Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 46: Poster: Machine Learning and Data Analytics
DY 46.4: Poster
Thursday, March 30, 2023, 13:00–16:00, P1
Neural-network based Monte Carlo Markov chain simulation of spin glasses — •Michael Engbers and Alexander K. Hartmann — Carl von Ossietzky University, Oldenburg, Germany
Spin glasses exhibit a complex equilibrium and non-equilibrium behavior at low temperatures. The reason is the existence of an energy landscape with many local minima and high barrieres. In computer simulations, this leads to long correlation times when investigating large systems. Due to this numerical hardness, the model has motivated the development of many new algorithmic approaches like generalized Wolff cluster algorithms, parallel tempering or genetic algorithms.
Recently, it has been shown that the application of generative neural networks can accelerate Monte Carlo simulations, also for simple spin models with apparently promising results.
Here, we use an autoregressive distribution estimator (NADE) to perform a Monte Carlo simulation of spin glasses [1]. We embedded the NADE into a Metropolis-Hastings Markov-chain approach, therefore ensuring detailed balance. We confirm previous results that the acceptance rates of the NADE approach surprisingly increase with decreasing temperature. Nevertheless, we show that crucial observables, such as the distribution of spin overlaps, indicate that this neural-network approach suffers from the lack of effective ergodicity.
[1] B. McNaughton, M.V. Milosević, A. Perali, and S. Pilati, Phys. Rev. E 101, 053312 (2020).