Preprint Proposes Manifold Framework to Benchmark Surface Ocean Carbon in SOCAT and GOBMs

The authors apply a geometry-based framework to SOCAT ship and mooring data against multiple GOBM ensembles. They compute manifold intrinsic dimension within standard feature spaces (temperature, salinity, nutrients, mixed-layer depth) and quantify information imbalance between observation and model manifolds. Results indicate GOBMs preserve relative variable importance but under-represent the effective dimensionality of real carbon variability, with largest discrepancies south of 50°S.

Surface ocean carbon directly controls the ocean's anthropogenic CO2 sink, which absorbs roughly 2.5 Gt C yr⁻¹. Systematic under-representation of manifold complexity therefore propagates into air-sea flux uncertainty and acidification projections used by IPCC assessments. The study also demonstrates that optimal representations improve distance-based machine-learning flux estimates and proposes two new weighted-ensemble metrics.

No temporal drift in representation quality appears across the 2000–2020 record, implying that model structural errors rather than data sparsity dominate. The work remains a preprint and lacks independent validation on withheld observational platforms or coupled Earth-system models.

Next steps include testing the metrics on CMIP7-era biogeochemistry output and incorporating them into operational carbon budget ensembles within two years.