Regensburg 2025 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

DY: Fachverband Dynamik und Statistische Physik

DY 28: Poster: Machine Learning, Data Science

DY 28.4: Poster

Wednesday, March 19, 2025, 15:00–18:00, P4

Stability of Machine-Learned Interatomic Potentials in Molecular Dynamics Simulations for Complex Organic Crystals — •Martin Tritthart, Florian Linder, Lukas Legenstein, Florian Unterkofler, Martin Klotz, and Egbert Zojer — Institute of Solid State Physics, Graz University of Technology, NAWI Graz, AUSTRIA

Understanding thermal conductivity and mechanical stability is crucial for several applications of organic semiconductors (OSCs) and metal-organic frameworks (MOFs). Molecular dynamics (MD) simulations are commonly used to deduce these properties, and in recent years, machine Learned interatomic potentials (MLIPs) have been employed to enhance accuracy compared to classical force fields. MLIPs are orders of magnitude faster than ab initio methods and can achieve close to DFT accuracy. Their accuracy,however, heavily depends on the quality of the training data. Incorrect predictions of forces and energies for atomic configurations outside the training dataset can accumulate in the MD trajectory, potentially even leading to disintegration of the (molecular) building blocks. This, however, limits long-time simulations, which are necessary to investigate thermal transport properties. To mitigate these issues and to realize a robust MLIP, reliable uncertainty estimates are needed especially for atomic configurations outside the region of the phase space sampled during training. Structures characterized by high uncertainties can then be calculated by ab initio methods and incorporated into the training dataset. This approach is tested in my contribution for prototypical MOF and OSC materials.

Keywords: Molecular Dynamic; Ab initio; Interatomic Potential; Machien Leraning; Complex Organic Crysatals