Source code for mchammer.ensembles.wang_landau_ensemble

"""Definition of the Wang-Landau algorithm class."""

import random

from collections import OrderedDict
from typing import Dict, List, Optional, Tuple, Union

import numpy as np

from ase import Atoms

from .. import WangLandauDataContainer
from ..calculators.base_calculator import BaseCalculator
from .thermodynamic_base_ensemble import BaseEnsemble
from icet.input_output.logging_tools import logger

logger = logger.getChild('wang_landau_ensemble')


[docs] class WangLandauEnsemble(BaseEnsemble): r"""Instances of this class allow one to sample a system using the Wang-Landau (WL) algorithm, see Phys. Rev. Lett. **86**, 2050 (2001) [WanLan01a]_. The WL algorithm enables one to acquire the density of states (DOS) as a function of energy, from which one can readily calculate many thermodynamic observables as a function of temperature. To this end, the WL algorithm accumulates both the microcanonical entropy :math:`S(E)` and a histogram :math:`H(E)` on an energy grid with a predefined spacing (:attr:`energy_spacing`). The algorithm is initialized as follows. #. Generate an initial configuration. #. Initialize counters for the microcanonical entropy :math:`S(E)` and the histogram :math:`H(E)` to zero. #. Set the fill factor :math:`f=1`. It then proceeds as follows. #. Propose a new configuration (see :attr:`trial_move`). #. Accept or reject the new configuration with probability .. math:: P = \min \{ 1, \, \exp [ S(E_\mathrm{new}) - S(E_\mathrm{cur}) ] \}, where :math:`E_\mathrm{cur}` and :math:`E_\mathrm{new}` are the energies of the current and new configurations, respectively. #. Update the microcanonical entropy :math:`S(E)\leftarrow S(E) + f` and histogram :math:`H(E) \leftarrow H(E) + 1` where :math:`E` is the energy of the system at the end of the move. #. Check the flatness of the histogram :math:`H(E)`. If :math:`H(E) > \chi \langle H(E)\rangle\,\forall E` reset the histogram :math:`H(E) = 0` and reduce the fill factor :math:`f \leftarrow f / 2`. The parameter :math:`\chi` is set via :attr:`flatness_limit`. #. If :math:`f` is smaller than :attr:`fill_factor_limit` terminate the loop, otherwise return to 1. The microcanonical entropy :math:`S(E)` and the histogram along with related information are written to the data container every time :math:`f` is updated. Using the density :math:`\rho(E) = \exp S(E)` one can then readily compute various thermodynamic quantities, including, e.g., the average energy: .. math:: \left<E\right> = \frac{\sum_E E \rho(E) \exp(-E / k_B T)}{ \sum_E \rho(E) \exp(-E / k_B T)} Parameters ---------- structure Atomic configuration to be used in the Wang-Landau simulation; also defines the initial occupation vector. calculator Calculator to be used for calculating potential changes. trial_move One can choose between two different trial moves for generating new configurations. In a ``'swap'`` move two sites are selected and their occupations are swapped. In a ``'flip'`` move one site is selected and its occupation is flipped to a different species. While ``'swap'`` moves conserve the concentrations of the species in the system, ``'flip'`` moves allow one in principle to sample the full composition space. energy_spacing Sets the bin size of the energy grid on which the microcanonical entropy :math:`S(E)`, and thus the density :math:`\exp S(E)`, is evaluated. The spacing should be small enough to capture the features of the density of states. Too small values will, however, render the convergence very tedious if not impossible. energy_limit_left Sets the lower limit of the energy range within which the microcanonical entropy :math:`S(E)` will be sampled. By default (``None``) no limit is imposed. Setting limits can be useful if only a part of the density of states is required. energy_limit_right Sets the upper limit of the energy range within which the microcanonical entropy :math:`S(E)` will be sampled. By default (``None``) no limit is imposed. Setting limits can be useful if only a part of the density of states is required. fill_factor_limit If the fill_factor :math:`f` falls below this value, the WL sampling loop is terminated. flatness_check_interval For computational efficiency the flatness condition is only evaluated every :attr:`flatness_check_interval`-th trial step. By default (``None``) :attr:`flatness_check_interval` is set to 1000 times the number of sites in :attr:`structure`, i.e., 