High resolution ocean models are needed to resolve circulation, hydrographic structure, and flow responses around seamounts, but their computational cost limits long simulations, repeated experiments, and applications over many sites. This study presents a hybrid neural model for reconstructing three dimensional ocean fields over complex seamount topography. The model combines a Fourier Neural Operator with a three dimensional U-Net. The Fourier Neural Operator represents broad spatial patterns and links distant parts of the flow field. The U-Net receives bathymetry, slope, curvature, an ocean mask, and information describing the terrain following grid, allowing it to represent local responses near summits, flanks, and passages. A separate learned weight combines the two model predictions for eastward velocity, northward velocity, vertical velocity, temperature anomaly, and salinity anomaly.

The dataset was produced from a high resolution ROMS simulation of a western Pacific region containing 25 seamount systems. Each sample includes all 50 terrain following layers and retains the horizontal extent of its seamount system. Coarse inputs were generated by Gaussian filtering, sixfold horizontal downsampling, and interpolation back to the original grid. The effective input resolution is about 11 km, while the target grid spacing is about 1.85 km in the zonal direction and 2.27 km in the meridional direction. Training and validation used samples from October, November, December, and January. The independent test set contains 120 August samples from four seamount systems that were not used during model development. The test therefore changes both seamount geometry and seasonal background.

Compared with interpolation of the coarse fields, the hybrid model reduced root mean square errors averaged over all 50 layers by 38.3% for horizontal speed, 36.1% for eastward velocity, 34.9% for northward velocity, 68.5% for temperature, and 44.1% for salinity. The error in vertical velocity changed by only 0.3%, showing that weak and intermittent vertical motion remains difficult to recover after strong spatial degradation. The model improved reconstruction at surface, intermediate, and near bottom layers, under weak and strong flow conditions, and across all four unseen seamount systems. Diagnostics derived from the reconstructed velocity fields also showed better agreement for kinetic energy, relative vorticity, divergence, strain rate, and horizontal power spectra. These results show that spectral modelling and explicit topographic information provide complementary benefits for three dimensional ocean downscaling. They also identify vertical velocity as the main remaining challenge for physically consistent reconstruction around complex seamounts.