_map.shape[0]) \
for vxx, vals in zip(variables, points.T)})
all_phase_data.append(pd.DataFrame(data_dict))
# all_phases_data now contains energy surface information for the system
return pd.concat(all_phase_data, axis=0, join='outer', \
ignore_index=True, verify_integrity=False)
def equilibrium(dbf, comps, phases, conditions, **kwargs):
"""
Calculate the equilibrium state of a system containing the specified
components and phases, under the specified conditions.
Model parameters are taken from 'dbf'.
Parameters
----------
dbf : Database
Thermodynamic database containing the relevant parameters.
comps : list
Names of components to consider in the calculation.
phases : list or dict
Names of phases to consider in the calculation.
conditions : dict or (list of dict)
StateVariables and their corresponding value.
verbose : bool, optional (Default: True)
Show progress of calculations.
grid_opts : dict, optional
Keyword arguments to pass to the initial grid routine.
Returns
-------
Structured equilibrium calculation.
Examples
--------
None yet.
"""
active_phases = unpack_phases(phases) or sorted(dbf.phases.keys())
comps = sorted(comps)
indep_vars = ['T', 'P']
grid_opts = kwargs.pop('grid_opts', dict())
verbose = kwargs.pop('verbose', True)
phase_records = dict()
callable_dict = kwargs.pop('callables', dict())
grad_callable_dict = kwargs.pop('grad_callables', dict())
points_dict = dict()
maximum_internal_dof = 0
# Construct models for each phase; prioritize user models
models = unpack_kwarg(kwargs.pop('model', Model), default_arg=Model)
if verbose:
print('Components:', ' '.join(comps))
print('Phases:', end=' ')
for name in active_phases:
mod = models[name]
if isinstance(mod, type):
models[name] = mod = mod(dbf, comps, name)
variables = sorted(mod.energy.atoms(v.StateVariable).union({key for key in conditions.keys() if key in [v.T, v.P]}), key=str)
site_fracs = sorted(mod.energy.atoms(v.SiteFraction), key=str)
maximum_internal_dof = max(maximum_internal_dof, len(site_fracs))
# Extra factor '1e-100...' is to work around an annoying broadcasting bug for zero gradient entries
models[name].models['_broadcaster'] = 1e-100 * Mul(*variables) ** 3
out = models[name].energy
if name not in callable_dict:
undefs = list(out.atoms(Symbol) - out.atoms(v.StateVariable))
for undef in undefs:
out = out.xreplace({undef: float(0)})
# callable_dict takes variables in a different order due to calculate() pecularities
callable_dict[name] = make_callable(out,
sorted((key for key in conditions.keys() if key in [v.T, v.P]),
key=str) + site_fracs)
if name not in grad_callable_dict:
grad_func = make_callable(Matrix([out]).jacobian(variables), variables)
else:
grad_func = grad_callable_dict[name]
# Adjust gradient by the approximate chemical potentials
plane_vars = sorted(models[name].energy.atoms(v.SiteFraction), key=str)
hyperplane = Add(*[v.MU(i)*mole_fraction(dbf.phases[name], comps, i)
for i in comps if i != 'VA'])
# Workaround an annoying bug with zero gradient entries
# This forces numerically zero entries to broadcast correctly
hyperplane += 1e-100 * Mul(*([v.MU(i) for i in comps if i != 'VA'] + plane_vars + [v.T, v.P])) ** 3
plane_grad = make_callable(Matrix([hyperplane]).jacobian(variables),
[v.MU(i) for i in comps if i != 'VA'] + plane_vars + [v.T, v.P])
plane_hess = make_callable(hessian(hyperplane, variables),
[v.MU(i) for i in comps if i != 'VA'] + plane_vars + [v.T, v.P])
phase_records[name.upper()] = PhaseRecord(variables=variables,
grad=grad_func,
plane_grad=plane_grad,
plane_hess=plane_hess)
if verbose:
print(name, end=' ')
if verbose:
print('[done]', end='\n')
conds = OrderedDict((key, unpack_condition(value)) for key, value in sorted(conditions.items(), key=str))
str_conds = OrderedDict((str(key), value) for key, value in conds.items())
