This PR implements two additional weights to be considered in the loss function of NNP optimizations.
Configurations can be weighted with a new keyword weight in the input.data file as proposed in #91.
Individual atomic force components can be weighted via the unused 6th column of the atom keyword in input.data.
The PR works as expected, but there are two aspects that require discussion:
In order to restore default behavior for existing data sets, atomic force weights of 0 are converted to 1.
Weights are only applied on the error and not in the Jacobian, as the latter one introduced deteriorating quality.
This PR implements two additional weights to be considered in the loss function of NNP optimizations.
Configurations can be weighted with a new keyword
weight
in the input.data file as proposed in #91. Individual atomic force components can be weighted via the unused 6th column of theatom
keyword in input.data.The PR works as expected, but there are two aspects that require discussion:
Resolves #91, resolves #121.