A source-target transfer learning framework with physics priors for subsurface chlorophyll reconstruction in the Northwest Pacific Ocean.

Three-dimensional (3D) chlorophyll-a (Chl-a) distributions provide a foundation for understanding upper-ocean ecosystem dynamics and biogeochemical variability. We developed the source-target coupled network (ST-CoupleNet), a physics-guided transfer learning framework that combines Earth system model simulations with observation-constrained reconstructions to estimate subsurface Chl-a in the Northwest Pacific Ocean. The framework first learns large-scale spatiotemporal vertical representations from long-term CMIP6 simulations and then adapts them using reconstructed fields constrained by BioGeoChemical-Argo (BGC-Argo) observations, reducing simulation biases while preserving regional variability. By integrating dual-attention, residual learning, and convolutional long short-term memory (ConvLSTM) architectures, ST-CoupleNet substantially improves reconstruction accuracy over a conventional convolutional neural network (CNN). This framework demonstrates how transferable knowledge from climate simulations can be combined with autonomous observations to generate continuous 3D Chl-a fields, enabling large-scale investigations of subsurface marine ecosystems and biogeochemical change.