candidate_generation.py 14.1 KB
Newer Older
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
1
2
3
4
import numpy as np
from abc import ABC, abstractmethod
from ase.data import covalent_radii
from ase.geometry import get_distances
5
from ase import Atoms
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
6
7
from ase.visualize import view

8
9
10
from utils import check_valid_bondlengths, get_min_distances_as_fraction_of_covalent

import warnings
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

class CandidateGenerator(ABC):
    """Baseclass for mutation and crossover operations as well
    as the startgenerator.

    Parameters:

    blmin: The minimum allowed distance between atoms in units of
    the covalent distance between atoms, where d_cov=r_cov_i+r_cov_j.
    
    blmax: The maximum allowed distance, in units of the covalent 
    distance, from a single isolated atom to the closest atom. If
    blmax=None, no constraint is enforced on isolated atoms.

    force_all_bonds_valid: If True all bondlengths are forced to
    be valid according to blmin/blmax. If False, only bondlengths 
    of atoms specified in bondlength checks during operations are
    tested. The specified atoms are typically the ones changed 
    during operations. Default is False, as True might cause
    problems with GOFEE, as GPR-relaxations and dual-steps might
    result in structures that does not obey blmin/blmax.
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
32
    """
33
34
    def __init__(self, blmin=0.7, blmax=1.4, constraints=None,
                 force_all_bonds_valid=False, *args, **kwargs):
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
35
36
        self.blmin = blmin
        self.blmax = blmax
37
        self.constraints = constraints
38
        self.force_all_bonds_valid = force_all_bonds_valid
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
39
40
        self.description = 'Unspecified'

41
42
43
44
45
    def check_valid_bondlengths(self, a, indices=None,
                                check_too_close=True, check_isolated=True):
        """ Method to check if bondlengths are valid according to blmin
        amd blmax.
        """
46
47
48
49
50
51
52
53
        if self.force_all_bonds_valid:
            # Check all bonds (mainly for testing)
            return check_valid_bondlengths(a, self.blmin, self.blmax+0.1,
                                           check_too_close=check_too_close,
                                           check_isolated=check_isolated)
        else:
            # Check only specified ones
            # (typically only for the atoms changed during operation)
54
55
            return check_valid_bondlengths(a, self.blmin, self.blmax+0.1,
                                           indices=indices,
56
57
                                           check_too_close=check_too_close,
                                           check_isolated=check_isolated)
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83

    def get_new_candidate(self, parents=None):
        """Standardized candidate generation method for all mutation
        and crossover operations.
        """
        # Check bondlengths
        if parents is not None:
            for i, parent in enumerate(parents):
                self.check_bondlengths(parent, f'SHORT BONDS IN PARENT {i}')

        for _ in range(5): # Make five tries
            a = self.operation(parents)
            if a is not None:
                a = self.finalize(a)
                break
        else:
            return None
        return a

    def train(self):
        """ Method to be implemented for the operations that rely on
        a Machine-Learned model to perform more informed/guided 
        mutation and crossover operations.
        """
        pass

Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
84
    @abstractmethod
85
    def operation(self):
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
86
87
        pass

88
    def finalize(self, a, a0=None, successfull=True):
89
90
91
92
93
94
        """ Method to finalize new candidates.
        """
        # Wrap positions
        a.wrap()

        # finalize description
95
96
        if successfull:
            description = self.description
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
97
        else:
98
            description = 'failed ' + self.description
99
100

        # Save description 
101
102
103
104
        try:
            a.info['key_value_pairs']['origin'] = description
        except:
            a.info['key_value_pairs'] = {'origin': description}
105
106
        
        self.check_bondlengths(a, 'SHORT BONDS AFTER OPPERATION')
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
107
108
        return a

109
110
111
112
113
114
115
116
117
118
119
120
    def check_bondlengths(self, a, warn_text):
            if self.force_all_bonds_valid:
                # Check all bonds
                valid_bondlengths = self.check_valid_bondlengths(a)
                assert valid_bondlengths, 'bondlengths are not valid'
            else:
                d_shortest_bond, index_shortest_bond = get_min_distances_as_fraction_of_covalent(a)
                if d_shortest_bond < self.blmin:
                    text = f"""{warn_text}:
                               Atom {index_shortest_bond} has bond with d={d_shortest_bond}d_covalent"""
                    warnings.warn(text)

121
122
123
124
125
126

class OperationSelector():
    """Class to produce new candidates by applying one of the 
    candidate generation operations which is supplied in the
    'operations'-list. The operations are drawn randomly according
    to the 'probabilities'-list.
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
    
    operations : "list" or "list of lists" of mutations/crossovers.

