The long term goal of this team is to build measurement based models for environmental systems. The current trend is to build more complex models in order to simulate the 'real' world, while a 'good' model is not necessarily the most complex model. On the contrary, it only needs that complexity, which is supported by experimental measurements. It has to be able to describe all significant variation in the data, without modelling the stochastic measurement uncertainties. These models are used for two purposes: (i) to extract the maximum amount of significant information out of noisy measurements; (ii) to make future predictions.
Suppose, for instance, that we have a climate model with a set of estimated parameters and an experimental record. Now, I would like to bring in an additional physical process, which would make the model more realistic. What will happen?
With every modelled process, some model parameters are associated. Tuning these parameters will match the extended model better on the record. So, our model has become more realistic. However, uncertainties are associated with these additional model parameters and these additional uncertainties will lower the prediction horizon of the model. So, a higher complexity could be a disadvantage for a model. Briefly, a first step in the search for a good climate model is the determination of its complexity. To reach this goal, the modeller needs the experimental data.
On the other hand of the spectrum, experimentalists are measuring climate/ ecological changes in the past. However, the past cannot be directly measured. Instead, proxies in sediments, biogenic carbonates, ice cores etc... are measured and consequently, the history is revealed as function of a distance, while we would like to have a time series. The latter can be introduced by comparing the measured record with a model, containing the time. The simplest models can be 14C-dating curves, but any model containing time can do the trick. So, at this stage, every experimentalist will need models and modellers and some evident questions arising are: which model to choose? How are the model parameters tuned (on a record, not yet dated or on a previously dated one)? How are errors propagating trough this interaction between models and measurements?
These two examples show that a horizontal orientated approach, with specialists in measuring and others in modelling, has its limits. For this reason, we have chosen to approach some important questions in environmental reconstruction and prediction in a vertical manner, where we follow the information stream from the bottom up, from the experimental set-up, the calibration of the instrument, dating the records, over tuning the model parameters, till the selection of the appropriate model. Such a quantitative formulation of the measurement problem is precisely what has been developed in the system identification community and our first short term goal is the introduction of this approach in the community of environmental sciences.