Source code for pyiron.atomistics.structure.neighbors

# coding: utf-8
# Copyright (c) Max-Planck-Institut für Eisenforschung GmbH - Computational Materials Design (CM) Department
# Distributed under the terms of "New BSD License", see the LICENSE file.

import numpy as np
from sklearn.cluster import AgglomerativeClustering
from scipy.sparse import coo_matrix
from scipy.special import gamma
from pyiron_base import Settings
from pyiron.atomistics.structure.analyse import get_average_of_unique_labels
import warnings

__author__ = "Joerg Neugebauer, Sam Waseda"
__copyright__ = (
    "Copyright 2020, Max-Planck-Institut für Eisenforschung GmbH - "
    "Computational Materials Design (CM) Department"
)
__version__ = "1.0"
__maintainer__ = "Sam Waseda"
__email__ = "waseda@mpie.de"
__status__ = "production"
__date__ = "Sep 1, 2017"

s = Settings()

[docs]class Tree: """ Class to get tree structure for the neighborhood information. Main attributes (do not modify them): - distances (numpy.ndarray/list): Distances to the neighbors of given positions - indices (numpy.ndarray/list): Indices of the neighbors of given positions - vecs (numpy.ndarray/list): Vectors to the neighbors of given positions Auxiliary attributes: - allow_ragged (bool): Whether to allow a ragged list of numpy arrays or not. This is relevant only if the cutoff length was specified, in which case the number of atoms for each position could differ. `allow_ragged = False` is computationally more efficient in most of the methods. When `allow_ragged = False`, the variables are filled as follows: `np.inf` in `distances`, `numpy.array([np.inf, np.inf, np.inf])` in `vecs`, `n_atoms+1` (or a larger value) in `indices` and -1 in `shells` - wrap_positions (bool): Whether to wrap back the positions entered by user in get_neighborhood etc. Since the information outside the original box is limited to a few layer, wrap_positions=False might miss some points without issuing an error. Furthermore, you can re-employ the original tree structure to get neighborhood information via get_indices, get_vectors, get_distances and get_neighborhood. The information delivered by get_neighborhood is the same as what can be obtained through the other three getters (except for the fact that get_neighborhood returns a Tree instance, while the others return numpy arrays), but since get_vectors anyway has to call get_distances and get_indices internally, the computational cost of get_neighborhood and get_vectors is the same. """ def __init__(self, ref_structure): self.distances = None self.vecs = None self.indices = None self._extended_positions = None self._wrapped_indices = None self._allow_ragged = False self._cell = None self._extended_indices = None self._ref_structure = ref_structure.copy() self.wrap_positions = False self._tree = None self.num_neighbors = None self.cutoff_radius = np.inf self._norm_order = 2 def __repr__(self): """ Returns: __repr__ """ to_return = ( "Main attributes:\n" + "- distances : Distances to the neighbors of given positions\n" + "- indices : Indices of the neighbors of given positions\n" ) if self.vecs is not None: to_return += "- vecs : Vectors to the neighbors of given positions\n" return to_return
[docs] def copy(self): new_neigh = Tree(self._ref_structure) new_neigh.distances = self.distances.copy() new_neigh.vecs = self.vecs.copy() new_neigh.indices = self.indices.copy() new_neigh._extended_positions = self._extended_positions new_neigh._wrapped_indices = self._wrapped_indices new_neigh._allow_ragged = self._allow_ragged new_neigh._cell = self._cell new_neigh._extended_indices = self._extended_indices new_neigh.wrap_positions = self.wrap_positions new_neigh._tree = self._tree new_neigh.num_neighbors = self.num_neighbors new_neigh.cutoff_radius = self.cutoff_radius new_neigh._norm_order = self._norm_order return new_neigh
