Optimised Morse transform of a Gaussian process feature space

Abstract

Morse projections are well-known in chemistry and allow one, within a Morse potential approximation, to redefine the potential in a simple quadratic form. The latter, being a non-linear transform, is also very helpful for machine learning methods as they improve the performance of models by projecting the feature space onto more well-suited coordinates. Usually, the Morse projection parameters are taken from numerical benchmarks. We investigate the effect of changing these parameters latter on the model learning, as well as using the machine learning method itself to make the parameters decision. We find that learning is not necessarily improved by the latter and that general Morse projections are extremely susceptible to changes in the training data.