Source code for qcelemental.covalent_radii

"""
Contains covalent radii
"""

import collections
from decimal import Decimal
from typing import Dict, Union

from .datum import Datum, print_variables
from .exceptions import DataUnavailableError
from .periodic_table import periodictable


[docs]class CovalentRadii: r"""Covalent radii sets. Parameters ---------- context : str {'ALVAREZ2008'} Origin of loaded data. Attributes ---------- cr : Dict[str, Datum] Each covalent radius is an entry in `cr`, where key is the "Fe"-cased element symbol if generic or symbol-prefixed label if specialized within element. The value is a Datum object with `lbl` the same as key, `units`, `data` value as Decimal object, and any uncertainty in the `comment` field. doi : str The DOI of the current context. name : str The name of the context ('ALVAREZ2008') native_units : str The units the original data was provided in. year : int The year the context was created. """ def __init__(self, context: str = "ALVAREZ2008"): self.cr: Dict[str, Datum] = collections.OrderedDict() from .data import alvarez_2008_covalent_radii if context == "ALVAREZ2008": self.doi = alvarez_2008_covalent_radii["doi"] self.native_units = alvarez_2008_covalent_radii["units"] # TypedDict wont be in until 3.8, have to ignore heterogeneous dicts for now for cr in alvarez_2008_covalent_radii["covalent_radii"]: # type: ignore self.cr[cr[0]] = Datum(cr[0], self.native_units, Decimal(cr[1]), comment=cr[2], doi=self.doi) else: raise KeyError(f"Context set as '{context}', " + "only contexts {'ALVAREZ2008', } are currently supported") self.name = context self.year = int(alvarez_2008_covalent_radii["date"][:4]) # type: ignore # Extra relationships aliases = [ ("C", "angstrom", self.cr["C_sp3"].data, "Largest (sp3) chosen for generic atom"), ("Mn", "angstrom", self.cr["Mn_highspin"].data, "Larger (high-spin) chosen for generic atom"), ("Fe", "angstrom", self.cr["Fe_highspin"].data, "Larger (high-spin) chosen for generic atom"), ("Co", "angstrom", self.cr["Co_highspin"].data, "Larger (high-spin) chosen for generic atom"), ] # yapf: disable # add alternate names to help QC programs for alias in aliases: ident, units, value, comment = alias self.cr[ident.capitalize()] = Datum(ident, units, value, comment=comment) def __str__(self) -> str: return "CovalentRadii(context='{}')".format(self.name)
[docs] def get( self, atom: Union[int, str], *, return_tuple: bool = False, units: str = "bohr", missing: float = None ) -> Union[float, "Datum"]: r""" Access a covalent radius for species `atom`. Parameters ---------- atom Identifier for element or nuclide, e.g., ``H``, ``D``, ``H2``, ``He``, ``hE4``. In general, one value recommended for each element; however, certain other exact labels may be available. ALVAREZ2008: C_sp3, C_sp2, C_sp, Mn_lowspin, Mn_highspin, Fe_lowspin, Fe_highspin, Co_lowspin, Co_highspin units Units of returned value. To return in native unit (ALVAREZ2008: angstrom), pass it explicitly. Only relevant for ``return_tuple=False`` since ``True`` returns underlying data structure with native units. missing How to handle when ``atom`` is valid but outside the available data range. When ``None``, raises DataUnavailableError. When a float, returns that float, so supply in ``units`` units. Supplying a float is a more compact assurance that a call will work over all the periodic table than the equivalent .. code-block:: python try: rad = qcel.covalentradii.get(atom) except qcel.DataUnavailableError: rad = 4.0 Only relevant for ``return_tuple=False``. return_tuple See below. Returns ------- float When ``return_tuple=False``, value of covalent radius. If multiple defined for element, returns largest. qcelemental.Datum When ``return_tuple=True``, Datum with units, description, uncertainty, and value of covalent radius as Decimal (preserving significant figures). If multiple defined for element, returns largest. Raises ------ NotAnElementError If `atom` cannot be resolved into an element or nuclide or label. DataUnavailableError If `atom` is a valid element or nuclide but not one for which a covalent radius is available and `missing=None`. """ if atom in self.cr.keys(): # catch extra labels like 'C_sp3' identifier = atom else: identifier = periodictable.to_E(atom) try: assert isinstance(identifier, str) # Should be string by now qca = self.cr[identifier] except KeyError as e: if missing is not None and return_tuple is False: return missing else: raise DataUnavailableError("covalent radius", identifier) from e if return_tuple: return qca else: return qca.to_units(units)
[docs] def string_representation(self) -> str: """Print name, value, and units of all covalent radii.""" return print_variables(self.cr)
[docs] def write_c_header(self, filename: str = "covrad.h", missing: float = 2.0) -> None: r"""Write C header file defining covalent radii array. Parameters ---------- filename File name for header. Note that changing this won't change the header guard. missing In order that the C array be atomic-number indexable and that it span the periodic table, this value is used anywhere data is missing. """ text = [ "#ifndef _qcelemental_covrad_h_", "#define _qcelemental_covrad_h_", "", "/* This file is autogenerated from the QCElemental python module */", "", "const double covalent_radii[] = {", ] for el in periodictable.E: try: qca = self.cr[el] text.append("{}, /*- [{}] {} {} -*/".format(qca.data, qca.units, qca.label, qca.comment)) except KeyError: text.append( "{:.2f}, /*- [{}] {} {} -*/".format( missing, self.native_units, el, "Default value for missing data" ) ) text.append("};") text.append("#endif /* header guard */") text.append("") with open(filename, "w") as handle: handle.write("\n".join(text)) print("File written ({}). Remember to add license and clang-format it.".format(filename))
# singleton covalentradii = CovalentRadii("ALVAREZ2008")