def gradb(direcs,
v='bo5',
bd='H-H',
nn=True,
bo_layer=[9, 2],
massages=1,
debd=False,
deba=True,
deang=True,
dft='siesta',
batch=50):
''' variables like: bo1_C-H, boc5_C rosi_C-H boc1
'''
v_ = v
v = v + '_' + bd
print('- grading ... ...')
ffield = 'ffield.json' if nn else 'ffield'
rn = MPNN(libfile=ffield,
direcs=direcs,
dft=dft,
nn=nn,
bo_layer=bo_layer,
massages=massages,
batch_size=batch,
pkl=True)
rn.initialize()
rn.session(learning_rate=3.0e-4, method='AdamOptimizer')
if nn:
bdlit = [
'bop', 'bop_si', 'bop_pi', 'bop_pp', 'F', 'bosi', 'bopi', 'bopp',
'powb', 'expb', 'sieng', 'EBD'
]
else:
bdlit = [
'bop', 'bop_si', 'bop_pi', 'bop_pp', 'f_1', 'f_2', 'f_3', 'f_4',
'f_5', 'bosi', 'bopi', 'bopp', 'powb', 'expb', 'sieng', 'EBD'
]
if debd:
bonds = rn.bonds
for b in bonds:
v = v_ + '_' + b
grad = rn.get_gradient(rn.Loss, rn.p[v])
text_ = '- the gradient of Loss/%s is ' % v
logger.info(text_ + str(grad))
if grad is None:
continue
if not np.isnan(grad):
continue
if rn.nbd[b] > 0:
grad = rn.get_gradient(rn.__dict__['EBD'][b], rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
('EBD' + '_' + b, v, str(grad)))
if not grad is None:
if np.isnan(grad):
for l in bdlit:
grad = rn.get_gradient(rn.__dict__[l][bd], rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
(l + '_' + b, v, str(grad)))
v = v_ + '_' + bd
if deba:
sl = ['EL', 'EOV', 'EUN']
alist = {
'EL': ['Delta_lp', 'Delta_e', 'explp'],
'EOV': ['Delta_lpcorr', 'Delta_lp', 'nlp', 'so', 'otrm1', 'otrm2'],
'EUN': [
'expeu1', 'expeu3', 'Delta_lpcorr', 'Delta_lp', 'Delta_e',
'explp', 'nlp'
]
}
for sp in rn.spec:
for l in sl:
if sp in rn.__dict__[l]:
grad = rn.get_gradient(rn.__dict__[l][sp], rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
(l + '_' + sp, v, str(grad)))
if not grad is None:
if np.isnan(grad):
for al in alist[l]:
grad = rn.get_gradient(rn.__dict__[al][sp],
rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
(al + '_' + sp, v, str(grad)))
if deang:
al = ['EANG', 'EPEN', 'ETC']
for ang in rn.angs:
# v = 'val1'+'_'+ang
if rn.nang[ang] > 0:
for l in al:
grad = rn.get_gradient(rn.__dict__[l][ang], rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
(l + '_' + ang, v, str(grad)))
tl = ['ETOR', 'Efcon']
for tor in rn.tors:
# v = 'tor2' # +'_'+tor
if rn.ntor[tor] > 0:
for l in tl:
grad = rn.get_gradient(rn.__dict__[l][tor], rn.p[v])
logger.info('- the gradient of %s/%s is: %s' %
(l + '_' + tor, v, str(grad)))
rn.sess.close()
def train_():
# this trajectory files can visulized by ase, i.e. use commond
# can see the configurations "ase gui C2H2-0.traj"
direcs = {
'C2H2-0': '../data/C2H2-0.traj',
'C2H30O130': '../data/C2H30O130.traj',
'C2H40': '../data/C2H40.traj',
'nm1-5': '../data/C1N1O2H3-0.traj',
'nmb': '../data/nmb.traj',
'h2osw': '../data/h2osw.traj',
'nm2l': '../data/nm2l.traj',
'C4N4O8H12-3-0': '../data/nml3-0.traj',
