Abstract
<jats:p>Despite the utility of machine-learned interatomic potentials (MLIPs), they can be difficult to train and use due to a lack of well-defined methods to assess their accuracy. This problem can be mitigated by combining traditional MLIPs with pair potentials that have machine learned coefficients. The use of pair potentials with strictly defined limiting behavior can help create a more reliable MLIP by restricting a general atomic neural network (ANN) to regions of the training set with lower variance in interatomic distances and total energies. This approach is demonstrated by simultaneously training both a short-range ANN as well as an ANN to fit charges for a Pauli repulsion potential. We show that this combination of ANNs significantly improves both the accuracy and stability of MLIPs across a wide range of materials, temperatures, and pressures. In particular, the addition of Pauli repulsion greatly improves the accuracy of potentials for use at high temperatures and high pressures, while also leading to a modest improvement in accuracy at ambient temperature and pressure conditions.</jats:p>