After a short break, the AMC Seminar is back. Prof. Phaedon-Stelios Koutsourelakis was appointed as AMC’s new Core Member. He is head of the Data-driven Materials Modeling group at the Department of Engineering Physics and Computation at the TUM School of Engineering and Design, and besides probabilistic coarse-grained models, he works on physics-integrated generative modeling as well as connecting quantum simulations, atomistic methods, and inverse materials design.

Koutsourelakis’ talk on “Energy-based coarse-graining without data” will introduces a data‑free, energy‑based generative method for coarse‑graining that learns directly from atomistic potentials to automatically discover variables, generate equilibrium samples, and accurately reconstruct molecular systems without requiring trajectories.

Date: Tuesday, December 9, 2025, 3:30 pm

Location: MIBE Lecture Hall

Abstract:

Coarse-grained (CG) models simplify molecular simulations, but conventional methods depend on expensive atomistic data and miss unsampled configurations. We present a fully data-free, generative approach to coarse-graining that learns directly from the all-atom potential, i.e. no trajectories are required. The model builds a structured latent space separating slow collective motions from fast local fluctuations, then maps these back to full atomic detail through a learnable bijection. Trained with an energy-based objective and adaptive tempering, it generates independent equilibrium samples and automatically discovers CG variables. Demonstrations on synthetic systems and alanine dipeptide show accurate reconstruction and full mode coverage, pointing to a new, data-free path for molecular coarse-graining and a built-in solution to the back-mapping problem.