gofee.py 18.3 KB
Newer Older
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
1
import numpy as np
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
2
3
import pickle
from os.path import isfile
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
4

5
6
7
8
9
10
from surrogate.gpr import GPR
from population import population

from ase import Atoms
from ase.io import read, write, Trajectory
from ase.calculators.singlepoint import SinglePointCalculator
11
from ase.calculators.dftb import Dftb
12
13

from parallel_utils import split, parallel_function_eval
14

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
15
16
from bfgslinesearch_constrained import relax

17
from candidate_operations.candidate_generation import CandidateGenerator
18
19
from candidate_operations.basic_mutations import RattleMutation

20
21
22
23
24
25
from mpi4py import MPI
world = MPI.COMM_WORLD

import traceback
import sys

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
26
27
from time import time

28
class GOFEE():
29
30
    """GOFEE global structure search method.
        
31
32
33
34
35
36
37
38
    structures: Atoms-object, list of Atoms-objects or None.
    In initial structures from which to start the sesarch.
    If None, the startgenerator must be supplied.
    If less than Ninit structures is supplied, the remaining
    ones are generated using the startgenerator or by rattling
    the supplied structures, depending on wether the
    startgenerator is supplied.

39
40
41
42
    calc: ASE calculator specifying the energy-expression
    with respect to which the atomic coordinates are
    globally optimized.

43
44
45
    gpr: The Gaussian Process Regression model used as the
    surrogate model for the Potential energy surface.
    
46
47
48
49
    startgenerator: Object used to generate initial random
    structures. Must be supplied if structures if structues=None.
    (This is the recommended way to initialize the search.)

50
51
    candidate_generator: OperationSelector object
    Object used to generate new candidates.
52

53
54
    trajectory: Trajectory object or str
    Name of trajectory to which all structures,
55
56
    evaluated during the search, is saved.

57
58
    kappa: float
    How much to weigh predicted uncertainty in acquisition
59
60
    function.

61
62
    max_steps: int
    Number of search steps.
63

64
65
    Ninit: int
    Number of initial structures. If len(structures) <
66
67
68
69
    Ninit, the remaining structures are generated using the
    startgenerator (if supplied) or by rattling the supplied
    'structures'.

70
    max_relax_dist: float
71
    Max distance that an atom is allowed to move during
72
73
    surrogate relaxation (in units of covalent distance).

74
75
    Ncandidates: int
    Number of new cancidate structures generated and
76
    surrogate-relaxed in each search iteration.
77

78
79
    population_size: int
    Size of population.
80

81
82
83
84
85
86
87
88
89
90
    dualpoint: boolean
    Whether to use dualpoint evaluation or not.

    min_certainty: float
    Max predicted uncertainty allowed for structures to be
    considdered for evaluation. (in units of the maximum possible
    uncertainty.)

    restart: str
    Filename for restart file.  Default value is *None*.
91
    """
92
93
94
95
96
    def __init__(self, structures=None,
                 calc=None,
                 gpr=None,
                 startgenerator=None,
                 candidate_generator=None,
97
                 trajectory='structures.traj',
98
                 kappa=2,
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
99
                 max_steps=200,
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
100
                 Ninit=10,
101
                 max_relax_dist=3.5,
102
103
104
                 Ncandidates=30,
                 population_size=5,
                 dualpoint=True,
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
105
                 min_certainty=0.7,
106
                 restart=None):
107
        
108
109
110
111
        if structures is None:
            assert startgenerator is not None
            self.structures = None
        else:
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
            if isinstance(structures, Atoms):
                self.structures = [structures]
            elif isinstance(structures, list):
                assert isinstance(structures[0], Atoms)
                self.structures = structures
            elif isinstance(structures, str):
                self.structures = read(structures, index=':')
        
        if calc is not None:
            self.calc = calc
        else:
            assert structures is not None
            calc = structures[0].get_calculator()
            assert calc is not None and not isinstance(calc, SinglePointCalculator)
            print('Using calculator from supplied structure(s)')
            self.calc = calc

129
        if startgenerator is None:
130
131
            assert structures is not None
            self.startgenerator = None
132
133
        else:
            self.startgenerator = startgenerator
134

135
136
137
138
139
140
141
142
143
        if startgenerator is not None:
            self.n_to_optimize = len(self.startgenerator.stoichiometry)
        else:
            self.n_to_optimize = len(self.structures[0])
            for constraint in self.structures[0].constraints:
                if isinstance(constraint, FixAtoms):
                    indices_fixed = constraint.get_indices()
                    self.n_to_optimize -= len(indices_fixed)
                    break
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
144
        
