Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On-the-Fly Machine-Learned Force Fields for High-Fidelity Polymer Glass Transition Simulations

Ashutosh Srivastava, Sakshi Agarwal, Shivank Shukla, Harikrishna Sahu, Rampi Ramprasad

TL;DR

This work tackles the long-standing challenge of predicting polymer glass transition temperatures ($T_g$) with first-principles accuracy for large, disordered systems. It introduces an adaptive on-the-fly force-field learning framework that builds robust MLFFs from roughly 1000 AIMD configurations and uses Bayesian uncertainty to trigger additional first-principles calculations only when needed. Applied to twelve diverse polymers, the method achieves $T_g$ values in excellent agreement with experiments while delivering about a $10^6$× speedup over AIMD and scaling approximately as $N^{0.90}$ with system size. This approach establishes a general, transferable, and scalable route to accurate thermophysical predictions in complex polymeric materials at quantum-mechanical fidelity, enabling large-scale studies of Tg, density, and thermal expansion across chemistries and architectures.

Abstract

Predicting polymer glass transition temperatures (Tg) with first-principles fidelity has long remained out of reach, as cooling multi-thousand-atom systems over a broad temperature range at acceptable rates exceeds the computational limits of ab initio molecular dynamics (AIMD). Here we employ a hybrid scheme that merges AIMD with accelerated on-the-fly (OTF) machine-learned force-field (MLFF) construction, enabling Tg prediction at quantum-mechanical accuracy with near-classical computational cost. The OTF protocol to construct MLFFs adaptively triggers first-principles calculations only when newly encountered configurations lie outside the current model's domain of confidence, allowing robust, parameter-free MLFFs to be built from merely 1000 AIMD-sampled configurations per polymer. These MLFFs are then utilized to perform long-time cooling simulations on amorphous supercells containing several thousand atoms. Applied across twelve polymers spanning aromatic, aliphatic, heteroatomic, and branched chemistries, the method yields predictions in excellent accord with experiment while reducing computational cost by approximately six orders of magnitude relative to AIMD. This work establishes a new paradigm for predictive polymer modeling, demonstrating that OTF-MLFFs provide a generalizable, accurate, and scalable route to simulating the thermophysical behavior of complex disordered materials at quantum-mechanical fidelity.

On-the-Fly Machine-Learned Force Fields for High-Fidelity Polymer Glass Transition Simulations

TL;DR

This work tackles the long-standing challenge of predicting polymer glass transition temperatures () with first-principles accuracy for large, disordered systems. It introduces an adaptive on-the-fly force-field learning framework that builds robust MLFFs from roughly 1000 AIMD configurations and uses Bayesian uncertainty to trigger additional first-principles calculations only when needed. Applied to twelve diverse polymers, the method achieves values in excellent agreement with experiments while delivering about a × speedup over AIMD and scaling approximately as with system size. This approach establishes a general, transferable, and scalable route to accurate thermophysical predictions in complex polymeric materials at quantum-mechanical fidelity, enabling large-scale studies of Tg, density, and thermal expansion across chemistries and architectures.

Abstract

Predicting polymer glass transition temperatures (Tg) with first-principles fidelity has long remained out of reach, as cooling multi-thousand-atom systems over a broad temperature range at acceptable rates exceeds the computational limits of ab initio molecular dynamics (AIMD). Here we employ a hybrid scheme that merges AIMD with accelerated on-the-fly (OTF) machine-learned force-field (MLFF) construction, enabling Tg prediction at quantum-mechanical accuracy with near-classical computational cost. The OTF protocol to construct MLFFs adaptively triggers first-principles calculations only when newly encountered configurations lie outside the current model's domain of confidence, allowing robust, parameter-free MLFFs to be built from merely 1000 AIMD-sampled configurations per polymer. These MLFFs are then utilized to perform long-time cooling simulations on amorphous supercells containing several thousand atoms. Applied across twelve polymers spanning aromatic, aliphatic, heteroatomic, and branched chemistries, the method yields predictions in excellent accord with experiment while reducing computational cost by approximately six orders of magnitude relative to AIMD. This work establishes a new paradigm for predictive polymer modeling, demonstrating that OTF-MLFFs provide a generalizable, accurate, and scalable route to simulating the thermophysical behavior of complex disordered materials at quantum-mechanical fidelity.
Paper Structure (5 sections, 1 equation, 5 figures)

This paper contains 5 sections, 1 equation, 5 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Discussion
  4. Conclusion
  5. Methods

Figures (5)

  • Figure 1: Adaptive on-the-fly machine-learning force-field (OTF-MLFF) framework for polymer thermophysical prediction. (a) Stepwise protocol for generating MLFFs, beginning with initial AIMD equilibration, Bayesian error estimation, automated dataset expansion, and real-time force-field refinement. (b) Workflow for predicting polymer glass transition temperature ($T_g$) and subsequent long-time ML-driven cooling trajectories enabling high-resolution volume–temperature (V-T) analysis to obtain density ($\rho$), $T_g$, and volume thermal expansion coefficient ($\alpha_V$).
  • Figure 2: Thermodynamic behavior and computational performance of machine-learned force-field simulations across system sizes. (a) Density as a function of simulation time for different integration step sizes ($\Delta$t = 0.3, 0.4, and 0.5 fs) at 600 K. (b) Summary of timestep selection, number AIMD enabled OTF-MLFF steps required for a 200 ps trajectory, configurations sampled, and total CPU time (hours). (c–f) Volume–temperature relations for polyethylene systems containing 206, 728, 1648, and 5562 atoms constructed from multiple chains of varying lengths.
  • Figure 3: Volume–temperature behaviour of the polymers considered in this study. The corresponding polymer structures are shown in each panel. Red and blue markers with shaded gray regions denote the average volumes and standard deviations obtained from the final 50 ps of the production trajectory. Dashed green lines with shaded orange bands indicate the calculated glass-transition temperatures ($T_g$) and their 66.67% confidence intervals. Experimental average $T_g$ values are indicated by red dotted lines, and the shaded blue region indicates the experimental range of the reported $T_g$ values for comparison.
  • Figure 4: Comparison of OTF-MLFF approach predictions with reference data. (a) Densities at 300 K and (b) glass-transition temperatures ($T_g$) for the polymers investigated, benchmarked against experimental measurements and classical force-field results (PCFF and GAFF2). For the last polymer (poly(N-methyl-2-vinylbenzamide)), PCFF and GAFF2 simulations were not possible due to the unavailability of classical force-field parameters.
  • Figure 5: Scaling behaviour and computational benchmarking of the simulations. (a) log-log plot of simulation time with system size ($N$), (b) effective exponent ($p_{eff}$) as a function of $N$, (c) simulation time per atom versus $N$, and (d) Comparison of total simulation times for OTF-MLFF, the Polymer Consistent Force Field (PCFF), the General AMBER Force Field (GAFF2), and ab initio molecular dynamics (AIMD).