Experiment

EXP 03: Large ensemble testbed

Tests whether reconstruction biases are robust across model structures and internal variability.

Overview

EXP 02 assessed sfCO₂ reconstruction methods using five GOBMs. This experiment extends that approach to multiple ESMs and multiple ensemble members per ESM. The question is whether the bias of each sfCO₂ reconstruction method is consistent across model structures and internal variability.

Lead Author

Manuscript

To be added

Experiment Information

Status:In preparation
DOI:To be added
Testbed: 18 ensemble members from 6 CMIP6 ESMs:
  • CanESM5 (Swart et al., 2019)
  • CanESM5-CanOE (Swart et al., 2019)
  • CESM2 (Danabasoglu, 2019)
  • CESM2-WACCM (Danabasoglu, 2019)
  • MPI-ESM1-2-LR (Wieners et al., 2019)
  • UKESM1-0-LL (Tang et al., 2019; Good et al., 2019)
Three ensemble members per ESM.
sfCO₂ reconstruction methods: 5 sfCO₂ reconstruction methods:
  • CarboScope (Rödenbeck et al., 2014, 2022)
  • CMEMS-LSCE-v1 (Chau et al., 2022)1
  • LDEO-ΔResidual (Heimdal et al., 2026)2
  • UExP-FNN-U (Watson et al., 2020; Ford et al., 2024)
  • VLIZ-pySOMFFN (based on Landschützer et al., 2016)3
The sfCO₂ reconstruction methods may differ methodologically from the versions submitted to GCB 2024 and those used in Experiment 2. Combined with the different testbed period and SOCAT version, comparisons with Experiment 2 should be interpreted with caution.
1 Lower spatial resolution (1°) version of the product used for the GCB, with ensembles of 50 instead of 100 members.
2 Not included in the GCB, where LDEO-HPD is used (Gloege et al., 2022; Bennington et al., 2022). LDEO-ΔResidual uses an eXtreme Gradient Boosting regression with ΔpCO₂-Residual as target variable, which shows improved skill (Heimdal et al., 2026).
3 Python reimplementation of VLIZ-SOMFFN with extended hyperparameter tuning. Not yet published; this testbed provides a first assessment of its skill.
SOCAT mask used:SOCATv2024
Design:EXP 03 evaluates the reconstruction bias of five sfCO₂ reconstruction methods using a testbed of six CMIP6 ESMs. Eighteen ESM ensemble members are subsampled to the spatiotemporal coverage of SOCATv2024, and sfCO₂ is reconstructed globally over 1983–2023 with each sfCO₂ reconstruction method. By providing multiple ensemble members, the ESMs capture a range of possible sfCO₂ states. This experiment expands on Experiment 2, where the skill of the sfCO₂ reconstruction methods is assessed using five GOBMs as a testbed. For Experiment 3 we ask: is the reconstruction bias of each sfCO₂ reconstruction method consistent and robust across model structures and internal variability?