Atomic-scale simulation of a bismuth liquid-metal ion source using a universal machine-learning interatomic potential
Myoung Choul Choi, Heejin Lim, Sangwon HyunLiquid-metal ion sources (LMISs) are central to focused-ion-beam technology, yet a predictive atomic-scale model has been prevented by the absence of transferable interatomic potentials for the emitter metals. Here, we show that the universal machine-learning interatomic potential MACE-MP-0, applied without any bismuth-specific fine-tuning, delivers a parameter-free atomic-scale description of a bismuth LMIS. Molecular-dynamics simulations reproduce, within experimental uncertainty, the liquid-Bi pair distribution function (first peak at 3.275 Å, 0.2% from neutron/x-ray experiment, with the full first- and second-shell profile recovered) and the self-diffusion coefficient (1.88 × 10−9 m2 s−1, +8% against first-principles molecular dynamics, bracketed by a Stokes–Einstein estimate), and give a sharp field-evaporation threshold at 6.0 V Å−1. The threshold exceeds the tabulated evaporation field of bismuth by a factor of about three, the same ratio reported for bond-breaking molecular dynamics of Cu, Ni, and W, so the model reproduces a known open discrepancy of this class of simulation rather than escaping it. A 3363-atom nanotip exhibits a sharp stable-to-unstable emission transition whose 18-fold rate jump follows the field-induced collapse of the evaporation barrier; the corresponding emission currents (0.9–112 μA) match the operating range of real Bi-LMIS. Above the threshold, the evaporation rate is linear in field (R2 = 0.9998 over 8–20 V Å−1) rather than exponential, and the seed-to-seed scatter falls from 21% at the threshold to below 1% above it, marking the transition from thermally activated to barrierless, supply-limited emission. The simulated time-of-flight mass spectrum is Bi+-dominated, consistent with a post-field-ionization charge-state analysis, while the underestimated Bi3+ yield is attributed to gas-phase aromatic stabilization outside the neutral model. The results establish foundation potentials as a transferable, physically faithful, and low-cost route to modeling high-field metal emission.