Understanding twist-disorder of polytetrafluoroethylene (PTFE) chains using neural network potential molecular dynamics

Molecular dynamics (MD) calculations are used to characterize the temperature-dependent evolution of helicity and twist disorder in polytetrafluoroethylene (PTFE) chains. The MD calculations employ machine-learned neural network potentials (NNPs) trained on density functional theory data. In particular, NNPs were derived using three functionals: PBE-D3, r2SCAN, and r2SCAN-rVV10. As temperature increases, for PBE-D3 and r2SCAN-rVV10, we observe gradual unwinding from the equilibrium helix to less twisted, increasingly disordered conformations, with distinct discontinuous transitions near 250 K (PBE-D3) and 410 K (r2SCAN-rVV10). Notably, the PBE-D3 NNP captures these transitions and associated structural changes most accurately, whereas the r2SCAN-rVV10 NNP shows similar phase behavior with signs of overbinding. In contrast, the NNP trained on r2SCAN suppresses the unwinding transition and produces tighter helical structures with increasing temperature, underscoring the strong dependence on the choice of functional. The onset of chain rotational motion is accompanied by diffusion along the chain axis, supporting the Klug-Franklin hypothesis of screw-like disorder. Observed helical unwinding occurs after the emergence of conformational disorder and correlates with the onset of rotational and translational mobility, revealing a progression from structural to dynamical disorder upon heating. Computed PTFE melt densities highlight long-range van der Waals (vdW) interactions in the reference data used to train the NNP, motivating the incorporation of long-range vdW energies into the training scheme via an analytical Ewald summation for damped dispersion energies. Our MD results provide molecular-scale insights into the structural and dynamical behavior of PTFE, especially the thermally induced twist disorder, and demonstrate the utility of our approach for exploring other fluorocarbon macromolecules.