'C4N4O8H12-5-0': '../data/nml5-0.traj',
'C4N4O8H12-5-1': '../data/nml5-1.traj',
'C4N4O8H12-6-0': '../data/nml6-0.traj',
'C4N4O8H12-6-1': '../data/nml6-1.traj',
'C4N4O8H12-7-1': '../data/nml7-1.traj',
'C4N4O8H12-8-0': '../data/nml8-0.traj',
'C4N4O8H12-9-0': '../data/nml9-0.traj',
'C4N4O8H12-11': '../data/nml11.traj',
'C4N4O8H12-10-0': '../data/nml10-0.traj',
'C4N4O8H12-10-1': '../data/nml10-1.traj',
'C4N4O8H12-12-1': '../data/nml12-1.traj',
'C4N4O8H12-13-0': '../data/nml13-0.traj',
'C4N4O8H12-140': '../data/nml14-0.traj',
'C4N4O8H12-141': '../data/nml14-1.traj',
'nm1-CN': '../data/nm1-CN.traj',
'c2h6-0': '../data/c2h6-1-0.traj',
'c2h6-1': '../data/c2h6-1-1.traj',
'c2h6-2': '../data/c2h6-1-2.traj',
'c2h6-3': '../data/c2h6-1-3.traj',
'c2h6-4': '../data/c2h6-1-4.traj',
'c2h4-0': '../data/c2h4-1-0.traj',
'c2h4-1': '../data/c2h4-1-1.traj',
'c2h4-2': '../data/c2h4-1-2.traj',
'c2h4-3': '../data/c2h4-1-3.traj',
'c2h4-CC': '../data/c2h4-CC.traj',
'c2h2-CC': '../data/c2h2-CC.traj',
'c2h6-CC': '../data/c2h6-CC.traj',
'c2h2-1-0': '../data/c2h2-1-0.traj',
'c2h2-1-1': '../data/c2h2-1-1.traj',
'c2h2-1-2': '../data/c2h2-1-2.traj',
'c2h2-1-3': '../data/c2h2-1-3.traj',
}
mn = MPNN(libfile='ffield.json',
direcs=direcs,
dft='siesta',
weight={
'c2h6-CC': 50.0,
'others': 2.0
},
messages=1,
bo_layer=[4, 1],
bf_layer=[12, 3],
be_layer=[9, 2],
EnergyFunction=3,
MessageFunction=1,
pkl=False,
batch_size=50,
losFunc='n2',
convergence=0.98)
mn.run(learning_rate=1.0e-4, step=100, print_step=10, writelib=100)
mn.close()
def e(traj='siesta.traj', batch_size=50, nn=True):
ffield = 'ffield.json' if nn else 'ffield'
images = Trajectory(traj)
e = []
eb, el, eo, eu = [], [], [], []
ea, ep, etc = [], [], []
et, ef = [], []
ev, eh, ec = [], [], []
from irff.irff import IRFF
ir = IRFF(atoms=images[0], libfile=ffield, nn=nn)
for i, atoms in enumerate(images):
ir.get_potential_energy(atoms)
e.append(ir.E.numpy())
eb.append(ir.Ebond.numpy())
el.append(ir.Elone.numpy())
eo.append(ir.Eover.numpy())
eu.append(ir.Eunder.numpy())
ea.append(ir.Eang.numpy())
ep.append(ir.Epen.numpy())
et.append(ir.Etor.numpy())
ef.append(ir.Efcon.numpy())
etc.append(ir.Etcon.numpy())
ev.append(ir.Evdw.numpy())
eh.append(ir.Ehb.numpy())
ec.append(ir.Ecoul.numpy())
from irff.mpnn import MPNN
rn = MPNN(libfile=ffield,
direcs={'tmp': traj},
dft='siesta',
opt=[],
optword='nocoul',
batch_size=batch_size,
atomic=True,
clip_op=False,
InitCheck=False,
nn=nn,
bo_layer=[9, 2],
pkl=False)
molecules = rn.initialize()
rn.session(learning_rate=1.0e-10, method='AdamOptimizer')
mol = 'tmp'
e_ = rn.get_value(rn.E[mol])
eb_ = rn.get_value(rn.ebond[mol])
el_ = rn.get_value(rn.elone[mol])
eu_ = rn.get_value(rn.eunder[mol])
eo_ = rn.get_value(rn.eover[mol])
ea_ = rn.get_value(rn.eang[mol])
ep_ = rn.get_value(rn.epen[mol])
etc_ = rn.get_value(rn.tconj[mol])
