Machine learning simulations uncover how plastic transforms into hard carbon

Seventy-five years after Rosalind Franklin’s foundational research into non-graphitising carbons, chemists at the University of Oxford have resolved the decade-old mystery of how hard carbon forms atom by atom

Published in Chemical Science, the study utilises machine learning-driven atomistic simulations to model the thermal breakdown of polyvinylidene chloride (PVDC), revealing the structural mechanisms that prevent hard carbon from crystallising into ordered graphite.

The insights provide a computational framework for designing customised anode materials for next-generation sodium-ion batteries.

Atomic mechanisms and structural non-graphitisability

Hard carbon is a critical component for sodium-ion battery anodes, offering a low-cost, earth-abundant alternative to conventional lithium-ion technologies. Unlike graphite, which consists of flat, stacked sheets of hexagonal carbon rings, hard carbon possesses a curved, disordered structure that retains its amorphous nature even under extreme heat.

By combining quantum-mechanical accuracy with high-speed machine learning algorithms, the Oxford team mapped the step-by-step decomposition of the PVDC polymer precursor:

• Dechlorination and hydrogen stripping:

  • Thermal degradation strips hydrogen and chlorine atoms from the PVDC chain, exposing highly reactive carbon sites.

• Cross-linking networks:

  • The active sites link neighbouring polymer fragments into a densely interconnected three-dimensional carbon matrix.

• Curvature-inducing motifs:

  • Instead of forming pure six-membered hexagonal rings (the building blocks of flat graphite), the reacting network incorporates non-hexagonal five- and seven-membered carbon rings.

• Geometric distortion:

  • As pentagons allow a hexagonal pattern to curve into a spherical football, these non-hexagonal rings warp the carbon lattice, preventing planar alignment of graphitic layers.

Research Team and Computational Methodology

The study was led by DPhil researcher Litong Wu, alongside former students Zitong Wu and Dr Zakariya El-Machachi, under the supervision of Professor Volker Deringer at Oxford’s Department of Chemistry.

Overcoming computational barriers

Traditional quantum-mechanical methods are too computationally expensive to simulate high-temperature chemical reactions across thousands of disordered atoms over extended timescales. Machine-learned interatomic potentials solved this by predicting atomic forces at quantum fidelity while accelerating simulation speeds.

Simulating precursor breakdown at atomic resolution lets researchers link initial chemical compositions directly to the final electrochemical performance of battery anodes.