def find_swaps():
all_atoms = []
e_ref = []
db = connect(db_name)
for row in db.select(converged=True):
all_atoms.append(row.toatoms())
final_struct_id = row.final_struct_id
final = db.get(id=final_struct_id)
e_ref.append(final.energy)
extractor = MCTrajectoryExtractor()
swaps = extractor.find_swaps(all_atoms)
with open(eci_file, 'r') as infile:
eci = json.load(infile)
e_pred = []
setting = settingFromJSON("almgsixSettings.json")
e_pred = {}
for swap in swaps:
print(swap)
for s in swap:
if s in e_pred.keys():
continue
atoms = attach_calculator(setting, atoms=all_atoms[s], eci=eci)
atoms.numbers = all_atoms[s].numbers
e_pred[s] = atoms.get_potential_energy()
for k, v in e_pred.items():
print(k, v, e_ref[k])
dev = extractor.swap_energy_deviations(swaps, e_pred, e_ref)
extractor.plot_swap_deviation(dev)
plt.show()
def show_fit():
data = np.loadtxt('data/almgsiX_data.csv', delimiter=',')
energy = data[:, -1]
energy -= np.mean(energy)
X = data[:, :-1]
singlets = data[:, 1:4]
settings = settingFromJSON('almgsixSettings.json')
concs = singlets2conc(settings.basis_functions, singlets)
coeff_all = np.loadtxt("data/coeff_all.txt")
coeff_alrich = np.loadtxt("data/coeff_alrich.txt")
coeff_al_low = np.loadtxt("data/coeff_al_low.txt")
c_al = np.array([v['Al'] for v in concs])
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
pred_all = X.dot(coeff_all)
pred_alrich = X.dot(coeff_alrich)
pred_al_low = X.dot(coeff_al_low)
weights = interpolator(c_al)
pred_sum = pred_alrich*weights + pred_al_low*(1.0 - weights)
print(energy - pred_all)
ax.plot(c_al, pred_all - energy, 'o', mfc='none')
ax.plot(c_al, pred_alrich - energy, 'x')
ax.plot(c_al, pred_al_low - energy, 'v', mfc='none')
ax.plot(c_al, pred_sum - energy, '^', mfc='none')
plt.show()
def get_gs():
atoms = bulk('Al', cubic=True) * (5, 1, 1)
setting = settingFromJSON(FNAME)
with open(ECI_FILE, 'r') as infile:
eci = json.load(infile)
atoms = attach_calculator(setting=setting, atoms=atoms, eci=eci)
setting.set_active_template(atoms=atoms)
cluster = setting.cluster_list.get_by_name("c2_d0000_0")[0]
cnst = PairConstraint(['X', 'X'], cluster, setting.trans_matrix, atoms)
symbols = ['Al', 'Mg', 'Si', 'X']
numX = 0
removed = False
for atom in atoms:
new_symb = choice(symbols)
atom.symbol = new_symb
if new_symb == 'X':
numX += 1
if numX == 1 and not removed:
symbols.remove('X')
removed = True
T = [1000, 800, 600, 400, 200, 100, 50]
for temp in T:
mc = Montecarlo(atoms, temp)
mc.add_constraint(cnst)
mc.run(steps=10 * len(atoms))
return atoms
def fit_eci():
data = np.loadtxt('data/almgsiX_data.csv', delimiter=',')
energy = data[:, -1]
X = data[:, :-1]
#energy -= np.mean(energy)
indices = list(range(X.shape[0]))
shuffle(indices)
N = int(0.8*len(indices))
train = indices[:N]
test = indices[N:]
X_test = X[test, :]
energy_test = energy[test]
# Remove test data
X = X[train, :]
energy = energy[train]
singlets = data[:, 1:4]
settings = settingFromJSON('almgsixSettings.json')
concs = singlets2conc(settings.basis_functions, singlets)
conc_array = np.array([[v for v in x.values()] for x in concs])
indicators = conc_array[:, :-1]
coeffs = []
weights = []
for types in product(range(0, 2), repeat=3):
multiplicator = np.ones(data.shape[0])
for i, t in enumerate(types):
if t == 0:
multiplicator *= interpolator(indicators[:, i])
else:
multiplicator *= (1.0-interpolator(indicators[:, i]))
mdata = np.zeros_like(X)
for i in range(X.shape[1]):
mdata[:, i] = X[:, i]*multiplicator[train]
model = LassoCV(eps=1e-6, cv=5)
#model = RidgeCV(alphas=np.logspace(-12, -5, 50))
energy_fit = energy*multiplicator[train]
model.fit(mdata, energy_fit)
coeffs.append(model.coef_)
weights.append(multiplicator)
pred = np.zeros_like(energy)
pred_test = np.zeros_like(energy_test)
print(X_test.shape)
for i in range(len(coeffs)):
pred += X.dot(coeffs[i])*weights[i][train]
pred_test += X_test.dot(coeffs[i])*weights[i][test]
rmse = np.sqrt(np.mean((energy - pred)**2))
rmse_test = np.sqrt(np.mean((energy_test - pred_test)**2))
print("MSE: {} meV/atom".format(rmse*1000.0))
print("MSE test: {} meV/atom".format(rmse_test*1000.0))
def structure_impact():
eci_file = 'data/almgsiX_eci.json'
with open(eci_file, 'r') as infile:
eci = json.load(infile)
print(eci)
settings = settingFromJSON('almgsixSettings.json')
print(settings.basis_functions)
dE = 0.011 # 11 meV/atom
ecis = np.array([[v for k, v in eci.items() if k.startswith(('c1', 'c2', 'c3', 'c4'))]])
x, _, _, _ = np.linalg.lstsq(ecis, [dE])
print(np.sum(np.abs(x)))
print(dE/np.max(np.abs(ecis)))
print(dE/np.median(np.abs(ecis)))
def explore():
eci_file = 'data/almgsiX_eci.json'
with open(eci_file, 'r') as infile:
eci = json.load(infile)
settings = settingFromJSON('almgsixSettings.json')
atoms = settings.atoms*(5, 5, 5)
atoms = attach_calculator(setting=settings, atoms=atoms, eci=eci)
num_structs = 10*len(atoms)
energies = []
symbs = ['Al', 'Mg', 'Si', 'X']
now = time.time()
for i in range(num_structs):
if time.time() - now > 30:
print("{} of {}".format(i, num_structs))
now = time.time()
idx = np.random.randint(0, high=len(atoms))
symb = choice(symbs)
atoms[idx].symbol = symb
energies.append(atoms.get_potential_energy()/len(atoms))
np.savetxt("data/energies.txt", energies)
from clease import settingFromJSON, CorrFunction
from ase.db import connect
from clease.basis_function import BinaryLinear
setting = settingFromJSON("almgsixSettings3.json")
db = connect(setting.db_name)
db.delete_external_table('polynomial_cf')
db.delete_external_table('binary_linear_cf')
bf = BinaryLinear(setting.unique_elements, reduntant_element='X')
print(setting.unique_elements)
print(bf.unique_elements)
setting.basis_func_type = bf
cf = CorrFunction(setting)
cf.reconfigure_db_entries()
def main():
setting = settingFromJSON("almgsixSettings.json")
setting.db_name = 'almgsiX_dft.db'
new_struct = NewStructures(setting)
new_struct.insert_structures(traj_init="data/restricted_gs_5x1x1.traj")