Including physics-informed atomization constraints in neural networks for reactive chemistry

Machine learning interatomic potentials (MLIPs) have emerged as powerful tools for investigating atomistic systems with high accuracy and relatively low computational cost. However, a common challenge with current MLIPs is their limited ability to accurately predict the relative energies of systems containing isolated or nearly isolated atoms, which often appear in various chemical processes. To address this limitation, we present a technique for modifying existing neural network architectures in a minimal way which accounts for the energies of isolated atoms and the atomization energy (AE) of a system. Using this technique, we build a model architecture we call HIP-NN-AE, an enhanced version of the Hierarchically Interacting Particle Neural Network (HIP-NN). Our results demonstrate that the HIP-NN-AE model significantly outperforms the previous HIP-NN model in multiple scenarios involving isolated atoms, such as complete atomization and bond dissociation processes. We also show that this innovation does not compromise the model performance on other tasks, including barrier height and conformational energy variation. The HIP-NN-AE model thus offers a robust solution to the challenges posed by isolated atoms in energy prediction tasks.