Teach the bonds, then test the sheet.
For C₂N, a learned potential brings heat transport closer to a first-principles reference and changes the predicted failure process. Adding just 2% vacancies then exposes a weakness that early elastic stretching largely conceals.
International Journal of Heat and Mass Transfer · 2021 · S Arabha, A Rajabpour
A sheet with holes, and a question about its atoms
C₂N is a porous carbon–nitrogen sheet with promising electronic and mechanical properties. Predicting its thermal behaviour is difficult because the pores scatter heat-carrying phonons, and predictions depend strongly on how atomic interactions are represented.
Can a machine-learning potential trained on first-principles data describe both C₂N’s heat conduction and its failure under tensile loading better than a classical potential?
Build the model before trusting the prediction
The authors passively fitted a machine-learning potential to ab initio molecular-dynamics configurations spanning a broad temperature range. They checked its phonon behaviour against density-functional calculations, ran non-equilibrium molecular dynamics for length-dependent thermal conductivity, and simulated uniaxial tension with and without vacancies.
Ab initio molecular-dynamics trajectories from 50 to 1000 K, including strained configurations, supply the passive training set for a moment tensor potential. The authors compare its phonon dispersion with DFT, then use the potential in non-equilibrium heat-flow calculations and uniaxial tension. Results are compared with Tersoff and cited first-principles calculations. Conductivity is extrapolated to a long-sheet limit; it is not the value of every finite specimen.
Key findings
Heat-transport validation
The learned-model estimate is close to the cited DFT-BTE reference, while Tersoff gives 63.3 W/m·K. Phonon modes supply a supporting physical test.
Tensile response
The learned potential predicts 390 ± 3 GPa stiffness and 42 ± 4 GPa ultimate strength, with a different post-yield response from Tersoff.
Defect sensitivity
The modest stiffness penalty conceals a substantially larger change in the failure response.
A reliable model has to follow more than one property
The paper makes a case for validating an atomistic model against both heat-transport and mechanical data: a potential that performs well for one property may not capture the other or the effects of defects.
For this sheet, the learned potential moves the heat-flow and stiffness predictions closer to the cited first-principles references and permits a study of vacancy-driven failure. The evidence also makes clear why conductivity alone is an incomplete test of a material model.