1000 Monte Carlo sweeps. flatness_limit The histogram :math:`H(E)` is deemed sufficiently flat if :math:`H(E) > \chi \left<H(E)\right>\,\forall E`. :attr:`flatness_limit` sets the parameter :math:`\chi`. window_search_penalty If :attr:`energy_limit_left` and/or :attr:`energy_limit_right` have been set, a modified acceptance probability, :math:`P=\min\{1,\,\exp[C_\mathrm{WSP}(d_\mathrm{new}- d_\mathrm{cur})]\}`, will be used until a configuration is found within the interval of interest. This parameter, specifically, corresponds to :math:`C_\mathrm{WSP}`, which controls how strongly moves that lead to an increase in the distance, i.e. difference in energy divided by the energy spacing, to the energy window (:math:`d_\mathrm{new}> d_\mathrm{cur}`) should be penalized. A higher value leads to a lower acceptance probability for such moves. user_tag Human-readable tag for ensemble. Default: ``None``. dc_filename Name of file the data container associated with the ensemble will be written to. If the file exists it will be read, the data container will be appended, and the file will be updated/overwritten. random_seed Seed for the random number generator used in the Monte Carlo simulation. ensemble_data_write_interval Interval at which data is written to the data container. This includes for example the current value of the calculator (i.e., usually the energy) as well as ensemble specific fields such as temperature or the number of atoms of different species. Default: Number of sites in the :attr:`structure`. data_container_write_period Period in units of seconds at which the data container is written to file. Writing periodically to file provides both a way to examine the progress of the simulation and to back up the data. Default: 600 s. trajectory_write_interval Interval at which the current occupation vector of the atomic configuration is written to the data container. Default: Number of sites in the :attr:`structure`. sublattice_probabilities Probability for picking a sublattice when doing a random swap. The list must contain as many elements as there are sublattices and it needs to sum up to 1. Example ------- The following snippet illustrates how to carry out a Wang-Landau simulation. For the purpose of demonstration, the parameters of the cluster expansion are set to obtain a two-dimensional square Ising model, one of the systems studied in the original work by Wang and Landau:: >>> from ase import Atoms >>> from icet import ClusterExpansion, ClusterSpace >>> from mchammer.calculators import ClusterExpansionCalculator >>> from mchammer.ensembles import WangLandauEnsemble >>> # prepare cluster expansion >>> prim = Atoms('Au', positions=[[0, 0, 0]], cell=[1, 1, 10], pbc=True) >>> cs = ClusterSpace(prim, cutoffs=[1.1], chemical_symbols=['Ag', 'Au']) >>> ce = ClusterExpansion(cs, [0, 0, 2]) >>> # prepare initial configuration >>> structure = prim.repeat((4, 4, 1)) >>> for k in range(8): ... structure[k].symbol = 'Ag' >>> # set up and run Wang-Landau simulation >>> calculator = ClusterExpansionCalculator(structure, ce) >>> mc = WangLandauEnsemble(structure=structure, ... calculator=calculator, ... energy_spacing=1, ... dc_filename='ising_2d_run.dc') >>> mc.run(number_of_trial_steps=len(structure)*100) >>> # Note: in practice one requires many more steps """ def __init__(self, structure: Atoms, calculator: BaseCalculator, energy_spacing: float, energy_limit_left: float = None, energy_limit_right: float = None, trial_move: str = 'swap', fill_factor_limit: float = 1e-6, flatness_check_interval: int = None, flatness_limit: float = 0.8, window_search_penalty: float = 2.0, user_tag: str = None, dc_filename: str = None, data_container: str = None, random_seed: int = None, data_container_write_period: float = 600, ensemble_data_write_interval: int = None, trajectory_write_interval: int = None, sublattice_probabilities: List[float] = None) -> None: # set trial move if trial_move == 'swap': self.do_move = self._do_swap self._get_sublattice_probabilities = self._get_swap_sublattice_probabilities elif trial_move == 'flip': self.do_move = self._do_flip self._get_sublattice_probabilities = self._get_flip_sublattice_probabilities else: raise ValueError('Invalid value for trial_move: {}.' ' Must be either "swap" or "flip".'