@property def norm_order(self): """ Norm to use for the neighborhood search and shell recognition. The definition follows the conventional Lp norm (cf. https://en.wikipedia.org/wiki/Lp_space). This is still an experimental feature and for anything other than norm_order=2, there is no guarantee that this works flawlessly. """ return self._norm_order @norm_order.setter def norm_order(self, value): raise ValueError( 'norm_order cannot be changed after initialization. Re-initialize the Neighbor class' + ' with the correct norm_order value' ) def _get_max_length(self, ref_vector=None): if ref_vector is None: ref_vector = self.distances if (ref_vector is None or len(ref_vector)==0 or not hasattr(ref_vector[0], '__len__')): return None return max(len(dd[dd<np.inf]) for dd in ref_vector) def _fill(self, value, filler=np.inf): max_length = self._get_max_length() if max_length is None: return value arr = np.zeros((len(value), max_length)+value[0].shape[1:], dtype=type(filler)) arr.fill(filler) for ii, vv in enumerate(value): arr[ii,:len(vv)] = vv return arr def _contract(self, value, ref_vector=None): if self._get_max_length(ref_vector=ref_vector) is None: return value return [vv[:np.sum(dist<np.inf)] for vv, dist in zip(value, self.distances)] @property def allow_ragged(self): """ Whether to allow ragged list of distancs/vectors/indices or fill empty slots with numpy.inf to get rectangular arrays """ return self._allow_ragged @allow_ragged.setter def allow_ragged(self, new_bool): if not isinstance(new_bool, bool): raise ValueError('allow_ragged must be a boolean') if self._allow_ragged == new_bool: return self._allow_ragged = new_bool if new_bool: self.distances = self._contract(self.distances) if self.vecs is not None: self.vecs = self._contract(self.vecs) self.indices = self._contract(self.indices) else: self.distances = self._fill(self.distances) self.indices = self._fill(self.indices, filler=len(self._ref_structure)) if self.vecs is not None: self.vecs = self._fill(self.vecs) def _get_extended_positions(self): if self._extended_positions is None: return self._ref_structure.positions return self._extended_positions def _get_wrapped_indices(self): if self._wrapped_indices is None: return np.arange(len(self._ref_structure.positions)) return self._wrapped_indices def _get_wrapped_positions(self, positions): if not self.wrap_positions: return np.asarray(positions) x = np.array(positions).copy() cell = self._ref_structure.cell x_scale = np.dot(x, np.linalg.inv(cell))+1.0e-12 x[...,self._ref_structure.pbc] -= np.dot(np.floor(x_scale), cell)[...,self._ref_structure.pbc] return x def _get_distances_and_indices( self, positions=None, allow_ragged=None, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2, ): if positions is None: if allow_ragged is None: return self.distances, self.indices if allow_ragged: return (self._contract(self.distances), self._contract(self.indices)) return (self._fill(self.distances), self._fill(self.indices, filler=len(self._ref_structure))) num_neighbors = self._estimate_num_neighbors( num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, ) if len(self._get_extended_positions()) < num_neighbors and cutoff_radius==np.inf: raise ValueError( 'num_neighbors too large - make width_buffer larger and/or make ' + 'num_neighbors smaller' ) distances, indices = self._tree.query( self._get_wrapped_positions(positions), k=num_neighbors, distance_upper_bound=cutoff_radius, p=self.norm_order, ) if cutoff_radius<np.inf and np.any(distances.T[-1]<np.inf): warnings.warn( 'Number of neighbors found within the cutoff_radius is equal to (estimated) ' + 'num_neighbors. Increase num_neighbors (or set it to None) or ' + 'width_buffer to find all neighbors within cutoff_radius.' ) self._extended_indices = indices.copy() indices[distances<np.inf] = self._get_wrapped_indices()[indices[distances<np.inf]] if allow_ragged is None: allow_ragged = self.allow_ragged if allow_ragged: return self._contract(distances, distances), self._contract(indices, distances) return distances, indices
