Machine Learning-Assisted Hybrid ReaxFF Simulations.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Traditional reactive force field (ReaxFF) molecular dynamics (MD) simulations are limited by computational cost for long time scales and large systems.
• Simulating complex chemical phenomena, such as polymer cross-linking, often requires methods that overcome these limitations.

Purpose of the Study:

• To develop and validate a novel machine learning (ML)-assisted Hybrid ReaxFF simulation method (Hybrid/Reax).
• To enable simulations of phenomena requiring longer time scales and/or larger systems than accessible with standard ReaxFF.
• To demonstrate the method's efficiency in simulating polymer cross-linking reactions.

Main Methods:

• Alternating reactive and non-reactive MD simulations assisted by ML models.
• Employing a specialized tracking tool to accelerate chemical reactions during reactive stages.
• Utilizing ML models to predict non-reactive force field parameters based on updated bond topology between simulation stages.
• Applying the Hybrid/Reax method to simulate the cross-linking of polyethylene (PE) analogue with dicumyl peroxide (DCP).

Main Results:

• Successfully simulated cross-linking reactions in a polyethylene matrix analogue (decane) with dicumyl peroxide (DCP).
• Achieved over 20 million MD steps on a system of 4660 atoms, observing cross-linking in over half of the initial 80 PE molecules.
• Completed the simulation in under 48 hours on a single Xeon processor, a task estimated to take ~1 month with pure ReaxFF MD.

Conclusions:

• The Hybrid/Reax method significantly accelerates MD simulations for complex chemical reactions.
• This ML-assisted approach makes previously computationally prohibitive simulations feasible.
• Hybrid/Reax offers a powerful tool for studying polymer cross-linking and other phenomena requiring extended simulation times.