Write a PREreview

Foundation machine-learning interatomic potentials (MLIPs) promise cheap, near-DFT energetics, but their reliability for defect kinetics in magnetic transition-metal sulfides is largely untested. We benchmark MACE-MP-0 (medium) and CHGNet-v0.3.0, zero-shot, against plane-wave DFT (Quantum ESPRESSO) for vacancy-anchored proton migration across four iron sulfides (pyrite, marcasite, mackinawite, greigite), plus a pentlandite structural-motif diagnostic. The proton is modelled as a neutral H0 defect in dry bulk supercells, a benchmark proxy rather than an interface calculation. A single unified PBE (U = 0) protocol gives the DFT reference barriers: pyrite V_S2 dimer hop 95 meV and V_Fe + S–H 268 meV; mackinawite 43 meV; marcasite 208 meV; greigite 1.86 eV (ferrimagnetic). We report the first harmonic zero-point corrections for *bulk proton migrationthe stiff reactant S–H stretch becomes the imaginary saddle mode—lowering rate-relevant barriers by 66–94 meV and rendering mackinawite effectively barrierless. The foundation potentials fail in four distinct, diagnosable ways: a flat or wrong-topology PES (pyrite V_S2), [Fe4S4] cubane motif collapse (pentlandite), unphysical energies and non-convergence on the inverse spinel (greigite), and false-positive band-collapse warnings (marcasite); where they converge (pyrite V_Fe) they underbind DFT by 1.2–1.5×. Greigite’s barrier places bulk H migration in the kinetically-forbidden regime (τ ≈ 90 Gyr), excluding inverse-spinel Fe–S phases from proton conduction. Foundation MLIPs are reliable only as topology screeners, with mandatory pre-flight checks, and cannot replace DFT for quantitative barriers in magnetic, defect-rich sulfides. An ASE IDPP minimum-image-convention fix is contributed upstream; the spin-polarized pentlandite barrier is a companion study.