[docs] def get_indices( self, positions=None, allow_ragged=None, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2, ): """ Get current indices or neighbor indices for given positions using the same Tree structure used for the instantiation of the Neighbor class. This function should not be used if the structure changed in the meantime. If `positions` is None and `allow_ragged` is None, this function returns the same content as `indices`. Args: positions (list/numpy.ndarray/None): Positions around which neighborhood vectors are to be computed (None to get current vectors) allow_ragged (bool): Whether to allow ragged list of arrays or rectangular numpy.ndarray filled with np.inf for values outside cutoff_radius num_neighbors (int/None): Number of neighboring atoms to calculate vectors for. Ignored if `positions` is None. cutoff_radius (float): cutoff radius. Ignored if `positions` is None. width_buffer (float): Buffer length for the estimation of num_neighbors. Ignored if `positions` is None. Returns: (list/numpy.ndarray) list (if allow_ragged=True) or numpy.ndarray (otherwise) of neighbor indices """ return self._get_distances_and_indices( positions=positions, allow_ragged=allow_ragged, num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, )[1]
[docs] def get_distances( self, positions=None, allow_ragged=None, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2, ): """ Get current positions or neighbor positions for given positions using the same Tree structure used for the instantiation of the Neighbor class. This function should not be used if the structure changed in the meantime. If `positions` is None and `allow_ragged` is None, this function returns the same content as `positions`. Args: positions (list/numpy.ndarray/None): Positions around which neighborhood vectors are to be computed (None to get current vectors) allow_ragged (bool): Whether to allow ragged list of arrays or rectangular numpy.ndarray filled with np.inf for values outside cutoff_radius num_neighbors (int/None): Number of neighboring atoms to calculate vectors for. Ignored if `positions` is None. cutoff_radius (float): cutoff radius. Ignored if `positions` is None. width_buffer (float): Buffer length for the estimation of num_neighbors. Ignored if `positions` is None. Returns: (list/numpy.ndarray) list (if allow_ragged=True) or numpy.ndarray (otherwise) of neighbor distances """ return self._get_distances_and_indices( positions=positions, allow_ragged=allow_ragged, num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, )[0]
[docs] def get_vectors( self, positions=None, allow_ragged=False, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2, ): """ Get current vectors or neighbor vectors for given positions using the same Tree structure used for the instantiation of the Neighbor class. This function should not be used if the structure changed in the meantime. If `positions` is None and `allow_ragged` is None, this function returns the same content as `vecs`. Args: positions (list/numpy.ndarray/None): Positions around which neighborhood vectors are to be computed (None to get current vectors) allow_ragged (bool): Whether to allow ragged list of arrays or rectangular numpy.ndarray filled with np.inf for values outside cutoff_radius num_neighbors (int/None): Number of neighboring atoms to calculate vectors for. Ignored if `positions` is None. cutoff_radius (float): cutoff radius. Ignored if `positions` is None. width_buffer (float): Buffer length for the estimation of num_neighbors. Ignored if `positions` is None. Returns: (list/numpy.ndarray) list (if allow_ragged=True) or numpy.ndarray (otherwise) of neighbor vectors """ return self._get_vectors( positions=positions, allow_ragged=allow_ragged, num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, )
def _get_vectors( self, positions=None, allow_ragged=None, num_neighbors=None, cutoff_radius=np.inf, distances=None, indices=None, width_buffer=1.2, ): if positions is not None: if distances is None or indices is None: distances, indices = self._get_distances_and_indices( positions=positions, allow_ragged=False, num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, ) vectors = np.zeros(distances.shape+(3,)) vectors -= self._get_wrapped_positions(positions).reshape(distances.shape[:-1]+(1, 3)) vectors[distances<np.inf] += self._get_extended_positions()[ self._extended_indices[distances<np.inf] ] vectors[distances==np.inf] = np.array(3*[np.inf]) if self._cell is not None: vectors[distances<np.inf] -= self._cell*np.rint(vectors[distances<np.inf]/self._cell) elif self.vecs is not None: vectors = self.vecs else: raise AssertionError( 'vectors not created