Speaker
Description
Path Integral Monte Carlo (PIMC) is an exact and powerful framework to describe nuclear quantum mechanical properties of different systems with complex electronic structures at finite temperatures. However, those computations come with a slow convergence rate, also a high variance strongly impacts the accuracy of computed observables.
To overcome those challenges, we introduced a generative neural network, more specifically a Variational Autoencoder (VAE) to enhance sampling efficiency by reproduction and generation of imaginary-time configurations, whereas those configurations shall be consistent with the PIMC statistics. By the use of the VAE, we are able to map the high-dimensional imaginary-time configurations onto a high-dimensional latent space, which contains the underlying statistical structure of the PIMC configurations. From the latent space we can sample and generate new configurations.
By the use of the VAE we are able to enable parallel generation of configurations, also we can integrate symmetry-aware architectures to ensure physical properties like translational or rotational invariance.
