"""ASE LAMMPS Calculator Library Version"""
import ctypes
import numpy as np
from numpy.linalg import norm
from ase.calculators.calculator import Calculator
from ase.data import (atomic_numbers as ase_atomic_numbers,
chemical_symbols as ase_chemical_symbols,
atomic_masses as ase_atomic_masses)
from ase.calculators.lammps import convert
from ase.geometry import wrap_positions
# TODO
# 1. should we make a new lammps object each time ?
# 4. need a routine to get the model back from lammps
# 5. if we send a command to lmps directly then the calculator does
# not know about it and the energy could be wrong.
# 6. do we need a subroutine generator that converts a lammps string
# into a python function that can be called
# 8. make matscipy as fallback
# 9. keep_alive not needed with no system changes
# 10. it may be a good idea to unify the cell handling with the one found in
# lammpsrun.py
# this one may be moved to some more generic place
def is_upper_triangular(arr, atol=1e-8):
"""test for upper triangular matrix based on numpy"""
# must be (n x n) matrix
assert len(arr.shape) == 2
assert arr.shape[0] == arr.shape[1]
return np.allclose(np.tril(arr, k=-1), 0., atol=atol) and \
np.all(np.diag(arr) >= 0.0)
def convert_cell(ase_cell):
"""
Convert a parallelepiped (forming right hand basis)
to lower triangular matrix LAMMPS can accept. This
function transposes cell matrix so the bases are column vectors
"""
cell = ase_cell.T
if not is_upper_triangular(cell):
# rotate bases into triangular matrix
tri_mat = np.zeros((3, 3))
A = cell[:, 0]
B = cell[:, 1]
C = cell[:, 2]
tri_mat[0, 0] = norm(A)
Ahat = A / norm(A)
AxBhat = np.cross(A, B) / norm(np.cross(A, B))
tri_mat[0, 1] = np.dot(B, Ahat)
tri_mat[1, 1] = norm(np.cross(Ahat, B))
tri_mat[0, 2] = np.dot(C, Ahat)
tri_mat[1, 2] = np.dot(C, np.cross(AxBhat, Ahat))
tri_mat[2, 2] = norm(np.dot(C, AxBhat))
# create and save the transformation for coordinates
volume = np.linalg.det(ase_cell)
trans = np.array([np.cross(B, C), np.cross(C, A), np.cross(A, B)])
trans /= volume
coord_transform = np.dot(tri_mat, trans)
return tri_mat, coord_transform
else:
return cell, None
[docs]class LAMMPSlib(Calculator):
r"""
**Introduction**
LAMMPSlib is an interface and calculator for LAMMPS_. LAMMPSlib uses
the python interface that comes with LAMMPS to solve an atoms model
for energy, atom forces and cell stress. This calculator creates a
'.lmp' object which is a running lammps program, so further commands
can be sent to this object executed until it is explicitly closed. Any
additional variables calculated by lammps can also be extracted. This
is still experimental code.
**Arguments**
======================= ======================================================
Keyword Description
======================= ======================================================
``lmpcmds`` list of strings of LAMMPS commands. You need to supply
enough to define the potential to be used e.g.
["pair_style eam/alloy",
"pair_coeff * * potentials/NiAlH_jea.eam.alloy Ni Al"]
``atom_types`` dictionary of ``atomic_symbol :lammps_atom_type``
pairs, e.g. ``{'Cu':1}`` to bind copper to lammps
atom type 1. If <None>, autocreated by assigning
lammps atom types in order that they appear in the
first used atoms object.
``atom_type_masses`` dictionary of ``atomic_symbol :mass`` pairs, e.g.
``{'Cu':63.546}`` to optionally assign masses that
override default ase.data.atomic_masses. Note that
since unit conversion is done automatically in this
module, these quantities must be given in the
standard ase mass units (g/mol)
``log_file`` string
path to the desired LAMMPS log file
``lammps_header`` string to use for lammps setup. Default is to use
metal units and simple atom simulation.
lammps_header=['units metal',
'atom_style atomic',
'atom_modify map array sort 0 0'])
``amendments`` extra list of strings of LAMMPS commands to be run
post initialization. (Use: Initialization amendments)
e.g.