indep_vals = list([float(x) for x in np.atleast_1d(val)]
for key, val in str_conds.items() if key in indep_vars)
components = [x for x in sorted(comps) if not x.startswith('VA')]
# 'calculate' accepts conditions through its keyword arguments
grid_opts.update({key: value for key, value in str_conds.items() if key in indep_vars})
if 'pdens' not in grid_opts:
grid_opts['pdens'] = 10
coord_dict = str_conds.copy()
coord_dict['vertex'] = np.arange(len(components))
grid_shape = np.meshgrid(*coord_dict.values(),
indexing='ij', sparse=False)[0].shape
coord_dict['component'] = components
if verbose:
print('Computing initial grid', end=' ')
grid = calculate(dbf, comps, active_phases, output='GM',
model=models, callables=callable_dict, fake_points=True, **grid_opts)
if verbose:
print('[{0} points, {1}]'.format(len(grid.points), sizeof_fmt(grid.nbytes)), end='\n')
properties = xray.Dataset({'NP': (list(str_conds.keys()) + ['vertex'],
np.empty(grid_shape)),
'GM': (list(str_conds.keys()),
np.empty(grid_shape[:-1])),
'MU': (list(str_conds.keys()) + ['component'],
np.empty(grid_shape)),
'points': (list(str_conds.keys()) + ['vertex'],
np.empty(grid_shape, dtype=np.int))
},
coords=coord_dict,
attrs={'iterations': 1},
)
# Store the potentials from the previous iteration
current_potentials = properties.MU.copy()
for iteration in range(MAX_ITERATIONS):
if verbose:
print('Computing convex hull [iteration {}]'.format(properties.attrs['iterations']))
# lower_convex_hull will modify properties
lower_convex_hull(grid, properties)
progress = np.abs(current_potentials - properties.MU).max().values
if verbose:
print('progress', progress)
if progress < MIN_PROGRESS:
if verbose:
print('Convergence achieved')
break
current_potentials[...] = properties.MU.values
if verbose:
print('Refining convex hull')
# Insert extra dimensions for non-T,P conditions so GM broadcasts correctly
energy_broadcast_shape = grid.GM.values.shape[:len(indep_vals)] + \
(1,) * (len(str_conds) - len(indep_vals)) + (grid.GM.values.shape[-1],)
driving_forces = np.einsum('...i,...i',
properties.MU.values[..., np.newaxis, :],
grid.X.values[np.index_exp[...] +
(np.newaxis,) * (len(str_conds) - len(indep_vals)) +
np.index_exp[:, :]]) - \
grid.GM.values.view().reshape(energy_broadcast_shape)
for name in active_phases:
dof = len(models[name].energy.atoms(v.SiteFraction))
current_phase_indices = (grid.Phase.values == name).reshape(energy_broadcast_shape[:-1] + (-1,))
# Broadcast to capture all conditions
current_phase_indices = np.broadcast_arrays(current_phase_indices,
np.empty(driving_forces.shape))[0]
# This reshape is safe as long as phases have the same number of points at all indep. conditions
current_phase_driving_forces = driving_forces[current_phase_indices].reshape(
current_phase_indices.shape[:-1] + (-1,))
# Note: This works as long as all points are in the same phase order for all T, P
current_site_fractions = grid.Y.values[..., current_phase_indices[(0,) * len(str_conds)], :]
if np.sum(current_site_fractions[(0,) * len(indep_vals)][..., :dof]) == dof:
# All site fractions are 1, aka zero internal degrees of freedom
# Impossible to refine these points, so skip this phase
points_dict[name] = current_site_fractions[(0,) * len(indep_vals)][..., :dof]
continue
# Find the N points with largest driving force for a given set of conditions
# Remember that driving force has a sign, so we want the "most positive" values
# N is the number of components, in this context
# N points define a 'best simplex' for every set of conditions
# We also need to restrict ourselves to one phase at a time
trial_indices = np.argpartition(current_phase_driving_forces,
-len(components), axis=-1)[..., -len(components):]
trial_indices = trial_indices.ravel()
