Artificial intelligence for modelling diffusion/advection flows in a porous medium [KIMoDA]

In the context of safety assessments for deep geological repositories, hydrogeological processes – in particular the migration of dissolved substances and radionuclides in porous and fractured media – are modelled numerically. AI-supported and hybrid modelling approaches have the potential to improve computational efficiency, calibration and model comparisons. At the same time, they are particularly sensitive in the context of final repositories because their results could be incorporated into safety-relevant assessments that must meet high standards of transparency, traceability, reproducibility and trust.

The KIMoDa project therefore investigates the opportunities and limitations of AI-based methods in comparison with conventional numerical models. Technically, AI and hybrid approaches are examined for selected transport processes in relevant host rocks and compared with standardised reference cases using numerical reference simulations, including PFLOTRAN. The assessment covers model quality, robustness, scalability, and possibilities for verification and interpretation.

The focus of the Öko-Institut’s contribution is on the socio-technical risks of the potential use of AI-based models in long-term safety analyses and in the context of the site selection process. The study examines how characteristics of AI systems – such as data dependency, limited explainability, bias, non-determinism, overfitting or limited transferability – interact with institutional practices, validation routines, data governance and communication requirements. This may give rise to risks of misinterpretation, inappropriate attribution of trust, reduced verifiability or unclear responsibilities. These issues are assessed in the light of key principles of the Site Selection Act (StandAG), such as traceability, reproducibility, participation, accountability and precaution, under conditions of profound uncertainty.

The approach involves a systematic review of the state of the art, the development of a socio-technical risk analysis, the definition of technical reference cases and evaluation metrics, and a comparison of AI-based, hybrid and conventional modelling approaches. In addition, explainable AI approaches, sensitivity analyses and appropriate visualisations are utilised to improve the auditability and communicability of the results. Finally, the findings will be synthesised from both technical and governance perspectives to identify the prerequisites and limitations of a trustworthy, traceable and publicly credible use of AI-supported modelling approaches in safety-related assessment processes.