Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 5: Machine Learning in Dynamics and Statistical Physics I
DY 5.3: Vortrag
Montag, 18. März 2024, 10:00–10:15, BH-N 243
Sampling free energies with deep generative models — •Maximilian Schebek1, Michele Invernizzi2, Frank Noé1,2,3,4, and Jutta Rogal5,1 — 1Department of Physics, Freie Universität Berlin, 14195 Berlin, Germany — 2Department of Mathematics and Computer Science, Freie Universität Berlin, 14195 Berlin, Germany — 3Department of Chemistry, Rice University, 77005 Houston, Texas, USA — 4AI4Science, Microsoft Research, 10178 Berlin, Germany — 5Department of Chemistry, New York University, New York, NY 10003, USA
Evaluating free energy differences is a computationally demanding task, that requires a rigorous sampling of the phase space. Here, we train generative machine learning models based on normalizing flows to map between probability distributions of condensed phase systems at different thermodynamic conditions. Using the trained model, uncorrelated configurations can easily be generated. The model architecture incorporates permutation invariance and periodic boundary conditions, which improves convergence and enables the treatment of solid and liquid systems on the same footing. Training the flow model in a conditional way allows us to assess free energy differences over a wide range of temperatures and pressures, needed to evaluate the relative stability of different phases and reconstruct phase diagrams. The developed approach is applied to determine the coexistence line between liquid and solid as well as two different solid phases of a Lennard-Jones system. Our results are in excellent agreement with state-of-the-art methods, while the computational costs are significantly reduced.
Keywords: Statistical mechanics; Deep Learning; Generative models; Free energy; Phase diagrams