statevar_indices = np.unravel_index(np.arange(np.multiply.reduce(properties.GM.values.shape + (len(components),))),
properties.GM.values.shape + (len(components),))[:len(indep_vals)]
points = current_site_fractions[np.index_exp[statevar_indices + (trial_indices,)]]
points.shape = properties.points.shape[:-1] + (-1, maximum_internal_dof)
# The Y arrays have been padded, so we should slice off the padding
points = points[..., :dof]
# Workaround for derivative issues at endmembers
points[points == 0.] = MIN_SITE_FRACTION
if len(points) == 0:
if name in points_dict:
del points_dict[name]
# No nearly stable points: skip this phase
continue
num_vars = len(phase_records[name].variables)
plane_grad = phase_records[name].plane_grad
plane_hess = phase_records[name].plane_hess
statevar_grid = np.meshgrid(*itertools.chain(indep_vals), sparse=True, indexing='xy')
# TODO: A more sophisticated treatment of constraints
num_constraints = len(indep_vals) + len(dbf.phases[name].sublattices)
constraint_jac = np.zeros((num_constraints, num_vars))
# Independent variables are always fixed (in this limited implementation)
for idx in range(len(indep_vals)):
constraint_jac[idx, idx] = 1
# This is for site fraction balance constraints
var_idx = len(indep_vals)
for idx in range(len(dbf.phases[name].sublattices)):
active_in_subl = set(dbf.phases[name].constituents[idx]).intersection(comps)
constraint_jac[len(indep_vals) + idx,
var_idx:var_idx + len(active_in_subl)] = 1
var_idx += len(active_in_subl)
grad = phase_records[name].grad(*itertools.chain(statevar_grid, points.T))
if grad.dtype == 'object':
# Workaround a bug in zero gradient entries
grad_zeros = np.zeros(points.T.shape[1:], dtype=np.float)
for i in np.arange(grad.shape[0]):
if isinstance(grad[i], int):
grad[i] = grad_zeros
grad = np.array(grad.tolist(), dtype=np.float)
bcasts = np.broadcast_arrays(*itertools.chain(properties.MU.values.T, points.T))
cast_grad = -plane_grad(*itertools.chain(bcasts, [0], [0]))
cast_grad = cast_grad.T + grad.T
grad = cast_grad
grad.shape = grad.shape[:-1] # Remove extraneous dimension
# This Hessian is an approximation updated using the BFGS method
# See Nocedal and Wright, ch.3, p. 198
# Initialize as identity matrix
hess = broadcast_to(np.eye(num_vars), grad.shape + (grad.shape[-1],)).copy()
newton_iteration = 0
while newton_iteration < MAX_NEWTON_ITERATIONS:
e_matrix = np.linalg.inv(hess)
dy_unconstrained = -np.einsum('...ij,...j->...i', e_matrix, grad)
proj_matrix = np.dot(e_matrix, constraint_jac.T)
inv_matrix = np.rollaxis(np.dot(constraint_jac, proj_matrix), 0, -1)
inv_term = np.linalg.inv(inv_matrix)
first_term = np.einsum('...ij,...jk->...ik', proj_matrix, inv_term)
# Normally a term for the residual here
# We only choose starting points which obey the constraints, so r = 0
cons_summation = np.einsum('...i,...ji->...j', dy_unconstrained, constraint_jac)
cons_correction = np.einsum('...ij,...j->...i', first_term, cons_summation)
dy_constrained = dy_unconstrained - cons_correction
# TODO: Support for adaptive changing independent variable steps
new_direction = dy_constrained[..., len(indep_vals):]
# Backtracking line search
new_points = points + INITIAL_STEP_SIZE * new_direction
alpha = np.full(new_points.shape[:-1], INITIAL_STEP_SIZE, dtype=np.float)
negative_points = np.any(new_points < 0., axis=-1)
while np.any(negative_points):
alpha[negative_points] *= 0.1
new_points = points + alpha[..., np.newaxis] * new_direction
negative_points = np.any(new_points < 0., axis=-1)
# If we made "near" zero progress on any points, don't update the Hessian until
# we've rebuilt the convex hull
# Nocedal and Wright recommend against skipping Hessian updates
# They recommend using a damped update approach, pp. 538-539 of their book
# TODO: Check the projected gradient norm, not the step length
if np.any(np.max(np.abs(alpha[..., np.newaxis] * new_direction), axis=-1) < MIN_STEP_LENGTH):