    probabilities : probability for each of the mutations/crossovers
    in operations. Must have the same dimensions as operations.
    """
    def __init__(self, probabilities, operations):
        cond1 = isinstance(operations[0], list)
        cond2 = isinstance(probabilities[0], list)
        if not cond1 and not cond2:
            operations = [operations]
            probabilities = [probabilities]
        element_count_operations = [len(op_list) for op_list in operations]
        element_count_probabilities = [len(prob_list)
                                       for prob_list in probabilities]
142
        assert element_count_operations == element_count_probabilities, 'the two lists must have the same shape'
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
        self.operations = operations
        self.rho = [np.cumsum(prob_list) for prob_list in probabilities]

    def __get_index__(self, rho):
        """Draw from the cumulative probalility distribution, rho,
        to return the index of which operation to use"""
        v = np.random.random() * rho[-1]
        for i in range(len(rho)):
            if rho[i] > v:
                return i
        
    def get_new_candidate(self, parents):
        """Generate new candidate by applying a randomly drawn
        operation on the structures. This is done successively for
        each list of operations, if multiple are present.
        """
        for op_list, rho_list in zip(self.operations, self.rho):
160
161
162
163
164
165
166
167
168
            for i_trial in range(5): # Do five trials
                to_use = self.__get_index__(rho_list)
                anew = op_list[to_use].get_new_candidate(parents)
                if anew is not None:
                    parents[0] = anew
                    break
            else:
                anew = parents[0]
                anew = op_list[to_use].finalize(anew, successfull=False)
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
169
170
        return anew

171
172
173
174
175
176
177
178
    def train(self, data):
        """ Method to train all trainable operations in
        self.operations.
        """
        for oplist in self.operations:
            for operation in oplist:
                operation.train(data)

179
180
181
182
183
184
185
186
187
188
189
190

def random_pos(box):
    """ Returns a random position within the box
         described by the input box. """
    p0 = box[0].astype(float)
    vspan = box[1]
    r = np.random.random((1, len(vspan)))
    pos = p0.copy()
    for i in range(len(vspan)):
        pos += vspan[i] * r[0, i]
    return pos

191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
class OperationConstraint():
    """ Class used to enforce constraints on the positions of
    atoms in mutation and crossover operations.

    Parameters:

    box: Box in which atoms are allowed to be placed. It should
    have the form [] [p0, vspan] where 'p0' is the position of
    the box corner and 'vspan' is a matrix containing the three
    spanning vectors.

    xlim: On the form [xmin, xmax], specifying, in the x-direction, 
    the lower and upper limit of the region atoms can be moved 
    within.

    ylim, zlim: See xlim.
    """
    def __init__(self, box=None, xlim=None, ylim=None, zlim=None):
        self.box = box
        self.xlim = xlim
        self.ylim = ylim
        self.zlim = zlim

    def check_if_valid(self, positions):
        """ Returns whether positions are valid under the 
        constraints or not.
        """
218
219
220
221
222
        if np.ndim(positions) == 1:
            pos = positions.reshape(-1,3)
        else:
            pos = positions
        
223
224
        if self.box is not None:
            pass
225
226
227
        if self.xlim is not None:
            if (np.any(pos[:,0] < self.xlim[0]) or 
                    np.any(pos[:,0] > self.xlim[1])):
228
                return False
229
230
231
        if self.ylim is not None:
            if (np.any(pos[:,1] < self.ylim[0]) or 
                    np.any(pos[:,1] > self.ylim[1])):
232
                return False
233
234
235
        if self.zlim is not None:
            if (np.any(pos[:,2] < self.zlim[0]) or 
                    np.any(pos[:,2] > self.zlim[1])):
236
237
                return False

238
239
240
        return True


241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
class StartGenerator(CandidateGenerator):
    """ Class used to generate random initial candidates.
    The candidates are generated by iteratively adding in
    one atom at a time within the box described.

    Parameters:

    slab: The atoms object describing the super cell to
    optimize within. Can be an empty cell or a cell 
    containing the atoms of a slab.

    stoichiometry: A list of atomic numbers for the atoms
    that are placed on top of the slab (if one is present).

    box_to_place_in: The box within which atoms are placed. The box
    should be on the form [p0, vspan] where 'p0' is the position of
    the box corner and 'vspan' is a matrix containing the three
    spanning vectors.

    blmin: The minimum allowed distance between atoms in units of
    the covalent distance between atoms, where d_cov=r_cov_i+r_cov_j.
    
    blmax: The maximum allowed distance, in units of the covalent 
    distance, from a single isolated atom to the closest atom. If
    blmax=None, no constraint is enforced on isolated atoms.

