NNP - Neural Network Potential (n2p2)¶
Description¶
High-dimensional neural network potentials (Behler-Parrinello type) write the energy of each atom as the output of a per-species neural network whose inputs are atom-centered symmetry functions. exaStamp evaluates them through the n2p2 library. The model is read from an n2p2 directory (input.nn, scaling.data, weights.XXX.data).
Build
Configure exaStamp with:
-DEXASTAMP_MLIP_N2P2_BUILD=ON-DEXASTAMP_MLIP_N2P2_ROOT_DIR=<path>(default/usr/local/n2p2)
The directory must contain the n2p2 headers in include/ and the lib/libnnp.so and lib/libnnpif.so libraries. libnnpif must provide the exaStamp interface class (InterfaceExastamp). If the directory does not exist, the plugin is silently disabled.
Warning
The n2p2 interface runs on CPU only and is not covered by the regression test suite.
Operator¶
Syntax
n2p2_force:
parameters:
dir: <string>
showew: <bool>
resetew: <bool>
showewsum: <int>
maxew: <int>
cflength: <float>
cfenergy: <float>
cutoff: <float>
dump_out: <bool>
Parameters
parameters.dir: string, required # n2p2 model directory (input.nn, scaling.data, weights.*.data).
parameters.showew: bool, required # Print extrapolation warnings (symmetry functions outside the training range).
parameters.resetew: bool, required # Reset the extrapolation warning counter at each step.
parameters.showewsum: int, required # Print a summary of extrapolation warnings every showewsum steps (0 = never).
parameters.maxew: int, required # Abort after this many extrapolation warnings.
parameters.cflength: float, required # Length conversion factor from Angstrom to the model length unit.
parameters.cfenergy: float, required # Energy conversion factor from eV to the model energy unit.
parameters.cutoff: float, required # Cutoff radius in Angstrom (plain number, no unit). Must be at least the largest symmetry-function cutoff of the model.
parameters.dump_out: bool, required # Enable n2p2's own output on stdout.
All parameters are required.