Local Correlation-Based Machine Learning Enables Coupled-Cluster-Level Energetics for Thousand-Atom Molecular Systems

Data-intensive molecular workflows increasingly require consistent energies for many molecules and configurations. High-accuracy data are much scarcer because the small energy differences governing conformational preferences and molecular recognition can require an accurate treatment of electron correlation. Coupled-cluster singles and doubles with perturbative triples [CCSD(T)] provides a widely used benchmark, but repeated calculations become prohibitive for medium-to-large molecular systems. Machine learning has enabled rapid coupled-cluster-quality predictions for small molecules, particularly by learning high-level corrections to lower-cost calculations, yet its extension to large systems remains limited by expensive full-system labels and increasingly complex electronic representations. Here we develop cluster-in-molecule machine learning (CIM-ML), which learns orbital-pair corrections from CIM resolution-of-the-identity second-order Møller-Plesset perturbation theory (CIM-RI-MP2) to CIM domain-based local pair natural orbital CCSD(T) [CIM-DLPNO-CCSD(T)]. Fixed CIM locality criteria define bounded learning environments, and the CIM energy expression assembles local predictions into an extensive total energy. A model trained on one (H2O)128 conformer transfers to additional conformers and (H2O)237, reproducing conformational-energy separations within approximately 1 kcal mol-1. For a 1,027-atom protein-ligand complex, training on 500 local clusters reduces the interaction-energy mean absolute error from 12.92 to 3.98 kcal mol-1. CIM-ML thus provides a domain-adaptable route for reusing high-level correlation information across large molecular ensembles.