ÖPG-CMD Joint Meeting 2026

Complex Configurations with Quantum Accuracy: Dynamics of Molecular Liquids from a General-Purpose Machine-Learned Potential

Sep 23, 2026, 11:15 AM

15m

HS 12.11 (University of Graz)

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

Speaker

Dr Florian Brünig (University of Luxembourg)

Description

Liquids provide a uniquely stringent test for atomistic modeling: experimentally accessible observables such as structure factors, diffusion coefficients, and vibrational spectra emerge from a delicate interplay of quantum-mechanical intermolecular interactions, collective fluctuations, and nuclear quantum effects over extended length and time scales. Still, these properties remain challenging to predict from first principles because accurate liquid simulations require both quantum-level interaction models and extensive statistical sampling. While conventional force fields are often parametrized directly against experimental data, limiting their value as predictive models, transferable machine-learning force fields offer the possibility of genuinely bottom-up simulations with experiment serving as an independent benchmark.

Here, we present recent developments of SO3LR [1], a machine-learned interatomic potential that combines the efficient SO3krates architecture [2] for short-range interactions with physically motivated long-range electrostatics and dispersion, pretrained on a broad dataset of molecular complexes. With JAX acceleration, SO3LR makes nanosecond-scale simulations of thousands of atoms feasible at low cost. Leveraging this efficiency and stability, we perform extensive molecular and path-integral dynamics of representative liquids, yielding converged structural, dynamical, and vibrational observables with realistic conformational sampling and full anharmonicity.

Direct comparison with experimental vibrational spectra and other liquid-phase data allows us to evaluate the predictive accuracy of the underlying quantum description and disentangle intrinsic limitations of classical force fields. Our results demonstrate that transferable machine-learning force fields can bridge the gap between empirical models and ab initio methods, enabling predictive simulations of condensed-phase systems with quantum-mechanical accuracy at experimentally relevant scales.

[1] Kabylda, A. et al. JACS 147, 33723–33734 (2025).

[2] Frank, J.T. et al. Nat Commun 15, 6539 (2024).