viscosity calculation not converged?

[ltalirz]

Composition: CaO 15.0 - Na2O 18.0 - SiO2 67.0

3k atoms

extrapolated T4: 869C

The 1227C case at high lag times looks a bit suspicious to me (?) but converged at least in principle.

10k atoms

extrapolated T4: 911C

The 1227C line does not look like it plateaued...

[Atilaac]

I assume the composition is in mol%, right?

[ltalirz]

Yes, always mol%

[Atilaac]

I checked locally the temperature is too low. The SACF did not decay to 0, and even after it decays, we need a simulation length at least 10 times that autocorrelation time. I will extend it to continue the run if no convergence is found.

[ltalirz]

Great, thanks!

[Atilaac]

@ltalirz I rewrote almost the whole viscosity workflow to remove the need to define a cutoff for the SACF. The new work flow uses another method that is better suited than Green-Kubo method but in the same time mathematically equivalent. I also added option to run multiple simulations with different velocity seeds for statistical averaging and also convergence loops similar to what Marcel have in the CTE workflow are available now. This new approach is from what I read in the literature, far superior than any other method despite its old origins (1960s).

I will submit the PR later today I just wanted to let you know in advance. I would also be grateful if you can run tests using the new workflow using the sample you mentioned above, I can not use large samples on my laptop.

[ltalirz]

I rewrote almost the whole viscosity workflow to remove the need to define a cutoff for the SACF.

Nice!

I also added option to run multiple simulations with different velocity seeds for statistical averaging and also convergence loops similar to what Marcel have in the CTE workflow are available now.

That sounds great!

We should at some point think together about our strategy for parallelization and write it down - both at the "top" level (structure generation and everything that follows) and then, where necessary, in lower levels like the viscosity.

This new approach is from what I read in the literature, far superior than any other method despite its old origins (1960s).

Ok, please present the ideas in the meeting later today.

I would also be grateful if you can run tests using the new workflow using the sample you mentioned above

It should not be a problem to run things over the weekend; we should just pick the parameters such that the calculations don't take until monday ;-)

[ltalirz]

Since 07be49e is merged, is there anything I need to do - should I change the API to use the new method?

[Atilaac]

If you want to use that helfand method you need to use it in the api. However, even if you don't, this should not break the old behavior.