TL;DR
This paper shows that even vastly different soil carbon models can produce similar global soil organic carbon (SOC) predictions when constrained by real-world data. They developed a novel data assimilation approach (PRODA) that harmonizes model outputs, suggesting that observational data is key to reducing uncertainty in climate projections.
What Problem Does This Paper Address?
Climate models rely on biogeochemical models to simulate how soil organic carbon (SOC) cycles, which is crucial for understanding climate feedback. However, these models often produce widely divergent results due to differences in their underlying structure and parameter values. This uncertainty hinders our ability to accurately predict future climate change. This paper tackles this problem by investigating whether data assimilation can reconcile the predictions of structurally distinct models, thereby reducing inter-model uncertainty.
Methods
The authors employed a PROcess-guided deep learning and DAta-driven modeling (PRODA) approach. This method combines:
- Bayesian data assimilation (MCMC): Used at a site-level to estimate model parameters by comparing model simulations to observed SOC data.
- Deep learning (Neural Networks): Used to generalize the site-level optimized parameters to a global scale, creating spatially continuous parameter maps.
This PRODA approach was applied to two distinct soil carbon models:
- CLM5: A microbial implicit model using first-order kinetics.
- COMPAS: A microbial explicit model using Michaelis-Menten kinetics.
Datasets
The study utilized a large dataset of over 50,000 globally distributed SOC vertical profiles. These profiles were sourced from:
- World Soil Information Service (WoSIS) (2019 snapshot)
- Northern Circumpolar Soil Carbon Database (NCSCD)
- Data from a previous study (Mishra et al., 2020)
Key Insights and Ideas
- Data is King: The study powerfully demonstrates that observational data can significantly reduce uncertainty, even between models with fundamentally different internal structures.
- Convergence is Achievable: Despite CLM5 and COMPAS having distinct ways of representing microbial processes and carbon transfer, PRODA optimization led them to similar global SOC storage and spatial patterns.
- First-Order Kinetics Suffice (Globally): For global SOC storage and spatial patterns, the simpler first-order kinetics used in CLM5 proved as effective as the more complex Michaelis-Menten kinetics in COMPAS, especially after data assimilation. This suggests that for large-scale predictions, the added complexity of Michaelis-Menten might not be necessary if parameters are well-constrained.
- Model Structure Still Matters (Locally): While global SOC predictions converged, some finer-scale processes like vertical transport and carbon input allocation remained less constrained and showed divergent patterns between the models, highlighting the need for more specific data for these components.
Key Results
- After PRODA optimization, CLM5 and COMPAS showed similar global SOC storage and spatial patterns.
- The correlation between the simulated spatial patterns of SOC by CLM5 and COMPAS increased from -0.026 (ad hoc parameterization) to 0.51 (PRODA optimized).
- PRODA-optimized CLM5 and COMPAS simulated global SOC storage in the top meter to be 1469 Pg C and 1507 Pg C, respectively.
- The study found that Michaelis-Menten kinetics in COMPAS could be approximated by multiplicative kinetics (effectively first-order with respect to substrate) at the global scale after data assimilation, as the Michaelis constants were much larger than substrate concentrations.
Limitations and Future Work
- Data Scarcity for Specific Processes: The paper acknowledges that parameters related to vertical transport and carbon input allocation were less constrained by the available SOC profile data. This suggests a need for more targeted observational data to improve predictions for these components.
- Equifinality Problem: The identifiability of some parameters remains weak, leading to the "equifinality problem" where different parameter values can produce similar SOC storage results. This challenges the prediction of future SOC changes under climate change.
- Temporal Dynamics: The study focused on steady-state SOC. Whether these converged models can predict dynamic SOC changes at different temporal scales remains an open question.
- Broader Data Inclusion: The authors suggest that incorporating a wider range of observational data (e.g., decomposition rates, radiocarbon data, microbial traits, vegetation, and hydrology dynamics) would further improve model predictive ability and inform model development.
- Exploring Other Nonlinear Kinetics: The study only investigated one form of nonlinear kinetics (Michaelis-Menten). Further research is needed to explore other nonlinear kinetic theories and their performance at different scales.
Summary updated