def __add_candidate__(self, a):
""" Adds a single candidate to the population. """
# check if the structure is too low in raw score
raw_score_a = get_raw_score(a)
raw_score_worst = get_raw_score(self.pop[-1])
if raw_score_a < raw_score_worst \
and len(self.pop) == self.pop_size:
return
# check if the new candidate should
# replace a similar structure in the population
for (i, b) in enumerate(self.pop):
if self.comparator.looks_like(a, b):
if get_raw_score(b) < raw_score_a:
del self.pop[i]
a.info['looks_like'] = count_looks_like(a,
self.all_cand,
self.comparator)
self.pop.append(a)
self.pop.sort(key=lambda x: get_raw_score(x),
reverse=True)
return
# the new candidate needs to be added, so remove the highest
# energy one
if len(self.pop) == self.pop_size:
del self.pop[-1]
# add the new candidate
a.info['looks_like'] = count_looks_like(a,
self.all_cand,
self.comparator)
self.pop.append(a)
self.pop.sort(key=lambda x: get_raw_score(x), reverse=True)
def __get_fitness__(self, candidates):
"""Input should be sorted according to raw_score."""
max_s = get_raw_score(candidates[0])
min_s = get_raw_score(candidates[-1])
T = min_s - max_s
shared_fit = []
for c in candidates:
sc = get_raw_score(c)
obj_fit = 0.5 * (1. - tanh(2. * (sc - max_s) / T - 1.))
m = 1.
ck = c.info['key_value_pairs'][self.comp_key]
for other in candidates:
if other != c:
name = tuple(sorted([c.info['confid'],
other.info['confid']]))
if name not in self.sh_cache:
ok = other.info['key_value_pairs'][self.comp_key]
d = abs(ck - ok)
if d < self.dt:
v = 1 - (d / self.dt)**self.alpha_sh
self.sh_cache[name] = v
else:
self.sh_cache[name] = 0
m += self.sh_cache[name]
shf = (obj_fit ** self.fit_scaling) / m
shared_fit.append(shf)
return shared_fit
def __get_fitness__(self, indecies, with_history=True):
"""Calculates the fitness using the formula from
L.B. Vilhelmsen et al., JACS, 2012, 134 (30), pp 12807-12816
Sign change on the fitness compared to the formulation in the
abovementioned paper due to maximizing raw_score instead of
minimizing energy. (Set raw_score=-energy to optimize the energy)
"""
scores = [get_raw_score(x) for x in self.pop]
min_s = min(scores)
max_s = max(scores)
T = min_s - max_s
if isinstance(indecies, int):
indecies = [indecies]
f = [
0.5 * (1. - tanh(2. * (scores[i] - max_s) / T - 1.))
for i in indecies
]
if with_history:
M = [float(self.pop[i].info['n_paired']) for i in indecies]
L = [float(self.pop[i].info['looks_like']) for i in indecies]
f = [
f[i] * 1. / sqrt(1. + M[i]) * 1. / sqrt(1. + L[i])
for i in range(len(f))
]
return f
def __initialize_pop__(self):
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
if len(all_cand) > 0:
shared_fit = self.__get_fitness__(all_cand)
all_sorted = list(zip(*sorted(zip(shared_fit, all_cand),
reverse=True)))[1]
# Fill up the population with the self.pop_size most stable
# unique candidates.
i = 0
while i < len(all_sorted) and len(self.pop) < self.pop_size:
c = all_sorted[i]
i += 1
eq = False
for a in self.pop:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
self.pop.append(c)
for a in self.pop:
a.info['looks_like'] = count_looks_like(a, all_cand,
self.comparator)
self.all_cand = all_cand
def __initialize_pop__(self):
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
if len(all_cand) > 0:
shared_fit = self.__get_fitness__(all_cand)
all_sorted = list(
zip(*sorted(zip(shared_fit, all_cand), reverse=True)))[1]
# Fill up the population with the self.pop_size most stable
# unique candidates.
i = 0
while i < len(all_sorted) and len(self.pop) < self.pop_size:
c = all_sorted[i]
i += 1
eq = False
for a in self.pop:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
self.pop.append(c)
for a in self.pop:
a.info['looks_like'] = count_looks_like(
a, all_cand, self.comparator)
self.all_cand = all_cand
def penalize(t):
# penalize explosion:
raw_score = get_raw_score(t)
max_volume_per_atom = 50.
if t.get_volume() / len(t) >= max_volume_per_atom:
raw_score -= 1e9
set_raw_score(t, raw_score)
def run_ga(n_to_test):
"""
This method specifies how to run the GA once the
initial random structures have been stored in godb.db.
