Year: 2026
Type: preprint
Abstract: Foundation models for atomistic simulations offer broadly applicable potentialenergy surfaces that can either be applied in a zero-shot manner (without domainspecific retraining) or fine-tuned to the task at hand. Their accuracy is typically judged by energy and force errors on near-equilibrium data, yet it remains largely unclear whether these metrics translate into reliable free-energy landscapes. Here, enhanced sampling can be a particularly demanding test: simulations are driven far from equilibrium into sparsely trained regions and held there over long trajectories, so that small systematic errors in forces or relative energies can accumulate and distort predicted populations, transition pathways, and barriers. We assess several molecular foundation models trained on different dataset sizes and architectures (including models with explicit long-range electrostatics) in prototypical metadynamics and umbrella sampling simulations. Our results show that recent foundation models can reproduce simple gas-phase free-energy profiles with near-reference accuracy, while more complex multidimensional and solvated free-energy landscapes remain challenging. However, assessing their performance in these systems is further complicated by limitations in current sampling protocols and the lack of well-defined reference data. These findings highlight the need for more rigorous sampling protocols and reliable benchmarks to systematically assess the accuracy of future molecular foundation models in free-energy calculations.
Source: ChemRxiv
Institution: Fritz Haber Institute of the Max Planck Society
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222631181