Closed bastonero closed 11 months ago
Thanks @bastonero! Does this one fix https://github.com/aiidateam/aiida-quantumespresso/issues/637?
I remember I was testing this when @giovannipizzi first opened the issue (and then got distracted, as usual), and the damped was very effective for relax
, but not-so-effective for vc-relax
. The strategy here is already quite good, but we may consider immediately switching to the damped algorithm for calculation == relax
.
Thanks @mbercx !
Thanks @bastonero! Does this one fix #637?
Yes, that was the idea (I forgot the number of that issue, thanks for linking).
I remember I was testing this when @giovannipizzi first opened the issue (and then got distracted, as usual), and the damped was very effective for
relax
, but not-so-effective forvc-relax
. The strategy here is already quite good, but we may consider immediately switching to the damped algorithm forcalculation == relax
.
Ok, amazing, then it's also tested! We can then proceed as you suggest.
@sphuber now working with the new logic suggested by @mbercx . I also added for the relax
case the fire
algorithm`. I think it's new, we should make sure from which version of QE it's implemented. But now we just support the latest 3 versions correct?
But now we just support the latest 3 versions correct?
I think we only support v6.6 and newer, but maybe for the next release we already even planned to drop support for v6 completely. @mbercx ?
Looking into QE history it seems the fire
algorithm is there since a while, so it should be a safe addition.
I think we only support v6.6 and newer, but maybe for the next release we already even planned to drop support for v6 completely. @mbercx ?
According to our compatibility guarantees, we support the last 3 minor releases, as well as older versions for up to two years. As QE v6.8 was released in July 2021:
We can also drop support for that in the next aiida-quantumespresso
release.
Nevertheless, I didn't understand whether this fire
algorithm is something brand new or it was introduced years ago. I cannot find on their GitLab changelog when this was introduced. Maybe it is already quite old stuff, so no need to drop support (yet)
Sometimes the BFGS algorithm for ionic minimizaiton fails, and the current handler simply restart from scratch. This might work, but here we try to improve upon this simplistic solution, trying first to lower the trusted radius, and then trying to switch algorithm.