Bridging DFT and ML with Density-Corrected SCAN: Chemical Accuracy for Water from Clusters to Bulk

Semilocal functionals such as SCAN suffer from density-driven delocalization errors that bias hydrogen bonding and reactivity. I show that applying the density-corrected DFT framework to SCAN (DC-SCAN) exposes small functional-driven errors and elevates SCAN to CCSD(T) accuracy for water. DC-SCAN quantitatively reproduces interaction energies for (H₂O)_n clusters and, via a data-driven many-body formalism, yields liquid structure in agreement with experiment. Building on this accuracy, we train a deep neural network (DNN) potential on DC-r²SCAN data to access the long time and length scales needed for rare events, accounting for nuclear quantum effects through path integrals. The resulting DNN@DC-r²SCAN simulations resolve the free-energy landscape of water autoionization, predict pKw = 13.71 at 300 K, and show how the Grotthuss mechanism stabilizes solvent-separated ion pairs while lowering the recombination barrier to ~2 kcal/mol. Extending the same DC-SCAN + DNN approach to water monolayers in sub-nanometer slit pores, we find that confinement markedly suppresses autoionization as hydroxide hypercoordination becomes frustrated, reorientation is hindered, and multihop Grotthuss transport is curtailed. Together, these results illustrate how correcting the density in SCAN and coupling it to ML potentials delivers CCSD(T)-level accuracy from clusters to bulk, enabling predictive simulations of aqueous chemistry central to electrochemistry, nanofluidics, and catalysis.