Experiment

EXP 02: Multiple Hindcast Testbed

‘Assess how sfCO₂ reconstruction biases (mean state and trend) vary across methods and GOBM testbeds and whether method-specific biases persist or are partially compensated within the multi-method ensemble.

Overview

sfCO₂ reconstruction methods are trained on sparse sfCO₂ observations, so their skill in unsampled regions cannot be tested against the real ocean. This experiment uses GOBMs as a known model truth to assess how well each sfCO₂ reconstruction method reproduces sfCO₂, FCO₂ and the ocean carbon sink.

Lead Author

Manuscript

Alizée Roobaert, Daniel J. Ford, Mohanan G. Sreeush, et al. Reconciling ocean carbon uptake estimates: a multi-model assessment of surface ocean CO₂ reconstruction skill within the SOCOMv2 initiative. ESS Open Archive, 20 May 2026.

Experiment Information

Status:In review
DOI:10.22541/essoar.15003624/v1
Testbed: 5 GOBMs:
  • FESOM2.1-REcoM3 (Gürses et al., 2023)
  • CESM-ETHZ (Doney et al., 2009)
  • NEMO-PISCES (Aumont et al., 2015)
  • MRI-ESM2-2 (Nakano et al., 2011; Urakawa et al., 2020; Sakamoto et al., 2023)
  • NorESM1-OCv1.2 (Schwinger et al., 2016)1
Model output from GCB 2023 (Friedlingstein et al., 2023). Together, these GOBMs span a range of circulation schemes, biogeochemical complexity and parameterizations, to test whether the skill of the sfCO₂ reconstruction methods is robust across GOBMs.
1 GCB 2024 version, identical to GCB 2023 except that sfCO₂ is referenced to moist air instead of dry air, as required for a consistent recalculation of FCO₂.
sfCO₂ reconstruction methods: 9 sfCO₂ reconstruction methods:
  • VLIZ-SOMFFN (Landschützer et al., 2016)
  • CarboScope (Rödenbeck et al., 2014, 2022)
  • JMA-MLR (Iida et al., 2021)
  • OceanSODA-ETHZv2 (Gregor et al., 2024)
  • UExP-FNN-U (Watson et al., 2020; Ford et al., 2024)
  • NIES-ML3 (Zeng et al., 2022)
  • CSIR-ML6 (Gregor et al., 2019)
  • CMEMS-LSCE-v1 (Chau et al., 2022)2
  • LDEO pCO2-Residual (Bennington et al., 2022)3
2 Lower spatial resolution version of CMEMS-FFNNv2, which is used for the GCB.
3 Used instead of LDEO-HPD, the version used for the GCB (Gloege et al., 2022). LDEO-HPD relies on GOBM priors, which would introduce the model truth into the sfCO₂ reconstruction. LDEO pCO2-Residual uses a similar eXtreme Gradient Boosting regression and achieves comparable validation performance.
SOCAT mask used:SOCATv2023
Design:EXP 02 uses a multi-model testbed to evaluate nine sfCO₂ reconstruction methods under sparse observational constraints and assess how sfCO₂ reconstruction biases propagate into FCO₂. Monthly sfCO₂ fields from five GOBMs are sampled according to monthly gridded 1° x 1° SOCATv2023, reconstructed, and compared with complete model fields. A harmonized FCO₂ protocol isolates differences arising from reconstructed sfCO₂.