Reweighting Free Energy Profiles with Universal Machine Learning Interatomic Potentials for Fast Consensus Building

Universal machine learning interatomic potentials (MLIPs) have made density functional theory (DFT)-level molecular dynamics highly affordable. However, their rapid proliferation makes it prohibitively expensive to benchmark every new model against the potential of mean force (PMF), the bridge between microscopic atomic fluctuations and macroscopic thermodynamic and kinetic observables. Reweighting the trajectory of a single “source” MLIP onto a suite of target potentials offers a cheaper route, but it is hindered by the statistical collapse of exponential free energy perturbation when the phase-space overlap between potentials is low. We show that for system sizes of practical interest, this collapse is essentially unavoidable, and we quantify it with an information-theoretic diagnostic, the reweighting entropy, that indicates when reweighting can be trusted. To sidestep the collapse, we introduce a systematic, scalable reweighting framework spanning exponential reweighting, the Gaussian approximation, and the energy-only estimator. When applied to the intramolecular radical cyclization of hexenyl (17 atoms), all three estimators perform well and reproduce the target ωB97M-V/def2-TZVP PMF. However, when extended to Li+ transport through a water-solvated zeolite (601 atoms), only the energy-only estimator recovers the PMFs of a diverse suite of MLIPs from a single source simulation by discarding the noisy entropic contribution to the free energy correction. An independent, brute-force umbrella sampling of one of the most divergent targets (MACE-MATPES-r2SCAN) for this system confirms it: beyond the shape of the PMF, both the reaction and activation free energies are recovered within chemical accuracy. The same analysis reveals that the studied MLIPs partition into two clusters, one of plain generalized gradient approximation (GGA) models (PBE, PBE-sol) and one of meta-GGA and/or dispersion-corrected models (PBE-D3, r2SCAN, r2SCAN-D4), reflecting differences in their training data and DFT reference. The resulting protocol thus offers an affordable and reliable route to cross-model consensus on materials properties, even when exact reweighting is statistically intractable.