Converged reference interaction energies for 100-atom noncovalent complexes: reconciling CCSD(T) and diffusion Monte Carlo

• The discrepancy. Al-Hamdani et al. (Nat. Commun. 2021; arXiv:2009.08927) found that LNO-CCSD(T) and DMC interaction energies for large supramolecular complexes differ by up to about 8 kcal/mol; the coronene dimer is the best-known case. Basis-set incompleteness and local approximations were examined and did not remove it.
• Argument that (T) is at fault. Schäfer, Irmler, Gallo and Grüneis (arXiv:2407.01442, preprint at the time of writing; check for the published version) report that they pinpoint the source of the discrepancy in the perturbative treatment of triples in CCSD(T) for large, highly polarisable systems and propose modified coupled-cluster methods that improve noncovalent interaction energies on the hundred-atom scale. Related earlier work (Masios et al., Phys. Rev. Lett. 131, 186401, 2023) analysed how the (T) correction diverges in the limit of large polarisable systems and proposed a repair.
• Argument that CCSD(T) is adequate. Lambie, Kats, Usyvat and Alavi (arXiv:2411.13986) use a Pariser–Parr–Pople model Hamiltonian with CCSDTQ and CCSDT(Q) benchmarks for stacks up to the dibenzocoronene dimer and find no sign that CCSD(T) overestimates the interaction energy; they therefore place the burden on the DMC side or on other approximations. Model-Hamiltonian conclusions may not transfer to the all-electron case.
• Status. Not settled: as of October 2026 the two readings above coexist in the literature and no all-electron higher-order coupled-cluster or independent QMC benchmark that the community regards as decisive has been established for the largest complexes. Recent crystal-lattice work (Shi et al., arXiv:2603.02180, preprint) uses correlated wavefunction theory with a many-body expansion for molecular crystals and assumes CCSD(T)-type reliability for polarisable organics.

Reference energies for dispersion-bound systems calibrate DFT dispersion corrections, semi-empirical methods and machine-learned potentials, and decide crystal polymorph rankings, host–guest binding and π-stacking in drug design. A resolved method choice would also tell practitioners when CCSD(T) can be trusted for nanoscale π systems.

Peer-reviewed benchmark values for at least five complexes of 100 or more atoms with an uncertainty of at most 0.5 kcal/mol, supported by agreement between two independent high-level approaches (for example higher-order coupled cluster with an extrapolated or corrected triples treatment and a converged-nodes QMC or AFQMC calculation), together with an explanation of the sign and size of the earlier disagreement.