Wujie Wang1 Rafael Gomez-Bombarelli1

1, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States

Molecular dynamics (MD) simulations provide key insights into the microscopic behavior of materials and, as a predictive tool, allow theory-driven design of functional materials. However, because of the large temporal and spatial scales involved in thermodynamic and kinetic phenomena in materials, atomistic simulations are often computationally infeasible. Coarse graining methods are widely used to tackle the challenge of expensive atomistic simulations. Current coarse graining methods require intensive manual tuning to determine both the coarse graining rules and the force field parametrization [1]. Inspired by recent works that achieve deep learning based accelerated calculation in Density Functional Theory [2], molecular kinetics [3] and free energy landscape calculations [4], we propose an alternative data-driven framework utilizing auto-encoders to map the atomistic trajectory into a latent space of coarse grained “super-atoms”. By training on data from atomistic molecular dynamics trajectories or electronic structure calculations, this approach adopts stochastic gradient optimization of the coarse-graining rules, the force field parametrization and the up-resolution rules. Thus, it can provide parameters for large scale coarse-grained calculations and also map the coarse-grained trajectory back to dynamic information with atomistic details. Through this approach, systematic coarse graining pipelines can be built for fast molecular dynamics simulations and high-throughput predictions of the thermodynamics and kinetics of materials.

[1]WG Noid, Perspective: Coarse-grained models for biomolecular systems WG Noid The Journal of Chemical Physics 139, 090901 (2013)
[2]Chmiela, S. et al. Machine learning of accurate energy-conserving molecular force fields. Sci. Adv. 3, e1603015 (2017).
[3] Wehmeyer, C. & Noé, F. Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics. (2017).
[4]Schneider, E., Dai, L., Topper, R. Q., Drechsel-Grau, C. & Tuckerman, M. E. Stochastic Neural Network Approach
for Learning High-Dimensional Free Energy Surfaces. Phys. Rev. Lett. 119, 150601 (2017).