nvalchemi.training.distillation.evaluation.NonConservativeResidual#

class nvalchemi.training.distillation.evaluation.NonConservativeResidual(*, num_probes, amplitude, segments, loop_work_mean_abs, loop_work_max_abs, force_floor, force_floor_max, force_rms, relative_floor, relative_floor_max)[source]#

Non-conservative component of a teacher’s force field.

The non-conservative residual is the part of a teacher’s force field that no conservative student can fit. A force field decomposes as \(F = -\nabla E + F_{\perp}\). A student that predicts forces as the gradient of an energy represents only the first term, whose work around any closed path vanishes. A closed-path integral of the teacher’s field therefore measures \(F_{\perp}\) alone. Every force here is a root-mean-square per-atom magnitude in eV/A. Compare force_floor with the forces_rmse of AccuracyMetrics accordingly: forces_rmse is a per-component figure, smaller by sqrt(3) for an isotropic error.

Parameters:
  • num_probes (int)

  • amplitude (float)

  • segments (int)

  • loop_work_mean_abs (float)

  • loop_work_max_abs (float)

  • force_floor (float)

  • force_floor_max (float)

  • force_rms (float)

  • relative_floor (float)

  • relative_floor_max (float)

num_probes#

Closed loops integrated in total, num_loops per graph.

Type:

int

amplitude#

Loop side length in A, as a root-mean-square per-atom displacement.

Type:

float

segments#

Midpoint-rule samples per side.

Type:

int

loop_work_mean_abs, loop_work_max_abs

Mean and maximum absolute work around one loop, in eV.

Type:

float

force_floor, force_floor_max

Mean and maximum lower bound the loop work places on a conservative student’s root-mean-square per-atom force error, in eV/A.

Type:

float

force_rms#

Root-mean-square teacher force at the loop centers over every atom.

Type:

float

relative_floor, relative_floor_max

Mean and maximum of each probe’s floor divided by the root-mean-square teacher force of its own graph. A batch that mixes force scales therefore reports a figure between its graphs’ ratios. A graph at equilibrium is divided by a small clamp rather than by zero.

Type:

float

model_config = {'extra': 'forbid', 'frozen': True}#

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].