NEP

NEP(path) reads a neuroevolution potential of version 4 in the nep.txt format of GPUMD, without (nep4) or with (nep4_zbl) a ZBL repulsion, and computes every property.

from mdapy.calculators import NEP

potential = NEP("nep.txt")
result = potential.compute(system, ("energy", "forces", "virial"))
descriptors = potential.descriptors(system)  # (N, dim)

Attribute	Type	Meaning
path	str	the potential file
elements	tuple of str	the elements of the potential, in the order of the file
cutoff	float64	the largest radial or angular cutoff in Å

Method	Returns
compute(system, properties, *, neighbors=None, progress=None, cancel=None)	as every calculator, with the progress= and cancel= of Parallelism
descriptors(system, *, neighbors=None, progress=None, cancel=None)	(N, dim) descriptors of each atom, scaled as the network sees them: the radial ones, then the angular ones by \(l\) and \(n\), then the four-body and five-body ones by \(n\)

The atoms need an element column; elements the potential does not have raise InvalidParameter. Other models of GPUMD (NEP3, NEP5, charge, dipole and polarizability models) and malformed files raise FileFormatError.

The model is that of GPUMD (Fan et al. (2022)), version 4 (Song et al. (2024)): radial descriptors from Chebyshev functions of the distance, angular ones from real spherical harmonics up to \(l_\text{max}\), with optional four-body and five-body invariants, each scaled by its factor in the file and fed to one hidden layer of \(\tanh\) neurons per element, which gives the energy \(E_i\) of atom \(i\). The cutoff of a pair is the mean of the cutoffs of its two elements. With ZBL, each pair within the angular cutoff adds the screened Coulomb repulsion of Ziegler et al. (1985), \(V(r) = \frac{k_e Z_i Z_j}{r}\,\phi(r/a)\, f_s(r)\), half to each atom; the switch \(f_s\) is 1 below an inner cutoff and falls as a cosine to 0 at an outer cutoff, as the zbl line of the file chooses:

zbl line	screening \(\phi\)	inner cutoff	outer cutoff
zbl r1 r2	universal ZBL	r1	r2
zbl r1 r2 f	universal ZBL	0	the smaller of r2 and f times the sum of the two covalent radii GPUMD uses
zbl 0 0	from a table at the end of the file, per pair of elements	from the table	from the table

As in GPUMD, the derivative of \(E_i\) by the vector to neighbor \(j\) adds to the atomic virial of atom \(j\), so atomic virials are those of GPUMD; EAM shares each pair's virial equally between its two atoms, as LAMMPS does. Both sum to the same virial.

Please cite

Fan et al. (2022) and Song et al. (2024); with ZBL, also Ziegler et al. (1985); and the paper of the potential.