    cluster: If True atoms are required to be placed within
    blmin*d_cov of one of the other atoms to be placed. If
    False the atoms in the slab are also included.
    """
    def __init__(self, slab, stoichiometry, box_to_place_in,
272
273
274
                 cluster=False, description='StartGenerator',
                 *args, **kwargs):
        CandidateGenerator.__init__(self, *args, **kwargs)
275
276
277
278
279
280
        self.slab = slab
        self.stoichiometry = stoichiometry
        self.box = box_to_place_in
        self.cluster = cluster
        self.description = description

281
    def operation(self, parents=None):
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
        a = self.make_structure()
        return a

    def make_structure(self):
        """ Generates a new random structure """
        Nslab = len(self.slab)
        Ntop = len(self.stoichiometry)
        num = np.random.permutation(self.stoichiometry)

        for i_trials in range(1000):
            a = self.slab.copy()
            for i in range(Ntop):
                pos_found = False
                for _ in range(300):
                    # Place new atom
                    posi = random_pos(self.box)
                    a += Atoms([num[i]], posi.reshape(1,3))

                    # Check if position of new atom is valid
                    not_too_close = self.check_valid_bondlengths(a, indices=[Nslab+i],
                                                          check_too_close=True,
                                                          check_isolated=False)
                    if len(a) == 1:  # The first atom
                        not_isolated = True
                    else:
                        if self.cluster:  # Check isolation excluding slab atoms.
                            not_isolated = self.check_valid_bondlengths(a[Nslab:], indices=[Nslab+i],
                                                                        check_too_close=False,
                                                                        check_isolated=True)
                        else:  # All atoms.
                            not_isolated = self.check_valid_bondlengths(a, indices=[Nslab+i],
                                                                        check_too_close=False,
                                                                        check_isolated=True)
                    valid_bondlengths = not_too_close and not_isolated
                    if not valid_bondlengths:
                        del a[-1]
                    else:
                        pos_found = True
                        break
                if not pos_found:
                    break
            if pos_found:
                break
        if i_trials == 999 and not pos_found:
            raise RuntimeError('StartGenerator: No valid structure was produced in 1000 trials.')
        else:
            return a
                
    
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
331
332
333
334
335
if __name__ == '__main__':
    from ase.io import read
    from ase.visualize import view

    from candidate_operations.basic_mutations import RattleMutation, RattleMutation2, PermutationMutation
336
337

    print(0.7*2*covalent_radii[1], 1.3*2*covalent_radii[1])
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
338
    
339
    np.random.seed(7)
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
340
341
342
343
    
    #a = read('/home/mkb/DFT/gpLEA/Si3x3/ref/gm_unrelaxed_done.traj', index='0')
    #a = read('si3x3.traj', index='0')
    #a = read('c6h6.traj', index='0')
344
345
346
347
348
349
350
351
352
353
    traj = read('c6h6_init.traj', index=':')
    #a = read('sn2o3.traj', index='0')
    #slab = read('slab_sn2o3.traj', index='0')

    """
    stoichiometry = 6*[50] + 10*[8]
    c = slab.get_cell()
    c[2,2] = 3.3
    p0 = np.array([0,0,14])
    box = [p0, c]
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
354
    
355
356
357
358
359
    sg = StartGenerator(slab, stoichiometry, box)
    """

    a = traj[0]
    rattle = RattleMutation(n_top=len(a), Nrattle=3, rattle_range=2)
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
360
361
362
    rattle2 = RattleMutation2(n_top=16, Nrattle=0.1)
    permut = PermutationMutation(n_top=16, Npermute=2)

363
364
    candidategenerator = OperationSelector([1], [rattle])
    #candidategenerator = CandidateGenerator([0., 1., 0.], [rattle, rattle2, permut])
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
365
366
    #candidategenerator = CandidateGenerator([[1],[1]], [[rattle2], [permut]])

367
368
369
370
371
372
373
    """
    for a in traj:
        vb = rattle.check_valid_bondlengths(a)
        print(vb)
    """

    traj_rattle = []
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
374
    for i in range(100):
375
376
377
378
379
        for j, a in enumerate(traj[13:14]):
            print('i =', i, 'j =', j)
            a0 = a.copy()
            anew = candidategenerator.get_new_candidate([a0,a0])
            traj_rattle += [a0, anew]
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
380

381
    view(traj_rattle)
Malthe Kjær Bisbo's avatar
update  
Malthe Kjær Bisbo committed
382
383
384
385
    """
    a_mut = rattle.get_new_candidate([a])
    view([a,a_mut])
    """