Source code for ase.md.contour_exploration

import numpy as np
from ase.optimize.optimize import Dynamics


def subtract_projection(a, b):
    '''returns new vector that removes vector a's projection vector b. Is
    also equivalent to the vector rejection.'''
    aout = a - np.vdot(a, b) / np.vdot(b, b) * b
    return aout


def normalize(a):
    '''Makes a unit vector out of a vector'''
    return a / np.linalg.norm(a)


[docs]class ContourExploration(Dynamics): def __init__(self, atoms, maxstep=0.5, parallel_drift=0.1, energy_target=None, angle_limit=20, potentiostat_step_scale=None, remove_translation=False, use_frenet_serret=True, initialization_step_scale=1e-2, use_target_shift=True, target_shift_previous_steps=10, use_tangent_curvature=False, rng=np.random, force_consistent=None, trajectory=None, logfile=None, append_trajectory=False, loginterval=1): """Contour Exploration object. Parameters: atoms: Atoms object The Atoms object to operate on. Atomic velocities are required for the method. If the atoms object does not contain velocities, random ones will be applied. maxstep: float Used to set the maximum distance an atom can move per iteration (default value is 0.5 Å). parallel_drift: float The fraction of the update step that is parallel to the contour but in a random direction. Used to break symmetries. energy_target: float The total system potential energy for that the potentiostat attepts to maintain. (defaults the initial potential energy) angle_limit: float or None Limits the stepsize to a maximum change of direction angle using the curvature. Gives a scale-free means of tuning the stepsize on the fly. Typically less than 30 degrees gives reasonable results but lower angle limits result in higher potentiostatic accuracy. Units of degrees. (default 20°) potentiostat_step_scale: float or None Scales the size of the potentiostat step. The potentiostat step is determined by linear extrapolation from the current potential energy to the target_energy with the current forces. A potentiostat_step_scale > 1.0 overcorrects and < 1.0 undercorrects. By default, a simple heuristic is used to selected the valued based on the parallel_drift. (default None) remove_translation: boolean When True, the net momentum is removed at each step. Improves potentiostatic accuracy slightly for bulk systems but should not be used with constraints. (default False) use_frenet_serret: Bool Controls whether or not the Taylor expansion of the Frenet-Serret formulas for curved path extrapolation are used. Required for using angle_limit based step scalling. (default True) initialization_step_scale: float Controls the scale of the initial step as a multiple of maxstep. (default 1e-2) use_target_shift: boolean Enables shifting of the potentiostat target to compensate for systematic undercorrection or overcorrection by the potentiostat. Uses the average of the *target_shift_previous_steps* to prevent coupled occilations. (default True) target_shift_previous_steps: int The number of pevious steps to average when using use_target_shift. (default 10) use_tangent_curvature: boolean Use the velocity unit tangent rather than the contour normals from forces to compute the curvature. Usually not as accurate. (default False) rng: a random number generator Lets users control the random number generator for the parallel_drift vector. (default numpy.random) force_consistent: boolean (default None) trajectory: Trajectory object or str (optional) Attach trajectory object. If *trajectory* is a string a Trajectory will be constructed. Default: None. logfile: file object or str (optional) If *logfile* is a string, a file with that name will be opened. Use '-' for stdout. Default: None. loginterval: int (optional) Only write a log line for every *loginterval* time steps. Default: 1 append_trajectory: boolean Defaults to False, which causes the trajectory file to be overwriten each time the dynamics is restarted from scratch. If True, the new structures are appended to the trajectory file instead. """ if potentiostat_step_scale is None: # a heuristic guess since most systems will overshoot when there is # drift self.potentiostat_step_scale = 1.1 + 0.6 * parallel_drift else: self.potentiostat_step_scale = potentiostat_step_scale self.rng = rng self.remove_translation = remove_translation self.use_frenet_serret = use_frenet_serret self.force_consistent = force_consistent self.use_tangent_curvature = use_tangent_curvature self.initialization_step_scale = initialization_step_scale self.maxstep = maxstep self.angle_limit = angle_limit self.parallel_drift = parallel_drift self.use_target_shift = use_target_shift # These will be populated once self.step() is called, but could be set # after instantiating with ce = ContourExploration(...) like so: # ce.Nold = Nold # ce.r_old = atoms_old.get_positions() # ce.Told = Told # to resume a previous contour trajectory. self.T = None self.Told = None self.N = None self.Nold = None self.r_old = None self.r = None if energy_target is None: self.energy_target = atoms.get_potential_energy( force_consistent=self.force_consistent) else: self.energy_target = energy_target # Initizing the previous steps at the target energy slows # target_shifting untill the system has had # 'target_shift_previous_steps' steps to equilibrate and should prevent # occilations. These need to be initialized before the initialize_old # step to prevent a crash self.previous_energies = np.full(target_shift_previous_steps, self.energy_target) # these first two are purely for logging, # auto scaling will still occur # and curvature will still be found if use_frenet_serret == True self.step_size = 0.0 self.curvature = 0 # loginterval exists for the MolecularDynamics class but not for # the more general Dynamics class Dynamics.__init__(self, atoms, logfile, trajectory, # loginterval, append_trajectory=append_trajectory, ) # we need velocities or NaNs will be produced, # if none are provided we make random ones velocities = self.atoms.get_velocities() if np.linalg.norm(velocities) < 1e-6: # we have to pass dimension since atoms are not yet stored atoms.set_velocities(self.rand_vect()) # Required stuff for Dynamics def todict(self): return {'type': 'contour-exploration', 'dyn-type': self.__class__.__name__, 'stepsize': self.step_size} def run(self, steps=50): """ Call Dynamics.run and adjust max_steps """ self.max_steps = steps + self.nsteps return Dynamics.run(self) def log(self): if self.logfile is not None: # name = self.__class__.__name__ if self.nsteps == 0: args = ( "Step", "Energy_Target", "Energy", "Curvature", "Step_Size", "Energy_Deviation_per_atom") msg = "# %4s %15s %15s %12s %12s %15s\n" % args self.logfile.write(msg) e = self.atoms.get_potential_energy( force_consistent=self.force_consistent) dev_per_atom = (e - self.energy_target) / len(self.atoms) args = ( self.nsteps, self.energy_target, e, self.curvature, self.step_size, dev_per_atom) msg = "%6d %15.6f %15.6f %12.6f %12.6f %24.9f\n" % args self.logfile.write(msg) self.logfile.flush() def rand_vect(self): '''Returns a random (Natoms,3) vector''' vect = self.rng.rand(len(self.atoms), 3) - 0.5 return vect def create_drift_unit_vector(self, N, T): '''Creates a random drift unit vector with no projection on N or T and with out a net translation so systems don't wander''' drift = self.rand_vect() drift = subtract_projection(drift, N) drift = subtract_projection(drift, T) # removes net translation, so systems don't wander drift = drift - drift.sum(axis=0) / len(self.atoms) D = normalize(drift) return D def compute_step_contributions(self, potentiostat_step_size): '''Computes the orthogonal component sizes of the step so that the net step obeys the smaller of step_size or maxstep.''' if abs(potentiostat_step_size) < self.step_size: delta_s_perpendicular = potentiostat_step_size contour_step_size = np.sqrt( self.step_size**2 - potentiostat_step_size**2) delta_s_parallel = np.sqrt( 1 - self.parallel_drift**2) * contour_step_size delta_s_drift = contour_step_size * self.parallel_drift else: # in this case all priority goes to potentiostat terms delta_s_parallel = 0.0 delta_s_drift = 0.0 delta_s_perpendicular = np.sign( potentiostat_step_size) * self.step_size return delta_s_perpendicular, delta_s_parallel, delta_s_drift def _compute_update_without_fs(self, potentiostat_step_size, scale=1.0): '''Only uses the forces to compute an orthogonal update vector''' # Without the use of curvature there is no way to estimate the # limiting step size self.step_size = self.maxstep * scale delta_s_perpendicular, delta_s_parallel, delta_s_drift = \ self.compute_step_contributions( potentiostat_step_size) dr_perpendicular = self.N * delta_s_perpendicular dr_parallel = delta_s_parallel * self.T D = self.create_drift_unit_vector(self.N, self.T) dr_drift = D * delta_s_drift dr = dr_parallel + dr_drift + dr_perpendicular dr = self.step_size * normalize(dr) return dr def _compute_update_with_fs(self, potentiostat_step_size): '''Uses the Frenet–Serret formulas to perform curvature based extrapolation to compute the update vector''' # this should keep the dr clear of the constraints # by using the actual change, not a velocity vector delta_r = self.r - self.rold delta_s = np.linalg.norm(delta_r) # approximation of delta_s we use this incase an adaptive step_size # algo get used delta_T = self.T - self.Told delta_N = self.N - self.Nold dTds = delta_T / delta_s dNds = delta_N / delta_s if self.use_tangent_curvature: curvature = np.linalg.norm(dTds) # on a perfect trajectory, the normal can be computed this way, # But the normal should always be tied to forces # N = dTds / curvature else: # normals are better since they are fixed to the reality of # forces. I see smaller forces and energy errors in bulk systems # using the normals for curvature curvature = np.linalg.norm(dNds) self.curvature = curvature if self.angle_limit is not None: phi = np.pi / 180 * self.angle_limit self.step_size = np.sqrt(2 - 2 * np.cos(phi)) / curvature self.step_size = min(self.step_size, self.maxstep) # now we can compute a safe step delta_s_perpendicular, delta_s_parallel, delta_s_drift = \ self.compute_step_contributions( potentiostat_step_size) N_guess = self.N + dNds * delta_s_parallel T_guess = self.T + dTds * delta_s_parallel # the extrapolation is good at keeping N_guess and T_guess # orthogonal but not normalized: N_guess = normalize(N_guess) T_guess = normalize(T_guess) dr_perpendicular = delta_s_perpendicular * (N_guess) dr_parallel = delta_s_parallel * self.T * \ (1 - (delta_s_parallel * curvature)**2 / 6.0) \ + self.N * (curvature / 2.0) * delta_s_parallel**2 D = self.create_drift_unit_vector(N_guess, T_guess) dr_drift = D * delta_s_drift # combine the components dr = dr_perpendicular + dr_parallel + dr_drift dr = self.step_size * normalize(dr) # because we guess our orthonormalization directions, # we renormalize to ensure a correct step size return dr def update_previous_energies(self, energy): '''Updates the energy history in self.previous_energies to include the current energy.''' # np.roll shifts the values to keep nice sequential ordering. self.previous_energies = np.roll(self.previous_energies, 1) self.previous_energies[0] = energy def compute_potentiostat_step_size(self, forces, energy): '''Computes the potentiostat step size by linear extrapolation of the potential energy using the forces. The step size can be positive or negative depending on whether or not the energy is too high or too low. ''' if self.use_target_shift: target_shift = self.energy_target - np.mean(self.previous_energies) else: target_shift = 0.0 # deltaU is the potential error that will be corrected for deltaU = energy - (self.energy_target + target_shift) f_norm = np.linalg.norm(forces) # can be positive or negative potentiostat_step_size = (deltaU / f_norm) * \ self.potentiostat_step_scale return potentiostat_step_size def step(self, f=None): atoms = self.atoms if f is None: f = atoms.get_forces() # get the velocity vector and old kinetic energy for momentum rescaling velocities = atoms.get_velocities() KEold = atoms.get_kinetic_energy() energy = atoms.get_potential_energy( force_consistent=self.force_consistent) self.update_previous_energies(energy) potentiostat_step_size = self.compute_potentiostat_step_size(f, energy) self.N = normalize(f) self.r = atoms.get_positions() # remove velocity projection on forces v_parallel = subtract_projection(velocities, self.N) self.T = normalize(v_parallel) if self.use_frenet_serret: if self.Nold is not None and self.Told is not None: dr = self._compute_update_with_fs(potentiostat_step_size) else: # we must have the old positions and vectors for an FS step # if we don't, we can only do a small step dr = self._compute_update_without_fs( potentiostat_step_size, scale=self.initialization_step_scale) else: # of course we can run less accuratly without FS. dr = self._compute_update_without_fs(potentiostat_step_size) # now that dr is done, we check if there is translation if self.remove_translation: net_motion = dr.sum(axis=0) / len(atoms) # print(net_motion) dr = dr - net_motion dr_unit = dr / np.linalg.norm(dr) dr = dr_unit * self.step_size # save old positions before update self.Nold = self.N self.rold = self.r self.Told = self.T # if we have constraints then this will do the first part of the # RATTLE algorithm: # If we can avoid using momenta, this will be simpler. masses = atoms.get_masses()[:, np.newaxis] atoms.set_positions(self.r + dr) new_momenta = (atoms.get_positions() - self.r) * masses # / self.dt # We need to store the momenta on the atoms before calculating # the forces, as in a parallel Asap calculation atoms may # migrate during force calculations, and the momenta need to # migrate along with the atoms. atoms.set_momenta(new_momenta, apply_constraint=False) # Now we get the new forces! f = atoms.get_forces(md=True) # I don't really know if removing md=True from above will break # compatibility with RATTLE, leaving it alone for now. f_constrained = atoms.get_forces() # but this projection needs the forces to be consistent with the # constraints. We have to set the new velocities perpendicular so they # get logged properly in the trajectory files. vnew = subtract_projection(atoms.get_velocities(), f_constrained) # using the md = True forces like this: # vnew = subtract_projection(atoms.get_velocities(), f) # will not work with constraints atoms.set_velocities(vnew) # rescaling momentum to maintain constant kinetic energy. KEnew = atoms.get_kinetic_energy() Ms = np.sqrt(KEold / KEnew) # Ms = Momentum_scale atoms.set_momenta(Ms * atoms.get_momenta()) # Normally this would be the second part of RATTLE # will be done here like this: # atoms.set_momenta(atoms.get_momenta() + 0.5 * self.dt * f) return f