Source code for ase_ga.element_mutations

# fmt: off

"""Mutation classes, that mutate the elements in the supplied
atoms objects."""
import numpy as np

from ase.data import atomic_numbers
from ase_ga.offspring_creator import OffspringCreator


[docs] def chunks(line, n): """split a list into smaller chunks""" return [line[i:i + n] for i in range(0, len(line), n)]
[docs] class ElementMutation(OffspringCreator): """The base class for all operators where the elements of the atoms objects are mutated""" def __init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=1, rng=np.random): OffspringCreator.__init__(self, verbose, num_muts=num_muts, rng=rng) if not isinstance(element_pool[0], (list, np.ndarray)): self.element_pools = [element_pool] else: self.element_pools = element_pool if max_diff_elements is None: self.max_diff_elements = [1e6 for _ in self.element_pools] elif isinstance(max_diff_elements, int): self.max_diff_elements = [max_diff_elements] else: self.max_diff_elements = max_diff_elements assert len(self.max_diff_elements) == len(self.element_pools) if min_percentage_elements is None: self.min_percentage_elements = [0 for _ in self.element_pools] elif isinstance(min_percentage_elements, (int, float)): self.min_percentage_elements = [min_percentage_elements] else: self.min_percentage_elements = min_percentage_elements assert len(self.min_percentage_elements) == len(self.element_pools) self.min_inputs = 1
[docs] def get_new_individual(self, parents): raise NotImplementedError
[docs] def get_mutation_index_list_and_choices(self, atoms): """Returns a list of the indices that are going to be mutated and a list of possible elements to mutate to. The lists obey the criteria set in the initialization. """ itbm_ok = False while not itbm_ok: itbm = self.rng.choice(range(len(atoms))) # index to be mutated itbm_ok = True for i, e in enumerate(self.element_pools): if atoms[itbm].symbol in e: elems = e[:] elems_in, indices_in = zip(*[(a.symbol, a.index) for a in atoms if a.symbol in elems]) max_diff_elem = self.max_diff_elements[i] min_percent_elem = self.min_percentage_elements[i] if min_percent_elem == 0: min_percent_elem = 1. / len(elems_in) break else: itbm_ok = False # Check that itbm obeys min/max criteria diff_elems_in = len(set(elems_in)) if diff_elems_in == max_diff_elem: # No more different elements allowed -> one element mutation ltbm = [] # list to be mutated for i in range(len(atoms)): if atoms[i].symbol == atoms[itbm].symbol: ltbm.append(i) else: # Fewer or too many different elements already if self.verbose: print(int(min_percent_elem * len(elems_in)), min_percent_elem, len(elems_in)) all_chunks = chunks(indices_in, int(min_percent_elem * len(elems_in))) itbm_num_of_elems = 0 for a in atoms: if a.index == itbm: break if a.symbol in elems: itbm_num_of_elems += 1 ltbm = all_chunks[itbm_num_of_elems // (int(min_percent_elem * len(elems_in))) - 1] elems.remove(atoms[itbm].symbol) return ltbm, elems
[docs] class RandomElementMutation(ElementMutation): """Mutation that exchanges an element with a randomly chosen element from the supplied pool of elements If the individual consists of different groups of elements the element pool can be supplied as a list of lists Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] max_diff_elements: The maximum number of different elements in the individual. Default is infinite. If the elements are grouped max_diff_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. min_percentage_elements: The minimum percentage of any element in the individual. Default is any number is allowed. If the elements are grouped min_percentage_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. rng: Random number generator By default numpy.random. Example: element_pool=[[A,B,C,D],[x,y,z]], max_diff_elements=[3,2], min_percentage_elements=[.25, .5] An individual could be "D,B,B,C,x,x,x,x,z,z,z,z" """ def __init__(self, element_pool, max_diff_elements=None, min_percentage_elements=None, verbose=False, num_muts=1, rng=np.random): ElementMutation.__init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'RandomElementMutation'
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] ltbm, choices = self.get_mutation_index_list_and_choices(f) new_element = self.rng.choice(choices) for a in f: if a.index in ltbm: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), self.descriptor + ': Parent {}'.format(f.info['confid']))
