Implementing other thermostats (e.g. Andersen or Langevin)

[EqualAPriori]

Hi all, thank you for the really great work.

I was curious if people have considered adding the Andersen or Langevin thermostats. The Berendsen thermostat is known to not generate a canonical ensemble, because energy fluctuations are not captured correctly.

If there's broader interest in this, I'd like to help out with the efforts. However, I'm a bit newer to the PySCF code and could use some pointers on how to best implement such a thermostat.

Cheers!

[matthew-hennefarth]

I think it would be nice to have other thermostats. You can checkout the integrators file to see how things are currently coded up. Specifically, you should just need to derive a new integrator from _Integrator and override the necessary methods, namely: _next which returns the next Frame with the new geometry, velocity, time step, and energy.

For example, the Velocity Verlet _next and the Berendson _next do basically the same thing, but the Berendson also scales the velocity.

The main thing is verifying the implementation of new thermostats to existing codes (namely ASE would be a good one if the thermostat exists within it).

[susilehtola]

I suggest improving interfaciability with MD codes over reimplementing MD in PySCF...

[EqualAPriori]

Thanks for the notes -- I saw that ASE has a langevin integrator, will try to get that working.

[EqualAPriori]

Hi all, could I get some pointers on how PySCF integrates with ASE?

ASE doesn't have PySCF listed as a supported calculator, and this PySCF doc uses another package, Siesta, instead to interface with ASE.

Is this generally the case that PySCF has to integrate ASE via Siesta, or is there a more direct integration?

[matthew-hennefarth]

Is this generally the case that PySCF has to integrate ASE via Siesta, or is there a more direct integration?

No, you should be able to directly integrate PySCF with ASE without Siesta.

There is no official interface between ASE and PySCF that I am aware of (hence why PySCF is not a supported calculator for ASE). There is a possible script/framework in #624 but it might be broken (or only broken for the GPU implementation in gpu4pyscf repo).

[tberkel]

Writing a PySCF calculator for ASE is on the to-do list. I wrote one a while ago for personal use, but it needs to be cleaned up and probably made more feature-complete. Personally, I'd like to get an ASE calculator implemented and then remove all the ASE dependencies in pyscf/pbc/tools/.

I agree with @susilehtola that PySCF should really just be a force engine for other MD packages.