Non-covalent Interactions at cm$^{-1}$ Accuracy: Data Efficient Physics-Informed Distillation for Machine Learning Interatomic Potentials
Foundation models in atomistic machine learning encode interaction physics across diverse atomic environments, but whether that structure can be transferred when building specialist potentials at quantum-chemical accuracy remains open. Here we show that knowledge distillation from a pretrained universal machine-learning interatomic potential (MLIP), followed by coupled-cluster fine-tuning with single and double excitations and perturbative triples [CCSD(T)], transfers not only low-cost labels but a physically meaningful prior on interaction length scales, anisotropy, and the repulsive-dispersive balance, which CCSD(T) data then sharpens to quantum-chemical accuracy. For He--benzene, fine-tuning with 30% of the CCSD(T) data outperforms direct training using the full 80%; a 60% reduction in the high-fidelity compute budget. A symmetry-adapted perturbation theory (SAPT)-informed adaptive short-range/long-range architecture further lowers the validation MAE from 0.75 1/cm to 0.49 1/cm. Across a circumarene series of polycyclic aromatic hydrocarbons (PAHs), swapping the MLIP teacher under an otherwise identical pipeline changes the coronene error by an order of magnitude while leaving the larger PAHs stable, direct evidence that distillation transfers physical structure, not labels alone. Together, these results identify the choice of pretrained teacher as a primary design axis for data-efficient quantum-chemical-accuracy potentials, alongside architecture and training protocol.