(103d) Systematic Control of Collective Variables Learned from Variational Autoencoders

Variational autoencoders (VAEs) are rapidly gaining popularity within molecular simulation for discovering low-dimensional, or latent, representations, which are critical for both analyzing and accelerating simulations. However, it remains unclear how the information a VAE learns is connected to its probabilistic structure, and, in turn, its loss function. Previous studies have focused on feature engineering, ad hoc modifications to loss functions, or adjustment of the prior to enforce desirable latent space properties. By applying priors of effectively arbitrary flexibility via normalizing flows, we focus instead on how adjusting the structure of the decoding model impacts the learned latent coordinate. We systematically adjust the power and complexity of the decoding distribution, observing that this has a significant impact on the structure of the latent space. By also varying weights on separate terms within each VAE loss function, we show that the level of detail encoded can be further tuned. We introduce a suite of broadly applicable metrics for generically assessing latent space structure and usage in order to rapidly characterize the utility of the learned collective variable without the need for in-depth prior knowledge of the system under study. By examining a wide variety of toy systems and molecules, we find that a meaningful latent space will only be learned if an appropriate balance is found between the complexity of the system, the power of the decoder, and the effective regularization strength. Our work provides practical guidance for utilizing VAEs to extract varying resolutions of low-dimensional information from molecular simulations.