jonaslb / psiesta

Python bindings for "Siesta as a subroutine" DFT

PSiesta: Siesta as a Python library

This repository contains Python bindings for the Siesta as a subroutine functionality in the Siesta density functional theory software. PSiesta does not simply execute the Siesta executable, it has Siesta built-in as a library. That means you can use it similarly to how you would GPAW, ie. with MPI in the Python code. You must still specify your Siesta-options in an fdf-file (it would be nice if we had a Python abstraction instead, but we don't have such a thing yet). There is also an ASE interface.

Example

The following myscript.py could be executed with mpirun -n 8 python3 -m mpi4py myscript.py:

from mpi4py import MPI
import sisl as si
from ase.optimize import QuasiNewton
from psiesta import FilePSiesta
from psiesta.ase import AseFilePSiesta

# mpi rank
rank = MPI.COMM_WORLD.Get_rank()

# Create the geometry
geom = si.geom.graphene()

# Make the calculator object.
# Arguments:       main_fdf,      working_dir,       label,      geometry, comm=MPI.COMM_WORLD
calc = FilePSiesta("options.fdf", "working/siesta/", "graphene", geometry=geom)

# Run siesta_forces for the given geometry.
energy, forces, stress = calc.run(geom)
if rank == 0:
    H = calc.read_hamiltonian()
    print(f"Energy: {energy}.")
    print(f"Fermi energy: {calc.get_fermi_energy()}")

# Mutate the geometry and run again
geom.xyz[0, 0] += 0.05
new_results = calc.run(geom)  # new_results is a namedtuple

# Write dH for later use
if rank == 0:
    H2 = calc.read_hamiltonian()
    dH = H2-H
    dH.write("my-deltaH.nc")

# Other example: Optimize with ASE:
atoms = geom.toASE()
atoms.set_calculator(AseFilePSiesta(
    # main_fdf, working_dir, label, geometry=None, comm=MPI.COMM_WORLD, atoms_converter=si.Geometry.fromASE
    "options.fdf", "workingdir/siesta/", "aseopt", geometry=geom
))
atoms.rattle(stdev=0.03)
opt = QuasiNewton(atoms, traj="optimize.traj")
opt.run(fmax=0.01)

Siesta will run inside the Python processes. Relevant properties, other than those returned directly, can be read from the output files in the calculation directory in-between runs. It is recommended to use sisl for this. There are some shortcuts for the hamiltonian (as shown above), as well as density matrices and fermi energy.

You should use a Siesta version later than the git master as of 2020-06-10, as a few fixes regarding library operation were merged at this point.

Obtaining source, building and installing

You can obtain the source by simply cloning this repository. To build, you must have a properly set up arch.make for Siesta in your Obj-dir, and you must have at least compiled Siesta there (see the note above for a patched Siesta). You can then run OBJ=/my/custom/siesta/Obj/ python3 setup.py install [--user] [--prefix=<prefix>] (or use build instead of install) to build PSiesta. The setup.py-file makes use of Siesta's own Makefile (which includes your arch.make) in combination with --dry-run to extract the compilation and link arguments. It should work for both intel and gnu compilers, but be aware that LTO can complicate things, and ensure that any external libraries that are used (eg. flook) are compiled with -fPIC.

On some platforms it is necessary to link more libraries than Siesta is otherwise compiled with. It is currently a little unclear why, but in one case I needed to use EXTRA_COMP_ARGS="-lmkl_avx512 -lmkl_def" (which the setup.py-file will recognize).

As noted above, you should use a Siesta version later than the git master as of 2020-06-10.

Behaviour

See also the SiestaSubroutine readme. In summary, the fsiesta module that this is based on copies all fdf and psf-files from your working directory <cwd> into <cwd>/<systemlabel> where <systemlabel> is the label you give. In that folder it will then start reading from <systemlabel>.fdf. Ensuring that this file exists is handled by the Python wrapper. It will also prepend some settings to your fdf-file: Notably MD.TypeOfRun forces is enforced to make Siesta accept given coordinates. It also sets the system label and configures the geometry.

Some limitations:

• Only a few properties can currently be fetched directly via the bindings. Other properties must be obtained via the output-files. Feel free to create an issue if you'd like something in particular built-in, or send a PR if you've implemented it already.
• You don't get an exception when eg. the fdf-file contains an error. Instead, the whole process dies. This is because on error, Siesta calls abort() to "helpfully" crash and spit out a stacktrace. TODO: Can we catch sigabrt and raise a Python exception with the stacktrace instead?