Lunch Talk: From Motion to Meaning: Learning the Statistical Structure of Atomic Interactions

From Motion to Meaning: Learning the Statistical Structure of Atomic Interactions

Join us for our Lunch Talk in November 2025. Stefan Chmiela from the Machine Learning for Molecular Simulation in Quantum Chemistry group will talk about “From Motion to Meaning: Learning the Statistical Structure of Atomic Interactions”. In this talk Stefan Chmiela will introduce generative, physics-informed models that learn the coarse-grained evolution of atomic systems, enabling fast, thermodynamically consistent simulations that capture slow dynamics and extend quantum-level accuracy to realistic materials and molecular environments.

During this Lunch Talk, Stefan Chmiela from BIFOLD will talk about “From Motion to Meaning: Learning the Statistical Structure of Atomic Interactions.” In his talk he introduces generative, physics-informed models that learn the coarse-grained evolution of atomic systems, enabling fast, thermodynamically consistent simulations that capture slow dynamics and extend quantum-level accuracy to realistic materials and molecular environments.

The Lunch Talk takes place at the TU Berlin. For further information on the Lunch Talks and registration, contact Dr. Laura Wollenweber via email.

Abstract:  Atomistic simulations bridge quantum mechanics and observable phenomena, enabling predictive insight into the behavior of materials and molecules. At their foundation lie force fields: models that describe how atoms interact. Machine-learned force fields have transformed this field, extending quantum-level accuracy from idealized systems to realistic environments such as liquids, interfaces, and biomolecules.

However, the traditional paradigm of explicitly tracking atomic motion faces a fundamental bottleneck. Simulations must advance in femtosecond time steps to resolve high-frequency vibrations, making the exploration of slow dynamics and rare events in large systems computationally prohibitive.

To move beyond these constraints, models must capture the statistical structure of atomic interactions and the mechanisms by which equilibrium emerges. Recent generative and self-consistent modeling principles now make this possible.

This talk will present new paradigms for atomistic simulation that enable learning the effective evolution of atomic configurations over coarse-grained time intervals. These approaches accelerate sampling while preserving thermodynamic consistency and accurately representing slow kinetics. By uniting physical laws with generative modeling, they open the door to faster, physically rigorous simulations and a deeper understanding of complex molecular and materials systems.

BIO: Stefan Chmiela from the Machine Learning for Molecular Simulation in Quantum Chemistry group -- > Linkedin