break
# Workaround for derivative issues at endmembers
new_points[new_points == 0.] = 1e-16
# BFGS update to Hessian
new_grad = phase_records[name].grad(*itertools.chain(statevar_grid, new_points.T))
if new_grad.dtype == 'object':
# Workaround a bug in zero gradient entries
grad_zeros = np.zeros(new_points.T.shape[1:], dtype=np.float)
for i in np.arange(new_grad.shape[0]):
if isinstance(new_grad[i], int):
new_grad[i] = grad_zeros
new_grad = np.array(new_grad.tolist(), dtype=np.float)
bcasts = np.broadcast_arrays(*itertools.chain(properties.MU.values.T, new_points.T))
cast_grad = -plane_grad(*itertools.chain(bcasts, [0], [0]))
cast_grad = cast_grad.T + new_grad.T
new_grad = cast_grad
new_grad.shape = new_grad.shape[:-1] # Remove extraneous dimension
# Notation used here consistent with Nocedal and Wright
s_k = np.empty(points.shape[:-1] + (points.shape[-1] + len(indep_vals),))
# Zero out independent variable changes for now
s_k[..., :len(indep_vals)] = 0
s_k[..., len(indep_vals):] = new_points - points
y_k = new_grad - grad
s_s_term = np.einsum('...j,...k->...jk', s_k, s_k)
s_b_s_term = np.einsum('...i,...ij,...j', s_k, hess, s_k)
y_y_y_s_term = np.einsum('...j,...k->...jk', y_k, y_k) / \
np.einsum('...i,...i', y_k, s_k)[..., np.newaxis, np.newaxis]
update = np.einsum('...ij,...jk,...kl->...il', hess, s_s_term, hess) / \
s_b_s_term[..., np.newaxis, np.newaxis] + y_y_y_s_term
hess = hess - update
cast_hess = -plane_hess(*itertools.chain(bcasts, [0], [0])).T + hess
hess = -cast_hess #TODO: Why does this fix things?
# TODO: Verify that the chosen step lengths reduce the energy
points = new_points
grad = new_grad
newton_iteration += 1
new_points = new_points.reshape(new_points.shape[:len(indep_vals)] + (-1, new_points.shape[-1]))
new_points = np.concatenate((current_site_fractions[..., :dof], new_points), axis=-2)
points_dict[name] = new_points
if verbose:
print('Rebuilding grid', end=' ')
grid = calculate(dbf, comps, active_phases, output='GM',
model=models, callables=callable_dict,
fake_points=True, points=points_dict, **grid_opts)
if verbose:
print('[{0} points, {1}]'.format(len(grid.points), sizeof_fmt(grid.nbytes)), end='\n')
properties.attrs['iterations'] += 1
# One last call to ensure 'properties' and 'grid' are consistent with one another
lower_convex_hull(grid, properties)
ravelled_X_view = grid['X'].values.view().reshape(-1, grid['X'].values.shape[-1])
ravelled_Y_view = grid['Y'].values.view().reshape(-1, grid['Y'].values.shape[-1])
ravelled_Phase_view = grid['Phase'].values.view().reshape(-1)
# Copy final point values from the grid and drop the index array
# For some reason direct construction doesn't work. We have to create empty and then assign.
properties['X'] = xray.DataArray(np.empty_like(ravelled_X_view[properties['points'].values]),
dims=properties['points'].dims + ('component',))
properties['X'].values[...] = ravelled_X_view[properties['points'].values]
properties['Y'] = xray.DataArray(np.empty_like(ravelled_Y_view[properties['points'].values]),
dims=properties['points'].dims + ('internal_dof',))
properties['Y'].values[...] = ravelled_Y_view[properties['points'].values]
# TODO: What about invariant reactions? We should perform a final driving force calculation here.
# We can handle that in the same post-processing step where we identify single-phase regions.
properties['Phase'] = xray.DataArray(np.empty_like(ravelled_Phase_view[properties['points'].values]),
dims=properties['points'].dims)
properties['Phase'].values[...] = ravelled_Phase_view[properties['points'].values]
del properties['points']
return properties
