Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 33: Machine Learning in Dynamics and Statistical Physics I
DY 33.6: Talk
Thursday, March 20, 2025, 10:45–11:00, H47
Sampling rare events with neural networks: Machine learning the density of states — •Moritz Riedel1, Johannes Zierenberg2, and Martin Weigel1 — 1Institute of Physics, Technische Universität Chemnitz, 09107 Chemnitz, Germany — 2Max Planck Institute for Dynamics and Self-Organization, 37077 Göttingen, Germany
Neural networks can be trained to generate samples from the Boltzmann distribution of many-particle systems. If suitable architectures such as normalizing flows or variational autoregressive networks are chosen, exact generation weights are known and hence present biases can be corrected for. Still, such networks typically struggle to learn and reproduce configurations from the full range of configuration space since effects such as mode collapse occur. For the simulation of rare events and suppressed states accessible in generalized frameworks such as the multicanonical ensemble such broad exploration is crucial. Here, we show how a combination of variational autoregressive network and autoencoder allows for a systematic exploration of configuration space in spin models, during which the network is able to learn the density of states. We demonstrate the efficacy of the approach in the Potts system in the strong first-order regime.
Keywords: Multicanonical ensemble; Autoencoder; Variational autoregressive network; Wang-Landau algorithm