"""
# Various initializations:
population_size = 10 # maximal size of the population
da = DataConnection('godb.db')
atom_numbers_to_optimize = da.get_atom_numbers_to_optimize() # = [14] * 7
n_to_optimize = len(atom_numbers_to_optimize) # = 7
# This defines how close the Si atoms are allowed to get
# in candidate structures generated by the genetic operators:
blmin = closest_distances_generator(atom_numbers_to_optimize,
ratio_of_covalent_radii=0.4)
# This is our OFPComparator instance which will be
# used to judge whether or not two structures are identical:
comparator = OFPComparator(n_top=None, dE=1.0, cos_dist_max=1e-3,
rcut=10., binwidth=0.05, pbc=[False]*3,
sigma=0.1, nsigma=4, recalculate=False)
# Defining a typical combination of genetic operators:
pairing = CutAndSplicePairing(da.get_slab(), n_to_optimize, blmin)
rattlemut = RattleMutation(blmin, n_to_optimize, rattle_prop=0.8,
rattle_strength=1.5)
operators = OperationSelector([2., 1.], [pairing, rattlemut])
# Relax the randomly generated initial candidates:
while da.get_number_of_unrelaxed_candidates() > 0:
a = da.get_an_unrelaxed_candidate()
a = relax_one(a)
da.add_relaxed_step(a)
# Create the population
population = Population(data_connection=da,
population_size=population_size,
comparator=comparator,
logfile='log.txt')
current_pop = population.get_current_population()
# Test n_to_test new candidates
for step in range(n_to_test):
print('Starting configuration number %d' % step, flush=True)
a3 = None
while a3 is None:
a1, a2 = population.get_two_candidates()
a3, description = operators.get_new_individual([a1, a2])
da.add_unrelaxed_candidate(a3, description=description)
a3 = relax_one(a3)
da.add_relaxed_step(a3)
population.update()
best = population.get_current_population()[0]
print('Highest raw score at this point: %.3f' % get_raw_score(best))
print('GA finished after step %d' % step)
write('all_candidates.traj', da.get_all_relaxed_candidates())
write('current_population.traj', population.get_current_population())
def __initialize_pop__(self):
""" Private method that initalizes the population when
the population is created. """
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
# all_cand.sort(key=lambda x: x.get_potential_energy())
if len(all_cand) > 0:
# Fill up the population with the self.pop_size most stable
# unique candidates.
ratings = []
best_raw = get_raw_score(all_cand[0])
i = 0
while i < len(all_cand):
c = all_cand[i]
i += 1
eq = False
for a in self.pop:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
if len(self.pop) < self.pop_size - self.bad_candidates:
self.pop.append(c)
else:
exp_fact = exp(get_raw_score(c) / best_raw)
ratings.append([c, (exp_fact - 1) * random()])
ratings.sort(key=itemgetter(1), reverse=True)
for i in range(self.bad_candidates):
self.pop.append(ratings[i][0])
for a in self.pop:
a.info['looks_like'] = count_looks_like(a, all_cand,
self.comparator)
self.all_cand = all_cand
self.__calc_participation__()
def __initialize_pop__(self):
""" Private method that initalizes the population when
the population is created. """
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
# all_cand.sort(key=lambda x: x.get_potential_energy())
if len(all_cand) > 0:
# Fill up the population with the self.pop_size most stable
# unique candidates.
ratings = []
best_raw = get_raw_score(all_cand[0])
i = 0
while i < len(all_cand):
c = all_cand[i]
i += 1
eq = False
for a in self.pop:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
if len(self.pop) < self.pop_size - self.bad_candidates:
self.pop.append(c)
else:
exp_fact = exp(get_raw_score(c) / best_raw)
ratings.append([c, (exp_fact - 1) * random()])
ratings.sort(key=itemgetter(1), reverse=True)
for i in range(self.bad_candidates):
self.pop.append(ratings[i][0])
for a in self.pop:
a.info['looks_like'] = count_looks_like(a, all_cand,
self.comparator)
self.all_cand = all_cand
self.__calc_participation__()
def singlepoint(t, kptdensity=1.5):
if get_raw_score(t) < -1e5:
return t
try:
calc = DftbPlusCalc(t, kpts=kptdensity, use_spline=True, read_chg=True)
t.set_calculator(calc)
E = t.get_potential_energy()
F = t.get_forces()
S = t.get_stress()
finalize(t, energy=E, forces=F, stress=S)
penalize(t)
except (RuntimeError, IOError):
print('Warning: problems with singlepoint recalculation')
finalize(t, energy=1e9, forces=None, stress=None)
return t
def get_all_relaxed_candidates_after_generation(self, gen):
""" Returns all candidates that have been relaxed up to
and including the specified generation
"""
q = 'relaxed=1,extinct=0,generation<={0}'
entries = self.c.select(q.format(gen))
trajs = []
for v in entries:
t = self.get_atoms(id=v.id)
t.info['confid'] = v.gaid
t.info['relax_id'] = v.id
trajs.append(t)
trajs.sort(key=lambda x: get_raw_score(x), reverse=True)
return trajs
def looks_like(self, a1, a2):
# Energy criterium
try:
dE = abs(a1.get_potential_energy() - a2.get_potential_energy())
except:
dE = abs(get_raw_score(a1) - get_raw_score(a2))
if dE >= self.dE:
return False
# Structure criterium
f1 = self.get_features(a1)
f2 = self.get_features(a2)
d1 = sum(f1.values(), [])
d2 = sum(f2.values(), [])
max_d = max(np.abs(np.array(d1) - np.array(d2)))
s = self.get_similarity(f1, f2)
# print s, max_d
if s > self.pair_cor_cum_diff or max_d > self.pair_cor_max:
return False
else:
return True
def get_all_relaxed_candidates_after_generation(self, gen):
""" Returns all candidates that have been relaxed up to
and including the specified generation
"""
q = 'relaxed=1,extinct=0,generation<={0}'
entries = self.c.select(q.format(gen))
trajs = []
for v in entries:
t = self.get_atoms(id=v.id)
t.info['confid'] = v.gaid
t.info['relax_id'] = v.id
trajs.append(t)
trajs.sort(key=lambda x: get_raw_score(x),
reverse=True)
return trajs
def singlepoint(t, kptdensity=3.5):
if get_raw_score(t) < -1e5:
return t
try:
calc = DftbPlusCalculator(t, kpts=kptdensity, use_spline=True,
maximum_angular_momenta={'Pd': 2, 'H': 0, 'O': 1})
t.set_calculator(calc)
E = t.get_potential_energy()
F = t.get_forces()
S = t.get_stress()
finalize(t, energy=E, forces=F, stress=S)
penalize(t)
except (IOError, TypeError, RuntimeError, UnboundLocalError) as err:
print(err)
print('Warning: problems with singlepoint recalculation')
finalize(t, energy=1e9, forces=None, stress=None)
return t
def __initialize_pop__(self):
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
if len(all_cand) > 0:
fitf = self.__get_fitness__(all_cand)
all_sorted = list(zip(fitf, all_cand))
all_sorted.sort(key=itemgetter(0), reverse=True)
sort_cand = []
for _, t2 in all_sorted:
sort_cand.append(t2)
all_sorted = sort_cand
# Fill up the population with the self.pop_size most stable
# unique candidates.
i = 0
while i < len(all_sorted) and len(self.pop) < self.pop_size:
c = all_sorted[i]
# Use variable_function to decide whether to run comparator
# if the function has been defined by the user. This does not
# need to be dependent on using the rank_data function.
if self.vf is not None:
c_vf = self.vf(c)
i += 1
eq = False
for a in self.pop:
if self.vf is not None:
a_vf = self.vf(a)
