Learning Thermal Response Forces: A Method for Extending the Thermodynamic Transferability of Coarse-Grained Models via Machine-Learning

TL;DR

This work proposes a novel and data-efficient means of learning temperature dependence into ML CG force-fields via training on the thermal response forces of the PMF and demonstrates how incorporating these terms into ML CG FFs confers significantly improved transferability for CG water models.

Abstract

Machine-learned (ML) coarse-grained (CG) models are a promising tool for significantly enhancing the efficiency of molecular simulations by systematically removing degrees of freedom while retaining fidelity to the underlying fine-grained model. The CG potential of mean force (PMF) is inherently dependent on thermodynamic conditions and, hence, a CG force-field (FF) which is trained at one thermodynamic state point is not necessarily accurate at another. We propose, in this work, a novel and data-efficient means of learning temperature dependence into ML CG force-fields via training on the thermal response forces of the PMF. We demonstrate how incorporating these terms into ML CG FFs confers significantly improved transferability for CG water models and demonstrate how this transferability enables accurate and predictive CG dynamics.

Figures (14)

• Figure 1: Parity plots for training of (a) PMF forces and (b) entropy forces. Forces are scaled by $T_0$ for dimensional equality between all quantities shown.
• Figure 2: RDF of CG models ($T_0 = 300$ K). RDFs of mapped AA and CG models at (a) trained temperature, (b) T = 250 K, and (c) T = 700 K are shown. (d) RDF error across simulated temperatures for all CG models.
• Figure 3: Change in local environment for mapped AA model relative to 300 K, assessed via the JS divergence between order parameter distributions at a given temperature and 300 K. Order parameters consist of coordination numbers within first and second solvation shell and tetrahedral order parameter.
• Figure 4: (a) Diffusion behavior across temperature for all models are shown. (b) Linear fit of the log ratio of diffusion constants between mapped AA and TCG2 model as a function of $T$ is shown.
• Figure S1: ADF behavior of developed CG models ($T_0 = 300$ K). ADFs of mapped AA and CG models at (a) trained temperature, (b) T = 250 K, and (c) T = 700 K are shown. (d) ADF error across simulated temperatures for all CG models.