Machine learning interatomic potentials (MLIPs) provide an efficient andpreciseframework for atomistic simulations of complex chemical systems. Despitetheirstrengths, further enhancements are needed to expand their predictivescopeacrossdiverse chemical phenomena. This work presents an advanced equivariantgraphneural network-based MLIP framework that incorporates charge- andspin-awarecapabilities, enabling accurate predictions of energies, forces, atomic charges,andatomic spin moments in chemical systems. We introduce a novel spin-dependentcharge equilibration (QEq) method, which extends their applicability tospin-polarizedsystems. The spin-polarized QEq-enhanced MACE models were thoroughlyevaluatedagainst density functional theory (DFT) reference datasets, demonstratingexceptionalpredictive performance for chemical systems, including organicmolecules,metalorganics, and 2D/3D periodic systems. A key highlight is the model’s abilitytoaccurately capture polaron distributions on the TiO2 (110) surface, showcasingitseffectiveness in modeling spin-related properties. By addressing both chargeandspindynamics, this approach broadens the potential of MLIPs to tackle morecomplexphenomena in a computationally efficient manner. This development marksasubstantial leap forward in the accuracy and versatility of MLIPs, pavingthewayformore reliable and comprehensive atomistic simulations in materials science, chemistry,and related fields.
Deqi Tang