# Only run comparator if the variable_function
# (self.vf) returns the same. If it returns something
# different the candidates are inherently different.
# This is done to speed up.
if a_vf == c_vf:
if self.comparator.looks_like(a, c):
eq = True
break
else:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
self.pop.append(c)
self.all_cand = all_cand
def __initialize_pop__(self):
# Get all relaxed candidates from the database
ue = self.use_extinct
all_cand = self.dc.get_all_relaxed_candidates(use_extinct=ue)
all_cand.sort(key=lambda x: get_raw_score(x), reverse=True)
if len(all_cand) > 0:
fitf = self.__get_fitness__(all_cand)
all_sorted = list(zip(fitf, all_cand))
all_sorted.sort(key=itemgetter(0), reverse=True)
sort_cand = []
for _, t2 in all_sorted:
sort_cand.append(t2)
all_sorted = sort_cand
# Fill up the population with the self.pop_size most stable
# unique candidates.
i = 0
while i < len(all_sorted) and len(self.pop) < self.pop_size:
c = all_sorted[i]
# Use variable_function to decide whether to run comparator
# if the function has been defined by the user. This does not
# need to be dependent on using the rank_data function.
if self.vf is not None:
c_vf = self.vf(c)
i += 1
eq = False
for a in self.pop:
if self.vf is not None:
a_vf = self.vf(a)
# Only run comparator if the variable_function
# (self.vf) returns the same. If it returns something
# different the candidates are inherently different.
# This is done to speed up.
if a_vf == c_vf:
if self.comparator.looks_like(a, c):
eq = True
break
else:
if self.comparator.looks_like(a, c):
eq = True
break
if not eq:
self.pop.append(c)
self.all_cand = all_cand
def __get_fitness__(self, indecies, with_history=True):
"""Calculates the fitness using the formula from
L.B. Vilhelmsen et al., JACS, 2012, 134 (30), pp 12807-12816
Sign change on the fitness compared to the formulation in the
abovementioned paper due to maximizing raw_score instead of
minimizing energy. (Set raw_score=-energy to optimize the energy)
"""
scores = [get_raw_score(x) for x in self.pop]
min_s = min(scores)
max_s = max(scores)
T = min_s - max_s
if isinstance(indecies, int):
indecies = [indecies]
f = [0.5 * (1. - tanh(2. * (scores[i] - max_s) / T - 1.))
for i in indecies]
if with_history:
M = [float(self.pop[i].info['n_paired']) for i in indecies]
L = [float(self.pop[i].info['looks_like']) for i in indecies]
f = [f[i] * 1. / sqrt(1. + M[i]) * 1. / sqrt(1. + L[i])
for i in range(len(f))]
return f
def converged(self):
cur_pop = self.pop.get_current_population()
if abs(get_raw_score(cur_pop[0]) - self.max_raw_score) <= self.eps:
return True
return False
syms = a2.get_chemical_symbols()
assert 'Ba' in syms
assert len(set(syms)) == 3
op = MoveUpMutation(cations, 1, 1.)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
assert 'Ba' not in syms
assert len(set(syms)) == 2
cations = ['Co', 'Ni', 'Cu']
a1 = Atoms('NiNiBrBr')
a1.info['confid'] = 1
op = MoveRightMutation(cations, 1, 1.)
a2, desc = op.get_new_individual([a1])
a2.info['confid'] = 2
syms = a2.get_chemical_symbols()
assert len(set(syms)) == 2
assert len([i for i in syms if i == 'Cu']) == 2
op = MoveLeftMutation(cations, 2, .5)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
from ase.ga import set_raw_score, get_raw_score
assert len(set(syms)) == 3
set_raw_score(a3, 5.0)
assert get_raw_score(a3) == 5.0
def looks_like(self, a1, a2):
d = abs(get_raw_score(a1) - get_raw_score(a2))
if d >= self.dist:
return False
else:
return True
def run_ga(n_to_test, kptdensity=3.5):
population_size = 20
da = DataConnection('godb.db')
atom_numbers_to_optimize = da.get_atom_numbers_to_optimize()
n_to_optimize = len(atom_numbers_to_optimize)
slab = da.get_slab()
all_atom_types = get_all_atom_types(slab, atom_numbers_to_optimize)
blmin = closest_distances_generator(all_atom_types, 0.05) # 0.5
# defining genetic operators:
mutation_probability = 0.75
pairing = CutAndSplicePairing(blmin,
p1=1.,
p2=0.,
minfrac=0.15,
use_tags=False)
cellbounds = CellBounds(
bounds={
'phi': [0.2 * 180., 0.8 * 180.],
'chi': [0.2 * 180., 0.8 * 180.],
'psi': [0.2 * 180., 0.8 * 180.]
