Dresden 2020 – scientific programme

The DPG Spring Meeting in Dresden had to be cancelled! Read more ...

Parts | Days | Selection | Search | Updates | Downloads | Help

SOE: Fachverband Physik sozio-ökonomischer Systeme

SOE 10: Data Analytics, Extreme Events, Nonlinear Stochastic Systems, and Networks (joint session DY/SOE)

SOE 10.3: Talk

Wednesday, March 18, 2020, 15:45–16:00, ZEU 118

Interpretable Embeddings from Molecular Simulations Using Gaussian Mixture Variational Autoencoders — •Yasemin Bozkurt Varolgunes^{1, 2}, Tristan Bereau¹, and Joseph F. Rudzinski¹ — ¹Max Planck Institute for Polymer Research, Mainz, Germany — ²Koc University, Istanbul, Turkey

Extracting insight from the molecular simulations data requires the identification of a few collective variables (CVs) whose corresponding low-dimensional free-energy landscape (FEL) retains the essential features of the underlying system. Autoencoders are powerful tools for dimensionality reduction, as they naturally force an information bottleneck. While variational autoencoders (VAEs) ensure continuity of the embedding by assuming a Gaussian prior, this is at odds with the multi-basin FELs that typically arise from the identification of meaningful CVs. Here, we incorporate this physical intuition into the prior by employing a Gaussian mixture variational autoencoder (GMVAE), which encourages the separation of metastable states within the embedding. The GMVAE performs dimensionality reduction and clustering within a single unified framework, and is capable of identifying the inherent dimensionality of the input data, in terms of the number of Gaussians required to categorize the data. We illustrate our approach on two toy models and a peptide, demonstrating the anti-clustering effect of the prior relative to standard VAEs. The resulting embeddings stand as appropriate representations for constructing Markov state models, highlighting the transferability of the dimensionality reduction from static equilibrium properties to dynamics.