["mass 1 58.6934"]
``post_changebox_cmds`` extra list of strings of LAMMPS commands to be run
after any LAMMPS 'change_box' command is performed by
the calculator. This is relevant because some
potentials either themselves depend on the geometry
and boundary conditions of the simulation box, or are
frequently coupled with other LAMMPS commands that
do, e.g. the 'buck/coul/long' pair style is often
used with the kspace_* commands, which are sensitive
to the periodicity of the simulation box.
``keep_alive`` Boolean
whether to keep the lammps routine alive for more
commands
======================= ======================================================
**Requirements**
To run this calculator you must have LAMMPS installed and compiled to
enable the python interface. See the LAMMPS manual.
If the following code runs then lammps is installed correctly.
>>> from lammps import lammps
>>> lmp = lammps()
The version of LAMMPS is also important. LAMMPSlib is suitable for
versions after approximately 2011. Prior to this the python interface
is slightly different from that used by LAMMPSlib. It is not difficult
to change to the earlier format.
**LAMMPS and LAMMPSlib**
The LAMMPS calculator is another calculator that uses LAMMPS (the
program) to calculate the energy by generating input files and running
a separate LAMMPS job to perform the analysis. The output data is then
read back into python. LAMMPSlib makes direct use of the LAMMPS (the
program) python interface. As well as directly running any LAMMPS
command line it allows the values of any of LAMMPS variables to be
extracted and returned to python.
**Example**
Provided that the respective potential file is in the working directory, one
can simply run (note that LAMMPS needs to be compiled to work with EAM
potentials)
::
from ase import Atom, Atoms
from ase.build import bulk
from ase.calculators.lammpslib import LAMMPSlib
cmds = ["pair_style eam/alloy",
"pair_coeff * * NiAlH_jea.eam.alloy Ni H"]
Ni = bulk('Ni', cubic=True)
H = Atom('H', position=Ni.cell.diagonal()/2)
NiH = Ni + H
lammps = LAMMPSlib(lmpcmds=cmds, log_file='test.log')
NiH.calc = lammps
print("Energy ", NiH.get_potential_energy())
**Implementation**
LAMMPS provides a set of python functions to allow execution of the
underlying C++ LAMMPS code. The functions used by the LAMMPSlib
interface are::
from lammps import lammps
lmp = lammps(cmd_args) # initiate LAMMPS object with command line args
lmp.scatter_atoms('x',1,3,positions) # atom coords to LAMMPS C array
lmp.command(cmd) # executes a one line cmd string
lmp.extract_variable(...) # extracts a per atom variable
lmp.extract_global(...) # extracts a global variable
lmp.close() # close the lammps object
For a single Ni atom model the following lammps file commands would be run
by invoking the get_potential_energy() method::
units metal
atom_style atomic
atom_modify map array sort 0 0
region cell prism 0 xhi 0 yhi 0 zhi xy xz yz units box
create_box 1 cell
create_atoms 1 single 0 0 0 units box
mass * 1.0
## user lmpcmds get executed here
pair_style eam/alloy
pair_coeff * * NiAlH_jea.eam.alloy Ni
## end of user lmmpcmds
run 0
where xhi, yhi and zhi are the lattice vector lengths and xy,
xz and yz are the tilt of the lattice vectors, all to be edited.
**Notes**
.. _LAMMPS: http://lammps.sandia.gov/
* Units: The default lammps_header sets the units to Angstrom and eV
and for compatibility with ASE Stress is in GPa.
* The global energy is currently extracted from LAMMPS using
extract_variable since lammps.lammps currently extract_global only
accepts the following ['dt', 'boxxlo', 'boxxhi', 'boxylo', 'boxyhi',
'boxzlo', 'boxzhi', 'natoms', 'nlocal'].
* If an error occurs while lammps is in control it will crash
Python. Check the output of the log file to find the lammps error.
