search

ExpHP / rsp2

phonons in rust
Apache License 2.0
2 stars 1 forks source link

Travis: Failure with 'lost atoms' on GNR #91

Closed ExpHP closed 5 years ago

ExpHP commented 5 years ago

https://travis-ci.com/ExpHP/rsp2/jobs/199836940

This is a bizarre failure. The non-MPI configuration succeeded! And the MPI configuration still only runs one processor, so it's not just a non-determinism thing.

All other tests succeed.

The specific error is lost atoms from lammps. Right before this happens, a debug message appears showing that it is looking at a step size of almost 100! A run on my own PC finds the same frequencies only up to 3 sig figs, and the first step size ends up being 0.028. This all suggests a slightly different structure is found after CG.

ExpHP commented 5 years ago

Note, the single negative mode found is the rotational acoustic mode. (thank goodness I finally added that animation feature!)

Such a large step size still seems surprising to me. Because the mode produced by the harmonic approximation does not actually rotate (it skews), the distance between atoms increases for larger step sizes, which should make them unfavorable.

Linesearch does begin with exponentially increasing step sizes, but I find it hard to believe that it could be skipping over the energy barrier!

One thing: The appropriate scale for alpha for the CG along eigenmodes has never been clear to me. Should this portion have larger or smaller step sizes than regular CG? I've always used the same parameters for both and it has always just "worked out," aside from this...

ExpHP commented 5 years ago

Actually...who cares?

I'm just going to be pragmatic here and add expected-non-translations: 1 to its config. This is why it's there, after all...

ExpHP commented 5 years ago

Fixed in dcccf930d85f27069cfdb6a297a3e7a96e2b96a6

The travis build has finally succeeded!!!