[docs] def mendeleiev_table(): r""" Returns the mendeleiev table as a python list of lists. Each cell contains either None or a pair (symbol, atomic number), or a list of pairs for the cells \* and \**. """ import re elems = 'HHeLiBeBCNOFNeNaMgAlSiPSClArKCaScTiVCrMnFeCoNiCuZnGaGeAsSeBrKrRb' elems += 'SrYZrNbMoTcRuRhPdAgCdInSnSbTeIXeCsBaLaCePrNdPmSmEuGdTbDyHoErTm' elems += 'YbLuHfTaWReOsIrPtAuHgTlPbBiPoAtRnFrRaAcThPaUNpPuAmCmBkCfEsFmMd' elems += 'NoLrRfDbSgBhHsMtDsRgUubUutUuqUupUuhUusUuo' L = [(e, i + 1) for (i, e) in enumerate(re.compile('[A-Z][a-z]*').findall(elems))] for i, j in ((88, 103), (56, 71)): L[i] = L[i:j] L[i + 1:] = L[j:] for i, j in ((12, 10), (4, 10), (1, 16)): L[i:i] = [None] * j return [L[18 * i:18 * (i + 1)] for i in range(7)]
[docs] def get_row_column(element): """Returns the row and column of the element in the periodic table. Note that Lanthanides and Actinides are defined to be group (column) 3 elements""" t = mendeleiev_table() en = (element, atomic_numbers[element]) for i in range(len(t)): for j in range(len(t[i])): if en == t[i][j]: return i, j elif isinstance(t[i][j], list): # Lanthanide or Actinide if en in t[i][j]: return i, 3
def get_periodic_table_distance(e1, e2): rc1 = np.array(get_row_column(e1)) rc2 = np.array(get_row_column(e2)) return sum(np.abs(rc1 - rc2))
[docs] class MoveDownMutation(ElementMutation): """ Mutation that exchanges an element with an element one step (or more steps if fewer is forbidden) down the same column in the periodic table. This mutation is introduced and used in: P. B. Jensen et al., Phys. Chem. Chem. Phys., 16, 36, 19732-19740 (2014) The idea behind is that elements close to each other in the periodic table is chemically similar, and therefore exhibit similar properties. An individual in the population is typically close to fittest possible, exchanging an element with a similar element will normally result in a slight increase (or decrease) in fitness. Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] max_diff_elements: The maximum number of different elements in the individual. Default is infinite. If the elements are grouped max_diff_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. min_percentage_elements: The minimum percentage of any element in the individual. Default is any number is allowed. If the elements are grouped min_percentage_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. rng: Random number generator By default numpy.random. Example: element_pool=[[A,B,C,D],[x,y,z]], max_diff_elements=[3,2], min_percentage_elements=[.25, .5] An individual could be "D,B,B,C,x,x,x,x,z,z,z,z" """ def __init__(self, element_pool, max_diff_elements=None, min_percentage_elements=None, verbose=False, num_muts=1, rng=np.random): ElementMutation.__init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'MoveDownMutation'
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] ltbm, choices = self.get_mutation_index_list_and_choices(f) # periodic table row, periodic table column ptrow, ptcol = get_row_column(f[ltbm[0]].symbol) popped = [] m = 0 for j in range(len(choices)): e = choices[j - m] row, column = get_row_column(e) if row <= ptrow or column != ptcol: # Throw away if above (lower numbered row) # or in a different column in the periodic table popped.append(choices.pop(j - m)) m += 1 used_descriptor = self.descriptor if len(choices) == 0: msg = '{0},{2} cannot be mutated by {1}, ' msg = msg.format(f.info['confid'], self.descriptor, f[ltbm[0]].symbol) msg += 'doing random mutation instead' if self.verbose: print(msg) used_descriptor = 'RandomElementMutation_from_{0}' used_descriptor = used_descriptor.format(self.descriptor) self.rng.shuffle(popped) choices = popped else: # Sorting the element that lie below and in the same column # in the periodic table so that the one closest below is first choices.sort(key=lambda x: get_row_column(x)[0]) new_element = choices[0] for a in f: if a.index in ltbm: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), used_descriptor + ': Parent {}'.format(f.info['confid']))
