Troubleshooting

Note

This page lists some common errors, warnings, and failures of the pARTn code. Some insight into what might be their cause is given, and how to go about solving them on your own. For any trouble not listed here, please post an issue to the GitLag issues page.

pARTn does not launch?

Look into Installation, E/F engine inputs, and make sure you have followed all steps correctly. Also make sure you use the appropriate versions of the E/F engines. Look up possible error messages related to the E/F engine in their respective reference manuals.

Crash without any message

You are possibly missing a crucial ARTn input parameter, please see List of ARTn input parameters page.

If the problem continues, please post an issue on the GitLab issues page.

Fail with a message

The ARTn algorithm can fail for several reasons. If the failure occured due to a known reason, there will be a message printed in the artn.out file, for example as:

Failure message: RESTART FILE DOES NOT EXIST

Which gives you an idea of what is happening. Depending on the message, there are several actions you can do to try to resolve the failure, they are listed below.

Error message:

  • RESTART FILE DOES NOT EXIST: you have desired a restart lrestart = .true. but do not provide the needed restart file artn.restart.

  • <variable_name> has unsupported value: self-explanatory, check List of ARTn input parameters

  • ENERGY EXCEEDS THE LIMIT: The energy difference Etot of the current configuration has exceeded the limit value, see etot_diff_limit. This usually happens when the push direction includes some (near-)collision of atoms, or the step size is much too large, or both.

  • BOX EXPLOSION: this error means the simulation box has exploded, which is identified by a very large movement of atoms coming from the E/F engine, it can happen in LAMMPS.

  • Atoms lost: Same as above.

  • Received a NaN value from engine: there is a problem with some array passed by the E/F engine to pARTn, look into the E/F calculation specification. If no apparent reason, then need to inspect by debugging the code/interface.

  • EIGENVALUE LOST: This is the most common failure when exploring the Potential Energy Surface. It singnifies that the structure is already outside of the basin (a negative lowest eigenvalue of the Hessian has already been found) for some number of steps, but then for some reason that eigenvalue goes back to positive (the value is “lost”). In this case the corresponding eigenvector cannot be trusted anymore to be pointing to the saddle point, thus the default solution is to abort the research.

There are several possible reasons for this, some numerical, and some physical:

  • due to the Lanczos diagonalization scheme, try tuning the lanczos_* parameters;

  • due to an abrupt change in the pushing vector when the structure passes the inflection line of the Hessian (lowest eigenvalue becomes negative), which causes the perpendicular relaxation to bring the structure back into the basin. This is mediated by the smoothing factor nsmooth (see nsmooth);

  • the loss of eigenvalue can sometimes be “corrected” by resetting the Lanczos algorithm with a random initial Lanczos vector, the number of times this resetting is allowed during a single research is given by the parameter nnewchance (see nnewchance). Using this method can however bring the structure very far from the initial structure, thus the saddle point and its adjacent minima can be completely irrelevant to what has been originally desired. It is thus recommended to keep nnewchance at some small integer value, e.g. nnewchance=2.