et_ = rn.get_value(rn.etor[mol])
ef_ = rn.get_value(rn.efcon[mol])
ev_ = rn.get_value(rn.evdw[mol])
ec_ = rn.get_value(rn.ecoul[mol])
eh_ = rn.get_value(rn.ehb[mol])
e_reax = {
'Energy': e,
'Ebond': eb,
'Eunger': eu,
'Eover': eo,
'Eang': ea,
'Epen': ep,
'Elone': el,
'Etcon': etc,
'Etor': et,
'Efcon': ef,
'Evdw': ev,
'Ecoul': ec,
'Ehbond': eh
}
e_irff = {
'Energy': e_,
'Ebond': eb_,
'Eunger': eu_,
'Eover': eo_,
'Eang': ea_,
'Epen': ep_,
'Elone': el_,
'Etcon': etc_,
'Etor': et_,
'Efcon': ef_,
'Evdw': ev_,
'Ecoul': ec_,
'Ehbond': eh_
}
for key in e_reax:
plt.figure()
plt.ylabel('%s (eV)' % key)
plt.xlabel('Step')
plt.plot(e_reax[key],
alpha=0.01,
linestyle='-.',
marker='o',
markerfacecolor='none',
markeredgewidth=1,
markeredgecolor='b',
markersize=4,
color='blue',
label='ReaxFF')
plt.plot(e_irff[key],
alpha=0.01,
linestyle=':',
marker='^',
markerfacecolor='none',
markeredgewidth=1,
markeredgecolor='red',
markersize=4,
color='red',
label='IRFF')
plt.legend(loc='best',
edgecolor='yellowgreen') # lower left upper right
plt.savefig('%s.eps' % key)
plt.close()
def db(gen='C3H7O1-0.traj', batch_size=50, nn=True, frame=7):
ffield = 'ffield.json' if nn else 'ffield'
# atoms = read(gen,index=1)
images = Trajectory(gen)
atoms = images[frame]
# his = TrajectoryWriter('tmp.traj',mode='w')
# calc = SinglePointCalculator(atoms,energy=0.0,free_energy=0.0)
# atoms.set_calculator(calc)
# his.write(atoms=atoms)
# his.close()
from irff.irff import IRFF
e, ei = [], []
ir = IRFF(atoms=atoms, libfile=ffield, nn=nn)
ir.get_potential_energy(atoms)
ei.append(ir.Ebond)
ebond = ir.ebond.numpy()
from irff.mpnn import MPNN
rn = MPNN(libfile=ffield,
direcs={'tmp': gen},
dft='siesta',
opt=[],
optword='nocoul',
batch_size=batch_size,
atomic=True,
clip_op=False,
InitCheck=False,
nn=nn,
bo_layer=[9, 2],
pkl=False)
molecules = rn.initialize()
rn.session(learning_rate=1.0e-10, method='AdamOptimizer')
mol = 'tmp'
er = rn.get_value(rn.ebond[mol])
blab = rn.lk.bdlab
bosi = ir.bosi.numpy()
bopsi = ir.bop_si.numpy()
boppi = ir.bop_pi.numpy()
boppp = ir.bop_pp.numpy()
bopow3 = ir.bopow3.numpy()
eterm3 = ir.eterm3.numpy()
bop = ir.bop.numpy()
bo = ir.bo0.numpy()
eterm1 = ir.eterm1.numpy()
bopow1 = ir.bopow1.numpy()
bodiv1 = ir.bodiv1.numpy()
r = ir.r.numpy()
# F = ir.F.numpy()
# Fi = ir.Fi.numpy()
# Fj = ir.Fj.numpy()
D_ = rn.get_value(rn.Deltap)
D = ir.Deltap.numpy()
for bd in rn.bonds:
for nb in range(rn.nbd[bd]):
ebd = rn.get_value(rn.EBD[bd])
mol_, i, j = blab[bd][nb]
bosi_ = rn.get_value(rn.bosi[bd])
bopsi_ = rn.get_value(rn.bop_si[bd])
boppi_ = rn.get_value(rn.bop_pi[bd])
boppp_ = rn.get_value(rn.bop_pp[bd])
bopow3_ = rn.get_value(rn.bopow3[bd])
eterm3_ = rn.get_value(rn.eterm3[bd])
bop_ = rn.get_value(rn.bop[bd])
bo_ = rn.get_value(rn.bo0[bd])
eterm1_ = rn.get_value(rn.eterm1[bd])
bopow1_ = rn.get_value(rn.bopow1[bd])
