Transferable Coarse-Grained Models: From Hydrocarbons to Polymers, and Backmapped by Machine Learning


TR Number



Journal Title

Journal ISSN

Volume Title


Virginia Tech


Coarse-grained (CG) molecular dynamics (MD) simulations have seen a wide range of applications from biomolecules, polymers to graphene and metals. In CG MD simulations, atomistic groups are represented by beads, which reduces the degrees of freedom in the systems and allows larger timesteps. Thus, large time and length scales could be achieved in CG MD simulations with inexpensive computational cost. The representative example of large time- and length-scale phenomena is the conformation transitions of single polymer chains as well as polymer chains in their architectures, self-assembly of biomaterials, etc. Polymers exist in many aspects of our life, for example, plastic packages, automobile parts, and even medical devices. However, the large chemical and structural diversity of polymers poses a challenge to the existing CG MD models due to their limited accuracy and transferabilities. In this regard, this dissertation has developed CG models of polymers on the basis of accurate and transferable hydrocarbon models, which are important components of the polymer backbone. CG hydrocarbon models were created with 2:1 and 3:1 mapping schemes and their force-field (FF) parameters were optimized by using particle swarm optimization (PSO). The newly developed CG hydrocarbon models could reproduce their experimental properties including density, enthalpy of vaporization, surface tension and self-diffusion coefficients very well. The cross interaction parameters between CG hydrocarbon and water models were also optimized by the PSO to repeat the experimental properties of Gibbs free energies and interfacial tensions. With the hydrocarbon models as the backbone, poly(acrylic acid) (PAA) and polystyrene (PS) models were constructed. Their side chains were represented by one COOH (carboxylic acid) and three BZ beads, respectively. Before testing the PAA and PS models, their monomer models, propionic acid and ethylbenzene, were created and validated, to confirm that the cross interactions between hydrocarbon and COOH beads, and between hydrocarbon and BZ beads could be accurately predicted by the Lorentz-Berthelot (LB) combining rules. Then the experimental properties, density of polymers at 300 K and glass transition temperatures, and the conformations of their all-atom models in solvent mixtures of water and dimethylformamide (DMF) were reproduced by the CG models. The CG PAA and PS models were further used to build the bottlebrush copolymers of PAA-PS and to predict the structures of PAA-PA in different compositions of binary solvents water/DMF. Although CG models are useful in understanding the phenomena at large time- or length- scales, atomistic information is lost. Backmapping is usually involved in reconstructing atomistic models from their CG models. Here, four machine learning (ML) algorithms, artificial neural networks (ANN), k-nearest neighbor (kNN), gaussian process regression (GPR), and random forest (RF) were developed to improve the accuracy of the backmapped all-atom structures. These optimized four ML models showed R2 scores of more than 0.99 when testing the backmapping against four representative molecules: furan, benzene, naphthalene, graphene.



Molecular dynamics simulations, Coarse-graining, Soft materials, Machine learning