A machine learning enabled hybrid optimization framework for efficient coarse-graining of a model polymer

Abstract

This work presents a framework governing the development of an efficient, accurate, and transferable coarse-grained (CG) model of a polyether material. The framework combines bottom-up and top-down approaches of coarse-grained model parameters by integrating machine learning (ML) with optimization algorithms. In the bottom-up approach, bonded interactions of the CG model are optimized using deep neural networks (DNN), where atomistic bonded distributions are matched. In the top-down approach, optimization of nonbonded parameters is accomplished by reproducing the temperature-dependent experimental density. We demonstrate that developed framework addresses the thermodynamic consistency and transferability issues associated with the classical coarse-graining approaches. The efficiency and transferability of the CG model is demonstrated through accurate predictions of chain statistics, the limiting behavior of the glass transition temperature, diffusion, and stress relaxation, where none were included in the parametrization process. The accuracy of the predicted properties are evaluated in context of molecular theories and available experimental data.

Perspective: Advances, Challenges, and Insight for Predictive Coarse-Grained Models

“…While ML methods have proven useful for parametrizing top-down models to reproduce thermodynamic properties, − they have also recently been harnessed to parametrize physics-based potentials in a bottom-up fashion. For instance, Hajizadeh and co-workers employed a genetic algorithm to parametrize top-down nonbonded potentials that matched temperature-dependent density measurements, while employing an artificial neural network (ANN) to parametrize bottom-up bonded potentials based upon information from united atom polymer simulations . In this work, iterative simulations of trial CG models were avoided by training independent ANNs to predict the conformational and thermodynamic properties of the CG model as a function of the potential parameters.…”

“…It would also be highly desirable to develop a unified, rigorous bottom-up approach for simultaneously optimizing both the CG mapping and the interaction potential for self-consistency with the mapped ensemble. , It may be possible to employ information-theoretic ideas or ML tools to predictively assess this self-consistency without explicit simulations of the CG model. More pragmatically, it may be beneficial to consider simultaneously optimizing the CG mapping and potential not only for accuracy and transferability but also for computational efficiency and simplicity.…”

“…For instance, Hajizadeh and co-workers employed a genetic algorithm to parametrize top-down nonbonded potentials that matched temperaturedependent density measurements, while employing an artificial neural network (ANN) to parametrize bottom-up bonded potentials based upon information from united atom polymer simulations. 154 In this work, iterative simulations of trial CG models were avoided by training independent ANNs to predict the conformational and thermodynamic properties of the CG model as a function of the potential parameters. Similarly, Pavan and co-workers combined top-down nonbonded potentials from the Martini model with bottom-up bonded potentials that were optimized via particle swarm optimization (PSO).…”

Integrating Machine Learning in the Coarse-Grained Molecular Simulation of Polymers

“…Shireen at al. 161 have trained DNNs 61 to predict the nonbonded interaction parameters for a CG model of poly(tetramethylene oxide), targeting both structural correlations extracted from atomistic MD simulations and temperature-dependent density values, in a hybrid bottomup/top-down framework. Ghasemi and Yazdani 162 employed Support Vector Regression 61 and PSO 153 to calibrate the parameters of a CG force field for polyvinyl chloride (PVC) to reproduce the density and stress−strain response of the material.…”

“…Bejagam et al studied the conformations of poly­( N -isopropylacrylamide) in solution using PSO, while Durumeric et al applied the Adversarial Residual CG method to parametrize CG models for two oligopeptides in water, using relative entropy minimization. Shireen at al . have trained DNNs to predict the nonbonded interaction parameters for a CG model of poly­(tetramethylene oxide), targeting both structural correlations extracted from atomistic MD simulations and temperature-dependent density values, in a hybrid bottom-up/top-down framework.…”