[Paper Review] Variational Coarse-Graining for Molecular Dynamics

Wujie Wang, Rafael Gómez‐Bombarelli|arXiv (Cornell University)|Dec 6, 2018

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper introduces Autograin, an end-to-end deep learning framework that jointly optimizes coarse-graining mapping and potential energy function using auto-encoders and force-matching. It enables automated, differentiable learning of coarse-grained variables and Hamiltonians, achieving accurate long-timescale simulations on model systems including single molecules and periodic bulk phases.

ABSTRACT

Molecular dynamics simulations provide theoretical insight into the microscopic behavior of materials in condensed phase and, as a predictive tool, enable computational design of new compounds. However, because of the large temporal and spatial scales involved in thermodynamic and kinetic phenomena in materials, atomistic simulations are often computationally unfeasible. Coarse-graining methods allow simulating larger systems, by reducing the dimensionality of the simulation, and propagating longer timesteps, by averaging out fast motions. Coarse-graining involves two coupled learning problems; defining the mapping from an all-atom to a reduced representation, and parametrizing a Hamiltonian over coarse-grained coordinates. Multiple statistical mechanics approaches have addressed the latter, but the former is generally a hand-tuned process based on chemical intuition. Here we present Autograin, an optimization framework based on auto-encoders to learn both tasks simultaneously. Autograin automatically learns the optimal mapping between all-atom and reduced representation, using the reconstruction loss to facilitate the learning of coarse-grained variables. In addition, a force-matching method is applied to variationally determine the coarse-grained potential energy function. This procedure is tested on a number of model systems including single-molecule and bulk-phase periodic simulations.

Motivation & Objective

• To address the computational infeasibility of atomistic molecular dynamics simulations at thermodynamic and kinetic timescales.
• To overcome the reliance on hand-tuned, chemically intuitive mappings in traditional coarse-graining methods.
• To develop a unified, differentiable framework that jointly learns the coarse-grained representation and its corresponding potential energy function.
• To enable accurate, long-timescale simulations of condensed-phase materials through automated, variational coarse-graining.

Proposed method

• Uses an auto-encoder architecture to learn a differentiable, nonlinear mapping from all-atom to coarse-grained configurations.
• Minimizes reconstruction loss to ensure the coarse-grained representation preserves essential structural and dynamic features of the all-atom system.
• Applies a variational force-matching method to learn the coarse-grained potential energy function by minimizing the difference between coarse-grained and all-atom forces.
• Optimizes both the mapping and potential simultaneously via backpropagation through the entire computational graph.
• Validates the framework on single-molecule and periodic bulk systems, including systems with complex interactions and long relaxation times.

Experimental results

Research questions

• RQ1 Can an end-to-end deep learning framework jointly optimize coarse-grained representation and potential energy function more effectively than hand-tuned methods?
• RQ2 How well can auto-encoder-based representations preserve physical and structural features of all-atom systems in diverse molecular environments?
• RQ3 To what extent does variational force-matching improve the accuracy of coarse-grained force predictions compared to standard approaches?
• RQ4 Can the framework enable stable and accurate long-timescale simulations on complex systems like periodic bulk phases?

Key findings

• Autograin successfully learns physically meaningful coarse-grained variables without relying on chemical intuition, as evidenced by reconstruction fidelity and physical consistency.
• The framework achieves accurate force predictions through variational force-matching, enabling stable integration into molecular dynamics engines.
• Simulations using the learned coarse-grained models reproduce long-timescale dynamics and structural properties of the original all-atom systems.
• The method demonstrates robustness across diverse systems, including single-molecule and periodic bulk phases, with consistent performance across different material types.