})
strainmut = StrainMutation(blmin,
stddev=0.7,
cellbounds=cellbounds,
use_tags=False)
blmin_soft = closest_distances_generator(all_atom_types, 0.1)
softmut = SoftMutation(blmin_soft, bounds=[2., 5.], use_tags=False)
rattlemut = RattleMutation(blmin,
n_to_optimize,
rattle_prop=0.8,
rattle_strength=2.5,
use_tags=False)
mutations = OperationSelector([4., 4., 2], [softmut, strainmut, rattlemut])
if True:
# recalculate raw scores
structures = da.get_all_relaxed_candidates()
for atoms in structures:
atoms = singlepoint(atoms, kptdensity=kptdensity)
da.c.delete([atoms.info['relax_id']])
if 'data' not in atoms.info:
atoms.info['data'] = {}
da.add_relaxed_step(atoms)
print('Finished recalculating raw scores')
# relaxing the initial candidates:
while da.get_number_of_unrelaxed_candidates() > 0:
a = da.get_an_unrelaxed_candidate()
a.wrap()
a = relax_one(a, kptdensity=kptdensity)
da.add_relaxed_step(a)
# create the population
population = Population(data_connection=da,
population_size=population_size,
comparator=comparator,
logfile='log.txt')
current_pop = population.get_current_population()
strainmut.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
pairing.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
# Test n_to_test new candidates
ga_raw_scores = []
step = 0
for step in range(n_to_test):
print('Starting configuration number %d' % step, flush=True)
clock = time()
a3 = None
r = random()
if r > mutation_probability:
while a3 is None:
a1, a2 = population.get_two_candidates()
a3, desc = pairing.get_new_individual([a1, a2])
else:
while a3 is None:
a1 = population.get_one_candidate()
a3, desc = mutations.get_new_individual([a1])
dt = time() - clock
op = 'pairing' if r > mutation_probability else 'mutating'
print('Time for %s candidate(s): %.3f' % (op, dt), flush=True)
a3.wrap()
da.add_unrelaxed_candidate(a3, description=desc)
a3 = relax_one(a3, kptdensity=kptdensity)
da.add_relaxed_step(a3)
# Various updates:
population.update()
current_pop = population.get_current_population()
if step % 10 == 0:
strainmut.update_scaling_volume(current_pop,
w_adapt=0.5,
n_adapt=4)
pairing.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
write('current_population.traj', current_pop)
# Print out information for easy analysis/plotting afterwards:
if r > mutation_probability:
print('Step %d %s %.3f %.3f %.3f' % (step, desc,\
get_raw_score(a1), get_raw_score(a2), get_raw_score(a3)))
else:
print('Step %d %s %.3f %.3f' % (step, desc,\
get_raw_score(a1), get_raw_score(a3)))
print('Step %d highest raw score in pop: %.3f' % \
(step, get_raw_score(current_pop[0])))
ga_raw_scores.append(get_raw_score(a3))
print('Step %d highest raw score generated by GA: %.3f' % \
(step, max(ga_raw_scores)))
emin = population.pop[0].get_potential_energy()
print('GA finished after step %d' % step)
print('Lowest energy = %8.3f eV' % emin, flush=True)
write('all_candidates.traj', da.get_all_relaxed_candidates())
write('current_population.traj', population.get_current_population())
def run_ga(n_to_test, kptdensity=None):
''' This method specifies how to run the GA once the
initial random structures have been stored in godb.db.
