TYC Junior Research Fellowship visitor talk: Extending Machine Learning Models Beyond Energy and Forces

TYC Junior Research Fellowship talk: Extending Machine Learning Models Beyond Energy and Forces Share on X

Nils Gönnheimer, University of Bayreuth

Abstract: The development of machine‑learning interatomic potentials (MLIPs) has revolutionized computational chemistry by combining the accuracy of first‑principles methods with the computational speed of empirical force fields. Many important properties, such as heat capacities, vibrational spectra, dielectric responses and optical activities, require either higher‑order derivatives (e.g. Hessians) or direct learning of non‑scalar quantities beyond energies and forces. In the first part of the talk, Hessian matrix evaluation is addressed: most MLIPs lack analytical second derivatives and must resort to costly, error‑prone finite differences, whereas implementing automatic‑differentiation (AD) Hessians within the equivariant MACE framework delivers both efficiency and numerical stability. In the second part, MACE‑μ‑α, a polarizability‑and‑dipole model built on the same equivariant architecture, is trained directly on molecular dipole moments and polarizability tensors, enabling accurate prediction of both infrared absorption and Raman scattering intensities. Together, these advances form a unified, beyond‑scalar MLIP platform for comprehensive spectroscopic characterization and rapid multi‑property prediction of complex materials.

Add to calendar:

Google Calendar

iCalendar