From Classical Force Fields to Neural Networks – Two Case Studies in Condensed-Phase Modeling

by Jonas Böhm (University of Bordeaux)

Thursday 3 Apr 2025, 09:00 → 10:00 Europe/Zurich

ODRA/111

As part of my PhD candidacy, I present two contrasting approaches to computationally efficient condensed-phase simulations, demonstrated through distinct case studies: one using a well-established method (classical force fields) and the other leveraging an emerging technique (deep learning). Both approaches seek to improve modeling efficiency over purely ab initio methods.
The first case study investigates the microscopic structure of liquid water-methanol mixtures using molecular dynamics (MD) simulations with a classical force field. The goal is to analyze structural peculiarities at low methanol concentrations and their potential correlation with the macroscopic partial molar volume anomaly of methanol, which exhibits a minimum at low methanol mole fractions. Through a detailed analysis of the hydrogen bond network, this study provides insights into structural features that may underlie this anomaly.
In contrast, the second case study investigates machine learning interatomic potentials (MLIPs) as a bridge between classical and ab initio calculations. MLIPs are applied to examine the structural stability and properties of Li₃YCl₆₋ₓBrₓ, demonstrating their potential as a computationally efficient alternative to density functional theory (DFT). The results not only provide a blueprint for guiding experimentalists toward promising material compositions but also emphasize the advantages of MLIPs in accelerating simulations while maintaining accuracy.
Although these studies focus on different physical systems, both highlight innovative computational strategies that enhance simulation efficiency and predictive power, underscoring the transformative impact of machine learning in this field.