Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 22: Data-driven Materials Science: Big Data and Workflows III
MM 22.9: Talk
Wednesday, March 11, 2026, 12:30–12:45, SCH/A216
MACE-µ-α: A Foundation Model for Molecular Dipole Moments and Polarizabilities — •Nils Gönnheimer1,2, Venkat Kapil3, Karsten Reuter2, and Johannes T. Margraf1,2 — 1Universität Bayreuth — 2Fritz-Haber-Institut der MPG — 3University College London
Machine-learning interatomic potentials (MLIPs) have had a strong impact on computational chemistry, physics, and materials science in recent years by filling the accuracy gap between first-principles methods and classical force fields, at a fraction of the computational cost of the former. MLIPs are so far typically limited to predicting energies and forces, however, while other properties traditionally obtained from first-principles calculations have remained less accessible. Here, equivariant neural network architectures have led to enormous progress, as they allow the prediction of vectorial and tensorial properties on the same footing as energies and forces.
Here, we present the MACE-µ-α architecture for predicting dielectric properties based on the MACE MLIP framework. Trained on over 1.6 million organic systems, the corresponding foundation model allows the accurate prediction of molecular dipole moments and polarizabilities, as well as Raman and IR spectra (when combined with an MLIP). Notably, despite being trained on gas-phase molecules and clusters, the model also shows transferability to condensed systems such as molecular crystals.
Keywords: Foundation model; Raman spectroscopy; IR spectroscopy