Source code for ase.utils.structure_comparator

"""Determine symmetry equivalence of two structures.
Based on the recipe from Comput. Phys. Commun. 183, 690-697 (2012)."""
from collections import Counter
from itertools import combinations, product, filterfalse

import numpy as np
from scipy.spatial import cKDTree as KDTree

from ase import Atom, Atoms
from ase.build.tools import niggli_reduce


def normalize(cell):
    for i in range(3):
        cell[i] /= np.linalg.norm(cell[i])


class SpgLibNotFoundError(Exception):
    """Raised if SPG lib is not found when needed."""

    def __init__(self, msg):
        super(SpgLibNotFoundError, self).__init__(msg)


[docs]class SymmetryEquivalenceCheck: """Compare two structures to determine if they are symmetry equivalent. Based on the recipe from Comput. Phys. Commun. 183, 690-697 (2012). Parameters: angle_tol: float angle tolerance for the lattice vectors in degrees ltol: float relative tolerance for the length of the lattice vectors (per atom) stol: float position tolerance for the site comparison in units of (V/N)^(1/3) (average length between atoms) vol_tol: float volume tolerance in angstrom cubed to compare the volumes of the two structures scale_volume: bool if True the volumes of the two structures are scaled to be equal to_primitive: bool if True the structures are reduced to their primitive cells note that this feature requires spglib to installed Examples: >>> from ase.build import bulk >>> from ase.utils.structure_comparator import SymmetryEquivalenceCheck >>> comp = SymmetryEquivalenceCheck() Compare a cell with a rotated version >>> a = bulk('Al', orthorhombic=True) >>> b = a.copy() >>> b.rotate(60, 'x', rotate_cell=True) >>> comp.compare(a, b) True Transform to the primitive cell and then compare >>> pa = bulk('Al') >>> comp.compare(a, pa) False >>> comp = SymmetryEquivalenceCheck(to_primitive=True) >>> comp.compare(a, pa) True Compare one structure with a list of other structures >>> import numpy as np >>> from ase import Atoms >>> s1 = Atoms('H3', positions=[[0.5, 0.5, 0], ... [0.5, 1.5, 0], ... [1.5, 1.5, 0]], ... cell=[2, 2, 2], pbc=True) >>> comp = SymmetryEquivalenceCheck(stol=0.068) >>> s2_list = [] >>> for d in np.linspace(0.1, 1.0, 5): ... s2 = s1.copy() ... s2.positions[0] += [d, 0, 0] ... s2_list.append(s2) >>> comp.compare(s1, s2_list[:-1]) False >>> comp.compare(s1, s2_list) True """ def __init__(self, angle_tol=1.0, ltol=0.05, stol=0.05, vol_tol=0.1, scale_volume=False, to_primitive=False): self.angle_tol = angle_tol * np.pi / 180.0 # convert to radians self.scale_volume = scale_volume self.stol = stol self.ltol = ltol self.vol_tol = vol_tol self.position_tolerance = 0.0 self.to_primitive = to_primitive # Variables to be used in the compare function self.s1 = None self.s2 = None self.expanded_s1 = None self.expanded_s2 = None self.least_freq_element = None def _niggli_reduce(self, atoms): """Reduce to niggli cells. Reduce the atoms to niggli cells, then rotates the niggli cells to the so called "standard" orientation with one lattice vector along the x-axis and a second vector in the xy plane. """ niggli_reduce(atoms) self._standarize_cell(atoms) def _standarize_cell(self, atoms): """Rotate the first vector such that it points along the x-axis. Then rotate around the first vector so the second vector is in the xy plane. """ # Rotate first vector to x axis cell = atoms.get_cell().T total_rot_mat = np.eye(3) v1 = cell[:, 0] l1 = np.sqrt(v1[0]**2 + v1[2]**2) angle = np.abs(np.arcsin(v1[2] / l1)) if (v1[0] < 0.0 and v1[2] > 0.0): angle = np.pi - angle elif (v1[0] < 0.0 and v1[2] < 0.0): angle = np.pi + angle elif (v1[0] > 0.0 and v1[2] < 0.0): angle = -angle ca = np.cos(angle) sa = np.sin(angle) rotmat = np.array([[ca, 0.0, sa], [0.0, 1.0, 0.0], [-sa, 0.0, ca]]) total_rot_mat = rotmat.dot(total_rot_mat) cell = rotmat.dot(cell) v1 = cell[:, 0] l1 = np.sqrt(v1[0]**2 + v1[1]**2) angle = np.abs(np.arcsin(v1[1] / l1)) if (v1[0] < 0.0 and v1[1] > 0.0): angle = np.pi - angle elif (v1[0] < 0.0 and v1[1] < 0.0): angle = np.pi + angle elif (v1[0] > 0.0 and v1[1] < 0.0): angle = -angle ca = np.cos(angle) sa = np.sin(angle) rotmat = np.array([[ca, sa, 0.0], [-sa, ca, 0.0], [0.0, 0.0, 1.0]]) total_rot_mat = rotmat.dot(total_rot_mat) cell = rotmat.dot(cell) # Rotate around x axis such that the second vector is in the xy plane v2 = cell[:, 1] l2 = np.sqrt(v2[1]**2 + v2[2]**2) angle = np.abs(np.arcsin(v2[2] / l2)) if (v2[1] < 0.0 and v2[2] > 0.0): angle = np.pi - angle elif (v2[1] < 0.0 and v2[2] < 0.0): angle = np.pi + angle elif (v2[1] > 0.0 and v2[2] < 0.0): angle = -angle ca = np.cos(angle) sa = np.sin(angle) rotmat = np.array([[1.0, 0.0, 0.0], [0.0, ca, sa], [0.0, -sa, ca]]) total_rot_mat = rotmat.dot(total_rot_mat) cell = rotmat.dot(cell) atoms.set_cell(cell.T) atoms.set_positions(total_rot_mat.dot(atoms.get_positions().T).T) atoms.wrap(pbc=[1, 1, 1]) return atoms def _get_element_count(self, struct): """Count the number of elements in each of the structures.""" return Counter(struct.numbers) def _get_angles(self, cell): """Get the internal angles of the unit cell.""" cell = cell.copy() normalize(cell) dot = cell.dot(cell.T) # Extract only the relevant dot products dot = [dot[0, 1], dot[0, 2], dot[1, 2]] # Return angles return np.arccos(dot) def _has_same_elements(self): """Check if two structures have same elements.""" elem1 = self._get_element_count(self.s1) return elem1 == self._get_element_count(self.s2) def _has_same_angles(self): """Check that the Niggli unit vectors has the same internal angles.""" ang1 = np.sort(self._get_angles(self.s1.get_cell())) ang2 = np.sort(self._get_angles(self.s2.get_cell())) return np.allclose(ang1, ang2, rtol=0, atol=self.angle_tol) def _has_same_volume(self): vol1 = self.s1.get_volume() vol2 = self.s2.get_volume() return np.abs(vol1 - vol2) < self.vol_tol def _scale_volumes(self): """Scale the cell of s2 to have the same volume as s1.""" cell2 = self.s2.get_cell() # Get the volumes v2 = np.linalg.det(cell2) v1 = np.linalg.det(self.s1.get_cell()) # Scale the cells coordinate_scaling = (v1 / v2)**(1.0 / 3.0) cell2 *= coordinate_scaling self.s2.set_cell(cell2, scale_atoms=True)
[docs] def compare(self, s1, s2): """Compare the two structures. Return *True* if the two structures are equivalent, *False* otherwise. Parameters: s1: Atoms object. Transformation matrices are calculated based on this structure. s2: Atoms or list s1 can be compared to one structure or many structures supplied in a list. If s2 is a list it returns True if any structure in s2 matches s1, False otherwise. """ if self.to_primitive: s1 = self._reduce_to_primitive(s1) self._set_least_frequent_element(s1) self._least_frequent_element_to_origin(s1) self.s1 = s1.copy() vol = self.s1.get_volume() self.expanded_s1 = None s1_niggli_reduced = False if isinstance(s2, Atoms): # Just make it a list of length 1 s2 = [s2] matrices = None translations = None transposed_matrices = None for struct in s2: self.s2 = struct.copy() self.expanded_s2 = None if self.to_primitive: self.s2 = self._reduce_to_primitive(self.s2) # Compare number of elements in structures if len(self.s1) != len(self.s2): continue # Compare chemical formulae if not self._has_same_elements(): continue # Compare angles if not s1_niggli_reduced: self._niggli_reduce(self.s1) self._niggli_reduce(self.s2) if not