The CLM5 Parameter Perturbation Experiment

Land models simulate societally-relevant processes such as ecosystem dynamics, terrestrial hydrology, and agriculture with a high degree of complexity. Adequately predicting terrestrial processes such as climate-carbon feedbacks relies on assessing confidence in these models and their predictive capabilities while minimizing sources of model error. Full complexity land models include a large number of parameters that determine fluxes and states predicted by the model. We seek to understand how a single land model might vary in its projections, using various unique (but plausible) parameter sets. This led to the Community Land Model, version 5 (CLM5) Parameter Perturbation Experiment. To find these plausible parameter sets, we first ran approximately 2500 simulations of CLM5, varying over 200 parameters one-at-a-time, across 6 different forcing scenarios. To explore parameter interactions and work towards automated calibration, we then ran 500 simulations varying a subset of important parameters simultaneously using a Latin hypercube sampling technique. This work has yielded two valuable datasets which will soon be publicly available, prompting multiple offshoot projects. Current work includes leveraging novel machine learning methods to emulate CLM5 parameter response functions and calibrate the model, providing increased computational efficiency as well as objective methods to assess calibration results. By sampling from the plausible parameter space and running future climate simulations, we can estimate the contribution of parametric uncertainty on emergent features of the climate system such as the trajectory of the land carbon sink.