* If the are commands directly sent to the LAMMPS object this may
change the energy value of the model. However the calculator will not
know of it and still return the original energy value.
"""
implemented_properties = ['energy', 'forces', 'stress']
started = False
initialized = False
default_parameters = dict(
atom_types=None,
atom_type_masses=None,
log_file=None,
lammps_name='',
keep_alive=False,
lammps_header=['units metal',
'atom_style atomic',
'atom_modify map array sort 0 0'],
amendments=None,
post_changebox_cmds=None,
boundary=True,
create_box=True,
create_atoms=True,
read_molecular_info=False,
comm=None)
def __init__(self, *args, **kwargs):
Calculator.__init__(self, *args, **kwargs)
self.lmp = None
def __del__(self):
if self.started:
self.lmp.close()
self.started = False
self.lmp = None
def set_cell(self, atoms, change=False):
lammps_cell, self.coord_transform = convert_cell(atoms.get_cell())
xhi, xy, xz, _, yhi, yz, _, _, zhi = convert(
lammps_cell.flatten(order='C'), "distance", "ASE", self.units)
box_hi = [xhi, yhi, zhi]
if change:
cell_cmd = ('change_box all '
'x final 0 {} y final 0 {} z final 0 {} '
'xy final {} xz final {} yz final {} units box'
''.format(xhi, yhi, zhi, xy, xz, yz))
if self.parameters.post_changebox_cmds is not None:
for cmd in self.parameters.post_changebox_cmds:
self.lmp.command(cmd)
else:
# just in case we'll want to run with a funny shape box,
# and here command will only happen once, and before
# any calculation
if self.parameters.create_box:
self.lmp.command('box tilt large')
# Check if there are any indefinite boundaries. If so,
# shrink-wrapping will end up being used, but we want to
# define the LAMMPS region and box fairly tight around the
# atoms to avoid losing any
lammps_boundary_conditions = self.lammpsbc(atoms).split()
if 's' in lammps_boundary_conditions:
pos = atoms.get_positions()
if self.coord_transform is not None:
pos = np.dot(self.coord_transform, pos.transpose())
pos = pos.transpose()
posmin = np.amin(pos, axis=0)
posmax = np.amax(pos, axis=0)
for i in range(0, 3):
if lammps_boundary_conditions[i] == 's':
box_hi[i] = 1.05 * abs(posmax[i] - posmin[i])
cell_cmd = ('region cell prism '
'0 {} 0 {} 0 {} '
'{} {} {} units box'
''.format(*box_hi, xy, xz, yz))
self.lmp.command(cell_cmd)
def set_lammps_pos(self, atoms):
# Create local copy of positions that are wrapped along any periodic
# directions
cell = convert(atoms.cell, "distance", "ASE", self.units)
pos = convert(atoms.positions, "distance", "ASE", self.units)
# If necessary, transform the positions to new coordinate system
if self.coord_transform is not None:
pos = np.dot(pos, self.coord_transform.T)
cell = np.dot(cell, self.coord_transform.T)
# wrap only after scaling and rotating to reduce chances of
# lammps neighbor list bugs.
pos = wrap_positions(pos, cell, atoms.get_pbc())
# Convert ase position matrix to lammps-style position array
# contiguous in memory
lmp_positions = list(pos.ravel())
# Convert that lammps-style array into a C object
c_double_array = (ctypes.c_double * len(lmp_positions))
lmp_c_positions = c_double_array(*lmp_positions)
# self.lmp.put_coosrds(lmp_c_positions)
self.lmp.scatter_atoms('x', 1, 3, lmp_c_positions)
def calculate(self, atoms, properties, system_changes):
self.propagate(atoms, properties, system_changes, 0)
def propagate(self, atoms, properties, system_changes, n_steps, dt=None,
dt_not_real_time=False, velocity_field=None):
""""atoms: Atoms object
Contains positions, unit-cell, ...
properties: list of str
List of what needs to be calculated. Can be any combination
of 'energy', 'forces', 'stress', 'dipole', 'charges', 'magmom'
and 'magmoms'.
system_changes: list of str
List of what has changed since last calculation. Can be
any combination of these five: 'positions', 'numbers', 'cell',
'pbc', 'charges' and 'magmoms'.