bodiv1_ = rn.get_value(rn.bodiv1[bd])
rbd = rn.get_value(rn.rbd[bd])
# if abs(bo_[nb][0]-bo[i][j])>0.00001:
print(
'- %s %2d %2d:' % (bd, i, j),
'rbd %10.7f %10.7f' % (rbd[nb][frame], r[i][j]),
'bop %10.7f %10.7f' % (bop_[nb][frame], bop[i][j]),
'Di %10.7f %10.7f' % (D_[i][frame], D[i]),
'Dj %10.7f %10.7f' % (D_[j][frame], D[j]),
'bo %10.7f %10.7f' % (bo_[nb][frame], bo[i][j]),
# 'F %10.7f %10.7f' %(F_[nb][0],F[i][j]),
# 'Fi %10.7f %10.7f' %(Fi_[nb][0],Fi[i][j]),
# 'Fj %10.7f %10.7f' %(Fj_[nb][0],Fj[i][j]),
'ebond %10.7f %10.7f' % (ebd[nb][frame], ebond[i][j]))
rcbo = rn.get_value(rn.rc_bo)
eb = rn.get_value(rn.ebond[mol])
# print(rcbo)
print('\n- bond energy:', ir.Ebond.numpy(), eb[frame], end='\n')
def da(traj='opt.traj', batch_size=50, ef=3, nn=True, frame=44):
ffield = 'ffield.json' if nn else 'ffield'
images = Trajectory(traj)
atoms = images[frame]
his = TrajectoryWriter('tmp.traj', mode='w')
his.write(atoms=atoms)
# his.write(atoms=images[frame+1])
his.close()
ir = IRFF(atoms=atoms, libfile=ffield, nn=nn, autograd=False)
ir.calculate(atoms)
rn = MPNN(libfile=ffield,
direcs={'tmp': 'tmp.traj'},
dft='siesta',
opt=[],
optword='nocoul',
batch_size=batch_size,
atomic=True,
clip_op=False,
InitCheck=False,
nn=nn,
bo_layer=[9, 2],
EnergyFunction=ef,
pkl=False)
molecules = rn.initialize()
rn.session(learning_rate=1.0e-10, method='AdamOptimizer')
mol = 'tmp'
anglab = rn.lk.anglab
angs = []
for ang in ir.angs:
angs.append(list(ang))
ea_ = rn.get_value(rn.EANG)
ea = ir.eang.numpy()
f7_ = rn.get_value(rn.f_7)
f7 = ir.f_7.numpy()
f8_ = rn.get_value(rn.f_8)
f8 = ir.f_8.numpy()
expang_ = rn.get_value(rn.expang)
expang = ir.expang.numpy()
theta_ = rn.get_value(rn.theta)
theta = ir.theta.numpy()
theta0_ = rn.get_value(rn.theta0)
theta0 = ir.thet0.numpy()
sbo3_ = rn.get_value(rn.SBO3)
sbo3 = ir.SBO3.numpy()
fa = open('ang.txt', 'w')
for ang in rn.angs:
for a_ in range(rn.nang[ang]):
a = anglab[ang][a_][1:]
if a not in angs:
a.reverse()
i, j, k = a
if a in angs:
ai = angs.index(a)
print(
' * %2d%s-%2d%s-%2d%s:' %
(i, ir.atom_name[i], j, ir.atom_name[j], k,
ir.atom_name[k]),
#'eang: %10.8f %10.8f' %(ea_[ang][a_][0],ea[ai]),
'f7: %10.8f %10.8f' % (f7_[ang][a_][0], f7[ai]),
'f8: %10.8f %10.8f' % (f8_[ang][a_][0], f8[ai]),
'expang: %10.8f %10.8f' %
(expang_[ang][a_][0], expang[ai]),
'theta: %10.8f %10.8f' % (theta_[ang][a_][0], theta[ai]),
# 'sbo3: %10.8f %10.8f' %(sbo3_[ang][a_][0],sbo3[ai]),
'theta0: %10.8f %10.8f' %
(theta0_[ang][a_][0], theta0[ai]),
file=fa)
else:
print(
' * %2d%s-%2d%s-%2d%s:' %
(i, ir.atom_name[i], j, ir.atom_name[j], k,
ir.atom_name[k]),
#'eang: %10.8f' %(ea_[ang][a_][0]),
'f7: %10.8f' % (f7_[ang][a_][0]),
'f8: %10.8f' % (f8_[ang][a_][0]),
'expang: %10.8f' % (expang_[ang][a_][0]),
'expang: %10.8f' % (expang_[ang][a_][0]),
'theta: %10.8f' % (theta_[ang][a_][0]),
