Machine-learning interatomic potentials achieving CCSD(T) accuracy for van-der-Waals-dominated systems via Δ-learning

Abstract: Machine-learning interatomic potentials (MLIPs) enable large-scale atomistic simulations at moderate computational cost while retaining ab initio accuracy. However, most MLIPs are trained on density-functional theory (DFT), which often falls short of chemical accuracy (1 kcal/mol). Conversely, coupled-cluster methods, particularly CCSD(T), which includes single, double, and perturbative triple excitations, are considered the gold standard of computational chemistry but rarely applied to periodic systems due to their huge computational cost. Here we present a $\Delta$-learning workflow to produce interatomic potentials with CCSD(T) accuracy for periodic systems including van der Waals (vdW) interactions. The procedure combines a dispersion-corrected tight-binding baseline with an MLIP trained on the differences of the target CCSD(T) energies from the baseline. This $\Delta$-learning strategy enables training on compact molecular fragments while preserving transferability. The dispersion interactions are captured by including vdW-bound multimers in the training set; together with the vdW-aware tight-binding baseline, the formally local MLIP attains CCSD(T) accuracy for systems governed by long-range vdW forces. The resulting potential yields root-mean-square energy errors below 0.4 meV/atom on both training and test sets and reproduces electronic total atomization energies, bond lengths, harmonic vibrational frequencies, and inter-molecular interaction energies for benchmark molecular systems. We apply the method to a prototypical quasi-two-dimensional covalent organic framework (COF) composed of carbon and hydrogen. The COF structure, inter-layer binding energies, and hydrogen absorption are analyzed at CCSD(T) accuracy. Overall, the developed $\Delta$-learning approach opens a practical route to large-scale atomistic simulations that include vdW interactions with chemical accuracy.