A Divide-and-Conquer Embedded Correlated Wavefunction Transfer Learning Approach for Chemically Accurate Condensed-Phase Interatomic Potentials: Liquid Water as a Test Case

Abstract Combining the accuracy of correlated wavefunction theory with the efficiency of machine-learned interatomic potentials offers a promising route toward predictive condensed-phase dynamics. In our previous work [J. Chem. Theory Comput.2026, 22(10), 5174−5184, 10.1021/acs.jctc.6c00403.], we introduced an embedded correlated wavefunction transfer learning (ECW-TL) approach, which efficiently refines a machine-learned potential to correlated-wavefunction accuracy within a predefined region containing the chemistry of interest. Although effective for localized processes such as ion pairing in solution, this formulation is not directly applicable to systems whose behavior and properties emerge collectively and cannot be confined to a selected region. Here, we overcome this limitation by developing a divide-and-conquer formulation of ECW-TL that partitions the system into embedded subsystems, enabling correlated-wavefunction corrections to be learned throughout the entire simulation cell. We demonstrate the method using ambient liquid water as a test case because of its diverse, collectively fluctuating hydrogen-bonding environments. The resulting potential fine-tuned at the embedded-CCSD(T) level improves the structural, thermodynamic, and dynamical properties of liquid water relative to underlying DFT models, bringing calculated observables closer to experiment. These results suggest that divide-and-conquer ECW-TL provides a promising, scalable route toward accurate simulations of complex condensed-phase systems.