[docs] class MoveUpMutation(ElementMutation): """ Mutation that exchanges an element with an element one step (or more steps if fewer is forbidden) up the same column in the periodic table. This mutation is introduced and used in: P. B. Jensen et al., Phys. Chem. Chem. Phys., 16, 36, 19732-19740 (2014) See MoveDownMutation for the idea behind Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] max_diff_elements: The maximum number of different elements in the individual. Default is infinite. If the elements are grouped max_diff_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. min_percentage_elements: The minimum percentage of any element in the individual. Default is any number is allowed. If the elements are grouped min_percentage_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. rng: Random number generator By default numpy.random. Example: element_pool=[[A,B,C,D],[x,y,z]], max_diff_elements=[3,2], min_percentage_elements=[.25, .5] An individual could be "D,B,B,C,x,x,x,x,z,z,z,z" """ def __init__(self, element_pool, max_diff_elements=None, min_percentage_elements=None, verbose=False, num_muts=1, rng=np.random): ElementMutation.__init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'MoveUpMutation'
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] ltbm, choices = self.get_mutation_index_list_and_choices(f) # periodic table row, periodic table column ptrow, ptcol = get_row_column(f[ltbm[0]].symbol) popped = [] m = 0 for j in range(len(choices)): e = choices[j - m] row, column = get_row_column(e) if row >= ptrow or column != ptcol: # Throw away if below (higher numbered row) # or in a different column in the periodic table popped.append(choices.pop(j - m)) m += 1 used_descriptor = self.descriptor if len(choices) == 0: msg = '{0},{2} cannot be mutated by {1}, ' msg = msg.format(f.info['confid'], self.descriptor, f[ltbm[0]].symbol) msg += 'doing random mutation instead' if self.verbose: print(msg) used_descriptor = 'RandomElementMutation_from_{0}' used_descriptor = used_descriptor.format(self.descriptor) self.rng.shuffle(popped) choices = popped else: # Sorting the element that lie above and in the same column # in the periodic table so that the one closest above is first choices.sort(key=lambda x: get_row_column(x)[0], reverse=True) new_element = choices[0] for a in f: if a.index in ltbm: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), used_descriptor + ': Parent {}'.format(f.info['confid']))
[docs] class MoveRightMutation(ElementMutation): """ Mutation that exchanges an element with an element one step (or more steps if fewer is forbidden) to the right in the same row in the periodic table. This mutation is introduced and used in: P. B. Jensen et al., Phys. Chem. Chem. Phys., 16, 36, 19732-19740 (2014) See MoveDownMutation for the idea behind Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] max_diff_elements: The maximum number of different elements in the individual. Default is infinite. If the elements are grouped max_diff_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. min_percentage_elements: The minimum percentage of any element in the individual. Default is any number is allowed. If the elements are grouped min_percentage_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. rng: Random number generator By default numpy.random. Example: element_pool=[[A,B,C,D],[x,y,z]], max_diff_elements=[3,2], min_percentage_elements=[.25, .5] An individual could be "D,B,B,C,x,x,x,x,z,z,z,z" """ def __init__(self, element_pool, max_diff_elements=None, min_percentage_elements=None, verbose=False, num_muts=1, rng=np.random): ElementMutation.__init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'MoveRightMutation'
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] ltbm, choices = self.get_mutation_index_list_and_choices(f) # periodic table row, periodic table column ptrow, ptcol = get_row_column(f[ltbm[0]].symbol) popped = [] m = 0 for j in range(len(choices)): e = choices[j - m] row, column = get_row_column(e) if row != ptrow or column <= ptcol: # Throw away if to the left (a lower numbered column) # or in a different row in the periodic table popped.append(choices.pop(j - m)) m += 1 used_descriptor = self.descriptor if len(choices) == 0: msg = '{0},{2} cannot be mutated by {1}, ' msg = msg.format(f.info['confid'], self.descriptor, f[ltbm[0]].symbol) msg += 'doing random mutation instead' if self.verbose: print(msg) used_descriptor = 'RandomElementMutation_from_{0}' used_descriptor = used_descriptor.format(self.descriptor) self.rng.shuffle(popped) choices = popped else: # Sorting so the element closest to the right is first choices.sort(key=lambda x: get_row_column(x)[1]) new_element = choices[0] for a in f: if a.index in ltbm: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), used_descriptor + ': Parent {}'.format(f.info['confid']))