self._has_same_angles(): continue # Compare volumes if self.scale_volume: self._scale_volumes() if not self._has_same_volume(): continue if matrices is None: matrices = self._get_rotation_reflection_matrices() if matrices is None: continue if translations is None: translations = self._get_least_frequent_positions(self.s1) # After the candidate translation based on s1 has been computed # we need potentially to swap s1 and s2 for robust comparison self._least_frequent_element_to_origin(self.s2) switch = self._switch_reference_struct() if switch: # Remember the matrices and translations used before old_matrices = matrices old_translations = translations # If a s1 and s2 has been switched we need to use the # transposed version of the matrices to map atoms the # other way if transposed_matrices is None: transposed_matrices = np.transpose(matrices, axes=[0, 2, 1]) matrices = transposed_matrices translations = self._get_least_frequent_positions(self.s1) # Calculate tolerance on positions self.position_tolerance = \ self.stol * (vol / len(self.s2))**(1.0 / 3.0) if self._positions_match(matrices, translations): return True # Set the reference structure back to its original self.s1 = s1.copy() if switch: self.expanded_s1 = self.expanded_s2 matrices = old_matrices translations = old_translations return False
def _set_least_frequent_element(self, atoms): """Save the atomic number of the least frequent element.""" elem1 = self._get_element_count(atoms) self.least_freq_element = elem1.most_common()[-1][0] def _get_least_frequent_positions(self, atoms): """Get the positions of the least frequent element in atoms.""" pos = atoms.get_positions(wrap=True) return pos[atoms.numbers == self.least_freq_element] def _get_only_least_frequent_of(self, struct): """Get the atoms object with all other elements than the least frequent one removed. Wrap the positions to get everything in the cell.""" pos = struct.get_positions(wrap=True) indices = struct.numbers == self.least_freq_element least_freq_struct = struct[indices] least_freq_struct.set_positions(pos[indices]) return least_freq_struct def _switch_reference_struct(self): """There is an intrinsic assymetry in the system because one of the atoms are being expanded, while the other is not. This can cause the algorithm to return different result depending on which structure is passed first. We adopt the convention of using the atoms object having the fewest atoms in its expanded cell as the reference object. We return True if a switch of structures has been performed.""" # First expand the cells if self.expanded_s1 is None: self.expanded_s1 = self._expand(self.s1) if self.expanded_s2 is None: self.expanded_s2 = self._expand(self.s2) exp1 = self.expanded_s1 exp2 = self.expanded_s2 if len(exp1) < len(exp2): # s1 should be the reference structure # We have to swap s1 and s2 s1_temp = self.s1.copy() self.s1 = self.s2 self.s2 = s1_temp exp1_temp = self.expanded_s1.copy() self.expanded_s1 = self.expanded_s2 self.expanded_s2 = exp1_temp return True return False def _positions_match(self, rotation_reflection_matrices, translations): """Check if the position and elements match. Note that this function changes self.s1 and self.s2 to the rotation and translation that matches best. Hence, it is crucial that this function calls the element comparison, not the other way around. """ pos1_ref = self.s1.get_positions(wrap=True) # Get the expanded reference object exp2 = self.expanded_s2 # Build a KD tree to enable fast look-up of nearest neighbours tree = KDTree(exp2.get_positions()) for i in range(translations.shape[0]): # Translate pos1_trans = pos1_ref - translations[i] for matrix in rotation_reflection_matrices: # Rotate pos1 = matrix.dot(pos1_trans.T).T # Update the atoms positions self.s1.set_positions(pos1) self.s1.wrap(pbc=[1, 1, 1]) if self._elements_match(self.s1, exp2, tree): return True return False def _expand(self, ref_atoms, tol=0.0001): """If an atom is closer to a boundary than tol it is repeated at the opposite boundaries. This ensures that atoms having crossed the cell boundaries due to numerical noise are properly detected. The distance between a position and cell boundary is calculated as: dot(position, (b_vec x c_vec) / (|b_vec| |c_vec|) ), where x is the cross product. """ syms = ref_atoms.get_chemical_symbols() cell = ref_atoms.get_cell() positions = ref_atoms.get_positions(wrap=True) expanded_atoms = ref_atoms.copy() # Calculate normal vectors to the unit cell faces normal_vectors = np.array([np.cross(cell[1, :], cell[2, :]), np.cross(cell[0, :], cell[2, :]), np.cross(cell[0, :], cell[1, :])]) normalize(normal_vectors) # Get the distance to the unit cell faces from each atomic position pos2faces = np.abs(positions.dot(normal_vectors.T)) # And the opposite faces pos2oppofaces = np.abs(np.dot(positions - np.sum(cell, axis=0), normal_vectors.T)) for i, i2face in enumerate(pos2faces): # Append indices for positions close to the other faces # and convert to boolean array signifying if the position at # index i is close to the faces bordering origo (0, 1, 2) or # the opposite faces (3, 4, 5) i_close2face = np.append(i2face, pos2oppofaces[i]) < tol # For each position i.e. row it holds that # 1 x True -> close to face -> 1 extra atom at opposite face # 2 x True -> close to edge -> 3 extra atoms at opposite edges # 3 x True -> close to corner -> 7 extra atoms opposite corners # E.g. to add atoms at all corners we need to use the cell # vectors: (a, b, c, a + b, a + c, b + c, a + b + c), we use # itertools.combinations to get them all for j in range(sum(i_close2face)): for c in combinations(np.nonzero(i_close2face)[0], j + 1): # Get the displacement vectors by adding the corresponding # cell vectors, if the atom is close to an opposite face # i.e. k > 2 subtract the cell vector disp_vec = np.zeros(3) for k in c: disp_vec += cell[k % 3] * (int(k < 3) * 2 - 1) pos = positions[i] + disp_vec expanded_atoms.append(Atom(syms[i], position=pos)) return expanded_atoms def _equal_elements_in_array(self, arr): s = np.sort(arr) return np.any(s[1:] == s[:-1]) def _elements_match(self, s1, s2, kdtree): """Check if all the elements in s1 match the corresponding position in s2 NOTE: The unit cells may be in different octants Hence, try all cyclic permutations of x,y and z """ pos1 = s1.get_positions() for order in range(1): # Is the order still needed? pos_order = [order, (order + 1) % 3, (order + 2) % 3] pos = pos1[:, np.argsort(pos_order)] dists, closest_in_s2 = kdtree.query(pos) # Check if the elements are the same if not np.all(s2.numbers[closest_in_s2] == s1.numbers): return False # Check if any distance is too large if np.any(dists > self.position_tolerance): return False # Check for duplicates in what atom is closest if self._equal_elements_in_array(closest_in_s2): return False return True def _least_frequent_element_to_origin(self, atoms): """Put one of the least frequent elements at the origin.""" least_freq_pos = self._get_least_frequent_positions(atoms) cell_diag = np.sum(atoms.get_cell(), axis=0) d = least_freq_pos[0] - 1e-6 * cell_diag atoms.positions -= d atoms.wrap(pbc=[1, 1, 1]) def _get_rotation_reflection_matrices(self): """Compute candidates for the transformation matrix.""" atoms1_ref = self._get_only_least_frequent_of(self.s1) cell = self.s1.get_cell().T cell_diag = np.sum(cell, axis=1) angle_tol = self.angle_tol # Additional vector that is added to make sure that # there always is an atom at the origin delta_vec = 1E-6 * cell_diag # Store three reference vectors and their lengths ref_vec = self.s2.get_cell() ref_vec_lengths = np.linalg.norm(ref_vec, axis=1) # Compute ref vec angles # ref_angles are arranged as [angle12, angle13, angle23] ref_angles = np.array(self._get_angles(ref_vec)) large_angles = ref_angles > np.pi / 2.0 ref_angles[large_angles] = np.pi - ref_angles[large_angles] # Translate by one cell diagonal so that a central cell is # surrounded by cells in all directions sc_atom_search = atoms1_ref * (3, 3, 3) new_sc_pos = sc_atom_search.get_positions() new_sc_pos -= new_sc_pos[0] + cell_diag - delta_vec lengths = np.linalg.norm(new_sc_pos, axis=1) candidate_indices = [] rtol = self.ltol / len(self.s1) for k in range(3): correct_lengths_mask = np.isclose(lengths, ref_vec_lengths[k], rtol=rtol, atol=0) # The first vector is not interesting correct_lengths_mask[0] = False # If no trial vectors can be found (for any direction) # then the candidates are different and we return None if not np.any(correct_lengths_mask): return None candidate_indices.append(np.nonzero(correct_lengths_mask)[0]) # Now we calculate all relevant angles in one step. The relevant angles # are the ones made by the current candidates. We will have to keep # track of the indices in the angles matrix and the indices in the # position and length arrays. # Get all candidate indices (aci), only unique values aci = np.sort(list(set().union(*candidate_indices))) # Make a dictionary from original positions and lengths index to # index in angle matrix i2ang = dict(zip(aci, range(len(aci)))) # Calculate the dot product divided by the lengths: # cos(angle) = dot(vec1, vec2) / |vec1| |vec2| cosa = np.inner(new_sc_pos[aci], new_sc_pos[aci]) / np.outer(lengths[aci], lengths[aci]) # Make sure the inverse cosine will work cosa[cosa > 1] = 1 cosa[cosa < -1] = -1 angles = np.arccos(cosa) # Do trick for enantiomorphic structures angles[angles > np.pi / 2] = np.pi - angles[angles > np.pi / 2] # Check which angles match the reference angles # Test for all combinations on candidates. filterfalse makes sure # that there are no duplicate candidates. product is the same as # nested for loops. refined_candidate_list = [] for p in filterfalse(self._equal_elements_in_array, product(*candidate_indices)): a = np.array([angles[i2ang[p[0]], i2ang[p[1]]], angles[i2ang[p[0]], i2ang[p[2]]], angles[i2ang[p[1]], i2ang[p[2]]]]) if np.allclose(a, ref_angles, atol=angle_tol, rtol=0): refined_candidate_list.append(new_sc_pos[np.array(p)].T) # Get the rotation/reflection matrix [R] by: # [R] = [V][T]^-1, where [V] is the reference vectors and # [T] is the trial vectors # XXX What do we know about the length/shape of refined_candidate_list? if len(refined_candidate_list) == 0: return None elif len(refined_candidate_list) == 1: inverted_trial = 1.0 / refined_candidate_list else: inverted_trial = np.linalg.inv(refined_candidate_list) # Equivalent to np.matmul(ref_vec.T, inverted_trial) candidate_trans_mat = np.dot(ref_vec.T, inverted_trial.T).T return candidate_trans_mat def _reduce_to_primitive(self, structure): """Reduce the two structure to their primitive type""" try: import spglib except ImportError: raise SpgLibNotFoundError( "SpgLib is required if to_primitive=True") cell = (structure.get_cell()).tolist() pos = structure.get_scaled_positions().tolist() numbers = structure.get_atomic_numbers() cell, scaled_pos, numbers = spglib.standardize_cell( (cell, pos, numbers), to_primitive=True) atoms = Atoms( scaled_positions=scaled_pos, numbers=numbers, cell=cell, pbc=True) return atoms