yet -> put positions or reinitialize with t_vec=True' ) if allow_ragged==self.allow_ragged: return vectors if allow_ragged: return self._contract(vectors, distances) return self._fill(vectors) def _estimate_num_neighbors(self, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2): """ Args: num_neighbors (int): number of neighbors width_buffer (float): width of the layer to be added to account for pbc. cutoff_radius (float): self-explanatory Returns: Number of atoms required for a given cutoff """ if num_neighbors is None and cutoff_radius==np.inf and self.num_neighbors is None: raise ValueError('Specify num_neighbors or cutoff_radius') elif num_neighbors is None and self.num_neighbors is None: volume = self._ref_structure.get_volume(per_atom=True) width_buffer = 1+width_buffer width_buffer *= 8*gamma(1+1/self.norm_order)**3/gamma(1+3/self.norm_order) num_neighbors = max(14, int(width_buffer*cutoff_radius**3/volume)) elif num_neighbors is None: num_neighbors = self.num_neighbors if self.num_neighbors is None: self.num_neighbors = num_neighbors self.cutoff_radius = cutoff_radius if num_neighbors > self.num_neighbors: warnings.warn("Taking a larger search area after initialization has the risk of " + "missing neighborhood atoms") return num_neighbors def _estimate_width(self, num_neighbors=None, cutoff_radius=np.inf, width_buffer=1.2): """ Args: num_neighbors (int): number of neighbors width_buffer (float): width of the layer to be added to account for pbc. cutoff_radius (float): cutoff radius Returns: Width of layer required for the given number of atoms """ if num_neighbors is None and cutoff_radius==np.inf: raise ValueError('Define either num_neighbors or cutoff_radius') if all(self._ref_structure.pbc==False): return 0 elif cutoff_radius!=np.inf: return cutoff_radius pbc = self._ref_structure.pbc n = sum(pbc) prefactor = 2**n*gamma(1+1/self.norm_order)**2/gamma(1+n/self.norm_order) width = np.prod(( np.linalg.norm(self._ref_structure.cell, axis=-1, ord=self.norm_order)-np.ones(3) )*pbc+np.ones(3)) width *= prefactor*np.max([num_neighbors, 8])/len(self._ref_structure) cutoff_radius = width_buffer*width**(1/np.sum(pbc)) return cutoff_radius
[docs] def get_neighborhood( self, positions, num_neighbors=None, t_vec=True, cutoff_radius=np.inf, width_buffer=1.2, ): """ Get neighborhood information of `positions`. What it returns is in principle the same as `get_neighborhood` in `Atoms`. The only one difference is the reuse of the same Tree structure, which makes the algorithm more efficient, but will fail if the reference structure changed in the meantime. Args: position: Position in a box whose neighborhood information is analysed num_neighbors (int): Number of nearest neighbors t_vec (bool): True: compute distance vectors (pbc are taken into account) cutoff_radius (float): Upper bound of the distance to which the search is to be done width_buffer (float): Width of the layer to be added to account for pbc. Returns: pyiron.atomistics.structure.atoms.Tree: Neighbors instances with the neighbor indices, distances and vectors """ new_neigh = self.copy() return new_neigh._get_neighborhood( positions=positions, num_neighbors=num_neighbors, t_vec=t_vec, cutoff_radius=cutoff_radius, exclude_self=False, width_buffer=width_buffer, )
def _get_neighborhood( self, positions, num_neighbors=12, t_vec=True, cutoff_radius=np.inf, exclude_self=False, width_buffer=1.2, ): start_column = 0 if exclude_self: start_column = 1 if num_neighbors is not None: num_neighbors += 1 distances, indices = self._get_distances_and_indices( positions, num_neighbors=num_neighbors, cutoff_radius=cutoff_radius, width_buffer=width_buffer, ) if num_neighbors is not None: self.num_neighbors -= 1 max_column = np.sum(distances<np.inf, axis=-1).max() self.distances = distances[...,start_column:max_column] self.indices = indices[...,start_column:max_column] self._extended_indices = self._extended_indices[...,start_column:max_column] if t_vec: self.vecs = self._get_vectors( positions=positions, distances=self.distances, indices=self._extended_indices ) return self def _check_width(self, width, pbc): if any(pbc) and np.prod(self.distances.shape)>0 and self.vecs is not None: if np.linalg.norm( self._fill(self._contract(self.vecs), filler=0.0)[...,pbc], axis=-1, ord=self.norm_order ).max() > width: return True return False
