O1. Identify key model parameters and potential relationships with future climate projections
We tested the sensitivity of the different parameters linked to the carbon, water, energy and nitrogen cycles in the model. These sensitivity analyses helped to identify important parameters for later experiments and were documented as the first parts of Raoult et al. (2023) and Raoult et al. (2024a).
02. Reduce the uncertainty associated with the key parameters identified using DA techniques, and in situ and EO data, to create observationally constrained PDFs for each parameter;
To reduce the uncertainty in the key land surface model parameters, we investigated different Bayesian approaches to model calibration. In addition to classic data assimilation approaches used to calibrate land surface models of high complexity, we explored a method called History Matching. History Matching, a technique that commonly uses machine learning emulators to rule out unlikely parameter values, has been shown to be very effective in tuning the parameters of atmospheric models. We showed that it was a very promising approach to tuning land surface models. This work is documented in Raoult et al. (2024b). We also used a Markov-Chain Monte-Carlo approach to calibrate parameters of a two-pool model of substrate dependence in plant respiration, as documented in Jones et al. (2024).
O3. Investigate how calibration can be used to constrain future climate projections.
We investigated two novel ways to use observations to constrain future climate projections. The first was the combine local model calibration with the emergent constraint approach. Emergent constraints help connect what we observe today with predictions for the future. By using current observations and our understanding of physical processes, we can better predict long-term trends. In addition, calibration is a tool that uses current observations to reduce the uncertainty in the representation of physical processes in climate models. We explored the potential of combining these two techniques to reduce the uncertainty in future projections. This idea is illustrated in Raoult et al. (2023) where we used a relationship between the change in CO2 levels by the end of the century and a key model parameter (Topt) that was optimised using data on plant productivity and heat exchange. We also explored other emergent constraints, including one linking carbon budgets to global warming. This work is documented in Cox et al. (2024).
The second approach considered using manipulation experiments to ensure calibrated parameters capture different model responses under a changing climate. Calibration can reduce the uncertainty in model simulations. However, since calibration is performed against present-day observations, it does not automatically give us confidence in future projections. Manipulation experiments give us a unique look into how the ecosystem may respond to future environmental changes. We used data from the Free Air CO2 Enrichment (FACE) manipulation experiment, which provides data under ambient and elevated CO2 conditions. Calibrating against data under both conditions, we showed we were able to find parameters that captured vegetation response over different CO2 conditions. This work is documented in Raoult et al. (2024a).
Overall.
The three objectives of CLARION give a clear workflow that can be applied to any land surface model and terrestrial process impacting the climate projections one wishes to investigate. The success of this approach is highlighted in Raoult et al., (2023) and Raoult et al., (2024a). Nevertheless, given the increasingly high complexity of land surface models, the plethora of observational datasets now available, and the emergence of machine learning techniques, there is still work to be done as a community to tackle challenges in calibration and capitalise on new data and techniques. As such, the final output of CLARION is a review of the current state of land surface model parameter estimation (Raoult et al., in prep).
