Ctrl+K

Fitting GAP

Fitting GAP#

wfl.fig.gap.simple.run_gap_fit() is a lightweight wrapper for calling gap_fit program from python. It makes use of Workflow’s atomic structure handling to fit in with the rest of potential fitting infrastructure and makes use of ExPyRe for submitting just this fitting function as a (remotely) queued cluster job.

The function needs a fitting_dict that gets converted in the gap_fit command line arguments. The dictionary gets processed so:

  • booleans -> “T”/”F”

  • lists -> { v1 v2 v3 … }

  • {‘key1’:[v1, v2, v3], ‘key2’:[v1, v2, v3]} -> ‘{key1:v1:v2:v3:key2:v1:v2:v3}’

  • strings with spaces get enclosed in quotes

  • otherwise {key:val} -> key=val

  • asserts that mandatory parameters are given

  • descriptors are passed in pram_dict[‘_gap’], which is a list of dictionaries, one dictionary per descriptor

  • atoms_filename and at_file mustn’t be present in the dictionary, the file with fitting configs gets created from the input ConfigSet.

  • The default executable is gap_fit, but it can also be given as an argument or as WFL_GAP_FIT_COMMAND environment variable.

Normally, OMP_NUM_THREADS is set to 1, so that only Workflow’s parallelization is used and not OpenMP. For fitting GAP, which $isn’t parallelizable over atomic structures like other operations are, OMP_NUM_THREADS should be set to turn on the OpenMP parallelization. The number of OpenMP threads to to use for fitting GAP, but nothing else, is controlled via WFL_GAP_FIT_OMP_NUM_THREADS environment variable.

Below is an example dictionary and the corresponding gap_fit string (gap_fit executable is added later).

gap_fit_dict = {'default_sigma': [0.01, 0.1, 0.1, 0.0],
                'sparse_seprate_file': False,
                'core_ip_args': 'IP Glue',
                'core_param_file': '/test/path/test/file.xml',
                'config_type_sigma': 'isolated_atom:1e-05:0.0:0.0:0.0:funky_configs:0.1:0.3:0.0:0.0',
                '_gap': [
                    {'soap': True, 'l_max': 6, 'n_max': '12',
                        'cutoff': 3, 'delta': 1,
                        'covariance_type': 'dot_product', 'zeta': 4,
                        'n_sparse': 200, 'sparse_method': 'cur_points'},
                    {'soap': True, 'l_max': 4, 'n_max': 12, 'cutoff': 6,
                        'delta': 1, 'covariance_type': 'dot_product',
                        'zeta': 4, 'n_sparse': 100,
                        'sparse_method': 'cur_points',
                        'atom_gaussian_width': 0.3, 'add_species': False,
                        'n_species': 3, 'Z': 8, 'species_Z': [8, 1, 6]}]}

gap_fit_string = 'default_sigma={0.01 0.1 0.1 0.0} sparse_seprate_file=F ' \
                'core_ip_args="IP Glue" core_param_file=/test/path/test/file.xml ' \
                'config_type_sigma=isolated_atom:1e-05:0.0:0.0:0.0:' \
                'funky_configs:0.1:0.3:0.0:0.0 ' \
                'gap={ soap=T l_max=6 n_max=12 cutoff=3 delta=1 ' \
                'covariance_type=dot_product zeta=4 n_sparse=200 ' \
                'sparse_method=cur_points : soap=T l_max=4 n_max=12 cutoff=6 delta=1' \
                ' covariance_type=dot_product zeta=4 n_sparse=100 ' \
                'sparse_method=cur_points atom_gaussian_width=0.3 add_species=F ' \
                'n_species=3 Z=8 species_Z={{8 1 6}} }'