def step(self, step=None):
""" Does one simulation time step
Attributes:
ttime: The time taken in applying the thermostat steps.
"""
self.qtime = -time.time()
info("\nMD STEP %d" % step, verbosity.debug)
if step == 0:
info(" @GEOP: Initializing L-BFGS", verbosity.debug)
print self.d
self.d += dstrip(self.forces.f) / np.sqrt(np.dot(self.forces.f.flatten(), self.forces.f.flatten()))
self.old_x[:] = self.beads.q
self.old_u[:] = self.forces.pot
self.old_f[:] = self.forces.f
if len(self.fixatoms) > 0:
for dqb in self.old_f:
dqb[self.fixatoms * 3] = 0.0
dqb[self.fixatoms * 3 + 1] = 0.0
dqb[self.fixatoms * 3 + 2] = 0.0
# Reduce the dimensionality
masked_old_x = self.old_x[:, self.gm.fixatoms_mask]
masked_d = self.d[:, self.gm.fixatoms_mask]
# self.gm is reduced inside its __init__() and __call__() functions
masked_qlist = self.qlist[:, self.gm.fixatoms_mask]
masked_glist = self.glist[:, self.gm.fixatoms_mask]
fdf0 = (self.old_u, -self.old_f[:, self.gm.fixatoms_mask])
# We update everything within L_BFGS (and all other calls).
L_BFGS(masked_old_x, masked_d, self.gm, masked_qlist, masked_glist, fdf0,
self.big_step, self.ls_options["tolerance"] * self.tolerances["energy"],
self.ls_options["iter"], self.corrections, self.scale, step)
# Restore the dimensionality
self.d[:, self.gm.fixatoms_mask] = masked_d
self.qlist[:, self.gm.fixatoms_mask] = masked_qlist
self.glist[:, self.gm.fixatoms_mask] = masked_glist
else:
fdf0 = (self.old_u, -self.old_f)
# We update everything within L_BFGS (and all other calls).
L_BFGS(self.old_x, self.d, self.gm, self.qlist, self.glist, fdf0,
self.big_step, self.ls_options["tolerance"] * self.tolerances["energy"],
self.ls_options["iter"], self.corrections, self.scale, step)
info(" Number of force calls: %d" % (self.gm.fcount)); self.gm.fcount = 0
# Update positions and forces
self.beads.q = self.gm.dbeads.q
self.forces.transfer_forces(self.gm.dforces) # This forces the update of the forces
# Exit simulation step
d_x_max = np.amax(np.absolute(np.subtract(self.beads.q, self.old_x)))
self.exitstep(self.forces.pot, self.old_u, d_x_max)
def step(self, step=None):
""" Does one simulation time step
Attributes:
ttime: The time taken in applying the thermostat steps.
"""
self.qtime = -time.time()
info("\nMD STEP %d" % step, verbosity.debug)
if step == 0:
info(" @GEOP: Initializing L-BFGS", verbosity.debug)
self.d += depstrip(self.forces.f) / np.sqrt(
np.dot(self.forces.f.flatten(), self.forces.f.flatten()))
self.old_x[:] = self.beads.q
self.old_u[:] = self.forces.pot
self.old_f[:] = self.forces.f
if len(self.fixatoms) > 0:
for dqb in self.old_f:
dqb[self.fixatoms * 3] = 0.0
dqb[self.fixatoms * 3 + 1] = 0.0
dqb[self.fixatoms * 3 + 2] = 0.0
fdf0 = (self.old_u, -self.old_f)
#d_x,new_d, new_qlist, new_glist = L_BFGS(self.old_x,
# Note that the line above is not needed anymore because we update everything
# within L_BFGS (and all other calls).
L_BFGS(self.old_x, self.d, self.gm, self.qlist, self.glist, fdf0,
self.big_step,
self.ls_options["tolerance"] * self.tolerances["energy"],
self.ls_options["iter"], self.corrections, self.scale, step)
info(" Number of force calls: %d" % (self.gm.fcount))
self.gm.fcount = 0
#Update positions and forces
self.beads.q = self.gm.dbeads.q
self.forces.transfer_forces(
self.gm.dforces) #This forces the update of the forces
# Exit simulation step
d_x_max = np.amax(np.absolute(np.subtract(self.beads.q, self.old_x)))
self.exitstep(self.forces.pot, self.old_u, d_x_max)
def step(self, step=None):
"""Does one simulation time step."""
