Appa

Miscellaneous code I use in my research.

(Appa is a pet flying bison in the series Avatar: The Last Airbender. He speeds up the journey of the Avatar by flying him and his friends around.)

Icon by rufftoon on DeviantArt

Setting up LAMMPS simulations

The appa.lammps module sets up LAMMPS simulations with MACE (and could be extended for use with other interatomic potentials). Example:

from appa.lammps import AtomisticSimulation, ArrayJob
from ase.io import read

# Define atomic structure
atoms = read("path/to/myatoms.xyz")

# Setup a simulation
sim1 = AtomisticSimulation(atoms)
sim1.set_potential("path/to/my_potential.lammps.pt")
sim1.set_molecular_dynamics(temperature=330, timestep=0.0005)
sim1.set_run(n_steps=2000000)

# Setup another simulation
sim2 = AtomisticSimulation(atoms)
sim2.set_potential("path/to/another_potential.lammps.pt")
sim2.set_molecular_dynamics(temperature=330, timestep=0.0005)
sim2.set_run(n_steps=2000000)

sims = [sim1, sim2]

# Write input files to submit as a batch job
job = ArrayJob("results", sims)
job.write_inputs()
job.write_jobfile()

Then you can submit the two simulations (indices 0 to 1) as an array job:

sbatch jobfile.sh --array=0-1

Learning curves

The appa.learning_curves module contains tools to plot learning curves for machine learning interatomic potentials. In this context a learning curve plot shows the test error of the model against the training set size. For all data points the model should be fully trained. (These learning curves are not to be confused with test/train error vs. #epochs curves obtained from training one model). See also fig. 2e of this paper. Example:

from appa.learning_curves import LearningCurve 
import numpy as np

seeds = range(3)  # define seeds with which different models are trained
# (even better would be to subsample different training sets)
subsets = [10, 110, 610, 1610, 2879]  # define size of each training set

# load 'true' forces and energy on test set
dft_forces = np.load('data/test.f.npy')
dft_energy = np.load('data/test.e.npy')

mace_lc = LearningCurve()
for i, size in enumerate(subsets):
    mace_lc.add_training_set(
        n_training_samples=size,
        force_component_errors=[np.load(f"models/test-{i:d}-{j:d}.f.npy") - dft_forces for j in seeds],
        energy_per_atom_errors=[np.load(f"models/test-{i:d}-{j:d}.e.npy") - dft_energy for j in seeds],
    )

Then plotting:

import matplotlib.pyplot as plt
fig = plt.figure(figsize=(7,3))
ax_lc = fig.add_subplot(121)
ax_vi = fig.add_subplot(122)

style = dict(
    color="#f52f2f",
    label="mace",
    marker="o",
    markersize=7,
    fillstyle="full",
)

mace_lc.make_violin(
    ax_vi, 
    error_type="force_component",
    face_color=style["color"],
    n_subsampling=10000,
)

mace_lc.make_learningcurve(
    ax_lc,
    error_type="force_component",
    **style,
)

fig.tight_layout()
plt.show()