"""
if len(system_changes) == 0:
return
self.coord_transform = None
if not self.started:
self.start_lammps()
if not self.initialized:
self.initialise_lammps(atoms)
else: # still need to reset cell
# NOTE: The whole point of ``post_changebox_cmds`` is that they're
# executed after any call to LAMMPS' change_box command. Here, we
# rely on the fact that self.set_cell(), where we have currently
# placed the execution of ``post_changebox_cmds``, gets called
# after this initial change_box call.
# Apply only requested boundary condition changes. Note this needs
# to happen before the call to set_cell since 'change_box' will
# apply any shrink-wrapping *after* it's updated the cell
# dimensions
if 'pbc' in system_changes:
change_box_str = 'change_box all boundary {}'
change_box_cmd = change_box_str.format(self.lammpsbc(atoms))
self.lmp.command(change_box_cmd)
# Reset positions so that if they are crazy from last
# propagation, change_box (in set_cell()) won't hang.
# Could do this only after testing for crazy positions?
# Could also use scatter_atoms() to set values (requires
# MPI comm), or extra_atoms() to get pointers to local
# data structures to zero, but then we would have to be
# careful with parallelism.
self.lmp.command("set atom * x 0.0 y 0.0 z 0.0")
self.set_cell(atoms, change=True)
if self.parameters.atom_types is None:
raise NameError("atom_types are mandatory.")
do_rebuild = (not np.array_equal(atoms.numbers,
self.previous_atoms_numbers)
or ("numbers" in system_changes))
if not do_rebuild:
do_redo_atom_types = not np.array_equal(
atoms.numbers, self.previous_atoms_numbers)
else:
do_redo_atom_types = False
self.lmp.command('echo none') # don't echo the atom positions
if do_rebuild:
self.rebuild(atoms)
elif do_redo_atom_types:
self.redo_atom_types(atoms)
self.lmp.command('echo log') # switch back log
self.set_lammps_pos(atoms)
if self.parameters.amendments is not None:
for cmd in self.parameters.amendments:
self.lmp.command(cmd)
if n_steps > 0:
if velocity_field is None:
vel = convert(
atoms.get_velocities(),
"velocity",
"ASE",
self.units)
else:
# FIXME: Do we need to worry about converting to lammps units
# here?
vel = atoms.arrays[velocity_field]
# If necessary, transform the velocities to new coordinate system
if self.coord_transform is not None:
vel = np.dot(self.coord_transform, vel.T).T
# Convert ase velocities matrix to lammps-style velocities array
lmp_velocities = list(vel.ravel())
# Convert that lammps-style array into a C object
c_double_array = (ctypes.c_double * len(lmp_velocities))
lmp_c_velocities = c_double_array(*lmp_velocities)
self.lmp.scatter_atoms('v', 1, 3, lmp_c_velocities)
# Run for 0 time to calculate
if dt is not None:
if dt_not_real_time:
self.lmp.command('timestep %.30f' % dt)
else:
self.lmp.command('timestep %.30f' %
convert(dt, "time", "ASE", self.units))
self.lmp.command('run %d' % n_steps)
if n_steps > 0:
# TODO this must be slower than native copy, but why is it broken?
pos = np.array(
[x for x in self.lmp.gather_atoms("x", 1, 3)]).reshape(-1, 3)
if self.coord_transform is not None:
pos = np.dot(pos, self.coord_transform)
# Convert from LAMMPS units to ASE units
pos = convert(pos, "distance", self.units, "ASE")
atoms.set_positions(pos)
vel = np.array(
[v for v in self.lmp.gather_atoms("v", 1, 3)]).reshape(-1, 3)
if self.coord_transform is not None:
vel = np.dot(vel, self.coord_transform)
if velocity_field is None:
atoms.set_velocities(convert(vel, 'velocity', self.units,
'ASE'))
# Extract the forces and energy
self.results['energy'] = convert(
self.lmp.extract_variable('pe', None, 0),
"energy", self.units, "ASE"
)
self.results['free_energy'] = self.results['energy']
stress = np.empty(6)
stress_vars = ['pxx', 'pyy', 'pzz', 'pyz', 'pxz', 'pxy']
for i, var in enumerate(stress_vars):
stress[i] = self.lmp.extract_variable(var, None, 0)
stress_mat = np.zeros((3, 3))
stress_mat[0, 0] = stress[0]
stress_mat[1, 1] = stress[1]
stress_mat[2, 2] = stress[2]
stress_mat[1, 2] = stress[3]
stress_mat[2, 1] = stress[3]
stress_mat[0, 2] = stress[4]
stress_mat[2, 0] = stress[4]
stress_mat[0, 1] = stress[5]
stress_mat[1, 0] = stress[5]
if self.coord_transform is not None:
stress_mat = np.dot(self.coord_transform.T,
np.dot(stress_mat, self.coord_transform))
stress[0] = stress_mat[0, 0]
stress[1] = stress_mat[1, 1]
stress[2] = stress_mat[2, 2]
stress[3] = stress_mat[1, 2]
stress[4] = stress_mat[0, 2]
stress[5] = stress_mat[0, 1]
self.results['stress'] = convert(-stress, "pressure", self.units, "ASE")
# definitely yields atom-id ordered force array
f = convert(np.array(self.lmp.gather_atoms("f", 1, 3)).reshape(-1, 3),
"force", self.units, "ASE")
if self.coord_transform is not None:
self.results['forces'] = np.dot(f, self.coord_transform)
else:
self.results['forces'] = f.copy()
# otherwise check_state will always trigger a new calculation
self.atoms = atoms.copy()
if not self.parameters.keep_alive:
self.lmp.close()
def lammpsbc(self, atoms):
"""Determine LAMMPS boundary types based on ASE pbc settings. For
non-periodic dimensions, if the cell length is finite then
fixed BCs ('f') are used; if the cell length is approximately
zero, shrink-wrapped BCs ('s') are used."""
retval = ''
pbc = atoms.get_pbc()
if np.all(pbc):
retval = 'p p p'
else:
cell = atoms.get_cell()
for i in range(0, 3):
if pbc[i]:
retval += 'p '
else:
# See if we're using indefinite ASE boundaries along this
# direction
if np.linalg.norm(cell[i]) < np.finfo(cell[i][0]).tiny:
retval += 's '
else:
retval += 'f '
return retval.strip()
def rebuild(self, atoms):
try:
n_diff = len(atoms.numbers) - len(self.previous_atoms_numbers)
except Exception: # XXX Which kind of exception?
n_diff = len(atoms.numbers)
if n_diff > 0:
if any([("reax/c" in cmd) for cmd in self.parameters.lmpcmds]):
self.lmp.command("pair_style lj/cut 2.5")
self.lmp.command("pair_coeff * * 1 1")
for cmd in self.parameters.lmpcmds:
if (("pair_style" in cmd) or ("pair_coeff" in cmd) or
("qeq/reax" in cmd)):
self.lmp.command(cmd)
cmd = "create_atoms 1 random {} 1 NULL".format(n_diff)
self.lmp.command(cmd)
elif n_diff < 0:
cmd = "group delatoms id {}:{}".format(
len(atoms.numbers) + 1, len(self.previous_atoms_numbers))
self.lmp.command(cmd)
cmd = "delete_atoms group delatoms"
self.lmp.command(cmd)
self.redo_atom_types(atoms)
def redo_atom_types(self, atoms):
current_types = set(
(i + 1, self.parameters.atom_types[sym]) for i, sym
in enumerate(atoms.get_chemical_symbols()))
try:
previous_types = set(
(i + 1, self.parameters.atom_types[ase_chemical_symbols[Z]])
for i, Z in enumerate(self.previous_atoms_numbers))
except Exception: # XXX which kind of exception?
