Machine learning interatomic potentials (MLIPs) provide an efficient and promising framework for atomistic simulations of complex chemical systems. Despite their strengths, further enhancements are still needed to expand their predictive scope across diverse chemical phenomena. This work presents an advanced equivariant graph neural network-based MLIP framework that incorporates charge- and spin-aware capabilities, enabling accurate predictions of energies, forces, atomic charges, and atomic spin moments in chemical systems. We introduce a novel spin-dependent charge equilibration (QEq) method,^[1] which extends the applicability of the QEq method^[2] to spin-polarized systems. The spin-polarized QEq-enhanced MACE^[3] models were thoroughly evaluated against density functional theory (DFT) reference datasets, demonstrating very good predictive performance for chemical systems, including organic molecules, metalorganics, and 2D/3D periodic systems. A key highlight is the model’s ability to accurately capture polaron distributions on the TiO₂ (110) surface, showcasing its effectiveness in modeling spin-related properties. By addressing both charge and spin dynamics, this approach broadens the potential of MLIPs to tackle more complex phenomena in a computationally efficient manner. This development thus further promotes the accuracy and versatility of MLIPs, paving the way for more reliable and comprehensive atomistic simulations in materials science, chemistry, and related fields.

[1] D. Tang and S. Luber, under review. (preprint at https://doi.org/10.26434/chemrxiv.15002391/v1)

[2] A. K. Rappe and W. A. I. Goddard, J. Phys. Chem. 95, 3358 (1991).

[3] I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, arXiv:2206.07697.