Systematic control of collective variables learned from variational autoencoders.

Area of Science:

• Computational Chemistry
• Machine Learning in Molecular Dynamics

Background:

• Variational autoencoders (VAEs) are increasingly used in molecular simulations to derive low-dimensional latent representations.
• Understanding the link between VAEs' probabilistic structure, learned information, and loss functions is crucial but remains unclear.
• Prior research often relied on feature engineering or ad hoc modifications.

Purpose of the Study:

• To investigate how the decoding model's structure influences the learned latent coordinates in VAEs.
• To provide practical guidance for optimizing VAEs for molecular simulations.

Main Methods:

• Applied flexible priors using normalizing flows.
• Systematically varied the power and flexibility of the decoding distribution.
• Adjusted weights on individual VAE loss function terms.

Main Results:

• The structure of the decoding distribution significantly impacts the latent space.
• Learned latent coordinate structure is sensitive to decoding model flexibility.
• VAE loss function weighting allows for tuning the level of encoded detail.

Conclusions:

• Decoding model architecture is a key factor in VAE latent space properties for molecular simulations.
• VAE loss function tuning enables control over information resolution.
• This work offers practical insights for applying VAEs to molecular dynamics and Monte Carlo simulations.