Hybrid Physics–AI Ecosystem Simulations Improve Biogeochemical Predictions in Temperate Shelf Seas

Biogeochemical models form a core part of marine forecasting and climate projections, yet they suffer from persistent biases in predicting key ecosystem variables, creating challenges across regional and global scales. To address this, we developed an AI-augmented three-dimensional hybrid framework that integrates machine-learning corrections directly into a process-based model’s productivity engine at runtime, keeping mechanistic formulations central while deploying physics-constrained, data-driven AI adjustments around them. We explored two independent hybrid pathways: a satellite-trained primary production scale-factor and a physiology-informed parameter adjustment. Using a temperate shelf-sea as a testbed, we evaluated multi-year hybrid simulations against in situ, Argo, and satellite observations, as well as data-assimilative (DA) reanalysis and high-resolution simulations. Results show that the hybrid framework substantially reduced long-standing biases and outperformed reanalysis and high-resolution simulations across several metrics, including evaluation years and variables, not seen during AI training. This demonstrates that correcting ecosystem process representations while remaining mass-conservative can yield greater accuracy than increasing spatial resolution or relying entirely on continuous DA. Furthermore, because our AI components utilise globally available satellite and experimental datasets, our framework is potentially transferable across global shelf seas. This low-computational, interpretable approach could deliver an effective alternative for operational forecasting and long-term climate applications.