nvalchemi.training.distillation.evaluation.compare_radial_distributions#
- nvalchemi.training.distillation.evaluation.compare_radial_distributions(reference, candidate)[source]#
Score how far a candidate trajectory’s structure sits from a reference.
The headline number is the Jensen-Shannon divergence of the two normalized pair-distance histograms. It is symmetric, bounded in
[0, 1]with a base-2 logarithm, and finite where one histogram has an empty bin. A Kullback-Leibler divergence or a chi-squared distance is not finite there on RDF data. Theg(r)curves themselves are compared with an integrated absolute difference. The comparison is only as species-resolved as its curves. To gate chemical ordering, build one curve pair per species pair with thepairargument ofradial_distribution().- Parameters:
reference (RadialDistribution) – Reference-trajectory curve.
candidate (RadialDistribution) – Student-trajectory curve, binned identically and resolved to the same species pair.
- Returns:
The divergence and the two curve distances.
- Return type:
- Raises:
ValueError – If the two curves do not share bin edges, if they resolve different species, or if either accumulated no pairs at all.
Examples
>>> from nvalchemi.training.distillation.evaluation import ( ... compare_radial_distributions, ... radial_distribution, ... ) >>> match = compare_radial_distributions( ... radial_distribution(teacher_frames), ... radial_distribution(student_frames), ... )