Marine phytoplankton are responsible for approximately half of the planet’s net primary production (NPP) and are undergoing rapid change in response to shifting surface ocean heat budgets. Satellite remote sensing has provided nearly three decades of surface ocean color data, enabling us to infer phytoplankton distribution and improve NPP estimates in the surface mixed layer with unprecedented spatial coverage. However, phytoplankton below the ocean mixed layer contribute substantially to global NPP, yet this contribution remains poorly constrained. This project will leverage satellite data, trained and validated with direct observations provided by the Biogeochemical-Argo autonomous float network, to detect signatures of stratification, deep fluorescence and deep biomass maxima in the surface ocean through physics-informed deep learning approaches. The approach is to examine multiple models with varying levels of abstraction and interpretability, additionally validated against independent ship- and mooring-based regional time series datasets. This study will enable identification of important knowledge gaps in the formulation of mechanistic modelling approaches to derive estimates of phytoplankton biomass, nutrient limitation and NPP beneath the ocean surface. This study will also examine these phenomena in basins with sparse direct observation coverage. The project will develop a novel pipeline using convolutional LSTM coupled with spatial and temporal transformer blocks