[docs]class Neighbors(Tree): """ Class for storage of the neighbor information for a given atom based on the KDtree algorithm Main attributes (do not modify them): - distances (numpy.ndarray/list): Distances to the neighbors of all atoms - indices (numpy.ndarray/list): Indices of the neighbors of all atoms - vecs (numpy.ndarray/list): Vectors to the neighbors of all atoms Auxiliary attributes: - allow_ragged (bool): Whether to allow a ragged list of numpy arrays or not. This is relevant only if the cutoff length was specified, in which case the number of neighbor atoms for each atom could differ. `allow_ragged = False` is computationally more efficient in most of the methods. When `allow_ragged = False`, the variables are filled as follows: `np.inf` in `distances`, `numpy.array([np.inf, np.inf, np.inf])` in `vecs`, `n_atoms+1` (or a larger value) in `indices` and -1 in `shells` - wrap_positions (bool): Whether to wrap back the positions entered by user in get_neighborhood etc. Since the information outside the original box is limited to a few layer, wrap_positions=False might miss some points without issuing an error. Furthermore, you can re-employ the original tree structure to get neighborhood information via get_indices, get_vectors, get_distances and get_neighborhood. The information delivered by get_neighborhood is the same as what can be obtained through the other three getters (except for the fact that get_neighborhood returns a Tree instance, while the others return numpy arrays), but since get_vectors anyway has to call get_distances and get_indices internally, the computational cost of get_neighborhood and get_vectors is the same. """ def __init__(self, ref_structure, tolerance=2): super().__init__(ref_structure=ref_structure) self._shells = None self._tolerance = tolerance self._cluster_vecs = None self._cluster_dist = None def __repr__(self): """ Returns: __repr__ """ to_return = super().__repr__() return to_return.replace('given positions', 'each atom') @property def shells(self): """ Returns the cell numbers of each atom according to the distances """ if self._shells is None: self._shells = self.get_local_shells() if self.allow_ragged: return self._contract(self._shells) return self._shells
[docs] def get_local_shells(self, tolerance=None, cluster_by_distances=False, cluster_by_vecs=False): """ Set shell indices based on distances available to each atom. Clustering methods can be used at the same time, which will be useful at finite temperature results, but depending on how dispersed the atoms are, the algorithm could take some time. If the clustering method(-s) have already been launched before this function, it will use the results already available and does not execute the clustering method(-s) again. Args: tolerance (int): decimals in np.round for rounding up distances cluster_by_distances (bool): If True, `cluster_by_distances` is called first and the distances obtained from the clustered distances are used to calculate the shells. If cluster_by_vecs is True at the same time, `cluster_by_distances` will use the clustered vectors for its clustering algorithm. For more, see the DocString of `cluster_by_distances`. (default: False) cluster_by_vecs (bool): If True, `cluster_by_vectors` is called first and the distances obtained from the clustered vectors are used to calculate the shells. (default: False) Returns: shells (numpy.ndarray): shell indices """ allow_ragged = self.allow_ragged if allow_ragged: self.allow_ragged = False if tolerance is None: tolerance = self._tolerance if cluster_by_distances: if self._cluster_dist is None: self.cluster_by_distances(use_vecs=cluster_by_vecs) shells = np.array([np.unique(np.round(dist, decimals=tolerance), return_inverse=True)[1]+1 for dist in self._cluster_dist.cluster_centers_[self._cluster_dist.labels_] ]) shells[self._cluster_dist.labels_<0] = -1 shells = shells.reshape(self.indices.shape) elif cluster_by_vecs: if self._cluster_vecs is None: self.cluster_by_vecs() shells = np.array([np.unique(np.round(dist, decimals=tolerance), return_inverse=True)[1]+1 for dist in np.linalg.norm( self._cluster_vecs.cluster_centers_[self._cluster_vecs.labels_], axis=-1, ord=self.norm_order ) ]) shells[self._cluster_vecs.labels_<0] = -1 shells = shells.reshape(self.indices.shape) else: shells = [] for dist in self.distances: shells.append(np.unique(np.round(dist[dist<np.inf], decimals=tolerance), return_inverse=True)[1]+1) shells = self._fill(shells, filler=-1) self.allow_ragged = allow_ragged if allow_ragged: return self._contract(shells) return shells
