Copper fatigue simulations expose a blind spot in foundation potentials: NequIP's intrinsic stacking-fault energy is seven times lower than EAM's, and its cracks heal more often
Synopsis
In cyclic-loading simulations of copper, the NequIP foundation interatomic potential agrees broadly with an experimentally vetted EAM potential on near-equilibrium properties, yet predicts an intrinsic stacking-fault energy about seven times lower than EAM, so crack tips emit partial dislocations more readily and flaws heal more often under fatigue. The study runs 72 fatigue simulations per potential on a Griffith crack geometry across three tensile axes, 50 to 500 K and 2 to 5 GPa pre-stress, separating equilibrium properties from emergent behavior that requires external work to trigger.
Interpretation
Cyclic-loading simulations of copper show that an off-the-shelf foundation interatomic potential and a physically constructed empirical potential agree on near-equilibrium properties yet diverge on fatigue, an emergent behavior driven by rare events that external stress must trigger. Foundation machine-learned interatomic potentials are trained on large, diverse DFT datasets, but their validation rests on equilibrium states, which does not guarantee correct emergent behavior in large-scale molecular dynamics. The two agree closely on lattice constant, elastic constants, thermal expansion, surface energies and the unstable stacking-fault energy, which differs by only 1.8%, while NequIP's intrinsic stacking-fault energy is about seven times lower than EAM's.
That difference changes how a crack tip evolves under cyclic load: NequIP favors partial-dislocation emission that blunts or even heals the flaw instead of growing it into a through crack. Across 72 simulations per potential spanning three tensile axes, 50 to 500 K and 2 to 5 GPa pre-stress, 32 EAM and 28 NequIP runs grew through cracks, while 34 EAM and 13 NequIP runs stayed nominal. Flaw healing appeared in 6 EAM and 31 NequIP simulations; EAM healed only with the [100] tensile axis and showed no temperature dependence, whereas NequIP healed along all three tensile directions.
Per-atom energy distributions after loading locate the difference between the potentials in the high-energy region that thermal fluctuations cannot reach, where training data constrain the model least. In an equilibrium crystal the per-atom energy distribution is narrow and unimodal, with a width set by the heat capacity, whereas plastic deformation creates new high-energy local minima. At 50 K both models place at least a quarter of all atoms at energies an order of magnitude above thermal fluctuations, roughly 0.1 eV or more; NequIP shows a population increase just above the thermally accessible energy, while EAM shows a longer high-energy tail at all temperatures.
Perspective
The comparison pits the Mishin EAM potential for copper against the NequIP OAM-S foundation potential in a 50 nm by 50 nm by 4 nm three-dimensionally periodic cell containing a single Griffith-type planar flaw, cycled ten times under strain control at 0.5% amplitude and a 1 ns period. The intended reader is a researcher preparing large-scale molecular dynamics with foundation interatomic potentials; what is measured is the response of this crack geometry under this set of thermodynamic conditions, not a prediction of fracture mechanics.
The authors note that training sets are not designed to include the defect configurations that arise during deformation, so events such as dislocation nucleation at a crack tip may be sparsely represented in the training data. NequIP produces unphysical high-frequency energy fluctuations in low-density, under-coordinated configurations, and such extrapolation cannot be detected from the model's predictions during use. The authors also report simulation crashes from insufficient memory, which dynamic load balancing only partly mitigates, and advise users to run their own stability tests to set a restart checkpoint frequency.
