Machine learning traces how hard carbon acquires its bends

A polymer’s route into hard carbon can be followed bond by bond in an Oxford author preprint posted on 19 June. The simulations connect the breakdown of polyvinylidene chloride, or PVDC, with the curved network that emerges.[1]

Learning the neighbourhood around an atom

The study uses MACE, an architecture introduced in a 2022 methods paper. It represents atoms as a graph: each atom has an element and a position, and nearby atoms exchange numerical information. Its calculations retain directions as well as distances and combine information involving several neighbours. The model can therefore distinguish arrangements that a list of separate pair distances would describe inadequately.[2][1]

Geometry imposes rules. Rotate the entire arrangement and its energy should remain unchanged; directional information inside the network must rotate consistently. MACE builds those relationships into its calculations. It learns contributions to an arrangement’s energy from atomic neighbourhoods, providing a reusable description that can be evaluated as the atoms move. The learned object is a model of interactions, rather than a prewritten sequence of chemical reactions.[2]

How the bends appear

The Oxford team trained its model against quantum calculations, then simulated 3,528 atoms across 640 picoseconds. Chlorine detaches, takes hydrogen and leaves as hydrogen chloride. Reactive carbon sites join neighbouring chains. Rearrangement then produces a network containing five- and seven-membered rings alongside hexagons.[1]

Those persistent odd-sided rings help explain resistance to flat, ordered graphite. The simulation accelerates heating and calibrates density while removing gas; its timings describe that computational protocol.[1]

A route out of disorder

Earlier physical experiments make the word “resistance” more precise. In a 2020 study, Jason Fogg and colleagues first carbonised PVDC, then applied repeated short heating pulses reaching 3,000°C. After six pulses, X-ray diffraction indicated that roughly 90% of the sample was graphite. Electron microscopy and Raman measurements supported the change in ordering.[3]

The proposed explanation was selective removal of defects. Oxygen adsorbed between pulses might attack the non-hexagonal structures during the next heating cycle, allowing fragments to rearrange into stacked sheets. The authors described this as a mechanism still requiring an oxygen-controlled test. Their apparatus also limited samples to about 50 milligrams.[3]

The experiment shows that a material’s initial disorder and its later treatment belong in the same account. A carbon classified by its resistance to ordinary heating can behave differently when the heating sequence and chemical environment change.[3]

The Oxford authors next propose comparing PVDC with PVC, which forms soft carbon. That extension needs additional training and validation; battery performance was not tested.[1]

Report an error or suggest a correction