145
146
147
        if candidate_generator is not None:
            self.candidate_generator = candidate_generator
        else:
148
149
150
            rattle = RattleMutation(self.n_to_optimize,
                                    Nrattle=3,
                                    rattle_range=4)
151
            self.candidate_generator = CandidateGenerator([1.0],[rattle])
152
153
154
155
156
157

        # Define parallel communication
        self.comm = world.Dup()  # Important to avoid mpi-problems from call to ase.parallel in BFGS
        self.master = self.comm.rank == 0

        self.kappa = kappa
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
158
        self.max_steps = max_steps
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
159
        self.Ninit = Ninit
160
        self.max_relax_dist = max_relax_dist
161
162
163
        self.Ncandidates = Ncandidates
        self.dualpoint = dualpoint
        self.min_certainty = min_certainty
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
164
165
166
167
168
169
170
171
172
173
        self.restart = restart
        
        if isinstance(trajectory, str):
            self.trajectory = Trajectory(filename=trajectory, mode='a', master=self.master)
            if self.restart:
                self.traj_name = trajectory
        elif isinstance(trajectory, Trajectory):
            self.trajectory = trajectory
        else:
            assert trajectory is None
174

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
175
176
        if restart is None or not isfile(restart):
            self.initialize()
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
177
178
179
180
181
182
183
184

            if gpr is not None:
                self.gpr = gpr
            else:
                self.gpr = GPR(template_structure=self.structures[0])
            
            # Initialize population
            self.population = population(population_size=population_size, gpr=self.gpr, similarity2equal=0.9999)
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
185
186
187
188
189
        else:
            self.read()
            self.comm.barrier()

    def initialize(self):
190
        self.get_initial_structures()
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
191
        self.steps = 0
192
193

    def get_initial_structures(self):
194
195
196
197
198
199
200
201
202
        """Method to prepare the initial structures for the search.
        
        The method makes sure that there are atleast self.Ninit
        initial structures.
        These structures are first of all the potentially supplied
        structures. If more structures are required, these are
        generated using self.startgenerator (if supplied), otherwise
        they are generated by heavily rattling the supplied structures.
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
203
        
204
        # Collect potentially supplied structures.
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
205
        if self.structures is not None:
206
            for a in self.structures:
207
                a.info = {'origin': 'PreSupplied'}
208
209
        else:
            self.structures = []
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
210
        
211
        Nremaining = self.Ninit - len(self.structures)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
212
        
213
214
215
216
217
218
219
220
221
222
223
224
        if Nremaining > 0 and self.startgenerator is None:
            # Initialize rattle-mutation for all atoms.
            rattle = RattleMutation(self.n_to_optimize,
                                    Nrattle=self.n_to_optimize,
                                    rattle_range=2)

        # Generation of remaining initial-structures (up to self.Ninit).
        for i in range(Nremaining):
            if self.startgenerator is not None:
                a = self.startgenerator.get_new_candidate()
            else:
                # Perform two times rattle of all atoms.
225
                a0 = self.structures[i % len(self.structures)]
226
227
                a = rattle.get_new_candidate([a])
                a = rattle.get_new_candidate([a])
228
229
230
            self.structures.append(a)
                
    def evaluate_initial_structures(self):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
231
232
233
        """ Evaluate energies and forces of all initial structures
        (self.structures) that have not yet been evaluated.
        """
234
235
236
237
238
239
240
241
        structures_init = []
        for a in self.structures:
            calc = a.get_calculator()
            if isinstance(calc, SinglePointCalculator):
                if 'energy' in calc.results and 'forces' in calc.results:
                    # Write without evaluating.
                    structures_init.append(a)
                    continue
242
243
            a = self.evaluate(a)
            structures_init.append(a)
244

245
246
        for a in structures_init:
            self.write(a)
247
248
        self.gpr.memory.save_data(structures_init)
        self.population.add(structures_init)
249

250
    def run(self):
251
252
        """ Method to run the search.
        """
253
254
        if self.steps == 0:
            self.evaluate_initial_structures()
255

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
256
        while self.steps < self.max_steps:
257
            self.print_master('\n ### steps: {} ###\n'.format(self.steps))
258
259
260
261
262
            self.train_surrogate()
            self.update_population()
            relaxed_candidates = self.get_surrogate_relaxed_candidates()

            kappa = self.kappa
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
            a_add = []
            for _ in range(5):
                try:
                    anew = self.select_with_acquisition(relaxed_candidates, kappa)
                    anew = self.evaluate(anew)
                    a_add.append(anew)
                    if self.dualpoint:
                        adp = self.get_dualpoint(anew)
                        adp = self.evaluate(adp)
                        a_add.append(adp)
                    break
                except Exception as err:
                    kappa /=2
                    if self.master:
                        traceback.print_exc(file=sys.stderr)
            self.gpr.memory.save_data(a_add)
279
280

