Wrong unit of default b_tau in npt_langevin_init

[chlwjd1234]

In torch_sim.integrators.npt.npt_langevin_init, the default value of b_tau is currently defined as

if b_tau is None:
    b_tau = 1 / (1000 * dt)  # Default barostat time constant

However, b_tau is later used to construct the barostat mass as:

cell_masses = (n_atoms_per_system + 1) * batch_kT * b_tau * b_tau

Since cell_masses should be in dimension of (energy) · (time)² ,
it should be changed as

b_tau = 1000 * dt

[orionarcher]

@abhijeetgangan

[thomasloux]

There is indeed something weird here. More investigation needs to be done on it.
If I remember well, the implementation of NPT Langevin corresponds exactly to press/langevin in LAMMPS. LAMMPS has all of its integrators using parameters with time units, but from memory this is not the first way NPT Langevin has been described, so we end up with some weird units like these ones.
Referring the original PR that exactly discuss b_tau (in their case, Pdamp) with time units and also at some point it was redefined as 1/time.
lammps/lammps#3791

[thomasloux]

I also point to LAMMPS docs which gives some more information about the units and how Pdamp should be setup: https://docs.lammps.org/fix_press_langevin.html