self.ptime = 0.0
self.ttime = 0.0
self.qtime = -time.time()
info("\nMD STEP %d" % step, verbosity.debug)
if self.mode == "bfgs":
# BFGS Minimization
# Initialize approximate Hessian inverse to the identity and direction
# to the steepest descent direction
if step == 0: # or np.sqrt(np.dot(self.bfgsm.d, self.bfgsm.d)) == 0.0: <-- this part for restarting at claimed minimum (optional)
info(" @GEOP: Initializing BFGS", verbosity.debug)
self.bfgsm.d = depstrip(self.forces.f) / np.sqrt(np.dot(self.forces.f.flatten(), self.forces.f.flatten()))
self.bfgsm.xold = self.beads.q.copy()
# Current energy and forces
u0 = self.forces.pot.copy()
du0 = - self.forces.f
# Store previous forces
self.cg_old_f[:] = self.forces.f
# Do one iteration of BFGS, return new point, function value,
# move direction, and current Hessian to use for next iteration
self.beads.q, fx, self.bfgsm.d, self.invhessian = BFGS(self.beads.q,
self.bfgsm.d, self.bfgsm, fdf0=(u0, du0), invhessian=self.invhessian,
max_step=self.max_step, tol=self.ls_options["tolerance"],
itmax=self.ls_options["iter"])
# x = current position - previous position; use for exit tolerance
x = np.amax(np.absolute(np.subtract(self.beads.q, self.bfgsm.xold)))
# Store old position
self.bfgsm.xold[:] = self.beads.q
info(" @GEOP: Updating bead positions", verbosity.debug)
elif self.mode == "lbfgs":
# L-BFGS Minimization
# Initialize approximate Hessian inverse to the identity and direction
# to the steepest descent direction
# Initialize lists of previous positions and gradient
if step == 0: # or np.sqrt(np.dot(self.bfgsm.d, self.bfgsm.d)) == 0.0: <-- this part for restarting at claimed minimum (optional)
info(" @GEOP: Initializing L-BFGS", verbosity.debug)
self.bfgsm.d = depstrip(self.forces.f) / np.sqrt(np.dot(self.forces.f.flatten(), self.forces.f.flatten()))
self.bfgsm.xold = self.beads.q.copy()
self.qlist = np.zeros((self.corrections, len(self.beads.q.flatten())))
self.glist = np.zeros((self.corrections, len(self.beads.q.flatten())))
# Current energy and force
u0, du0 = (self.forces.pot.copy(), - self.forces.f)
# Store previous forces
self.cg_old_f[:] = self.forces.f.reshape(len(self.cg_old_f))
# Do one iteration of L-BFGS, return new point, function value,
# move direction, and current Hessian to use for next iteration
self.beads.q, fx, self.bfgsm.d, self.qlist, self.glist = L_BFGS(self.beads.q,
self.bfgsm.d, self.bfgsm, self.qlist, self.glist,
fdf0=(u0, du0), max_step=self.max_step, tol=self.ls_options["tolerance"],
itmax=self.ls_options["iter"],
m=self.corrections, k=step)
info(" @GEOP: Updated position list", verbosity.debug)
info(" @GEOP: Updated gradient list", verbosity.debug)
# x = current position - old position. Used for convergence tolerance
x = np.amax(np.absolute(np.subtract(self.beads.q, self.bfgsm.xold)))
# Store old position
self.bfgsm.xold[:] = self.beads.q
info(" @GEOP: Updated bead positions", verbosity.debug)
# Routine for steepest descent and conjugate gradient
else:
if (self.mode == "sd" or step == 0):
# Steepest descent minimization
# gradf1 = force at current atom position
# dq1 = direction of steepest descent
# dq1_unit = unit vector of dq1
gradf1 = dq1 = depstrip(self.forces.f)
# Move direction for steepest descent and 1st conjugate gradient step
dq1_unit = dq1 / np.sqrt(np.dot(gradf1.flatten(), gradf1.flatten()))
info(" @GEOP: Determined SD direction", verbosity.debug)
else:
