Uncertainty analysis tool as part of safety assessment framework: model-independent or model-tailored?

Safety assessment for a radioactive waste disposal facility is built on a systematic analysis of the long-term performance of natural and engineered barriers, the potential migration of radionuclides from the disposal facility, their movement in the environment and resulting radiation hazards.

Quantitative implementation of such kind of analysis requires an elaborated set of numerical models (thermo-mechanical, geochemical, groundwater flow and transport, etc.) that are realized in a variety of software tools.

It goes without saying that numerical models for such a complex system are associated with significant uncertainties of diverse origins: lack of the site-specific or material-specific data, natural variability of the host geological media, imperfect understanding of the underlying processes and so on.

The focus of our study is to provide uncertainty assessment, sensitivity analysis and calibration tools for the whole framework of numerical models involved in the safety assessment.

It became apparent on the way toward this goal that we need to balance between model-independent and model-tailored solutions. In addition to the expected diversity of input-output formats or objective functions for model calibration, we face limitations in the universality of the methods themselves.

For instance, the choice of global sensitivity analysis method is conditioned by model linearity, monotonicity, multimodality and asymmetry and of course its computational cost.

The selection process of the suitable optimization algorithm for calibration purposes is even more complicated because a universal optimization method is even theoretically impossible, and one algorithm can outperform another only if it is adjusted to the specific problem.

As a result, a sufficient list of sensitivity analysis methods includes correlation and regression analysis, multiple-start perturbation, variance-based and density-based methods. The set of calibration methods composed of methods with different search abilities including swarm intelligence, evolutionary and memetic algorithms, and their hybrids. The hybridization allows simultaneously benefit from exploration (global search) ability of one algorithm and exploitation (local search) power of another.

It is also worth mentioning that «unfortunate» results of sensitivity analysis or calibration may indicate the necessity of model revision. Examples of such indicators are low sensitivities to empirically significant parameters or optimal values of parameters close to the boundaries of the reasonable predefined range.

In light of the above uncertainty and sensitivity analysis and parameter calibration became not the model-independent final stage of numerical assessment, but an inseparable part of the model development routine.