'''
# Various initializations:
population_size = 10
da = DataConnection('godb.db')
atom_numbers_to_optimize = da.get_atom_numbers_to_optimize()
n_to_optimize = len(atom_numbers_to_optimize)
slab = da.get_slab()
all_atom_types = get_all_atom_types(slab, atom_numbers_to_optimize)
blmin = closest_distances_generator(all_atom_types,
ratio_of_covalent_radii=0.05)
# Defining the mix of genetic operators:
mutation_probability = 0.3333
pairing = CutAndSplicePairing(slab, n_to_optimize, blmin)
rattlemut = RattleMutation(blmin, n_to_optimize,
rattle_prop=0.8, rattle_strength=1.5)
mirrormut = MirrorMutation(blmin, n_to_optimize)
mutations = OperationSelector([1., 1.], [rattlemut, mirrormut])
if True:
# Recalculate raw scores of any relaxed candidates
# present in the godb.db database (only applies to
# iter007).
structures = da.get_all_relaxed_candidates()
for atoms in structures:
atoms = singlepoint(atoms)
da.c.delete([atoms.info['relax_id']])
if 'data' not in atoms.info:
atoms.info['data'] = {}
da.add_relaxed_step(atoms)
print('Finished recalculating raw scores')
# Relax the randomly generated initial candidates:
while da.get_number_of_unrelaxed_candidates() > 0:
a = da.get_an_unrelaxed_candidate()
a.wrap()
a = relax_one(a)
da.add_relaxed_step(a)
# Create the population
population = Population(data_connection=da,
population_size=population_size,
comparator=comparator,
logfile='log.txt')
current_pop = population.get_current_population()
# Test n_to_test new candidates
ga_raw_scores = []
step = 0
for step in range(n_to_test):
print('Starting configuration number %d' % step, flush=True)
clock = time()
a3 = None
r = random()
if r > mutation_probability:
while a3 is None:
a1, a2 = population.get_two_candidates()
a3, desc = pairing.get_new_individual([a1, a2])
else:
while a3 is None:
a1 = population.get_one_candidate()
a3, desc = mutations.get_new_individual([a1])
dt = time() - clock
op = 'pairing' if r > mutation_probability else 'mutating'
print('Time for %s candidate(s): %.3f' % (op, dt), flush=True)
a3.wrap()
da.add_unrelaxed_candidate(a3, description=desc)
a3 = relax_one(a3)
da.add_relaxed_step(a3)
# Various updates:
population.update()
current_pop = population.get_current_population()
write('current_population.traj', current_pop)
# Print out information for easy analysis/plotting afterwards:
if r > mutation_probability:
print('Step %d %s %.3f %.3f %.3f' % (step, desc,\
get_raw_score(a1), get_raw_score(a2), get_raw_score(a3)))
else:
print('Step %d %s %.3f %.3f' % (step, desc,\
get_raw_score(a1), get_raw_score(a3)))
print('Step %d highest raw score in pop: %.3f' % \
(step, get_raw_score(current_pop[0])))
ga_raw_scores.append(get_raw_score(a3))
print('Step %d highest raw score generated by GA: %.3f' % \
(step, max(ga_raw_scores)))
emin = population.pop[0].get_potential_energy()
print('GA finished after step %d' % step)
print('Lowest energy = %8.3f eV' % emin, flush=True)
write('all_candidates.traj', da.get_all_relaxed_candidates())
write('current_population.traj', population.get_current_population())
def test_element_operators(seed):
import numpy as np
from ase import Atoms
from ase.ga.element_crossovers import OnePointElementCrossover
# set up the random number generator
rng = np.random.RandomState(seed)
a1 = Atoms('SrSrSrBaClClClClBrBrBrBr')
a1.info['confid'] = 1
a2 = Atoms('CaCaMgBaFFFFFFFF')
a2.info['confid'] = 2
cations = ['Sr', 'Ba', 'Ca', 'Mg']
anions = ['Cl', 'F', 'Br']
op = OnePointElementCrossover([cations, anions], [3, 2], [.25, .5],
rng=rng)
a3, desc = op.get_new_individual([a1, a2])
syms = a3.get_chemical_symbols()
assert len(set([i for i in syms if i in cations])) < 4
assert len(set([i for i in syms if i in anions])) < 3
from ase.ga.element_mutations import RandomElementMutation
op = RandomElementMutation([cations, anions], [3, 2], [.25, .5], rng=rng)
a4, desc = op.get_new_individual([a1])
syms = a4.get_chemical_symbols()
