MACE-POLAR-1

MACE-POLAR-1 is a machine-learning interatomic potential (MLIP) and electrostatic foundation model for molecular chemistry, introduced by researchers at the University of Cambridge in a February 2026 arXiv preprint. It extends the widely used MACE architecture — an equivariant message-passing graph neural network — with an explicit, physically grounded treatment of long-range electrostatics and electrostatic induction. Most MLIPs rely on local atomic descriptors with a finite cutoff and therefore cannot capture long-range Coulomb interactions or charge transfer; MACE-POLAR-1 is designed to remedy this limitation.

The model combines local many-body geometric features with a non-self-consistent-field formalism that updates learnable charge and spin densities through polarizable iterations to model induction, followed by global charge equilibration via learnable Fukui functions that control total charge and total spin. This lets the model describe systems with varying charge and spin states while remaining computationally efficient, and yields interpretable, spin-resolved charge densities and responses to external fields.

MACE-POLAR-1 sits within the rapidly growing landscape of foundation-model MLIPs (such as the broader MACE-OFF and MACE-MP families) but distinguishes itself through explicit polarizable electrostatics, which the authors show is decisive for non-covalent interactions and protein-ligand binding.

#Key Features

• Polarizable electrostatics: Updates learnable charge and spin densities through polarizable iterations to model induction, going beyond fixed local descriptors.
• Charge equilibration via Fukui functions: Global charge equilibration with learnable Fukui functions controls total charge and total spin, enabling variable charge/spin states.
• MACE backbone: Builds on the equivariant MACE message-passing architecture, retaining its accuracy and efficiency for local many-body interactions.
• Strong non-covalent performance: Reports a roughly fourfold improvement over short-ranged models on protein-ligand interactions and sub-kcal/mol molecular crystal formation energies on X23-DMC.
• Interpretable outputs: Provides spin-resolved charge densities and field responses, supporting physical interpretation of predictions.

#Technical Details

MACE-POLAR-1 augments the MACE equivariant GNN with a non-self-consistent-field electrostatics module: local many-body features feed polarizable iterations that refine learnable charge and spin densities, followed by global charge equilibration through learnable Fukui functions. The model is trained on the OMol25 dataset of roughly 100 million hybrid DFT calculations. The authors report accuracy competitive with hybrid DFT across thermochemistry, reaction barriers, conformational energies, and transition-metal complexes, with the long-range electrostatic treatment driving large gains on non-covalent interactions and supramolecular complexes — including sub-kcal/mol crystal formation energies on X23-DMC and a fourfold improvement over short-ranged models on protein-ligand interactions. As a recent preprint, no public code or weight release is referenced in the paper.

#Applications

MACE-POLAR-1 targets computational molecular chemistry and drug discovery, where accurate treatment of electrostatics and induction is essential. Its ability to span small molecules through protein-ligand complexes makes it relevant for binding-energy estimation, conformational and reaction-energy calculations, molecular-crystal property prediction, and simulations of systems with variable charge or spin states.

#Impact

By embedding explicit polarizable electrostatics into a foundation-model MLIP, MACE-POLAR-1 addresses one of the most consequential shortcomings of local interatomic potentials and demonstrates measurable gains precisely where electrostatics dominate — non-covalent and protein-ligand interactions. If accompanied by an open release, it could broaden the reach of MACE-family models into electrostatics-sensitive biomolecular applications. As an unreviewed preprint without a referenced code release, its reported benchmark gains await independent reproduction and community evaluation.