[docs] def get_global_shells(self, tolerance=None, cluster_by_distances=False, cluster_by_vecs=False): """ Set shell indices based on all distances available in the system instead of setting them according to the local distances (in contrast to shells defined as an attribute in this class). Clustering methods can be used at the same time, which will be useful at finite temperature results, but depending on how dispersed the atoms are, the algorithm could take some time. If the clustering method(-s) have already been launched before this function, it will use the results already available and does not execute the clustering method(-s) again. Args: tolerance (int): decimals in np.round for rounding up distances (default: 2) cluster_by_distances (bool): If True, `cluster_by_distances` is called first and the distances obtained from the clustered distances are used to calculate the shells. If cluster_by_vecs is True at the same time, `cluster_by_distances` will use the clustered vectors for its clustering algorithm. For more, see the DocString of `cluster_by_distances`. (default: False) cluster_by_vecs (bool): If True, `cluster_by_vectors` is called first and the distances obtained from the clustered vectors are used to calculate the shells. (default: False) Returns: shells (numpy.ndarray): shell indices (cf. shells) """ allow_ragged = self.allow_ragged if allow_ragged: self.allow_ragged = False if tolerance is None: tolerance = self._tolerance if self.distances is None: raise ValueError('neighbors not set') distances = self.distances if cluster_by_distances: if self._cluster_dist is None: self.cluster_by_distances(use_vecs=cluster_by_vecs) distances = self._cluster_dist.cluster_centers_[ self._cluster_dist.labels_ ].reshape(self.distances.shape) distances[self._cluster_dist.labels_<0] = np.inf elif cluster_by_vecs: if self._cluster_vecs is None: self.cluster_by_vecs() distances = np.linalg.norm( self._cluster_vecs.cluster_centers_[self._cluster_vecs.labels_], axis=-1, ord=self.norm_order, ).reshape(self.distances.shape) distances[self._cluster_vecs.labels_<0] = np.inf dist_lst = np.unique(np.round(a=distances, decimals=tolerance)) shells = -np.ones_like(self.indices).astype(int) shells[distances<np.inf] = np.absolute( distances[distances<np.inf, np.newaxis]-dist_lst[np.newaxis, dist_lst<np.inf] ).argmin(axis=-1)+1 self.allow_ragged = allow_ragged if allow_ragged: return self._contract(shells) return shells
[docs] def get_shell_matrix( self, chemical_pair=None, cluster_by_distances=False, cluster_by_vecs=False ): """ Shell matrices for pairwise interaction. Note: The matrices are always symmetric, meaning if you use them as bilinear operators, you have to divide the results by 2. Args: chemical_pair (list): pair of chemical symbols (e.g. ['Fe', 'Ni']) Returns: list of sparse matrices for different shells Example: from pyiron import Project structure = Project('.').create_structure('Fe', 'bcc', 2.83).repeat(2) J = -0.1 # Ising parameter magmoms = 2*np.random.random((len(structure)), 3)-1 # Random magnetic moments between -1 and 1 neigh = structure.get_neighbors(num_neighbors=8) # Iron first shell shell_matrices = neigh.get_shell_matrix() print('Energy =', 0.5*J*magmoms.dot(shell_matrices[0].dot(matmoms))) """ pairs = np.stack((self.indices, np.ones_like(self.indices)*np.arange(len(self.indices))[:,np.newaxis], self.get_global_shells(cluster_by_distances=cluster_by_distances, cluster_by_vecs=cluster_by_vecs)-1), axis=-1 ).reshape(-1, 3) shell_max = np.max(pairs[:,-1])+1 if chemical_pair is not None: c = self._ref_structure.get_chemical_symbols() pairs = pairs[np.all(np.sort(c[pairs[:,:2]], axis=-1)==np.sort(chemical_pair), axis=-1)] shell_matrix = [] for ind in np.arange(shell_max): indices = pairs[ind==pairs[:,-1]] if len(indices)>0: ind_tmp = np.unique(indices[:,:-1], axis=0, return_counts=True) shell_matrix.append(coo_matrix((ind_tmp[1], (ind_tmp[0][:,0], ind_tmp[0][:,1])), shape=(len(self._ref_structure), len(self._ref_structure)) )) else: shell_matrix.append(coo_matrix((len(self._ref_structure), len(self._ref_structure)))) return shell_matrix