.format(trial_move)) # set default values that are system dependent if flatness_check_interval is None: flatness_check_interval = len(structure) * 1000 # parameters pertaining to construction of entropy and histogram self._energy_spacing = energy_spacing self._fill_factor_limit = fill_factor_limit self._flatness_check_interval = flatness_check_interval self._flatness_limit = flatness_limit # energy window self._window_search_penalty = window_search_penalty self._bin_left = self._get_bin_index(energy_limit_left) self._bin_right = self._get_bin_index(energy_limit_right) if self._bin_left is not None and \ self._bin_right is not None and self._bin_left >= self._bin_right: raise ValueError('Invalid energy window: left boundary ({}, {}) must be' ' smaller than right boundary ({}, {})' .format(energy_limit_left, self._bin_left, energy_limit_right, self._bin_right)) # ensemble parameters self._ensemble_parameters = {} self._ensemble_parameters['energy_spacing'] = energy_spacing self._ensemble_parameters['trial_move'] = trial_move self._ensemble_parameters['energy_limit_left'] = energy_limit_left self._ensemble_parameters['energy_limit_right'] = energy_limit_right # The following parameters are _intentionally excluded_ from # the ensemble_parameters dict as it would prevent users from # changing their values between restarts. The latter is advantageous # as these runs can require restarts and possibly parameter adjustments # to achieve convergence. # * fill_factor_limit # * flatness_check_interval # * flatness_limit # * entropy_write_frequency # * window_search_penalty # add species count to ensemble parameters symbols = set([symbol for sub in calculator.sublattices for symbol in sub.chemical_symbols]) for symbol in symbols: key = 'n_atoms_{}'.format(symbol) count = structure.get_chemical_symbols().count(symbol) self._ensemble_parameters[key] = count # set the convergence, which may be updated in case of a restart self._converged: bool = None # the constructor of the parent classes must be called *after* # the ensemble_parameters dict has been populated super().__init__( structure=structure, calculator=calculator, user_tag=user_tag, random_seed=random_seed, dc_filename=dc_filename, data_container=data_container, data_container_class=WangLandauDataContainer, data_container_write_period=data_container_write_period, ensemble_data_write_interval=ensemble_data_write_interval, trajectory_write_interval=trajectory_write_interval) # handle probabilities for swaps on different sublattices if sublattice_probabilities is None: self._sublattice_probabilities = self._get_sublattice_probabilities() else: self._sublattice_probabilities = sublattice_probabilities # initialize Wang-Landau algorithm; in the case of a restart # these quantities are read from the data container file; the # if-conditions prevent these values from being overwritten self._potential = self.calculator.calculate_total( occupations=self.configuration.occupations) self._reached_energy_window = self._inside_energy_window( self._get_bin_index(self._potential)) if not hasattr(self, '_fill_factor'): self._fill_factor = 1.0 if not hasattr(self, '_fill_factor_history'): if self._reached_energy_window: self._fill_factor_history = {self.step: self._fill_factor} else: self._fill_factor_history = {} if not hasattr(self, '_entropy_history'): self._entropy_history = {} if not hasattr(self, '_histogram'): self._histogram: Dict[int, int] = {} if not hasattr(self, '_entropy'): self._entropy: Dict[int, float] = {} @property def fill_factor(self) -> float: """ current value of fill factor """ return self._fill_factor @property def fill_factor_history(self) -> Dict[int, float]: """evolution of the fill factor in the Wang-Landau algorithm (key=MC trial step, value=fill factor) """ return self._fill_factor_history @property def converged(self) -> Optional[bool]: """ True if convergence has been achieved """ return self._converged @property def flatness_limit(self) -> float: r"""The histogram :math:`H(E)` is deemed sufficiently flat if :math:`H(E) > \chi \left<H(E)\right>\,\forall E` where :attr:`flatness_limit` sets the parameter :math:`\chi`. """ return self._flatness_limit @flatness_limit.setter def flatness_limit(self, new_value) -> None: self._flatness_limit = new_value self._converged = None @property def fill_factor_limit(self) -> float: """ If the fill factor :math:`f` falls below this value, the Wang-Landau sampling is terminated. """ return self._fill_factor_limit @fill_factor_limit.setter def fill_factor_limit(self, new_value) -> None: self._fill_factor_limit = new_value self._converged = None @property def flatness_check_interval(self) -> int: """ Number of MC trial steps between checking the flatness condition. """ return self._flatness_check_interval @flatness_check_interval.setter def flatness_check_interval(self, new_value: int) -> None: self._flatness_check_interval = new_value
