Application of Artificial Intelligence and Machine Learning in the Safety Assessment of Near-Surface Disposal of Category A Radioactive Waste

This thesis will investigate the use of state-of-the-art Python-based machine-learning methods to develop metamodels of the NIRAS safety assessment model for near-surface disposal of Category A radioactive waste [1, 2]. Metamodels, or surrogate models, can provide computationally efficient approximations of computationally expensive numerical models, enabling applications that require a large number of model evaluations [3]. The developed metamodels will therefore be investigated as a means of efficiently reproducing the NIRAS model for applications such as global sensitivity analysis and Monte Carlo simulations for uncertainty assessment.

A key challenge is achieving sufficient accuracy for a multi-input model with multiple time-dependent outputs. In addition to predicting radionuclide fluxes from the disposal facility, the metamodel should reproduce the evolution of activity within the different components of the disposal system. Different machine-learning approaches will be implemented and compared in terms of predictive accuracy, computational efficiency and robustness. Particular attention will be given to the validation of the metamodels against independent calculations from the original NIRAS model and to their practical applicability to safety-assessment calculations.

The thesis will also investigate uncertainty evaluation, including the propagation of input uncertainty through the metamodel and the potential uncertainty introduced by the metamodel approximation itself. Finally, accomplished and emerging methods for global sensitivity and uncertainty analysis will be investigated and applied to assess how machine-learning metamodels can facilitate these analyses for computationally demanding safety-assessment models [4, 5].

References