Accurately quantifying global internal tide (IT) energy conversion is crucial for parameterizing abyssal mixing in Earth system models (ESMs). However, mapping this conversion remains a formidable challenge. High-resolution ocean circulation models accurately resolve small-scale baroclinic tidal processes yet remain computationally intractable for long-duration global climate simulations. While conventional linear internal tide theory offers an efficient parameterization scheme for quantifying tidal energy conversion, it exhibits substantial inaccuracies over supercritical bathymetry driven by strong nonlinear wave coupling and robust high-mode scattering.
Here, we present a novel, physics-constrained deep learning framework, the ViT-ConvLSTM-Dual (VC-Dual), to parameterize the global semidiurnal IT energy conversion rate. Rather than relying on arbitrary statistical features, our model's inputs are strictly constrained to physical predictors governing tidal-topographic interactions (e.g., buoyancy frequencies, barotropic tidal ellipses, and high-resolution bathymetry). To address the distinct dynamical regimes and orders-of-magnitude disparities in energy fluxes, we employ a domain-specific dual-model strategy, training independent networks for marginal seas and the open ocean.

Benchmarked against tide-resolving HYCOM simulations, the VC-Dual framework yields unprecedented accuracy. Compared to traditional linear theory, our model reduces the root mean square error by ~93.6% in marginal seas and ~98.1% in the open ocean. It successfully resolves fine-scale conversion hotspots at shelf breaks and mid-ocean ridges where analytical methods fundamentally break down. Furthermore, comprehensive sensitivity analyses demonstrate that the VC-Dual model maintains robust predictive skill even when subjected to severe data imperfections, including high-level Gaussian noise, spatial data gaps, and coarse-resolution downsampling (simulating low-resolution ESM grids).

This computationally efficient and physically consistent parameterization bridges the gap between open-ocean linear regimes and nonlinear marginal seas. By providing reliable, high-fidelity global IT generation fields under realistic data constraints, the proposed framework offers a robust foundation for quantifying the global baroclinic energy budget and significantly improving ocean mixing representations in next-generation climate simulations.