[docs] def run(self, number_of_trial_steps: int): """ Samples the ensemble for the given number of trial steps. Parameters ---------- number_of_trial_steps Maximum number of MC trial steps to run in total. The run will terminate earlier if :attr:`fill_factor_limit` is reached. reset_step If ``True`` the MC trial step counter and the data container will be reset to zero and empty, respectively. Raises ------ TypeError If :attr:`number_of_trial_steps` is not an ``int``. """ if self.converged: logger.warning('Convergence has already been reached.') else: super().run(number_of_trial_steps)
def _terminate_sampling(self) -> bool: """Returns ``True`` if the Wang-Landau algorithm has converged. This is used in the :func:`run` method implemented in :class:`BaseEnsemble` to evaluate whether the sampling loop should be terminated. """ # N.B.: self._converged can be None if self._converged is not None: return self._converged else: return False def _restart_ensemble(self): """Restarts ensemble using the last state saved in the data container file. Note that this method does _not_ use the :attr:`last_state` property of the data container but rather uses the last data written to the data frame. """ super()._restart_ensemble() self._fill_factor = self.data_container._last_state['fill_factor'] self._fill_factor_history = self.data_container._last_state['fill_factor_history'] self._entropy_history = self.data_container._last_state['entropy_history'] self._histogram = self.data_container._last_state['histogram'] self._entropy = self.data_container._last_state['entropy'] histogram = np.array(list(self._histogram.values())) limit = self._flatness_limit * np.average(histogram) self._converged = (self._fill_factor <= self._fill_factor_limit ) & np.all(histogram >= limit)
[docs] def write_data_container(self, outfile: Union[str, bytes]): """Updates the last state of the Wang-Landau simulation and writes the data container to file. Parameters ---------- outfile File to which to write. """ self._data_container._update_last_state( last_step=self.step, occupations=self.configuration.occupations.tolist(), accepted_trials=self._accepted_trials, random_state=random.getstate(), fill_factor=self._fill_factor, fill_factor_history=self._fill_factor_history, entropy_history=self._entropy_history, histogram=OrderedDict(sorted(self._histogram.items())), entropy=OrderedDict(sorted(self._entropy.items()))) self.data_container.write(outfile)
def _acceptance_condition(self, potential_diff: float) -> bool: """Evaluates Metropolis acceptance criterion. Parameters ---------- potential_diff Change in the thermodynamic potential associated with the trial step. """ # acceptance/rejection step bin_old = self._get_bin_index(self._potential) bin_new = self._get_bin_index(self._potential + potential_diff) bin_cur = bin_old if self._allow_move(bin_cur, bin_new): S_cur = self._entropy.get(bin_cur, 0) S_new = self._entropy.get(bin_new, 0) delta = np.exp(S_cur - S_new) if delta >= 1 or delta >= self._next_random_number(): accept = True self._potential += potential_diff bin_cur = bin_new else: accept = False else: accept = False if not self._reached_energy_window: # check whether the target energy window has been reached self._reached_energy_window = self._inside_energy_window(bin_cur) # if the target window has been reached remove unused bins # from histogram and entropy counters if self._reached_energy_window: self._fill_factor_history[self.step] = self._fill_factor # flush data from data container except for initial step self._data_container._data_list = [self._data_container._data_list[0]] self._entropy = {k: self._entropy[k] for k in self._entropy if self._inside_energy_window(k)} self._histogram = {k: self._histogram[k] for k in self._histogram if self._inside_energy_window(k)} else: # then reconsider accept/reject based on whether we # approached the window or not dist_new = np.inf dist_old = np.inf if self._bin_left is not None: dist_new = min(dist_new, abs(bin_new - self._bin_left)) dist_old = min(dist_old, abs(bin_old - self._bin_left)) if self._bin_right is not None: dist_new = min(dist_new, abs(bin_new - self._bin_right)) dist_old = min(dist_old, abs(bin_old - self._bin_right)) assert dist_new < np.inf and dist_old < np.inf exp_dist = np.exp((dist_old - dist_new) * self._window_search_penalty) if exp_dist >= 1 or exp_dist >= self._next_random_number(): # should be accepted if not accept: # reset potential self._potential += potential_diff bin_cur = bin_new accept = True else: # should be rejected if accept: # reset potential self._potential -= potential_diff bin_cur = bin_old accept = False # update histograms and entropy counters self._update_entropy(bin_cur) return accept def _update_entropy(self, bin_cur: int) -> None: """Updates counters for histogram and entropy, checks histogram flatness, and updates fill factor if indicated. """ # update histogram and entropy self._entropy[bin_cur] = self._entropy.get(bin_cur, 0) + self._fill_factor self._histogram[bin_cur] = self._histogram.get(bin_cur, 0) + 1 # check flatness of histogram if self.step % self._flatness_check_interval == 0 and \ self.step > 0 and self._reached_energy_window: # shift entropy counter in order to avoid overflow entropy_ref = np.min(list(self._entropy.values())) for k in self._entropy: self._entropy[k] -= entropy_ref # check whether the Wang-Landau algorithm has converged histogram = np.array(list(self._histogram.values())) limit = self._flatness_limit * np.average(histogram) is_flat = np.all(histogram >= limit) self._converged = (self._fill_factor <= self._fill_factor_limit) & is_flat if is_flat and not self._converged: # update fill factor self._fill_factor /= 2 self._fill_factor_history[self.step] = self._fill_factor # update entropy history self._entropy_history[self.step] = OrderedDict( sorted(self._entropy.items())) # reset histogram self._histogram = dict.fromkeys(self._histogram, 0) def _get_bin_index(self, energy: float) -> Optional[int]: """ Returns bin index for histogram and entropy dictionaries. """ if energy is None or np.isnan(energy): return None return int(round(energy / self._energy_spacing)) def _allow_move(self, bin_cur: Optional[int], bin_new: int) -> bool: """Returns ``True`` if the current move is to be included in the accumulation of histogram and entropy. This logic has been moved into a separate function in order to enhance readability. """ if self._bin_left is None and self._bin_right is None: # no limits on energy window return True if self._bin_left is not None: if bin_cur < self._bin_left: # not yet in window (left limit) return True if bin_new < self._bin_left: # imposing left limit return False if self._bin_right is not None: if bin_cur > self._bin_right: # not yet in window (right limit) return True if bin_new > self._bin_right: # imposing right limit return False return True def _inside_energy_window(self, bin_k: int) -> bool: """Returns ``True`` if :attr:`bin_k` is inside the energy window specified for this simulation. """ if self._bin_left is not None and bin_k < self._bin_left: return False if self._bin_right is not None and bin_k > self._bin_right: return False return True def _do_trial_step(self): """ Carries out one Monte Carlo trial step. """ sublattice_index = self.get_random_sublattice_index(self._sublattice_probabilities) return self.do_move(sublattice_index=sublattice_index) def _do_swap(self, sublattice_index: int, allowed_species: List[int] = None) -> int: """Carries out a Monte Carlo trial that involves swapping the species on two sites. This method has been copied from :class:`ThermodynamicBaseEnsemble`. Parameters --------- sublattice_index Index of sublattice the swap will act on. allowed_species List of atomic numbers for allowed species. Returns ------- Returns 1 or 0 depending on if trial move was accepted or rejected. """ sites, species = self.configuration.get_swapped_state(sublattice_index, allowed_species) potential_diff = self._get_property_change(sites, species) if self._acceptance_condition(potential_diff): self.update_occupations(sites, species) return 1 return 0 def _do_flip(self, sublattice_index: int, allowed_species: List[int] = None) -> int: """Carries out one Monte Carlo trial step that involves flipping the species on one site. This method has been adapted from :class:`ThermodynamicBaseEnsemble`. Parameters --------- sublattice_index Index of sublattice the flip will act on. allowed_species List of atomic numbers for allowed species. Returns ------- Returns 1 or 0 depending on if trial move was accepted or rejected. """ index, species = self.configuration.get_flip_state(sublattice_index, allowed_species) potential_diff = self._get_property_change([index], [species]) if self._acceptance_condition(potential_diff): self.update_occupations([index], [species]) return 1 return 0 def _get_swap_sublattice_probabilities(self) -> List[float]: """Returns sublattice probabilities suitable for swaps. This method has been copied without modification from :class:`ThermodynamicBaseEnsemble`. """ sublattice_probabilities = [] for i, sl in enumerate(self.sublattices): if self.configuration.is_swap_possible(i): sublattice_probabilities.append(len(sl.indices)) else: sublattice_probabilities.append(0) norm = sum(sublattice_probabilities) if norm == 0: raise ValueError('No swaps are possible on any of the active sublattices.') sublattice_probabilities = [p / norm for p in sublattice_probabilities] return sublattice_probabilities def _get_flip_sublattice_probabilities(self) -> List[float]: """Returns the default sublattice probability which is based on the sizes of a sublattice. This method has been copied without modification from :class:`ThermodynamicBaseEnsemble`. """ sublattice_probabilities = [] for _, sl in enumerate(self.sublattices): if len(sl.chemical_symbols) > 1: sublattice_probabilities.append(len(sl.indices)) else: sublattice_probabilities.append(0) norm = sum(sublattice_probabilities) sublattice_probabilities = [p / norm for p in sublattice_probabilities] return sublattice_probabilities
[docs] def get_bins_for_parallel_simulations(n_bins: int, energy_spacing: float, minimum_energy: float, maximum_energy: float, overlap: int = 4, bin_size_exponent: float = 1.0) -> List[Tuple[float, float]]: """Generates a list of energy bins (lower and upper bound) suitable for parallel Wang-Landau simulations. For the latter, the energy range is split up into a several bins (:attr:`n_bins`). Each bin is then sampled in a separate Wang-Landau simulation. Once the density of states in the individual bins has been converged the total density of states can be constructed by patching the segments back together. To this end, one requires some over overlap between the segments (:attr:`overlap`). The function returns a list of tuples. Each tuple provides the lower (:attr:`energy_limit_left`) and upper (:attr:`energy_limit_right`) bound of one bin, which are then to be used to set :attr:`energy_limit_left` and :attr:`energy_limit_right` when initializing a :class:`WangLandauEnsemble` instance. Note ---- The left-most/right-most bin has no lower/upper bound (set to ``None``). Parameters ---------- n_bins Number of bins. energy_spacing Sets the bin size of the energy grid used by the Wang-Landau simulation, see :class:`WangLandauEnsemble` for details. minimum_energy An estimate for the lowest energy to be encountered in this system. maximum_energy An estimate for the highest energy to be encountered in this system. overlap Amount of overlap between bins in units of :attr:`energy_spacing`. bin_size_exponent This parameter allows one to generate a non-uniform distribution of bin sizes. If :attr:`bin_size_exponent` is smaller than one bins at the lower and upper end of the energy range (specified via :attr:`minimum_energy` and :attr:`maximum_energy`) will be shrunk relative to the bins in the middle of the energy range. In principle this can be used one to achieve a more even distribution of computational load between the individual Wang-Landau simulations. Note ---- This is an option for advanced users. Only use this keyword if you know what you are doing. """ limits = np.linspace(-1, 1, n_bins + 1) limits = np.sign(limits) * np.abs(limits) ** bin_size_exponent limits *= 0.5 * (maximum_energy - minimum_energy) limits += 0.5 * (maximum_energy + minimum_energy) limits[0], limits[-1] = None, None bounds = [] for k, (energy_limit_left, energy_limit_right) in enumerate(zip(limits[:-1], limits[1:])): if energy_limit_left is not None and energy_limit_right is not None and \ (energy_limit_right - energy_limit_left) / energy_spacing < 2 * overlap: raise ValueError('Energy window too small. min/max: {}/{}' .format(energy_limit_right, energy_limit_left) + ' Try decreasing n_bins ({}) and/or overlap ({}).' .format(n_bins, overlap)) if energy_limit_left is not None: energy_limit_left -= overlap * energy_spacing if energy_limit_right is not None: energy_limit_right += overlap * energy_spacing bounds.append((energy_limit_left, energy_limit_right)) return bounds