            # Add structure to population
281
282
            index_lowest = np.argmin([a.get_potential_energy() for a in a_add])
            self.population.add([a_add[index_lowest]])
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
283
284
285
286

            # Save search state
            self.dump((self.steps, self.population, np.random.get_state()))
            
287
288
            if self.master:
                print('anew pred:', anew.info['key_value_pairs']['Epred'], anew.info['key_value_pairs']['Epred_std'])
289
                print('E_true:', [a.get_potential_energy() for a in a_add])
290
                print('pop:', [a.get_potential_energy() for a in self.population.pop])
291
292

            self.steps += 1
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
            
    def get_dualpoint(self, a, lmax=0.10, Fmax_flat=5):
        """Returns dual-point structure, i.e. the original structure
        perturbed slightly along the forces.
        
        lmax: The atom with the largest force will be displaced by
        this distance
        
        Fmax_flat: maximum atomic displacement. is increased linearely
        with force until Fmax = Fmax_flat, after which it remains
        constant as lmax.
        """
        F = a.get_forces()
        a_dp = a.copy()

        # Calculate and set new positions
        Fmax = np.sqrt((F**2).sum(axis=1).max())
        pos_displace = lmax * F*min(1/Fmax_flat, 1/Fmax)
        pos_dp = a.positions + pos_displace
        a_dp.set_positions(pos_dp)
        return a_dp

    def print_master(self, *args):
        self.comm.barrier()
        if self.master:
            print(*args, flush=True)

    def get_surrogate_relaxed_candidates(self):
321
322
323
324
325
        """ Method supplying a number of surrogate-relaxed new
        candidates. The method combines the generation of new
        candidates with subsequent surrogate relaxation.
        The tasks are parrlelized over all avaliable cores.
        """
326
327
328
329
330
331
332
333
334
335
        Njobs = self.Ncandidates
        task_split = split(Njobs, self.comm.size)
        def func1():
            return [self.generate_candidate() for i in task_split[self.comm.rank]]
        candidates = parallel_function_eval(self.comm, func1)
        def func2():
            return [self.surrogate_relaxation(candidates[i], Fmax=0.1, steps=200, kappa=self.kappa)
                    for i in task_split[self.comm.rank]]
        relaxed_candidates = parallel_function_eval(self.comm, func2)
        relaxed_candidates = self.certainty_filter(relaxed_candidates)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
336
        relaxed_candidates = self.population.pop_MLrelaxed + relaxed_candidates
337
338
339
340
341
342
343
        
        if self.master:
            Epred = np.array([a.info['key_value_pairs']['Epred'] for a in relaxed_candidates])
            Epred_std = np.array([a.info['key_value_pairs']['Epred_std'] for a in relaxed_candidates])
            fitness = Epred - self.kappa*Epred_std
            print(np.c_[Epred, Epred_std, fitness])
        return relaxed_candidates
344
345
346
347
348
349
350
351
352
353
354
355
356

    def generate_candidate(self):
        """ Method to generate new candidate.
        """
        parents = self.population.get_structure_pair()
        a_mutated = self.candidate_generator.get_new_candidate(parents)
        return a_mutated

    def surrogate_relaxation(self, a, Fmax=0.1, steps=200, kappa=None):
        """ Method to carry out relaxations of new candidates in the
        surrogate potential.
        """
        calc = self.gpr.get_calculator(kappa)
357
        a_relaxed = relax(a, calc, max_relax_dist=self.max_relax_dist, Fmax=Fmax, steps_max=steps)
358
359
360
361
362
363
364
365
366
        # Evaluate uncertainty
        E, Estd = self.gpr.predict_energy(a_relaxed, eval_std=True)