[docs] def find_neighbors_by_vector(self, vector, return_deviation=False): """ Args: vector (list/np.ndarray): vector by which positions are translated (and neighbors are searched) return_deviation (bool): whether to return distance between the expect positions and real positions Returns: np.ndarray: list of id's for the specified translation Example: a_0 = 2.832 structure = pr.create_structure('Fe', 'bcc', a_0) id_list = structure.find_neighbors_by_vector([0, 0, a_0]) # In this example, you get a list of neighbor atom id's at z+=a_0 for each atom. # This is particularly powerful for SSA when the magnetic structure has to be translated # in each direction. """ z = np.zeros(len(self._ref_structure)*3).reshape(-1, 3) v = np.append(z[:,np.newaxis,:], self.vecs, axis=1) dist = np.linalg.norm(v-np.array(vector), axis=-1, ord=self.norm_order) indices = np.append(np.arange(len(self._ref_structure))[:,np.newaxis], self.indices, axis=1) if return_deviation: return indices[np.arange(len(dist)), np.argmin(dist, axis=-1)], np.min(dist, axis=-1) return indices[np.arange(len(dist)), np.argmin(dist, axis=-1)]
[docs] def cluster_by_vecs( self, distance_threshold=None, n_clusters=None, linkage='complete', affinity='euclidean' ): """ Method to group vectors which have similar values. This method should be used as a part of neigh.get_global_shells(cluster_by_vecs=True) or neigh.get_local_shells(cluster_by_vecs=True). However, in order to specify certain arguments (such as n_jobs or max_iter), it might help to have run this function before calling parent functions, as the data obtained with this function will be stored in the variable `_cluster_vecs` Args: distance_threshold (float/None): The linkage distance threshold above which, clusters will not be merged. (cf. sklearn.cluster.AgglomerativeClustering) n_clusters (int/None): The number of clusters to find. (cf. sklearn.cluster.AgglomerativeClustering) linkage (str): Which linkage criterion to use. The linkage criterion determines which distance to use between sets of observation. The algorithm will merge the pairs of cluster that minimize this criterion. (cf. sklearn.cluster.AgglomerativeClustering) affinity (str/callable): Metric used to compute the linkage. Can be `euclidean`, `l1`, `l2`, `manhattan`, `cosine`, or `precomputed`. If linkage is `ward`, only `euclidean` is accepted. """ allow_ragged = self.allow_ragged if allow_ragged: self.allow_ragged = False if distance_threshold is None and n_clusters is None: distance_threshold = np.min(self.distances) dr = self.vecs[self.distances<np.inf] self._cluster_vecs = AgglomerativeClustering( distance_threshold=distance_threshold, n_clusters=n_clusters, linkage=linkage, affinity=affinity, ).fit(dr) self._cluster_vecs.cluster_centers_ = get_average_of_unique_labels( self._cluster_vecs.labels_, dr ) new_labels = -np.ones_like(self.indices).astype(int) new_labels[self.distances<np.inf] = self._cluster_vecs.labels_ self._cluster_vecs.labels_ = new_labels self.allow_ragged = allow_ragged
[docs] def cluster_by_distances( self, distance_threshold=None, n_clusters=None, linkage='complete', affinity='euclidean', use_vecs=False, ): """ Method to group vectors which have similar lengths. This method should be used as a part of neigh.get_global_shells(cluster_by_vecs=True) or neigh.get_local_shells(cluster_by_distances=True). However, in order to specify certain arguments (such as n_jobs or max_iter), it might help to have run this function before calling parent functions, as the data obtained with this function will be stored in the variable `_cluster_distances` Args: distance_threshold (float/None): The linkage distance threshold above which, clusters will not be