[Paper Review] Multiscale Prediction of Polymer Relaxation Dynamics via Computational and Data-Driven Methods

Nguyen T. T. Duyen, Ngo T. Que+1 more|arXiv (Cornell University)|Jan 19, 2026

Material Dynamics and Properties0 citations

TL;DR

The paper integrates MD simulations, machine learning, and ECNLE theory to predict polymer glass transition dynamics and relaxation, using T_g inputs from simulations/ML to compute temperature-dependent relaxation times, fragility, and diffusion.

ABSTRACT

We present a multiscale modeling approach that integrates molecular dynamics simulations, machine learning, and the Elastically Collective Nonlinear Langevin Equation (ECNLE) theory to investigate the glass transition dynamics of polymer systems. The glass transition temperatures (Tg) of four representative polymers are estimated using simulations and machine learning model trained on experimental datasets. These predicted Tg values are used as inputs to the ECNLE theory to compute the temperature dependence of structural relaxation times and diffusion coefficients, and the dynamic fragility. The Tg values predicted from simulations show good quantitative agreement with experimental data. While machine learning tends to slightly overestimate Tg, the resulting dynamic fragility values remain close to experimental fragilities. Overall, ECNLE calculations using these inputs agree well with broadband dielectric spectroscopy results. Our integrated approach provides a practical and scalable tool for predicting the dynamic behavior of polymers, particularly in systems where experimental data are limited.

Motivation & Objective

• Motivate accurate prediction of glass transition temperature (T_g) and relaxation dynamics in polymers.
• Develop a multiscale framework combining molecular dynamics, machine learning, and elastically collective nonlinear Langevin equation (ECNLE) theory.
• Provide T_g inputs from simulations and ML to drive ECNLE predictions across temperatures.
• Validate predictions against experimental data and assess the influence of input accuracy on dynamics.

Proposed method

• Use MD to estimate T_g for PPS, PI, PPG, and PB via cooling and volume analysis.
• Train a Gaussian Process Regression (GPR) model with SMILES-derived fingerprints to predict T_g from a polymer dataset.
• Compute structural relaxation times and diffusion coefficients with ECNLE theory using T_g as thermal mapping input.
• Apply a thermal mapping that relates volume fraction to temperature via T ≈ T_g + (Φ_g − Φ)/(βΦ_0) to enable T-based predictions.
• Explore the non-universal coupling via a_c to adjust the elastic barrier in ECNLE and assess sensitivity of predictions.

Experimental results

Research questions

• RQ1 Can MD simulations and ML-predicted T_g values be used as inputs to ECNLE theory to predict the temperature dependence of polymer relaxation times and fragility?
• RQ2 How do MD and ML estimations of T_g compare to experimental data in terms of driving ECNLE predictions?
• RQ3 What is the impact of including a material-specific coupling parameter a_c on the accuracy of ECNLE-based predictions?
• RQ4 Is a simple thermal mapping sufficient for translating volume fractions to temperatures across different polymers?
• RQ5 Do ECNLE predictions using MD or ML T_g inputs reproduce diffusion behavior observed experimentally?

Key findings

• MD-derived T_g values show good quantitative agreement with experimental data for the studied polymers.
• Machine-learning T_g predictions overestimate some T_g values but still yield dynamic fragility close to experimental results.
• ECNLE calculations using experimental, MD, and ML T_g inputs largely reproduce the temperature dependence of structural relaxation times and diffusion coefficients.
• The dynamic fragility m calculated with MD or ML inputs remains within a reasonable range of experimental fragilities, with ML inputs sometimes offering closer agreement.
• A_c tuning affects the fit to experimental τ_α(T), indicating sensitivity to how collective elastic barriers are modeled within ECNLE.
• Diffusion coefficients computed via ECNLE align well with broadband dielectric spectroscopy results for selected polymers.