Force Descriptors

Force descriptors encode the atomic environment in a form suitable for learning the force component of a machine learning force field (MLFF). Each descriptor is projected onto a random unit vector to give a scalar training target, following the force-projection approach of Botu et al.

Command-line usage

generate_force_descriptors trajectory.xyz [OPTIONS]

The trajectory must contain force data (forces or momenta arrays).

Options

Option	Default	Description
`--fp-type` / `-f`	`Split2b3b_ss`	Fingerprint type (see below)
`--rc` / `-r`	`10.5`	Cutoff radius in Å for both 2-body and 3-body terms
`--n2b`	`20`	Number of 2-body eta functions
`--n3b`	`10`	Number of 3-body eta functions

Fingerprint types

BP2b — Behler-Parrinello 2-body

Force-projected two-body fingerprint. Efficient for simple systems.

generate_force_descriptors traj.xyz --fp-type BP2b --rc 8.0 --n2b 30

Split2b3b_ss — Split 2-body + 3-body (same species)

Combines a 2-body and a 3-body fingerprint restricted to same-species neighbour pairs. Recommended for multi-component systems.

generate_force_descriptors traj.xyz --fp-type Split2b3b_ss --rc 10.5 --n2b 20 --n3b 10

Output

Results are written to ./descriptors/force_descriptors.dat. Each row contains the fingerprint vector followed by the force projection value and the atomic symbol:

<fp_1>  <fp_2>  ...  <fp_N>  <Fv>  <symbol>

This file is used directly as input to the MLFF training step.