ÖPG-CMD Joint Meeting 2026

Using machine-learned interatomic potentials for accurate thermal conductivity predictions via non-equilibrium molecular dynamics

Sep 23, 2026, 5:15 PM

15m

HS 12.11 (University of Graz)

3) Contributed talk M21 - Recent Developments in Machine Learned Interatomic Potentials Mini-Colloquium

Speaker

Mr Florian Unterkofler (Institute of Solid State Physics, TU Graz)

Description

With the rise of machine-learned interatomic potentials, simulations have become an even more crucial tool for predicting material properties. We previously achieved accurate predictions of experimentally observed thermal conductivity of acenes, using system-specific, machine-learned Moment Tensor Potentials (MTPs) within a lattice dynamics approach.[1] To obtain a complementary real-space perspective, we now investigate whether comparable accuracy can be achieved using non-equilibrium molecular dynamics (NEMD).
Here, we present the workflow required to obtain accurate and reliable predictions when applying MTPs in NEMD simulations. We show that, due to the inherently stochastic nature of both MD and MTP training, a thorough statistical analysis of multiple simulations with different initial conditions and different realizations of the MTP is necessary. Furthermore, we highlight the importance of selecting appropriate training data to generate robust MTPs.
When these considerations are taken into account, we achieve an excellent agreement between experiments, lattice-dynamics, and NEMD results, with NEMD simulations providing tools to investigate heat-transport bottlenecks in real space.

[1] L. Legenstein et al., npj Comput Mater 11, 29 (2025)