[docs] class MoveLeftMutation(ElementMutation): """ Mutation that exchanges an element with an element one step (or more steps if fewer is forbidden) to the left in the same row in the periodic table. This mutation is introduced and used in: P. B. Jensen et al., Phys. Chem. Chem. Phys., 16, 36, 19732-19740 (2014) See MoveDownMutation for the idea behind Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] max_diff_elements: The maximum number of different elements in the individual. Default is infinite. If the elements are grouped max_diff_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. min_percentage_elements: The minimum percentage of any element in the individual. Default is any number is allowed. If the elements are grouped min_percentage_elements should be supplied as a list with each input corresponding to the elements specified in the same input in element_pool. rng: Random number generator By default numpy.random. Example: element_pool=[[A,B,C,D],[x,y,z]], max_diff_elements=[3,2], min_percentage_elements=[.25, .5] An individual could be "D,B,B,C,x,x,x,x,z,z,z,z" """ def __init__(self, element_pool, max_diff_elements=None, min_percentage_elements=None, verbose=False, num_muts=1, rng=np.random): ElementMutation.__init__(self, element_pool, max_diff_elements, min_percentage_elements, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'MoveLeftMutation'
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] ltbm, choices = self.get_mutation_index_list_and_choices(f) # periodic table row, periodic table column ptrow, ptcol = get_row_column(f[ltbm[0]].symbol) popped = [] m = 0 for j in range(len(choices)): e = choices[j - m] row, column = get_row_column(e) if row != ptrow or column >= ptcol: # Throw away if to the right (a higher numbered column) # or in a different row in the periodic table popped.append(choices.pop(j - m)) m += 1 used_descriptor = self.descriptor if len(choices) == 0: msg = '{0},{2} cannot be mutated by {1}, ' msg = msg.format(f.info['confid'], self.descriptor, f[ltbm[0]].symbol) msg += 'doing random mutation instead' if self.verbose: print(msg) used_descriptor = 'RandomElementMutation_from_{0}' used_descriptor = used_descriptor.format(self.descriptor) self.rng.shuffle(popped) choices = popped else: # Sorting so the element closest to the left is first choices.sort(key=lambda x: get_row_column(x)[1], reverse=True) new_element = choices[0] for a in f: if a.index in ltbm: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), used_descriptor + ':Parent {}'.format(f.info['confid']))
[docs] class FullElementMutation(OffspringCreator): """Mutation that exchanges an all elements of a certain type with another randomly chosen element from the supplied pool of elements. Any constraints on the mutation are inhereted from the original candidate. Parameters: element_pool: List of elements in the phase space. The elements can be grouped if the individual consist of different types of elements. The list should then be a list of lists e.g. [[list1], [list2]] rng: Random number generator By default numpy.random. """ def __init__(self, element_pool, verbose=False, num_muts=1, rng=np.random): OffspringCreator.__init__(self, verbose, num_muts=num_muts, rng=rng) self.descriptor = 'FullElementMutation' if not isinstance(element_pool[0], (list, np.ndarray)): self.element_pools = [element_pool] else: self.element_pools = element_pool
[docs] def get_new_individual(self, parents): f = parents[0] indi = self.initialize_individual(f) indi.info['data']['parents'] = [f.info['confid']] # Randomly choose an element to mutate in the current individual. old_element = self.rng.choice([a.symbol for a in f]) # Find the list containing the chosen element. By choosing a new # element from the same list, the percentages are not altered. for i in range(len(self.element_pools)): if old_element in self.element_pools[i]: lm = i not_val = True while not_val: new_element = self.rng.choice(self.element_pools[lm]) not_val = new_element == old_element for a in f: if a.symbol == old_element: a.symbol = new_element indi.append(a) return (self.finalize_individual(indi), self.descriptor + ': Parent {}'.format(f.info['confid']))