10.12.2012 Views

Euradwaste '08 - EU Bookshop - Europa

Euradwaste '08 - EU Bookshop - Europa

Euradwaste '08 - EU Bookshop - Europa

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

3. The sensitivity analysis benchmark<br />

Many PAMINA partners were interested in getting in depth knowledge and experience in the use of<br />

SA techniques. This was the reason to set a benchmark on SA techniques that could help them getting<br />

that objective. The SA benchmark has been designed as a two-step process. The first step is<br />

dedicated to analyse a set of mathematical functions most of whose sensitivity indices are well<br />

known. The second step consists in analysing a simplified, though representative, PA model. Finally,<br />

though not included as a part of the benchmark but as other tasks committed within PA-<br />

MINA, several partners will study their respective system models, or parts of them, using SA techniques.<br />

The first part of the benchmark consists in studying twelve mathematical models (seven mandatory,<br />

five voluntary) with different SA techniques (the techniques are chosen by each participant, attending<br />

to their respective interests). The first target is to provide the right simple framework to test and<br />

debug their respective SA tools. The second one is to start with rather simple models (three input<br />

parameter linear model) and to continue analysing models with some added complexity: increase<br />

the number of parameters, add nonlinearities, consider non-monotonic models, include periodicity,<br />

consider continuous models whose derivative does not exist at some given points, consider models<br />

with interactions, check the different capability to estimate accurately large and small sensitivity<br />

indices, etc. An additional target is to see the importance of the sample size in the accuracy of the<br />

sensitivity indices.<br />

The model under study in the second step of the benchmark reproduces the behaviour of a radioactive<br />

HLW repository and the contaminant disposed of. Only four radionuclides are considered in<br />

this model, 129 I and the decay chain 237 Np, 233 U and 229 Th. The repository is considered without any<br />

geometric complexity, just a point. Engineered barriers are modelled through a containment time<br />

during which there is no release. After such containment period, the contaminant starts releasing at<br />

a fractional constant rate (one rate for Iodine, a different one for the chain members). The contaminant<br />

is carried by groundwater through two consecutive geosphere layers to the biosphere, where it<br />

gets into a water stream from which exposed population take drinking water. This model has thirtythree<br />

inputs, twelve of which are affected by uncertainty. These model inputs are the initial inventory<br />

of each considered radionuclide, their decay rates (�), their dose conversion factors (�), and all<br />

the other inputs that characterise the physical-chemical properties of the near field, both geosphere<br />

layers and the biosphere. The complex input-output relation, characterised by strong interactions<br />

among input parameters, makes it a challenging model to test SA techniques.<br />

In this case, the target is twofold: firstly to compare different options within a given SA techniques<br />

(to study the added value of using more complex versions of a given technique – classical FAST<br />

versus extended FAST, first order regressions versus higher order regressions, the benefits of using<br />

transformations of inputs and outputs in the SA, the added value of and the problems that arise<br />

when estimating the effect of interactions and total sensitivity indices, etc.); secondly to crosscompare<br />

the results obtained using different techniques. This cross-comparison may be studied<br />

from two points of view: what differences arise when using techniques that target different objectives<br />

and why; i.e.: PCC and FAST, what differences arise when using techniques that target equal<br />

objectives; i.e.: FAST and Sobol’s indices. Additionally, the effect of the sample size is also in the<br />

focus of the second step of the benchmark.<br />

Many PAMINA partners are interested in bringing to their PA studies all the knowledge and experience<br />

acquired during the benchmark. PAMINA RTDC2 and RTDC4 offer a very good opportunity<br />

to test all those techniques in real PA models proposed by several participants. Table 1 shows a<br />

394

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!