SMIRNOFF: Using non-integer bond orders in interpolation

[mattwthompson]

The spec describes interpolating force constants between bond orders of integers, i.e. 1 and 2. It might be useful to allow interpolating between bond orders that aren't integers, i.e. having the force constant change with bond order with one slope between 1.0 and 1.5, but a different slope between 1.5 and 2.0. The spec does not disallow this, but if we wanted to use them, it would be useful to explicitly state that non-integer values are allows.

This was brought up at some point in the past few months by the team that was investigating WBOs - @SimonBoothroyd or @pavankum may be able to provide more context, including if this is still of scientific interest.

[pavankum]

Yeah, parameters with a steep slope due to lack of diverse training data tend to produce unphysical k values when extrapolated below 1, if there is flexibility to include non-integer bond orders that would be great so that we can divide into ranges of super-steep and not-so-steep slopes. Another approach discussed was to use a flat value below 1 and interpolate afterwards, may be these can be named as a different scheme, instead of using the tag linear may be step-linear or something similar that conveys it explicitly.

[davidlmobley]

This is interesting. Yes, I suggest explicitly allowing this.