Reactive Machine-Learned Potentials for Fluoropolymer Binders: Unified Physical and Chemical Property Validation

Abstract

Accurate atomistic simulation of polymer-bonded explosive (PBX) initiation demands a force field that simultaneously captures mechanical deformation, thermal response, and chemical bond scission-yet no such description has existed for fluoropolymer binders. Classical force fields reproduce equilibrium physical properties but cannot describe bond breaking, and ReaxFF contains no parameterization for F23series fluoropolymers, leaving binder degradation and explosive decomposition inaccessible to molecular simulation. Here we report reactive machine learning potentials (MLPs) for three fluoropolymer binders widely used in PBXs: F2311, F2313, and F2314. Trained on density functional theory (DFT) datasets via the deep potential framework with iterative active learning, the MLPs achieve ab initio accuracy at a fraction of the computational cost. Validation spans both physical and chemical regimes: radial distribution functions, equilibrium densities, glass transition temperatures (Tg), and elastic constants agree quantitatively with experiment and ab initio molecular dynamics benchmarks. A microscopic analysis of backbone torsional transition events further reveals a positive correlation between dihedral activation energy and Tg, providing atomic-scale mechanistic insight into their thermal-mechanical behavior. Critically, the MLPs faithfully reproduce DFT atomic forces throughout bond scission under uniaxial tension, with bond dissociation energies deviating only 0.14-0.15 eV from DFT reference values. This work fills a critical gap in reactive force field coverage for F23-series fluoropolymer binders and establishes a validated framework spanning physical and reactive property regimes, providing a foundation for future atomistic studies of PBX initiation.