Hello,
The quote below is in the FIRE documentation for LAMMPS. In the FIRE implementation in our MD code, this problem was simple to fix. Just don’t check for convergence until after the latency period! Per the original paper, the latency period, which I guess you define internally as I don’t see that the user can set it, is the soonest you should start to increase the time step and lower the alpha.
IMPORTANT NOTE: Etol should be set to 0.0 when using the quickmin or fire minimization styles. This is because they periodically reset velocities to 0.0 and take a zero-length step which will appear as a 0.0 energy change, stopping the minimizer if etol is finite.
Cheers,
Carolyn Phillips