active

Run on-the-fly active learning, based on committee uncertainty estimates over a group of supplied NEP potentials. It requires at least two potentials, and every potential must be a NEP potential. GPUMD stops with an input error otherwise. The first potential specified in run.in is the main potential. The main potential propagates the molecular dynamics simulation. With dump_observer in average mode, the average of the potentials propagates it instead.

The uncertainty \(\sigma_f\) is estimated as the maximum over the atoms \(i\) of the standard deviation of the force over the \(M\) models,

\[\sigma_f = \textrm{max}_i \sqrt{ \sigma_{i,x}^2 + \sigma_{i, y}^2 + \sigma_{i, z}^2 },\]

where

\[\sigma_{i,k}^2 = \frac{1}{M} \sum_{m=1}^{M} \left( F_{i,k}^{(m)} - \bar{F}_{i,k} \right)^2\]

is the variance of the \(k\) Cartesian component of the force on atom \(i\). Here, \(F_{i,k}^{(m)}\) is that component for model \(m\), and \(\bar{F}_{i,k}\) is its mean over the models. If the uncertainty exceeds the specified threshold, \(\sigma_f>\delta\), for a structure in a step of a molecular dynamics simulation, then that structure is appended to the file active.xyz in the extended XYZ format. Additionally, the simulation time \(t\) and \(\sigma_f\) are written to the file active.out whether or not \(\sigma_f>\delta\). A model with forces that are not finite gives a NaN uncertainty. If the uncertainty of any atom is NaN, \(\sigma_f\) is NaN and the structure is appended to active.xyz. The energy, virial and forces in active.xyz are those of the main potential alone. The stress field holds the pressure tensor in eV/ų, which adds the kinetic contribution of the velocities to the virial of the main potential. The forces exclude those that other keywords add to the atoms, such as add_force and add_efield. Checking the uncertainty leaves the trajectory of the molecular dynamics run unchanged.

active takes five arguments. The first four set the interval for uncertainty estimation and the per-atom quantities written to active.xyz. The fifth sets the threshold \(\delta\) in units of eV/Å.

Syntax

active <interval> <has_velocity> <has_force> <has_uncertainty> <threshold>

interval is the interval (number of steps) between checking the uncertainty. If set to 1, the uncertainty will be computed for every step of the MD simulation. has_velocity can be 1 or 0, which means the velocities will or will not be included in the exyz output. has_force can be 1 or 0, which means the forces will or will not be included in the exyz output. has_uncertainty can be 1 or 0, which means the per atom uncertainties will or will not be included in the exyz output. threshold is a non-negative float, and corresponds to the threshold \(\delta\) in units of eV/Å.

Examples

Example 1

To run on-the-fly active learning using a committee of 5 NEP potentials, checking if the uncertainty exceeds 0.01 eV/Å every tenth MD step, write:

potential nep0
potential nep1
potential nep2
potential nep3
potential nep4
...
active 10 1 1 1 0.01

before the run keyword. This will generate two output files, active.xyz and active.out.

Caveats

• This keyword is not propagating. That means, its effect will not be passed from one run to the next.

• If the system has exploded, unphysical structures may be saved since no upper bound is set on the uncertainty \(\sigma_f\). Ensure that the resulting structures in active.xyz are physical.