Surface Conditioning for Zero-Shot Multivariate Correction in a Marine Biogeochemistry Emulator

Bias correction in process-based models usually means iterative data assimilation with explicit covariance modelling. Here we show that the conditioning pathway of a pretrained ocean biogeochemical (BGC) water-column autoencoder can serve as a lightweight observation injection point, requiring no retraining or optimisation. The decoder is conditioned on surface variables via Feature-wise Linear Modulation (FiLM), originally designed to incorporate physical boundary forcings. We extend these inputs to include surface chlorophyll-a and replace the simulated field with satellite observations at inference time. The surface injection propagates into depth-resolved, multivariate corrections across the output variables in a trophically consistent manner, and evaluated against float-based observations, the conditioning improves accuracy at all depths.