merged. (cf. sklearn.cluster.AgglomerativeClustering) n_clusters (int/None): The number of clusters to find. (cf. sklearn.cluster.AgglomerativeClustering) linkage (str): Which linkage criterion to use. The linkage criterion determines which distance to use between sets of observation. The algorithm will merge the pairs of cluster that minimize this criterion. (cf. sklearn.cluster.AgglomerativeClustering) affinity (str/callable): Metric used to compute the linkage. Can be `euclidean`, `l1`, `l2`, `manhattan`, `cosine`, or `precomputed`. If linkage is `ward`, only `euclidean` is accepted. use_vecs (bool): Whether to form clusters for vecs beforehand. If true, the distances obtained from the clustered vectors is used for the distance clustering. Otherwise neigh.distances is used. """ allow_ragged = self.allow_ragged if allow_ragged: self.allow_ragged = False if distance_threshold is None: distance_threshold = 0.1*np.min(self.distances) dr = self.distances[self.distances<np.inf] if use_vecs: if self._cluster_vecs is None: self.cluster_by_vecs() labels_to_consider = self._cluster_vecs.labels_[self._cluster_vecs.labels_>=0] dr = np.linalg.norm( self._cluster_vecs.cluster_centers_[labels_to_consider], axis=-1, ord=self.norm_order ) self._cluster_dist = AgglomerativeClustering( distance_threshold=distance_threshold, n_clusters=n_clusters, linkage=linkage, affinity=affinity, ).fit(dr.reshape(-1, 1)) self._cluster_dist.cluster_centers_ = get_average_of_unique_labels( self._cluster_dist.labels_, dr ) new_labels = -np.ones_like(self.indices).astype(int) new_labels[self.distances<np.inf] = self._cluster_dist.labels_ self._cluster_dist.labels_ = new_labels self.allow_ragged = allow_ragged
[docs] def reset_clusters(self, vecs=True, distances=True): """ Method to reset clusters. Args: vecs (bool): Reset `_cluster_vecs` (cf. `cluster_by_vecs`) distances (bool): Reset `_cluster_distances` (cf. `cluster_by_distances`) """ if vecs: self._cluster_vecs = None if distances: self._cluster_distances = None
[docs] def cluster_analysis( self, id_list, return_cluster_sizes=False ): """ Args: id_list: return_cluster_sizes: Returns: """ self._cluster = [0] * len(self._ref_structure) c_count = 1 # element_list = self.get_atomic_numbers() for ia in id_list: # el0 = element_list[ia] nbrs = self.indices[ia] # print ("nbrs: ", ia, nbrs) if self._cluster[ia] == 0: self._cluster[ia] = c_count self.__probe_cluster(c_count, nbrs, id_list) c_count += 1 cluster = np.array(self._cluster) cluster_dict = { i_c: np.where(cluster == i_c)[0].tolist() for i_c in range(1, c_count) } if return_cluster_sizes: sizes = [self._cluster.count(i_c + 1) for i_c in range(c_count - 1)] return cluster_dict, sizes return cluster_dict # sizes
def __probe_cluster(self, c_count, neighbors, id_list): """ Args: c_count: neighbors: id_list: Returns: """ for nbr_id in neighbors: if self._cluster[nbr_id] == 0: if nbr_id in id_list: # TODO: check also for ordered structures self._cluster[nbr_id] = c_count nbrs = self.indices[nbr_id] self.__probe_cluster(c_count, nbrs, id_list) # TODO: combine with corresponding routine in plot3d
[docs] def get_bonds(self, radius=np.inf, max_shells=None, prec=0.1): """ Args: radius: max_shells: prec: minimum distance between any two clusters (if smaller considered to be single cluster) Returns: """ def get_cluster(dist_vec, ind_vec, prec=prec): ind_where = np.where(np.diff(dist_vec) > prec)[0] + 1 ind_vec_cl = [np.sort(group) for group in np.split(ind_vec, ind_where)] return ind_vec_cl dist = self.distances ind = self.indices el_list = self._ref_structure.get_chemical_symbols() ind_shell = [] for d, i in zip(dist, ind): id_list = get_cluster(d[d < radius], i[d < radius]) # print ("id: ", d[d<radius], id_list, dist_lst) ia_shells_dict = {} for i_shell_list in id_list: ia_shell_dict = {} for i_s in i_shell_list: el = el_list[i_s] if el not in ia_shell_dict: ia_shell_dict[el] = [] ia_shell_dict[el].append(i_s) for el, ia_lst in ia_shell_dict.items(): if el not in ia_shells_dict: ia_shells_dict[el] = [] if max_shells is not None: if len(ia_shells_dict[el]) + 1 > max_shells: continue ia_shells_dict[el].append(ia_lst) ind_shell.append(ia_shells_dict) return ind_shell