Machine-learned interatomic models (ML-IAM) have gained significant interest as a means of achieving “quantum accurate” atomistic simulations. However, in systems with strong dispersion interactions, these models are costly due to the need to describe both short- and long-range (e.g., bonded and non-bonded) interactions. In this work, we show that ML-IAM expense can be significantly reduced without compromising on accuracy by training models to only the short-ranged portion of the interaction range and including an empirical dispersion correction. We demonstrate that using this approach, we can generate efficient models using the Chebyshev Interaction Model for Efficient Simulation (ChIMES) to model structural, dynamic, and thermodynamic properties of nitrogen under extreme high temperature and high pressure conditions that are as accurate as those generated through the more expensive standard fitting scheme.
