Dresden 2026 – scientific programme

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 13: Charged Soft Matter, Polyelectrolytes and Ionic Liquids

CPP 13.5: Talk

Monday, March 9, 2026, 16:15–16:30, ZEU/0260

Machine-Learning Prediction of Ionic Conductivity from Short-Time Structural Data in Ionic Liquids — •David Bienek — Helmholtz-Institut Münster, Münster, Germany

Molecular dynamics (MD) simulations are a powerful tool for studying electrolyte materials, but the reliable calculation of ionic conductivities and transference numbers requires long trajectories. In this project, we explore whether machine-learning (ML) models can predict dynamic properties of ionic liquids (ILs) from structural information obtained in much shorter simulations. ILs are chosen as model systems because prior work has demonstrated links between structure and hydrodynamic behaviour. Their simple cation-anion composition also allows systematic generation of simulated systems. As structural descriptors, we employ radial distribution functions (RDFs), which are physically interpretable.

We test linear regression, random forests, and neural networks for predicting ionic conductivity. After feature engineering we find substantial improvements in accuracy. A peak-based representation proves particularly helpful for identifying structure-dynamics relations. Feature-importance analyses consistently indicate that the dominant information originates from the first coordination shell. Moreover, simpler models (random forest, linear regression) outperform neural networks, highlighting the necessity to use appropriate model complexity. Overall, our results show that ML can support the identification of structure-dynamics correlations in electrolytes and may help estimate transport properties from short MD trajectories.

Keywords: Ionic Liquids; Machine Learning; Molecular Dynamics Simulation