Add callback function for LAMMPS

[samwaseda]

callback is one of the features of LAMMPS which makes it extremely efficient while being still very interactive. In short, you can append a function that returns forces which are then added to the total forces calculated in LAMMPS. You can find more info on this page.

Example I: Add random forces

def random_forces(x, *args):
    return 0.1 * np.random.randn(*x.shape)

lmp = pr.create_job('Lammps', 'random_forces')
lmp.structure = pr.create.structure.bulk('Ni', cubic=True).repeat(3)
lmp.server.run_mode.interactive = True
lmp.set_callback(random_forces)
lmp.calc_md()
lmp.run()
lmp.interactive_close()

Example II: Push H in (1, 1, 1)-direction

class CallBack:
    def __init__(self, job):
        self.job = job

    def push_hydrogen(self, x, *args):
        f = np.zeros_like(self.job.structure.positions)
        f[self.job.structure.select_index('H')[0]] = np.array(3*[0.5])
        return f

lmp = pr.create_job('Lammps', 'push_hydrogen')
lmp.structure = pr.create.structure.bulk('Ni', cubic=True).repeat(3)
lmp.structure += pr.create.structure.atoms(
    elements=['H'], positions=[lmp.structure.analyse.get_voronoi_vertices()[0]]
)
callback = CallBack(lmp)
lmp.server.run_mode.interactive = True
lmp.set_callback(callback.push_hydrogen)
lmp.calc_md()
lmp.run()
lmp.interactive_close()

[coveralls]

Pull Request Test Coverage Report for Build 1546466930

• 15 of 18 (83.33%) changed or added relevant lines in 1 file are covered.
• 24 unchanged lines in 1 file lost coverage.
• Overall coverage increased (+0.08%) to 69.923%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
pyiron_atomistics/lammps/interactive.py	15	18	83.33%

Files with Coverage Reduction	New Missed Lines	%
pyiron_atomistics/atomistics/structure/factory.py	24	76.74%

Totals
Change from base Build 1531836487:	0.08%
Covered Lines:	11503
Relevant Lines:	16451

---

💛 - Coveralls

[liamhuber]

Super cool, I think this is absolutely very useful. One small syntax tweak requested in the code itself, and a whole bunch of (sometimes optional) documentation request.

Looking at it from 30k ft, the big issue I see is how to connect this to our serialization and reproducibility. As far as I can see, right now if you activate this functionality the final serialized job will give no indication of what was done, and so you'll just strangely get a trajectory whose forces and evolution don't align with your potential+thermo/barostatting. I guess this is already the case for jobs we mess with interactively. Since you explicitly list this as an "expert feature", I don't think this concern should stop this PR from merging, and I don't have a good solution myself right now anyhow -- but we should think about how to handle this...maybe some sort of my_callback = pr.atomistics.lammps.Callback(fnc=my_function_defined_elsewhere); lmp.set_callback(my_callback.fnc) (or lmp.set_callback(my_callback), where we check if the argument is an instance of our fancy new class and parse accordingly), and then handle serialization of the raw function code under the hood?

[liamhuber]

add example 1 to docstring

[samwaseda]

You mean example 2? (1 is already there)