The time scales accessible to molecular dynamics simulations is limited by the short integration time steps required for numerical stability, which frustrates comprehensive sampling of configurational phase space needed to compute converged thermodynamic averages, surmount free energy barriers, and simulate rare events. We have maintained a long interest in enhanced sampling techniques and, in particular, the development of methods for data-driven discovery and acceleration. We developed Molecular Enhanced Sampling with Autoencoders (MESA) to interleave successive rounds of variable discovery and enhanced sampling in high-variance CVs and Girsanov Reweighting Enhanced Sampling Technique (GREST) as a means to do the same for slow CVs by appealing to a dynamical reweighting formalism. We developed State-free Reversible VAMPnets (SRVs) as a means to learn slow CVs from simulation trajectories, and developed Latent Space Simulators (LSS) that use three back-to-back deep learning networks to (i) learn the slow collective variables of the system, (ii) propagate the dynamics within this slow latent space, and (iii) generatively reconstruct molecular configurations within a dynamically coarse-grained kinetic model that enables generation of ultra-long simulation trajectories at up to six orders of magnitude lower cost than molecular dynamics. We have made our methods freely available by contributing them to enhanced sampling libraries such as SSAGES and PLUMED.

We are pursuing the following projects in this theme:

• Application of latent space simulators to large and multi-molecular systems
• Data-driven CV discovery and enhanced sampling for condensed phase systems
• Backmapping of configurationally or dynamically coarse-grained systems to restore atomic resolution using denoising diffusion probability models (DDPM)
• Incorporation of memory effects / non-Markovian dynamics into dynamical propagators
• Adaptive sampling techniques for efficient model parameterization