# Conjugate gradient, Polak-Ribiere
# gradf1: force at current atom position
# gradf0: force at previous atom position
# dq1 = direction to move
# dq0 = previous direction
# dq1_unit = unit vector of dq1
gradf0 = self.cg_old_f
dq0 = self.cg_old_d
gradf1 = depstrip(self.forces.f)
beta = np.dot((gradf1.flatten() - gradf0.flatten()), gradf1.flatten()) / (np.dot(gradf0.flatten(), gradf0.flatten()))
dq1 = gradf1 + max(0.0, beta) * dq0
dq1_unit = dq1 / np.sqrt(np.dot(dq1.flatten(), dq1.flatten()))
info(" @GEOP: Determined CG direction", verbosity.debug)
# Store force and direction for next CG step
self.cg_old_d[:] = dq1
self.cg_old_f[:] = gradf1
if len(self.fixatoms) > 0:
for dqb in dq1_unit:
dqb[self.fixatoms*3] = 0.0
dqb[self.fixatoms*3+1] = 0.0
dqb[self.fixatoms*3+2] = 0.0
self.lm.set_dir(depstrip(self.beads.q), dq1_unit)
# Reuse initial value since we have energy and forces already
u0, du0 = (self.forces.pot.copy(), np.dot(depstrip(self.forces.f.flatten()), dq1_unit.flatten()))
# Do one SD/CG iteration; return positions and energy
(x, fx) = min_brent(self.lm, fdf0=(u0, du0), x0=0.0,
tol=self.ls_options["tolerance"],
itmax=self.ls_options["iter"], init_step=self.ls_options["step"])
# Automatically adapt the search step for the next iteration.
# Relaxes better with very small step --> multiply by factor of 0.1 or 0.01
self.ls_options["step"] = 0.1 * x * self.ls_options["adaptive"] + (1 - self.ls_options["adaptive"]) * self.ls_options["step"]
self.beads.q += dq1_unit * x
info(" @GEOP: Updated bead positions", verbosity.debug)
self.qtime += time.time()
# Determine conditions for converged relaxation
if ((fx - u0) / self.beads.natoms <= self.tolerances["energy"])\
and ((np.amax(np.absolute(self.forces.f)) <= self.tolerances["force"])
or (np.sqrt(np.dot(self.forces.f.flatten() - self.cg_old_f.flatten(),
self.forces.f.flatten() - self.cg_old_f.flatten())) == 0.0))\
and (x <= self.tolerances["position"]):
softexit.trigger("Geometry optimization converged. Exiting simulation")
def step(self, step=None):
"""Does one simulation time step."""
info("\nMD STEP %d" % step, verbosity.debug)
# Fetch spring constants
self.nebbfgsm.kappa = self.spring["kappa"]
self.neblm.kappa = self.spring["kappa"]
self.ptime = self.ttime = 0
self.qtime = -time.time()
if self.mode == "lbfgs":
# L-BFGS Minimization
# Initialize direction to the steepest descent direction
if step == 0: # or np.sqrt(np.dot(self.bfgsm.d, self.bfgsm.d)) == 0.0: <-- this part for restarting at claimed minimum
info(" @GEOP: Initializing L-BFGS", verbosity.debug)
fx, nebgrad = self.nebbfgsm(self.beads.q)
# Set direction to direction of NEB forces
self.nebbfgsm.d = -nebgrad
self.nebbfgsm.xold = self.beads.q.copy()
# Initialize lists of previous positions and gradients
self.qlist = np.zeros(
(self.corrections, len(self.beads.q.flatten())))
self.glist = np.zeros(
(self.corrections, len(self.beads.q.flatten())))
else:
fx, nebgrad = self.nebbfgsm(self.beads.q)
# Intial gradient and gradient modulus
u0, du0 = (fx, nebgrad)
# Store old force
self.old_f[:] = -nebgrad
# Do one iteration of L-BFGS and return positions, gradient modulus,
# direction, list of positions, list of gradients
# self.beads.q, fx, self.nebbfgsm.d, self.qlist, self.glist = L_BFGS(self.beads.q,
L_BFGS(self.beads.q,
self.nebbfgsm.d,
self.nebbfgsm,
self.qlist,
self.glist,
fdf0=(u0, du0),
big_step=self.big_step,
tol=self.ls_options["tolerance"],
itmax=self.ls_options["iter"],
m=self.corrections,
scale=self.scale,
k=step)
info(" @GEOP: Updated position list", verbosity.debug)
info(" @GEOP: Updated gradient list", verbosity.debug)