# 'sbo3: %10.8f' %(sbo3_[ang][a_][0]),
'theta0: %10.8f' % (theta0_[ang][a_][0]))
fa.close()
epen = rn.get_value(rn.epen)
eang = rn.get_value(rn.eang)
print(' * penalty energy:', ir.Epen.numpy(), epen[mol][0])
print(' * angel energy:', ir.Eang.numpy(), eang[mol][0])
def db(traj='opt.traj', batch_size=50, ef=3, nn=True, frame=40):
ffield = 'ffield.json' if nn else 'ffield'
images = Trajectory(traj)
atoms = images[frame]
e, ei = [], []
ir = IRFF(atoms=atoms, libfile=ffield, nn=nn, autograd=False)
ir.calculate(atoms)
ei.append(ir.Ebond)
ebond = ir.ebond # .numpy()
rn = MPNN(libfile=ffield,
direcs={'tmp': traj},
dft='siesta',
opt=[],
optword='nocoul',
batch_size=batch_size,
atomic=True,
clip_op=False,
InitCheck=False,
nn=nn,
bo_layer=[9, 2],
EnergyFunction=ef,
pkl=False)
molecules = rn.initialize()
rn.session(learning_rate=1.0e-10, method='AdamOptimizer')
mol = 'tmp'
er = rn.get_value(rn.ebond[mol])
blab = rn.lk.bdlab
bosi = ir.bosi.numpy()
bopsi = ir.bop_si.numpy()
boppi = ir.bop_pi.numpy()
boppp = ir.bop_pp.numpy()
bopow3 = ir.bopow3.numpy()
eterm3 = ir.eterm3.numpy()
bop = ir.bop.numpy()
bo = ir.bo0.numpy()
eterm1 = ir.eterm1.numpy()
bopow1 = ir.bopow1.numpy()
bodiv1 = ir.bodiv1.numpy()
r = ir.r.numpy()
D_ = rn.get_value(rn.Deltap)
D = ir.Deltap.numpy()
ebd_ = np.zeros([ir.natom, ir.natom])
for bd in rn.bonds:
for nb in range(rn.nbd[bd]):
ebd = rn.get_value(rn.EBD[bd])
mol_, i, j = blab[bd][nb]
bosi_ = rn.get_value(rn.bosi[bd])
bopsi_ = rn.get_value(rn.bop_si[bd])
boppi_ = rn.get_value(rn.bop_pi[bd])
boppp_ = rn.get_value(rn.bop_pp[bd])
bopow3_ = rn.get_value(rn.bopow3[bd])
eterm3_ = rn.get_value(rn.eterm3[bd])
bop_ = rn.get_value(rn.bop[bd])
bo_ = rn.get_value(rn.bo0[bd])
eterm1_ = rn.get_value(rn.eterm1[bd])
bopow1_ = rn.get_value(rn.bopow1[bd])
bodiv1_ = rn.get_value(rn.bodiv1[bd])
rbd = rn.get_value(rn.rbd[bd])
ebd_[i][j] = ebd[nb][frame]
ebd_[j][i] = ebd[nb][frame]
# if abs(bo_[nb][0]-bo[i][j])>0.00001:
print('- %s %2d %2d:' % (bd, i, j),
'rbd %10.7f %10.7f' % (rbd[nb][frame], r[i][j]),
'bop %10.7f %10.7f' % (bop_[nb][frame], bop[i][j]),
'Di %10.7f %10.7f' % (D_[i][frame], D[i]),
'Dj %10.7f %10.7f' % (D_[j][frame], D[j]),
'bo %10.7f %10.7f' % (bo_[nb][frame], bo[i][j]),
'ebond %10.7f %10.7f' % (ebd[nb][frame], ebond[i][j]))
rcbo = rn.get_value(rn.rc_bo)
eb = rn.get_value(rn.ebond[mol])
print('\n- bond energy:', ir.Ebond.numpy(), eb[frame], end='\n')
for i in range(ir.natom):
for j in range(ir.natom):
if abs(ebd_[i][j] - ebond[i][j]) > 0.00001:
print('- %s %2d %2d:' % (bd, i, j), 'rbd %10.7f' % (r[i][j]),
'bop %10.7f' % (bop[i][j]),
'Di %10.7f %10.7f' % (D_[i][frame], D[i]),
'Dj %10.7f %10.7f' % (D_[j][frame], D[j]),
'bo %10.7f' % (bo[i][j]),
'ebond %10.7f %10.7f' % (ebd_[i][j], ebond[i][j]))
# print(rcbo)
print('\n- bond energy:', ir.Ebond.numpy(), eb[frame], end='\n')