Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 3: Data-driven Materials Science: Big Data and Worksflows
MM 3.4: Vortrag
Montag, 17. März 2025, 11:00–11:15, H10
Beyond Numerical Hessians: Applications for Higher Order Derivatives in Machine Learning Interatomic Potentials — •Nils Gönnheimer1,2, Karsten Reuter1, and Johannes T. Margraf2 — 1University of Bayreuth — 2Fritz-Haber-Institut der MPG, Berlin
The development of machine learning interatomic potentials (MLIPs) has revolutionized computational chemistry by enhancing the accuracy of empirical force fields while retaining a large computational speed-up compared to first-principles calculations. Despite these advancements, calculating Hessian matrices remains challenging due to the lack of analytical second-order derivatives, necessitating the use of computationally expensive finite difference methods (which can lead to numerical instabilities because of rounding errors). Automatic differentiation (AD) offers a promising approach to reducing this computational effort and making the calculation of Hessian matrices more efficient and accurate. In this contribution, we discuss the implementation of AD Hessians in the equivariant MACE framework. This new methodology finds applications in screening the heat capacities of metal-organic frameworks (MOFs) and in the calculation of infrared (IR) spectra, which are an ubiquitous tool for molecular characterization.
Keywords: Hessian Matrix; Machine Learning Interatomic Potentials; Automatic Differenation; Vibrational Analysis; Heat Capacities