# x = current position - previous position. Use to determine converged minimization
x = np.amax(
np.absolute(np.subtract(self.beads.q, self.nebbfgsm.xold)))
# Store old positions
self.nebbfgsm.xold[:] = self.beads.q
self.beads.q = self.nebbfgsm.dbeads.q
self.forces.transfer_forces(self.nebbfgsm.dforces)
info(" @GEOP: Updated bead positions", verbosity.debug)
# Routine for steepest descent and conjugate gradient
# TODO: CURRENTLY DOES NOT WORK. MUST BE ELIMINATED OR DEBUGGED
else:
if (self.mode == "sd" or step == 0):
# Steepest descent minimization
# gradf1 = force at current atom position
# dq1 = direction of steepest descent
# dq1_unit = unit vector of dq1
nebgrad = self.neblm(self.beads.q)[0]
gradf1 = dq1 = -nebgrad
# Move direction for steepest descent and 1st conjugate gradient step
dq1_unit = dq1 / np.sqrt(
np.dot(gradf1.flatten(), gradf1.flatten()))
info(" @GEOP: Determined SD direction", verbosity.debug)
else:
# Conjugate gradient, Polak-Ribiere
# gradf1: force at current atom position
# gradf0: force at previous atom position
# dq1 = direction to move
# dq0 = previous direction
# dq1_unit = unit vector of dq1
gradf0 = self.old_f
dq0 = self.old_d
nebgrad = self.neblm(self.beads.q)[0]
gradf1 = -nebgrad
beta = np.dot((gradf1.flatten() - gradf0.flatten()),
gradf1.flatten()) / (np.dot(
gradf0.flatten(), gradf0.flatten()))
dq1 = gradf1 + max(0.0, beta) * dq0
dq1_unit = dq1 / np.sqrt(np.dot(dq1.flatten(), dq1.flatten()))
info(" @GEOP: Determined CG direction", verbosity.debug)
# Store force and direction for next CG step
self.old_d[:] = dq1
self.old_f[:] = gradf1
if len(self.fixatoms) > 0:
for dqb in dq1_unit:
dqb[self.fixatoms * 3] = 0.0
dqb[self.fixatoms * 3 + 1] = 0.0
dqb[self.fixatoms * 3 + 2] = 0.0
self.neblm.set_dir(dstrip(self.beads.q), dq1_unit)
# Reuse initial value since we have energy and forces already
u0 = np.dot(-nebgrad.flatten(), dq1_unit.flatten())
u0 = np.sqrt(np.dot(u0, u0))
(x, fx) = min_brent_neb(self.neblm,
fdf0=u0,
x0=0.0,
tol=self.ls_options["tolerance"],
itmax=self.ls_options["iter"],
init_step=self.ls_options["step"])
# Automatically adapt the search step for the next iteration.
# Relaxes better with very small step --> multiply by factor of 0.1 or 0.01
self.ls_options["step"] = 0.1 * x * self.ls_options["adaptive"] + (
1 - self.ls_options["adaptive"]) * self.ls_options["step"]
self.beads.q += dq1_unit * x
info(" @GEOP: Updated bead positions", verbosity.debug)
self.qtime += time.time()
# Determine conditions for converged relaxation
if ((fx - u0) / self.beads.natoms <= self.tolerances["energy"])\
and ((np.amax(np.absolute(self.forces.f)) <= self.tolerances["force"])
or (np.sqrt(np.dot(self.forces.f.flatten() - self.old_f.flatten(),
self.forces.f.flatten() - self.old_f.flatten())) == 0.0))\
and (x <= self.tolerances["position"]):
softexit.trigger(
"Geometry optimization converged. Exiting simulation")
else:
info(
" @GEOP: Not converged, deltaEnergy = %.8f, tol = %.8f" %
((fx - u0 / self.beads.natoms), self.tolerances["energy"]),
verbosity.debug)
info(
" @GEOP: Not converged, force = %.8f, tol = %f" % (np.amax(
np.absolute(self.forces.f)), self.tolerances["force"]),
verbosity.debug)
info(
" @GEOP: Not converged, deltaForce = %.8f, tol = 0.00000000" %
(np.sqrt(
np.dot(self.forces.f.flatten() - self.old_f.flatten(),
self.forces.f.flatten() - self.old_f.flatten()))),
verbosity.debug)
info(
" @GEOP: Not converged, deltaX = %.8f, tol = %.8f" %
(x, self.tolerances["position"]), verbosity.debug)