Organizers: Bettina Keller

Abstract: 

This minisymposium focuses on modern AI and machine learning methods for representing, understanding, and simulating complex dynamical systems, with particular relevance to the multiscale challenges addressed within CRC 1114. The session highlights methodological developments that bridge data-driven learning and mathematical modelling and capture essential structures in high-dimensional, stochastic, and multiscale dynamics.

Speakers: 

11:30h: Stefan Chmiela (TU Berlin)

Accelerating Molecular Dynamics with Implicit Machine Learning

Over the past decade, machine-learned force fields (MLFFs) have evolved from accurate interpolators for small molecules into scalable, general-purpose models that enable near-quantum-accurate molecular dynamics (MD) for systems well beyond their training data. Building on these advances, this talk presents two complementary directions we are pursuing that rethink MD by treating force prediction and trajectory integration as a unified computational problem.

The first direction exploits the temporal coherence of MD by reusing latent representations across consecutive simulation steps. We introduce implicit MLFFs defined by self-consistent fixed-point equations rather than explicit neural network layers, enabling warm-started force prediction along MD trajectories while preserving accuracy and atomistic resolution.

The second direction extends implicit modeling to trajectory integration. Learned Hamiltonian Flow Maps predict the mean phase-space evolution over longer time intervals, enabling stable updates beyond the timestep limits of classical integrators. A Mean Flow consistency condition allows these models to be trained directly from independent phase-space samples, making them compatible with the large-scale datasets already used to train modern general-purpose MLFFs.

Together, these two implicit formulations extend the scaling frontier of quantum-mechanically faithful molecular simulation, enabling longer trajectories and larger atomistic systems within fixed computational budgets.

12:00h: Hendrik H. Heenen (Fritz-Haber-Institut, Max-Planck-Gesellschaft)

From atomistic free energies to growth kinetics: ML-accelerated multiscale simulation of graphene growth on liquid copper

Predicting how thin films grow under realistic conditions is inherently a multiscale problem, with the governing chemistry residing at the atomistic scale while the experimental observables emerge over far longer time and length scales. Computational descriptions commonly bridge this gap by borrowing concepts from computational surface science. Growth is reduced to a small set of elementary processes on a static, crystalline substrate whose rates follow from harmonic transition-state theory, which in turn serve as input to coarse-grained kinetic models. However, these approximations break down when one of the participating phases resists such a simple description. Liquid copper catalysts, which promise large-scale high-quality graphene growth, epitomize such a case, where the disordered and dynamic substrate admits neither a fixed set of adsorption sites nor a harmonic treatment of the relevant processes. Here, we use machine-learned interatomic potentials trained to density functional theory within an active learning framework that autonomously samples the high-dimensional, disordered configurational space and renders large-scale free-energy sampling on the fluctuating liquid surface tractable. The carefully extracted and referenced free energies then act as the learned interface between scales, forming the basis of a simple mean-field microkinetic model, which rationalizes the experimentally measured growth kinetics as controlled by both precursor attachment and availability. The observed interplay between the atomic and meso scales further lets us rationalize why the liquid catalyst yields higher-quality graphene than solid Cu. The approach demonstrated here is applicable to a wide range of phenomena at solid and liquid catalytic interfaces, offering a general route from atomistic energetics to experimentally comparable growth and reaction kinetics.

12:30h: Feliks Nüske (MPI Magdeburg)

Consistent Projection of Langevin Dynamics: Preserving Thermodynamics and Kinetics in Coarse-Grained Models

Coarse graining (CG) is an important task for efficient modeling and simulation of complex multi-scale systems, such as the conformational dynamics of biomolecules. This work presents a projection-based coarse-graining formalism for general underdamped Langevin dynamics. Following the Zwanzig projection approach, we derive a closed-form expression for the coarse grained dynamics. In addition, we show how the generator Extended Dynamic Mode Decomposition (gEDMD) method, which was developed in the context of Koopman operator methods, can be used to model the CG dynamics and evaluate its kinetic properties, such as transition timescales. Finally, we combine our approach with thermodynamic interpolation (TI), a generative approach to transform samples between thermodynamic conditions, to extend the scope of the approach across thermodynamic states without repeated numerical simulations. Using a two-dimensional model system, we demonstrate that the proposed method allows to accurately capture the thermodynamic and kinetic properties of the full-space model.