assert len(set([i for i in syms if i in cations])) < 4
assert len(set([i for i in syms if i in anions])) < 3
op = RandomElementMutation(anions, 2, .5, rng=rng)
a4, desc = op.get_new_individual([a2])
syms = a4.get_chemical_symbols()
assert len(set([i for i in syms if i in anions])) == 2
from ase.ga.element_mutations import MoveDownMutation
from ase.ga.element_mutations import MoveUpMutation
from ase.ga.element_mutations import MoveRightMutation
from ase.ga.element_mutations import MoveLeftMutation
a1 = Atoms('SrSrClClClCl')
a1.info['confid'] = 1
op = MoveDownMutation(cations, 2, .5, rng=rng)
a2, desc = op.get_new_individual([a1])
a2.info['confid'] = 2
syms = a2.get_chemical_symbols()
assert 'Ba' in syms
assert len(set(syms)) == 3
op = MoveUpMutation(cations, 1, 1., rng=rng)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
assert 'Ba' not in syms
assert len(set(syms)) == 2
cations = ['Co', 'Ni', 'Cu']
a1 = Atoms('NiNiBrBr')
a1.info['confid'] = 1
op = MoveRightMutation(cations, 1, 1., rng=rng)
a2, desc = op.get_new_individual([a1])
a2.info['confid'] = 2
syms = a2.get_chemical_symbols()
assert len(set(syms)) == 2
assert len([i for i in syms if i == 'Cu']) == 2
op = MoveLeftMutation(cations, 2, .5, rng=rng)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
from ase.ga import set_raw_score, get_raw_score
assert len(set(syms)) == 3
set_raw_score(a3, 5.0)
assert get_raw_score(a3) == 5.0
a3, desc = operators.get_new_individual([a1, a2])
# Relax it and add to database
da.add_unrelaxed_candidate(a3, description=desc)
relax(a3)
da.add_relaxed_step(a3)
# Update the population
population.update()
current_pop = population.get_current_population()
write('current_population.traj', current_pop)
# Update the strain mutation and pairing operators
if step % 10 == 0:
strainmut.update_scaling_volume(current_pop, w_adapt=0.5,
n_adapt=4)
pairing.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
# Print out information for easier follow-up/analysis/plotting:
print('Step %d %s %.3f %.3f %.3f' % (step, desc,
get_raw_score(a1), get_raw_score(a2), get_raw_score(a3)))
print('Step %d highest raw score in pop: %.3f' %
(step, get_raw_score(current_pop[0])))
print('GA finished after step %d' % step)
hiscore = get_raw_score(current_pop[0])
print('Highest raw score = %8.4f eV' % hiscore)
write('all_candidates.traj', da.get_all_relaxed_candidates())
write('current_population.traj', current_pop)
syms = a2.get_chemical_symbols()
assert 'Ba' in syms
assert len(set(syms)) == 3
op = MoveUpMutation(cations, 1, 1.)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
assert 'Ba' not in syms
assert len(set(syms)) == 2
cations = ['Co', 'Ni', 'Cu']
a1 = Atoms('NiNiBrBr')
a1.info['confid'] = 1
op = MoveRightMutation(cations, 1, 1.)
a2, desc = op.get_new_individual([a1])
a2.info['confid'] = 2
syms = a2.get_chemical_symbols()
assert len(set(syms)) == 2
assert len([i for i in syms if i == 'Cu']) == 2
op = MoveLeftMutation(cations, 2, .5)
a3, desc = op.get_new_individual([a2])
syms = a3.get_chemical_symbols()
from ase.ga import set_raw_score, get_raw_score
assert len(set(syms)) == 3
set_raw_score(a3, 5.0)
assert get_raw_score(a3) == 5.0
def extract_best_unique(comparator, max_select=None, num_stddev=None,
score_limit=None, dbfile='best_unique.db'):
''' Writes a database containing the best unique structures
from a set of global optimization runs in the current
working directory. These runs must have written an
'all_candidates.traj' file containing all the structures
with their raw scores.
comparator: a class instance with suitable _compare_structure_
and looks_like methods for comparing two structures
max_select: upper bound on the number of best unique structures
to select. If None (default), no bound is enforced.
num_stddev: number of standard deviations relative to the
average score of all candidates, which is used to
pre-select only the more stable structures.
