MACE-MDP: A Foundation Model for Molecular Dipole Moments and Polarizabilities

Pre-trained foundation models for machine-learning interatomic potentials (MLIPs) have recently transformed atomistic modeling by enabling efficient zero-shot sampling of potential energy surfaces at near Density-functional theory (DFT) accuracy.However, such models are largely restricted to predicting energies and forces, while largescale models for electronic response properties remain scarce, despite their importance for predicting the spectroscopic and electrostatic behavior of molecules and materials.Here, we introduce MACE-MDP, a foundation model that jointly predicts molecular dipole moments and fully anisotropic polarizability tensors and can be seamlessly combined with existing MLIPs.To enable this, we extend the SPICE dataset with high-accuracy first-principles polarizabilities, creating the SPICE-α dataset, and further develop an equivariant message-passing architecture based on MACE to learn both response properties simultaneously.MACE-MDP achieves near first-principles accuracy for dipoles and polarizabilities of organic molecules while substantially reducing computational cost.We further demonstrate transferability from molecules and dimers to small oligomers, like trimers and molecular crystals through harmonic and anharmonic IR and Raman spectroscopy benchmarks, including polymorph identification in molecular crystals through comparison with experimental Raman spectra.These results establish the reliable zero-shot prediction of dielectric response properties and enable rapid first-principles-accurate vibrational spectroscopy, opening the door to applications such as molecular discovery and crystal polymorph identification.