2025 AIChE Annual Meeting

(33i) Ready-to-Use Polymerization Simulations Combining Universal Neural Network Potentials with Time-Dependent Bond Boosting

Authors

Shunsuke Tonogai, Preferred Networks
Yu Miyazaki, Preferred Networks
Akihide Hayashi, Preferred Networks
Polymer materials play crucial roles in various industrial sectors, including electronics, healthcare, energy storage, and structural materials. Their properties are primarily determined by molecular-level chemical reactions such as polymerization and cross-linking. However, directly observing and analyzing these reactions at the atomic scale is experimentally challenging. Therefore, molecular simulations have been effectively utilized to analyze these phenomena. Previously, methods combining pseudo-reaction models or ReaxFF with accelerated molecular dynamics (MD) have been proposed and applied to simulate polymerization reactions of various polymers, including epoxy, styrene, and polyamide. However, these methods face practical challenges such as the necessity for computationally intensive transition state search to determine reaction parameters (e.g., activation energies) and the need for detailed adjustment of reaction acceleration conditions for each individual bond.

In this study, we propose a novel simulation method that fundamentally overcomes these challenges by combining a newly developed time-dependent bond boost method with PFP, one of the universal neural network potentials (uNNPs). Unlike the conventional Bond Boost method [1], which requires complex parameter adjustments for each reaction, our method gradually increases the bias potential over time along the reaction pathway, eliminating the need for individual parameter tuning. Additionally, when the reaction is completed, the bias potential automatically returns to zero, transitioning seamlessly back to standard molecular dynamics simulation. This mechanism allows efficient acceleration of diverse reactions under consistent conditions. Furthermore, unlike metadynamics, our method does not leave residual bias on the stable structures formed after reaction completion, thus avoiding the unintended induction of reverse reactions. The adoption of uNNPs such as PFP significantly enhances the accuracy and versatility of simulations by eliminating the extensive parameter fitting required for traditional reactive force fields. uNNPs readily accommodate diverse chemical environments and elements typically challenging for conventional methods, providing improved computational efficiency suitable for larger-scale and longer-time simulations.

The effectiveness of the proposed method was validated through representative polymerization reactions. Specifically, for epoxy-amine curing reactions using bisphenol A-type epoxy resins with various amine-based curing agents, we successfully reproduced the experimentally known relative curing rates [2]. We also demonstrated applicability to other representative reaction systems, such as radical polymerization of vinyl monomers and titanium chloride-catalyzed coordination polymerization of olefins. These simulations required minimal adjustment of bond-specific parameters compared to conventional methods. Additionally, by employing PFP, we confirmed the capability to simulate systems containing elements difficult to handle with traditional reactive force fields without the need for force-field parameter refitting. Moreover, PFP exhibited faster computation speeds compared to other uNNPs, such as M3GNet and ANI, and supported larger system sizes. It also demonstrated a notable advantage in generating chemically realistic structures following accelerated MD simulations.

[1] A. Vashisth, C. Ashraf, W. Zhang, C.E. Bakis, and A.C.T. van Duin, J. Phys. Chem. A 122(32), 6633–6642 (2018).
[2] M. Pramanik, E.W. Fowler, and J.W. Rawlins, J. Coat. Technol. Res. 11(2), 143–157 (2014).