import ipopt
ipopt.setLoggingLevel(50)
import numpy as np
from collections import namedtuple
from pycalphad.variables import string_type
from pycalphad.core.constants import MIN_SITE_FRACTION
SolverResult = namedtuple('SolverResult', ['converged', 'x', 'chemical_potentials'])
[docs]class SolverBase(object):
""""Base class for solvers."""
ignore_convergence = False
[docs] def solve(self, prob):
"""
*Implement this method.*
Solve a non-linear problem
Parameters
----------
prob : pycalphad.core.problem.Problem
Returns
-------
pycalphad.core.solver.SolverResult
"""
raise NotImplementedError("A subclass of Solver must be implemented.")
[docs]class InteriorPointSolver(SolverBase):
"""
Standard solver class that uses IPOPT.
Attributes
----------
verbose : bool
If True, will print solver diagonstics. Defaults to False.
infeasibility_threshold : float
Dual infeasibility threshold used to tighten constraints and
attempt a second solve, if necessary. Defaults to 1e-4.
ipopt_options : dict
Dictionary of options to pass to IPOPT.
Methods
-------
solve
Solve a pycalphad.core.problem.Problem
apply_options
Encodes ipopt_options and applies them to problem
"""
def __init__(self, verbose=False, infeasibility_threshold=1e-4, **ipopt_options):
"""
Standard solver class that uses IPOPT.
Parameters
----------
verbose : bool
If True, will print solver diagonstics. Defaults to False.
infeasibility_threshold : float
Dual infeasibility threshold used to tighten constraints and
attempt a second solve, if necessary. Defaults to 1e-4.
ipopt_options : dict
See https://www.coin-or.org/Ipopt/documentation/node40.html for all options
"""
self.verbose = verbose
self.infeasibility_threshold = infeasibility_threshold
# set default options
self.ipopt_options = {
'max_iter': 200,
'print_level': 0,
# This option improves convergence when using L-BFGS
'limited_memory_max_history': 100,
'tol': 1e-1,
'constr_viol_tol': 1e-5,
'nlp_scaling_method': 'none',
'hessian_approximation': 'limited-memory'
}
if not self.verbose:
# suppress the "This program contains Ipopt" banner
self.ipopt_options['sb'] = ipopt_options.pop('sb', 'yes')
# update the default options with the passed options
self.ipopt_options.update(ipopt_options)
[docs] def apply_options(self, problem):
"""
Apply global options to the solver
Parameters
----------
problem : ipopt.problem
A problem object that will be solved
Notes
-----
Strings are encoded to byte strings.
"""
for option, value in self.ipopt_options.items():
if isinstance(value, string_type):
problem.addOption(option.encode(), value.encode())
else:
problem.addOption(option.encode(), value)
[docs] def solve(self, prob):
"""
Solve a non-linear problem
Parameters
----------
prob : pycalphad.core.problem.Problem
Returns
-------
SolverResult
"""
cur_conds = prob.conditions
comps = prob.pure_elements
nlp = ipopt.problem(
n=prob.num_vars,
m=prob.num_constraints,
problem_obj=prob,
lb=prob.xl,
ub=prob.xu,
cl=prob.cl,
cu=prob.cu
)
self.apply_options(nlp)
length_scale = np.min(np.abs(prob.cl))
length_scale = max(length_scale, 1e-9)
# Note: Using the ipopt derivative checker can be tricky at the edges of composition space
# It will not give valid results for the finite difference approximation
x, info = nlp.solve(prob.x0)
dual_inf = np.max(np.abs(info['mult_g']*info['g']))
if dual_inf > self.infeasibility_threshold:
if self.verbose:
print('Trying to improve poor solution')
# Constraints are getting tiny; need to be strict about bounds
if length_scale < 1e-6:
nlp.addOption(b'compl_inf_tol', 1e-3 * float(length_scale))
nlp.addOption(b'bound_relax_factor', MIN_SITE_FRACTION)
# This option ensures any bounds failures will fail "loudly"
# Otherwise we are liable to have subtle mass balance errors
nlp.addOption(b'honor_original_bounds', b'no')
else:
nlp.addOption(b'dual_inf_tol', self.infeasibility_threshold)
accurate_x, accurate_info = nlp.solve(x)
if accurate_info['status'] >= 0:
x, info = accurate_x, accurate_info
chemical_potentials = prob.chemical_potentials(x)
if info['status'] == -10:
# Not enough degrees of freedom; nothing to do
if len(prob.composition_sets) == 1:
converged = True
chemical_potentials[:] = prob.composition_sets[0].energy
else:
converged = False
elif info['status'] < 0:
if self.verbose:
print('Calculation Failed: ', cur_conds, info['status_msg'])
converged = False
else:
converged = True
if self.verbose:
print('Chemical Potentials', chemical_potentials)
print(info['mult_x_L'])
print(x)
print('Status:', info['status'], info['status_msg'])
return SolverResult(converged=converged, x=x, chemical_potentials=chemical_potentials)