A Physics-Constrained Synergetic Active- and Transfer-Learning Framework for Constructing a Spectroscopically Accurate Full-Dimensional Potential Energy Surface of H2O–N2 Complex

Abstract High-resolution spectroscopy of weakly bound molecular complexes requires theoretical predictions with uncertainties below 0.1 cm–1. Achieving this level of accuracy places stringent demands on the underlying full-dimensional potential energy surfaces (PESs). However, constructing such PESs remains challenging because flexible configuration space requires dense sampling and the ab initio energies used for fitting must meet spectroscopic accuracy requirements. Here, we develop a physics-constrained synergetic active- and transfer-learning framework (SynATL) that combines uncertainty-driven active learning, high-level refinement via transfer learning, and an explicit long-range representation to construct spectroscopically accurate full-dimensional PESs. SynATL was used to construct a full-dimensional PES of the H2O–N2 complex at the post-CCSD(T) level, achieving a weighted root-mean-square error (RMSE) of 0.815 cm–1 on the test set and sub-cm–1 RMSEs in the spectroscopically relevant low-energy and attractive regions. Applied to rovibrational calculations of H2O–N2 and D2O–N2, the resulting PES reproduces microwave transitions with a root-mean-square deviation (RMSD) of 1.6 × 10–4 cm–1, about 80% lower than previous theoretical values, and gives 10–3 cm–1 level band-specific RMSDs for 235 assigned infrared lines across the H2O–N2 bending/overtone and D2O–N2 stretching/bending bands. SynATL therefore provides an efficient and transferable framework for full-dimensional PESs for quantitative spectroscopy of weakly bound complexes.