nvalchemi.dynamics.NVTLangevin#

class nvalchemi.dynamics.NVTLangevin(model, dt, temperature, friction, random_seed=42, n_steps=None, hooks=None, convergence_hook=None, **kwargs)[source]#

NVT integrator via BAOAB Langevin dynamics.

Samples the canonical ensemble via a stochastic Ornstein-Uhlenbeck process that acts as an exact thermostat.

Parameters:
  • model (BaseModelMixin) – The neural network potential model.

  • dt (float or torch.Tensor) – Integration timestep in femtoseconds [M] or scalar.

  • temperature (float or torch.Tensor) – Target temperature in Kelvin [M] or scalar.

  • friction (float or torch.Tensor) – Langevin friction coefficient \(\gamma\) in 1/fs [M] or scalar. Controls thermostat coupling strength.

  • random_seed (int, optional) – Global RNG seed for the stochastic O step. Default 42.

  • n_steps (int, optional) – Total steps for run().

  • hooks (list[Hook], optional) – Initial hooks.

  • convergence_hook (ConvergenceHook or dict, optional) – Convergence criterion.

  • **kwargs – Forwarded to BaseDynamics.

__needs_keys__#

{"forces"}.

Type:

set[str]

__provides_keys__#

{"positions", "velocities"}.

Type:

set[str]

apply_per_system_params(params, batch)[source]#

Rebind each walker’s target temperature and rescale its velocities.

Langevin has no thermostat memory to transform — its noise amplitude is read from the per-system temperature every step — so the rebinding is the target update plus the velocity rescaling that keeps the kinetic energy consistent with the new target. Velocities scale by sqrt(T_new / T_old), in the same call, because the two together are the indivisible change.

Parameters:
  • params (Mapping[str, torch.Tensor]) – Must contain "temperature" in Kelvin per graph, shape [B].

  • batch (Batch) – The live batch; velocities is modified in place.

Raises:
  • RuntimeError – If called before the integrator state exists.

  • KeyError – If params names something this integrator cannot rebind.

Return type:

None

post_update(batch)[source]#

BAOAB post-force final B step.

Parameters:

batch (Batch) – Current batch; velocities updated in-place.

Return type:

None

pre_update(batch)[source]#

BAOAB pre-force half: B-A-O-A sequence.

Parameters:

batch (Batch) – Current batch; positions and velocities updated in-place.

Return type:

None

property random_seed: int#

Base seed of the thermostat noise, which every step offsets by step_count.