High throughput phonon calculations

[tisabe]

Hi all,
I am intending to compute phonon band structures of many (thousands) small (20 atoms) structures. For a small subset I have used the ASE calculator but this seems inefficient compared to the speed of MD runs with LAMMPS and MLIAP. With phonopy, I can generate supercells, but I only managed to do single force calculations in a LAMMPS script, and the overhead of a full run for a single structure also seems large. Loading positions in a LAMMPS loop seems tedious...
Has anyone managed a similar workflow, but efficiently? Are there other ways to go about this?
Another interesting possibility might be calculating the hessian matrix, but this would need more code to support this.

[tisabe]

I just found out about torch-sim. This looks really promising, and I am interested to find out the speedup of ASE vs. torch-sim vs. LAMMPS (ML-IAP).

[gabor1]

Well, only torch-sim will do the batched evaluation you are looking for.