previous_types = set()
for (i, i_type) in current_types - previous_types:
cmd = "set atom {} type {}".format(i, i_type)
self.lmp.command(cmd)
self.previous_atoms_numbers = atoms.numbers.copy()
def restart_lammps(self, atoms):
if self.started:
self.lmp.command("clear")
# hope there's no other state to be reset
self.started = False
self.initialized = False
self.previous_atoms_numbers = []
self.start_lammps()
self.initialise_lammps(atoms)
def start_lammps(self):
# Only import lammps when running a calculation
# so it is not required to use other parts of the
# module
from lammps import lammps
# start lammps process
if self.parameters.log_file is None:
cmd_args = ['-echo', 'log', '-log', 'none', '-screen', 'none',
'-nocite']
else:
cmd_args = ['-echo', 'log', '-log', self.parameters.log_file,
'-screen', 'none', '-nocite']
self.cmd_args = cmd_args
if self.lmp is None:
self.lmp = lammps(self.parameters.lammps_name, self.cmd_args,
comm=self.parameters.comm)
# Run header commands to set up lammps (units, etc.)
for cmd in self.parameters.lammps_header:
self.lmp.command(cmd)
for cmd in self.parameters.lammps_header:
if "units" in cmd:
self.units = cmd.split()[1]
if 'lammps_header_extra' in self.parameters:
if self.parameters.lammps_header_extra is not None:
for cmd in self.parameters.lammps_header_extra:
self.lmp.command(cmd)
self.started = True
def initialise_lammps(self, atoms):
# Initialising commands
if self.parameters.boundary:
# if the boundary command is in the supplied commands use that
# otherwise use atoms pbc
for cmd in self.parameters.lmpcmds:
if 'boundary' in cmd:
break
else:
self.lmp.command('boundary ' + self.lammpsbc(atoms))
# Initialize cell
self.set_cell(atoms, change=not self.parameters.create_box)
if self.parameters.atom_types is None:
# if None is given, create from atoms object in order of appearance
s = atoms.get_chemical_symbols()
_, idx = np.unique(s, return_index=True)
s_red = np.array(s)[np.sort(idx)].tolist()
self.parameters.atom_types = {j: i + 1 for i, j in enumerate(s_red)}
# Initialize box
if self.parameters.create_box:
# count number of known types
n_types = len(self.parameters.atom_types)
create_box_command = 'create_box {} cell'.format(n_types)
self.lmp.command(create_box_command)
# Initialize the atoms with their types
# positions do not matter here
if self.parameters.create_atoms:
self.lmp.command('echo none') # don't echo the atom positions
self.rebuild(atoms)
self.lmp.command('echo log') # turn back on
else:
self.previous_atoms_numbers = atoms.numbers.copy()
# execute the user commands
for cmd in self.parameters.lmpcmds:
self.lmp.command(cmd)
# Set masses after user commands, e.g. to override
# EAM-provided masses
for sym in self.parameters.atom_types:
if self.parameters.atom_type_masses is None:
mass = ase_atomic_masses[ase_atomic_numbers[sym]]
else:
mass = self.parameters.atom_type_masses[sym]
self.lmp.command('mass %d %.30f' % (
self.parameters.atom_types[sym],
convert(mass, "mass", "ASE", self.units)))
# Define force & energy variables for extraction
self.lmp.command('variable pxx equal pxx')
self.lmp.command('variable pyy equal pyy')
self.lmp.command('variable pzz equal pzz')
self.lmp.command('variable pxy equal pxy')
self.lmp.command('variable pxz equal pxz')
self.lmp.command('variable pyz equal pyz')
# I am not sure why we need this next line but LAMMPS will
# raise an error if it is not there. Perhaps it is needed to
# ensure the cell stresses are calculated
self.lmp.command('thermo_style custom pe pxx emol ecoul')
self.lmp.command('variable fx atom fx')
self.lmp.command('variable fy atom fy')
self.lmp.command('variable fz atom fz')
# do we need this if we extract from a global ?
self.lmp.command('variable pe equal pe')
self.lmp.command("neigh_modify delay 0 every 1 check yes")
self.initialized = True