Computer-based modeling tools are essential to support the design process of nuclear waste repositories in order to achieve the highest levels of safety via the understanding and prediction of the complex processes that take place. The problem to be simulated is a thermo-hydro-mechanical (THM) problem in porous media, which includes heat transport, flow of vapor and liquid water, and the stresses and strains that result of these transport phenomena. THM simulation tools like CODE_BRIGHT, one of the most renowned codes in the field, have been intensively tested against experiments in the last decades, but they are mainly based on serial algorithms designed for single-processor computers or relative small shared memory multiprocessor systems. Unfortunately, the computational requirements for nuclear repositories are huge and these codes are unable to provide accurate predictions at required levels of complexity. The final objective of the NuWaSim project was to provide an extremely scalable nuclear waste simulator that will efficiently run on the largest European supercomputers, and that will be eventually used by SME engineering consultants and nuclear agencies to design nuclear deep repositories with the highest levels of safety. To this end, we have developed a parallel version of some of the most relevant THM models available in CODE_BRIGT by exploiting the state-of-the-art numerical algorithms available in the FEMPAR scientific software framework. FEMPAR was selected as the starting point since it has shown perfect scalability up to half a million cores in JUQUEEN, one of the largest European supercomputers. The result of the NuWaSim project is a large-scale finite element solver prototype that has been able to solve large THM problems using up to the order of 3000 cores in the Mare Nostrum 4 supercomputer in Barcelona. The code is based on parallel iterative Krylov sub-space methods combined with large-scale domain decomposition preconditioners. In addition, an alternative discrtization approach based on immersed (or unfitted) Cartesian grids was exploded in order to reduce the mesh generation effort and improve code scalability in large-scale nuclear waste repositories with complex geometries.