Setting it to zero means all better-than-average
structures are considered for further selection.
Three standard deviations around the mean is used as
cutoff in determining the average, to exclude very
low-score outliers.
score_limit: as an alternative to num_stddev, this argument sets
the minimal raw score for structures to be included.
dbfile: name of the database where the final selection will be saved.
'''
db = connect(dbfile)
all_candidates = []
all_cand_dict = {}
for (dirpath, dirnames, filenames) in os.walk('.'):
if 'all_candidates.traj' in filenames and 'run' in dirpath:
print('Found run directory', dirpath)
candidates = read(dirpath + '/all_candidates.traj@:')
all_candidates.extend(candidates)
all_cand_dict[dirpath] = candidates
all_candidates.sort(key=lambda x: get_raw_score(x), reverse=True)
raw_scores = np.array([get_raw_score(atoms) for atoms in all_candidates])
std = np.std(raw_scores)
mean = raw_scores[len(raw_scores) // 2]
min_score = mean - 3 * std
izero = np.argmax(raw_scores < min_score)
if izero != 0:
raw_scores = raw_scores[:izero]
all_candidates = all_candidates[:izero]
average = np.mean(raw_scores)
std = np.std(raw_scores)
max_score = np.max(raw_scores)
min_score = np.min(raw_scores)
if num_stddev is not None:
cut_score = average - num_stddev * std
elif score_limit is not None:
cut_score = max_score - score_limit
else:
cut_score = min_score
print('Average = %.3f, Std. dev = %.3f' % (average, std))
print('N = %d before selecting unique structures' % len(all_candidates))
print('Max score = %.3f, min score = %.3f, cut score = %.3f' % \
(max_score, min_score, cut_score), flush=True)
args = []
for key, val in all_cand_dict.items():
raw_scores = np.array([get_raw_score(atoms) for atoms in val])
izero = np.argmax(raw_scores < cut_score)
if izero != 0:
val = val[:izero]
args.append([comparator, val, False])
po = mp.Pool(processes=None)
harvest = po.map(get_unique, args, chunksize=1)
po.close()
po.join()
all_candidates = [atoms for allcand in harvest for atoms in allcand]
all_candidates.sort(key=lambda x: get_raw_score(x), reverse=True)
raw_scores = [get_raw_score(atoms) for atoms in all_candidates]
best_unique = []
print('N_unique = %d before next selection round' % len(all_candidates),
flush=True)
best_unique = get_unique([comparator, all_candidates, True])
best_unique.sort(key=lambda x: get_raw_score(x), reverse=True)
N = len(best_unique)
print('N_unique = %d before further refinement' % N, flush=True)
if max_select is None or max_select >= N:
selection = range(N)
else:
selection = range(max_select)
print('Selected indices:', selection)
for i in selection:
atoms = best_unique[i]
raw_score = get_raw_score(atoms)
db.write(atoms, raw_score_from_ga=raw_score, gaid=i, relaxed=0)
print('N_unique = %d after final refinement' % len(selection))
return
def looks_like(self, a1, a2):
d = abs(get_raw_score(a1) - get_raw_score(a2))
if d >= self.dist:
return False
else:
return True
relax(a3, cellbounds=cellbounds)
da.add_relaxed_step(a3)
# If the relaxation has changed the cell parameters
# beyond the bounds we disregard it in the population
cell = a3.get_cell()
if not cellbounds.is_within_bounds(cell):
da.kill_candidate(a3.info['confid'])
# Update the population
population.update()
if step % 10 == 0:
# Update the scaling volumes of the strain mutation
# and the pairing operator based on the current
# best structures contained in the population
current_pop = population.get_current_population()
strainmut.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
pairing.update_scaling_volume(current_pop, w_adapt=0.5, n_adapt=4)
write('current_population.traj', current_pop)
print('GA finished after step %d' % step)
hiscore = get_raw_score(current_pop[0])
print('Highest raw score = %8.4f eV' % hiscore)
all_candidates = da.get_all_relaxed_candidates()
write('all_candidates.traj', all_candidates)
current_pop = population.get_current_population()
write('current_population.traj', current_pop)