CoupModel: Model Use, Calibration, and Validation

The CoupModel has been developed to represent a platform with various modules that can be linked together for the users specific application. This article provides a review of the model development, calibration procedures, and previous applications. For each application, the user can select different modules and specify how they should be linked. In the next step, appropriate inputs to run the model are specified based on the choice of modules. By offering water, heat, tracer, chloride, nitrogen, and carbon modules, the CoupModel allows the user to simulate a wide range of soil-plant-atmosphere interactions for any terrestrial ecosystem. The spatial distribution is lumped or distributed to any user-defined scale, and the temporal resolution is from minutes to hundreds of years. The platform allows the user to specify inputs as (1) forcing time series, (2) simple predefined patterns of variation by parameter functions, or (3) dynamic parameters that change value at specified dates during the simulation. Output variables from simulations can be compared with any independent measurement either as a time series or as a single value. The performance is expressed using conventional statistical indicators or as log likelihood sums. Simulations are made as single runs to represent a unique input or as a multiple series of simulations based on random or systematic sampling of parameter values. Parameters can also represent an object that is a collection of different parameters to represent a particular system, such as a soil profile. Two possible approaches may be used for calibration: Bayesian or generalized likelihood uncertainty estimation (GLUE). The former uses a Markov chain Monte Carlo (MCMC) method to sample among parameter values based on predefined error parameters for estimation of log likelihoods. Experience has shown that one can learn both how to improve measurements and how to understand the importance of model structural errors by using uncertainty-based calibration methods. The major advantage of GLUE compared to Bayesian calibration is that the learning process is simple and flexible. Bayesian calibration requires Gaussian distributions of residuals, which in many cases are not justified, and also requires careful consideration of error functions prior to the simulations. Although in some applications with a single output variable the Bayesian method is most suitable, for many multiple-criteria problems the flexibility of the GLUE approach is favored over the elegance of the Bayesian framework.