Improving stability and transferability of machine learned interatomic potentials using physically informed bounding potentials

While machine-learning techniques have shown great progress in advancing the frontier of accuracy and scope in interatomic potentials, they still suffer from a number of drawbacks. Principle among these is an inability to extrapolate outside of the training data which can result in poor transferability or stability issues limiting their usefulness outside of specific scenarios. This is in contrast to traditional potential formalisms such as the Embedded Atom Method (EAM), which have shown excellent transferability thanks to the physical intuition which motivated their creation. We introduce here a modification to the machine-learned Rapid Artificial Neural Network (RANN) formalism which uses an EAM potential to bound the prediction of the energies. This constrains the predicted energies outside the training space, resulting in more stable and transferable potentials. Using zinc as an example, we demonstrate the improved stability and show that this bounding potential improves the quality of the potential within the training data.