20 As described above, the Baltic Sea is a very complex marine system, with vertical stratification, significant horizontal density gradients and fresh water supplies. Furthermore, it is partially covered with ice, particularly the Northern parts in the Gulf of Bothnia and the Gulf of Finland, during some months each year (which a?ects not only deposition events taking place during winter but also has implications on water circulation itself). In spite of this, model results are consistent, even in the case of hydrodynamic models. The USEV model constitutes a very simple approach in which all these processes are neglected. In contrast, they are included in the complex THREETOX model. Therefore, it can be concluded that they do not play a significant role in the redistribution of contaminants within the Baltic Sea (see Figure 10). Of course, this may not be the case in a di?erent marine area. In addition, given the relatively short simulated times (5 years) and water residence time in the Baltic Sea (some 10–30 years [15]), exchanges of radionuclides with the Atlantic Ocean do not play a significant role. While the THREETOX and USEV models only include deposition of radionuclides released from the Chornobyl accident as a 137Cs source (added over a pre-Chornobyl accident background), Sellafield and La Hague releases are considered in the POSEIDON and NRPA models. From the intercomparison of model results and comparisons with observations shown in Figures 8 and 9, it is clear that Chornobyl fallout is the dominant source, as already noted in Section 2.3 above. Significant work has been carried out concerning multi-model applications [37, 38]. It has been claimed that, given a certain level of process understanding, di?erent model structures and parameter values can be equally acceptable. Traditionally it is supposed that an ‘ideal model’ exists, this being a unique model, inherent to nature. Thus, di?erent models are di?erent realizations of the ideal model in view of the specific applications for which they were developed. Consequently, a multi-model approach can be accepted if, and only if, the di?erent models are developed to solve problems of various types, for which di?erent realizations of the ideal model can be appropriate. However, it has been observed that this statement cannot be easily supported [37] and the results of the current work confirm this previous finding. Models with very different structures and parameters have been applied to the same environmental problem and no criteria can be found to decide which could be the most appropriate one. In this sense, it is interesting to point out that models may perform di?erently depending on the target variable. For instance, one model may predict radionuclide concentrations in bed sediments in good agreement with measurements, but it may provide not so good results for the water column. For another model, the situation may be the opposite. For a correct model comparison, the appropriate question needs to be ‘asked’ to each model. This is particularly relevant when box and hydrodynamic models are compared and it has been documented previously that “di?erent model approaches can lead to comparable results if these results are extracted in the correct way” [39]. Measurements of radionuclide concentrations in the Baltic Sea were already available within the framework of HELCOM10 when this study was carried out. Thus, a real blind-test exercise was not possible. However, as noted previously, no calibration was made for the POSEIDON, THREETOX and NRPA models. Only in the case of the USEV model was data on 137Cs inventories in the water column and seabed used to calibrate uptake/release processes (see Annex V below). Thus, the results of the present exercise have not been significantly influenced by the prior knowledge of data. 10 http://www.helcom.fi/