Machine-learned ACE models with charge equilibration in LAMMPS
Machine-learned interatomic potentials (ML-IAPs) are used to study many important physical and chemical phenomena, but often fall short for atomic systems with long-range interactions. For example, using an ML-IAP to study reactions in oxides is limited because of the large degree of charge transfer and long-range electrostatics. Methods that account for charge transfer and incorporate electrostatics into ML-IAPs are usually limited by cost and accuracy. Oxides and analogous systems represent a weak point in the modeling capabilities of ML-IAPs for these reasons. In LAMMPS, there is a spectrum of approaches one may take to incorporate charge in ML-IAPs. Some of these established methods are reviewed, and we highlight a novel approach that uses machine-learned models of electronegativity to accurately predict charge transfer during reactive MD simulations. The use of ML models of electronegativity in MD simulations poses a potential problem with stability of conservative ML-IAPs; it is facilitated by the implementation of a new extended-Lagrangian charge equilibration scheme that is highly stable. Some implementations for this approach may be used in LAMMPS, and the corresponding models may be trained in FitSNAP. The process is outlined for training and using atomic cluster expansion (ACE) models with this approach in LAMMPS. Results are presented for Uranium dioxide and water systems, and benefits of using ML models of electronegativity with the new extended-Lagrangian charge equilibration scheme are highlighted. These include quantitatively accurate predictions of driving forces for charge transfer in chemical reactions, improved predictions of properties, and improvements in unphysical behavior observed with previous approaches. For example, in water, using previous approaches could lead to hydrogen contributing more than 1 e worth of charge to a chemical bond and there may be excess charge transfer from one hydrogen to another that is very far away. Neither is physically correct, and the new approach in LAMMPS helps address some of these problems.