From Cluster to Bulk: Adaptation Approaches for Equivariant Foundation Potentials

Parameter-efficient adaptation offers a data-efficient route to specializing pretrained machine-learned interatomic potentials (MLIPs), but whether molecular-level gains transfer to reliable bulk behavior is unclear. We adapt two equivariant foundation potentials, MACE-OFF23 (M) and MACELES-OFF (which adds learned long-range electrostatics), using rank-32 LoRA trained on narrow (monomer--dimer) versus broad (monomer-through-hexamer) water-cluster coverage, and trace performance from held-out clusters to liquid density, radial structure, diffusion, and viscosity. Broad coverage cuts held-out force error on multi-neighbor clusters by 61--73\% and reduces mean density error from 17--19\% to 1.6--7.4\% in both foundations. These gains do not transfer reliably to dynamics: despite comparable force accuracy and structural improvement, the two broadly adapted models disagree in shear viscosity by a factor of 2.4, one near experiment and the other dominated by an anomalously persistent shear-stress relaxation mode. Molecular force RMSE and equilibrium structure are therefore insufficient criteria for validating an adapted MLIP's dynamical fidelity, and this transfer depends strongly on the pretrained foundation.