        # Save prediction in info-dict
        a_relaxed.info['key_value_pairs']['Epred'] = E
        a_relaxed.info['key_value_pairs']['Epred_std'] = Estd
        a_relaxed.info['key_value_pairs']['kappa'] = self.kappa
        
        return a_relaxed
367
368
        
    def certainty_filter(self, structures):
369
370
371
372
        """ Method to filter away the most uncertain surrogate-relaxed
        candidates, which might otherewise get picked for first-principles
        evaluation, based on the very high uncertainty alone.
        """
373
374
375
376
377
378
379
380
381
382
383
384
385
        certainty = np.array([a.info['key_value_pairs']['Epred_std']
                              for a in structures]) / np.sqrt(self.gpr.K0)
        min_certainty = self.min_certainty
        for _ in range(5):
            filt = certainty < min_certainty
            if np.sum(filt.astype(int)) > 0:
                structures = [structures[i] for i in range(len(filt)) if filt[i]]
                break
            else:
                min_certainty = min_certainty + (1-min_certainty)/2
        return structures

    def update_population(self):
386
387
388
        """ Method to update the population with the new pirst-principles
        evaluated structures.
        """
389
390
391
392
393
394
395
396
397
398
399
        Njobs = len(self.population.pop)
        task_split = split(Njobs, self.comm.size)
        func = lambda: [self.surrogate_relaxation(self.population.pop[i],
                                                  Fmax=0.001, steps=200, kappa=None)
                        for i in task_split[self.comm.rank]]
        self.population.pop_MLrelaxed = parallel_function_eval(self.comm, func)
        if self.master:
            print('ML-relaxed pop forces:\n',
                  [(a.get_forces()**2).sum(axis=1).max()**0.5 for a in self.population.pop_MLrelaxed])

    def train_surrogate(self):
400
401
402
403
404
405
        """ Method to train the surrogate model.
        The method only performs hyperparameter optimization every 
        ten training instance, as carrying out the hyperparameter
        optimization is significantly more expensive than the basic
        training.
        """
406
        # Train
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
407
        if self.steps < 50 or (self.steps % 10) == 0:
408
409
410
411
412
413
414
415
            self.gpr.optimize_hyperparameters(comm=self.comm)
        else:
            self.gpr.train()
        if self.master:
            print('kernel:', list(np.exp(self.gpr.kernel.theta)))
            print('lml:', self.gpr.lml)

    def select_with_acquisition(self, structures, kappa):
416
417
418
419
        """ Method to select single most "promizing" candidate 
        for first-principles evaluation according to the acquisition
        function min(E-kappa*std(E)).
        """
420
421
422
423
424
425
426
427
428
        Epred = np.array([a.info['key_value_pairs']['Epred']
                          for a in structures])
        Epred_std = np.array([a.info['key_value_pairs']['Epred_std']
                              for a in structures])
        acquisition = Epred - kappa*Epred_std
        index_select = np.argmin(acquisition)
        return structures[index_select]

    def evaluate(self, a):
429
430
431
        """ Method to evaluate the energy and forces of the selacted
        candidate.
        """
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
432
        a = self.comm.bcast(a, root=0)
433
434
435
        a.wrap()

        if isinstance(self.calc, Dftb):
436
            if self.master:
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
                try:
                    a.set_calculator(self.calc)
                    E = a.get_potential_energy()
                    F = a.get_forces()
                    results = {'energy': E, 'forces': F}
                    calc_sp = SinglePointCalculator(a, **results)
                    a.set_calculator(calc_sp)
                    success = True
                except:
                    print('dftb failed on master', flush=True)
                    success = False
            else:
                success = None
            success = self.comm.bcast(success, root=0)
            if success == False:
                write('fail.traj', a)
                print('Raising error rank', self.comm.rank, flush=True)
                raise RuntimeError('DFTB evaluation failed')
455
456
457
458
459
460
461
462
            a = self.comm.bcast(a, root=0)
        else:
            a.set_calculator(self.calc)
            E = a.get_potential_energy()
            F = a.get_forces()
            results = {'energy': E, 'forces': F}
            calc_sp = SinglePointCalculator(a, **results)
            a.set_calculator(calc_sp)
463
464
465
466
467
468

        self.write(a)

        return a

    def write(self, a):
469
470
        """ Method for writing new evaluated structures to file.
        """
471
472
473
        if self.trajectory is not None:
            self.trajectory.write(a)

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
474
    def dump(self, data):
475
476
477
        """ Method to save restart-file used if the search is
        restarted from some point in the search. 
        """
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
478
479
        if self.comm.rank == 0 and self.restart is not None:
            pickle.dump(data, open(self.restart, "wb"), protocol=2)
480

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
481
    def read(self):
482
483
484
        """ Method to restart a search from the restart-file and the
        trajectory-file containing all structures evaluated so far.
        """
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
485
486
487
488
        self.steps, self.population, random_state = pickle.load(open(self.restart, "rb"))
        np.random.set_state(random_state)
        training_structures = read(self.traj_name, index=':')
        self.gpr.memory.save_data(training_structures)